Some initial learnings:\n- Generating web app-based reports goes much better if there are explicit templates/pre-defined components for the model to use.\n- Claude can \u201cheal\u201d broken charts if you give it access to chart images and run a separate QA loop.
Would either feedback from the community or to hear from others that have tried similar things!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Agentic Data Analysis with Claude Code"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://rubenflamshepherd.com/articles/2026-03-09-agentic-data-analysis-with-claude-code"}},"_tags":["story","author_rubenflamshep","story_47325702","show_hn"],"author":"rubenflamshep","created_at":"2026-03-10T16:44:53Z","created_at_i":1773161093,"num_comments":0,"objectID":"47325702","points":6,"story_id":47325702,"story_text":"Hey HN, as a former data analyst, I\u2019ve been tooling around trying to get agents to do my old job. The result is this system that gets you maybe 80% of the way there. I think this is a good data point for what the current frontier models are capable of and where they are still lacking (in this case \u2014 hypothesis generation and general data intuition).
Some initial learnings:\n- Generating web app-based reports goes much better if there are explicit templates/pre-defined components for the model to use.\n- Claude can \u201cheal\u201d broken charts if you give it access to chart images and run a separate QA loop.
Would either feedback from the community or to hear from others that have tried similar things!","title":"Show HN: Agentic Data Analysis with Claude Code","updated_at":"2026-03-14T19:37:44Z","url":"https://rubenflamshepherd.com/articles/2026-03-09-agentic-data-analysis-with-claude-code"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ramtatatam"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hello fellow HN users!
I recently stumbled across the problem where I am given a list of lines and need to join them into shapes. Some lines may overlap with other lines, some lines might be very close (unwanted artifacts), some lines might produce shapes that are not closed and I would need to join beginning and end myself.
Is there any literature which would touch on some more general problem? My intuition tells me I probably should look into graphs theory but maybe there is more intuitive approach? Lines are coming from contour detection algorithm and they may be very very close each to other but not touch each other- human eye would see the shape easily, but from programmatic point of view I feel I'll have to figure out some thresholds that if lines are close enough I consider them part of shape."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Is there any CS concept for joining individual lines into shapes"}},"_tags":["story","author_ramtatatam","story_27122422","ask_hn"],"author":"ramtatatam","children":[27122821],"created_at":"2021-05-11T19:29:48Z","created_at_i":1620761388,"num_comments":3,"objectID":"27122422","points":4,"story_id":27122422,"story_text":"Hello fellow HN users!
I recently stumbled across the problem where I am given a list of lines and need to join them into shapes. Some lines may overlap with other lines, some lines might be very close (unwanted artifacts), some lines might produce shapes that are not closed and I would need to join beginning and end myself.
Is there any literature which would touch on some more general problem? My intuition tells me I probably should look into graphs theory but maybe there is more intuitive approach? Lines are coming from contour detection algorithm and they may be very very close each to other but not touch each other- human eye would see the shape easily, but from programmatic point of view I feel I'll have to figure out some thresholds that if lines are close enough I consider them part of shape.","title":"Ask HN: Is there any CS concept for joining individual lines into shapes","updated_at":"2024-09-20T08:31:28Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"dmbaggett"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"A friend in my network recently posted a well-known CS puzzle on Facebook:
You have 12 coins, one of which weighs more than the other 11. You have a balance, and wish to use it to find the heavier coin. You may only use the balance three times. How do you find the heavier coin?
This is a great puzzle for an early CS learning because it gives some intuition into the general strategy of divide-and-conquer. It also nicely extends to harder variants: can you do the same for 27 coins with three weighings? Is this the limit? Why? Can you prove your answer?
My question for HN is: can you recommend some puzzle books that focus on CS puzzles like this one that are suitable for a bright high schooler? (The bright high school in this case is my daughter.)"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: CS puzzle books for high schoolers?"}},"_tags":["story","author_dmbaggett","story_23379899","ask_hn"],"author":"dmbaggett","created_at":"2020-06-01T15:57:28Z","created_at_i":1591027048,"num_comments":0,"objectID":"23379899","points":1,"story_id":23379899,"story_text":"A friend in my network recently posted a well-known CS puzzle on Facebook:
You have 12 coins, one of which weighs more than the other 11. You have a balance, and wish to use it to find the heavier coin. You may only use the balance three times. How do you find the heavier coin?
This is a great puzzle for an early CS learning because it gives some intuition into the general strategy of divide-and-conquer. It also nicely extends to harder variants: can you do the same for 27 coins with three weighings? Is this the limit? Why? Can you prove your answer?
My question for HN is: can you recommend some puzzle books that focus on CS puzzles like this one that are suitable for a bright high schooler? (The bright high school in this case is my daughter.)","title":"Ask HN: CS puzzle books for high schoolers?","updated_at":"2024-09-20T06:19:12Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"lyavin"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Sources of Intuitions and Data on Artificial General Intelligence"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["intuition"],"value":"https://www.lesserwrong.com/posts/BibDWWeo37pzuZCmL/sources-of-intuitions-and-data-on-agi"}},"_tags":["story","author_lyavin","story_16448511"],"author":"lyavin","created_at":"2018-02-23T19:08:22Z","created_at_i":1519412902,"num_comments":0,"objectID":"16448511","points":2,"story_id":16448511,"title":"Sources of Intuitions and Data on Artificial General Intelligence","updated_at":"2024-09-20T02:04:25Z","url":"https://www.lesserwrong.com/posts/BibDWWeo37pzuZCmL/sources-of-intuitions-and-data-on-agi"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"11235813213455"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["general"],"value":"How Hand Dexterity Made Human General Intelligence"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"https://medium.com/intuitionmachine/hand-dexterity-and-human-general-intelligence-af5596e1d36d"}},"_tags":["story","author_11235813213455","story_38754679"],"author":"11235813213455","created_at":"2023-12-24T16:47:46Z","created_at_i":1703436466,"num_comments":0,"objectID":"38754679","points":3,"story_id":38754679,"title":"How Hand Dexterity Made Human General Intelligence","updated_at":"2024-09-20T16:01:31Z","url":"https://medium.com/intuitionmachine/hand-dexterity-and-human-general-intelligence-af5596e1d36d"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"chaitanyabaweja"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"A Century of General Relativity. Part I: History and Intuition"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"https://medium.com/swlh/a-century-of-general-relativity-part-i-history-and-intuition-27bf0c85535a"}},"_tags":["story","author_chaitanyabaweja","story_21090007"],"author":"chaitanyabaweja","created_at":"2019-09-27T09:32:21Z","created_at_i":1569576741,"num_comments":0,"objectID":"21090007","points":2,"story_id":21090007,"title":"A Century of General Relativity. Part I: History and Intuition","updated_at":"2024-09-20T05:02:48Z","url":"https://medium.com/swlh/a-century-of-general-relativity-part-i-history-and-intuition-27bf0c85535a"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"hsikka"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hey HN,
I am a graduate student doing ML research, and lately I've been thinking a lot about designing learning systems from the hardware through the software layers.
I have no experience with what is going at the processor levels, and I was wondering what prerequisite subjects or general curricula I should follow to learn and reason at these lower levels of abstraction.
To be clear, I'm doing this to build intuitions about new computational systems and how different chips, from ASIC to neuromorphic, may be designed.
Any resources or advice telling me I'm a fool is welcome!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: How to Self-Study Integrated Circuit Design?"}},"_tags":["story","author_hsikka","story_19890949","ask_hn"],"author":"hsikka","children":[19891117,19893447,19893875,19894412,19895605,19896024,19897059,19897283,19897698,19898030,19898132,19899501,19899861,19901273,19910562,19912343,19918930,19945473],"created_at":"2019-05-12T09:26:26Z","created_at_i":1557653186,"num_comments":39,"objectID":"19890949","points":178,"story_id":19890949,"story_text":"Hey HN,
I am a graduate student doing ML research, and lately I've been thinking a lot about designing learning systems from the hardware through the software layers.
I have no experience with what is going at the processor levels, and I was wondering what prerequisite subjects or general curricula I should follow to learn and reason at these lower levels of abstraction.
To be clear, I'm doing this to build intuitions about new computational systems and how different chips, from ASIC to neuromorphic, may be designed.
Any resources or advice telling me I'm a fool is welcome!","title":"Ask HN: How to Self-Study Integrated Circuit Design?","updated_at":"2026-02-23T19:26:40Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"thesephist"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hi HN! I've found this visualization tool immensely helpful over the years for getting an intuition for how an LLM "sees" some piece of text, and with a bit of elbow grease decided to move all compute to client side so I could make it publicly available.
I've found it particularly useful for
- Understanding exactly how repetition and patterns affect a small LM's ability to predict correctly
- Understanding different tokenization patterns and how it affects model output
- Getting a general sense of how "hard" different prediction tasks are for GPT-style models
Known problems (that I probably won't fix, since this was a kind of one-off project)
- Doesn't work well with Unicode grapheme clusters that are multiple GPT-2 tokens (e.g. emoji, smart quotes)
- Support for other models (maybe later?)"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Fully client-side GPT2 prediction visualizer"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://perplexity.vercel.app/"}},"_tags":["story","author_thesephist","story_37398812","show_hn"],"author":"thesephist","children":[37398858,37400608,37401157,37401383,37401721,37402196,37402682,37403022,37405847,37439687],"created_at":"2023-09-05T22:42:39Z","created_at_i":1693953759,"num_comments":11,"objectID":"37398812","points":153,"story_id":37398812,"story_text":"Hi HN! I've found this visualization tool immensely helpful over the years for getting an intuition for how an LLM "sees" some piece of text, and with a bit of elbow grease decided to move all compute to client side so I could make it publicly available.
I've found it particularly useful for
- Understanding exactly how repetition and patterns affect a small LM's ability to predict correctly
- Understanding different tokenization patterns and how it affects model output
- Getting a general sense of how "hard" different prediction tasks are for GPT-style models
Known problems (that I probably won't fix, since this was a kind of one-off project)
- Doesn't work well with Unicode grapheme clusters that are multiple GPT-2 tokens (e.g. emoji, smart quotes)
- Support for other models (maybe later?)","title":"Show HN: Fully client-side GPT2 prediction visualizer","updated_at":"2024-09-20T15:06:43Z","url":"https://perplexity.vercel.app/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"beefield"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"I am currently working in a young company (not a software company per se, but we need to e.g. do some IoT type things.) and have ended up having quite a lot of responsibility in developing/designing the information infrastructure for our company. Now, I do not see myself as a professional in most of these things, and I think at some point we would need to hire a person to step up our game. At this point I would like to keep this as a general question, so assuming I want to hire a developer that should be better than I am, how should I approach this problem? How do I distinguish candidates that can do the smooth talk from the ones that can walk the talk and understand the balance between that on the one hand you need to keep the technological debt in reins but in the other hand sometimes you just need to get sh*t done, even if that is ugly?
Or is this just an impossible task and best bet is to trust to dumb luck, intuition and fizzbuzz?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: How to hire a better software engineer than I am?"}},"_tags":["story","author_beefield","story_14898865","ask_hn"],"author":"beefield","children":[14900056,14901459,14902332,14903460],"created_at":"2017-08-01T09:00:17Z","created_at_i":1501578017,"num_comments":7,"objectID":"14898865","points":12,"story_id":14898865,"story_text":"I am currently working in a young company (not a software company per se, but we need to e.g. do some IoT type things.) and have ended up having quite a lot of responsibility in developing/designing the information infrastructure for our company. Now, I do not see myself as a professional in most of these things, and I think at some point we would need to hire a person to step up our game. At this point I would like to keep this as a general question, so assuming I want to hire a developer that should be better than I am, how should I approach this problem? How do I distinguish candidates that can do the smooth talk from the ones that can walk the talk and understand the balance between that on the one hand you need to keep the technological debt in reins but in the other hand sometimes you just need to get sh*t done, even if that is ugly?
Or is this just an impossible task and best bet is to trust to dumb luck, intuition and fizzbuzz?","title":"Ask HN: How to hire a better software engineer than I am?","updated_at":"2024-09-20T01:13:30Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"coderunner"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"There's a couple companies in the business of using a 3d simulation for training autonomous cars like Waymo, Cruise, Nvidia, and Applied Intuition. I don't quite understand their product though.
1. Are the trained object detectors in the simulation applied to real world data also or is only the part that makes decisions transferred to the real vehicle (e.g. it's safe to turn left here) while detectors trained on real world images of cars, people, etc. used?
2. Tangentially, I thought that in general detectors trained on computer generated images was not very applicable to real world images. eg training on a bunch of images of 3d modeled humans won't work well with testing on pictures of real humans. Is this not true?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Autonomous Cars Simulation"}},"_tags":["story","author_coderunner","story_19702040","ask_hn"],"author":"coderunner","children":[19702923,19703272,19704556],"created_at":"2019-04-19T18:57:06Z","created_at_i":1555700226,"num_comments":6,"objectID":"19702040","points":12,"story_id":19702040,"story_text":"There's a couple companies in the business of using a 3d simulation for training autonomous cars like Waymo, Cruise, Nvidia, and Applied Intuition. I don't quite understand their product though.
1. Are the trained object detectors in the simulation applied to real world data also or is only the part that makes decisions transferred to the real vehicle (e.g. it's safe to turn left here) while detectors trained on real world images of cars, people, etc. used?
2. Tangentially, I thought that in general detectors trained on computer generated images was not very applicable to real world images. eg training on a bunch of images of 3d modeled humans won't work well with testing on pictures of real humans. Is this not true?","title":"Ask HN: Autonomous Cars Simulation","updated_at":"2024-09-20T04:00:10Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"starchild3001"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"(A counterpoint to the prevailing narrative that singularity will be highly discontinuous and disruptive)
The concept of the technological singularity, a hypothetical future event where artificial superintelligence surpasses human intelligence and leads to unforeseeable changes in human civilization, has been a topic of fascination and concern for many years. However, there are several reasons to believe that the singularity may not be as disruptive or as significant as some predict. In this essay, we will explore five key reasons why the technological singularity might be a "big nothing."
1. Resistance to Adopting Superintelligence One of the main reasons why the singularity may not have a profound impact on daily life is that people will likely be unwilling to recognize and listen to a "so-called" superintelligent being. Humans have a natural tendency to be skeptical of authority figures, especially those that claim to have superior knowledge or abilities. Even if a superintelligent AI were to emerge, many people would likely question its credibility and resist its influence in their daily decision-making processes.
History has shown that people often prefer to rely on their own judgment and intuition rather than blindly following the advice of experts or authority figures. This tendency is likely to be even more pronounced when it comes to an artificial intelligence, as people may view it as a threat to their autonomy and way of life. As a result, the impact of superintelligence on society may be limited by people's willingness to accept and integrate its guidance into their lives.
2. Difficulty in Identifying Superintelligence Another reason why the singularity may not be as significant as some believe is that it will be very challenging to define and recognize superintelligence. Intelligence is a complex and multifaceted concept that encompasses a wide range of abilities, including reasoning, problem-solving, learning, and creativity. Even among humans, there is no universally accepted definition or measure of intelligence, and it is often difficult to compare the intelligence of individuals across different domains or contexts.
Given this complexity, it will be even more challenging to determine whether an artificial intelligence has truly achieved superintelligence. Even if an AI system demonstrates remarkable abilities in specific tasks or domains, it may not necessarily be considered superintelligent by everyone. There will likely be ongoing debates and disagreements among experts and the general public about whether a particular AI system qualifies as superintelligent, which could limit its impact and influence on society.
3. Limitations of Artificial Intelligence Fundamental differences between machine intelligence and human intelligence may permanently limit the scope and applicability of AI. This can be understood through analogies. For example planes can fly, but they aren't a replacement for birds. Similarly, submarines can swim, but they aren't a replacement for fish. Likewise a machine built to mimic human intelligence may never be a perfect replacement of human intelligence or intellect, due to structural incompatibilities and differences (biological organism vs silicon machine). The gap may remain this way for the foreseeable future.
Contemporary AI systems predominantly rely on narrow, domain-specific algorithms trained on vast datasets. They lack the general intelligence and versatility that humans possess, which enables us to learn from experience, apply knowledge across diverse domains, and navigate novel scenarios. The degree to which silicon machines can emulate human capabilities remains uncertain, even if they eventually surpass us in specific areas such as information retrieval, logical reasoning, textual Q&A, analysis, scientific research and discovery.
<To be continued>"},"title":{"matchLevel":"none","matchedWords":[],"value":"Why the Technological Singularity May Be a \"Big Nothing\""}},"_tags":["story","author_starchild3001","story_45154950","ask_hn"],"author":"starchild3001","children":[45154951,45155520,45157588,45157740,45159783,45160118,45161604],"created_at":"2025-09-07T02:48:03Z","created_at_i":1757213283,"num_comments":8,"objectID":"45154950","points":8,"story_id":45154950,"story_text":"(A counterpoint to the prevailing narrative that singularity will be highly discontinuous and disruptive)
The concept of the technological singularity, a hypothetical future event where artificial superintelligence surpasses human intelligence and leads to unforeseeable changes in human civilization, has been a topic of fascination and concern for many years. However, there are several reasons to believe that the singularity may not be as disruptive or as significant as some predict. In this essay, we will explore five key reasons why the technological singularity might be a "big nothing."
1. Resistance to Adopting Superintelligence One of the main reasons why the singularity may not have a profound impact on daily life is that people will likely be unwilling to recognize and listen to a "so-called" superintelligent being. Humans have a natural tendency to be skeptical of authority figures, especially those that claim to have superior knowledge or abilities. Even if a superintelligent AI were to emerge, many people would likely question its credibility and resist its influence in their daily decision-making processes.
History has shown that people often prefer to rely on their own judgment and intuition rather than blindly following the advice of experts or authority figures. This tendency is likely to be even more pronounced when it comes to an artificial intelligence, as people may view it as a threat to their autonomy and way of life. As a result, the impact of superintelligence on society may be limited by people's willingness to accept and integrate its guidance into their lives.
2. Difficulty in Identifying Superintelligence Another reason why the singularity may not be as significant as some believe is that it will be very challenging to define and recognize superintelligence. Intelligence is a complex and multifaceted concept that encompasses a wide range of abilities, including reasoning, problem-solving, learning, and creativity. Even among humans, there is no universally accepted definition or measure of intelligence, and it is often difficult to compare the intelligence of individuals across different domains or contexts.
Given this complexity, it will be even more challenging to determine whether an artificial intelligence has truly achieved superintelligence. Even if an AI system demonstrates remarkable abilities in specific tasks or domains, it may not necessarily be considered superintelligent by everyone. There will likely be ongoing debates and disagreements among experts and the general public about whether a particular AI system qualifies as superintelligent, which could limit its impact and influence on society.
3. Limitations of Artificial Intelligence Fundamental differences between machine intelligence and human intelligence may permanently limit the scope and applicability of AI. This can be understood through analogies. For example planes can fly, but they aren't a replacement for birds. Similarly, submarines can swim, but they aren't a replacement for fish. Likewise a machine built to mimic human intelligence may never be a perfect replacement of human intelligence or intellect, due to structural incompatibilities and differences (biological organism vs silicon machine). The gap may remain this way for the foreseeable future.
Contemporary AI systems predominantly rely on narrow, domain-specific algorithms trained on vast datasets. They lack the general intelligence and versatility that humans possess, which enables us to learn from experience, apply knowledge across diverse domains, and navigate novel scenarios. The degree to which silicon machines can emulate human capabilities remains uncertain, even if they eventually surpass us in specific areas such as information retrieval, logical reasoning, textual Q&A, analysis, scientific research and discovery.
<To be continued>","title":"Why the Technological Singularity May Be a \"Big Nothing\"","updated_at":"2026-03-05T22:40:17Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"SonicsLegs"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hey guys,
I'm a mid-twenties software engineer who plays a lot of video games in my spare time. Games of choice are mostly FPS titles (e.g. Overwatch, CS:GO) and I tend to be competitive when I play. In case it's relevant, I also have a particularly vicious case of ADHD.
The nature of the games I play make a headset effectively mandatory, as sound cues are particularly important. That being the case, I've noticed that, for some reason that I can't pin down, I tend to perform better overall in-game when I'm not wearing a headset. Specifically, when I don't wear a headset, I:
- Auto-pilot less (I think more about my objectives, how I want to execute, and what I'm doing in general)\n- Tense up less during "stressful situations" (I'm generally overall calmer)\n- Seem to aim/track targets better (likely a result of the above; my mouse arm is more relaxed)\n- Am overall more focused\n- Simply win more games
As a result, I frequently opt to put my headset around my neck and simply turn up the volume, however this is far from ideal.
Why would just wearing a headset impact my performance? Is this a known problem?
At first I thought it was a comfort thing, and tried other headsets to no avail. I'm now leaning towards it being psychological (sound makes the experience more immersive, the increased immersion makes the "virtually stressful" situations actually stressful, and the increased stress impedes my higher-level thinking), however that's just baseless intuition on my part.
I'd really appreciate any insights any of you have. :)"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Why do I perform better in video games when I don't wear a headset?"}},"_tags":["story","author_SonicsLegs","story_18696731","ask_hn"],"author":"SonicsLegs","children":[18696816,18696915,18696948,18698617,18700660,18705546,18721569],"created_at":"2018-12-17T03:29:47Z","created_at_i":1545017387,"num_comments":8,"objectID":"18696731","points":6,"story_id":18696731,"story_text":"Hey guys,
I'm a mid-twenties software engineer who plays a lot of video games in my spare time. Games of choice are mostly FPS titles (e.g. Overwatch, CS:GO) and I tend to be competitive when I play. In case it's relevant, I also have a particularly vicious case of ADHD.
The nature of the games I play make a headset effectively mandatory, as sound cues are particularly important. That being the case, I've noticed that, for some reason that I can't pin down, I tend to perform better overall in-game when I'm not wearing a headset. Specifically, when I don't wear a headset, I:
- Auto-pilot less (I think more about my objectives, how I want to execute, and what I'm doing in general)\n- Tense up less during "stressful situations" (I'm generally overall calmer)\n- Seem to aim/track targets better (likely a result of the above; my mouse arm is more relaxed)\n- Am overall more focused\n- Simply win more games
As a result, I frequently opt to put my headset around my neck and simply turn up the volume, however this is far from ideal.
Why would just wearing a headset impact my performance? Is this a known problem?
At first I thought it was a comfort thing, and tried other headsets to no avail. I'm now leaning towards it being psychological (sound makes the experience more immersive, the increased immersion makes the "virtually stressful" situations actually stressful, and the increased stress impedes my higher-level thinking), however that's just baseless intuition on my part.
I'd really appreciate any insights any of you have. :)","title":"Ask HN: Why do I perform better in video games when I don't wear a headset?","updated_at":"2024-09-20T03:35:28Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"haensi"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"My wife (staying at home) and I have a 2yo girl and a 7mo girl and raise them trilingually (English, Spanish, German).\nTo our surprise, the 2yo has a better pronunciation than monolingual peers.\nShe also has a remarkable memory and can remember names of places and people.\nMoreover, she discovers letters on signs and posters. Her favorite book is a visual dictionary. [3]\nIn general, our daughters can have access to all tools after a short introduction and under supervision.\nScissors? Hold em like this. Kitchen knife? Handle with care. China plates? Careful.
Most people in our neighborhood do not have a strategy on how to raise the children.\nThey rather trust in day care to take care of that or follow their intuition.
I feel a strong need to strategically work out an education.\nSo far, I have been studying the Montessori work, namely 'Montessori, the Science Behind the Genius' [2].
I am interested in the following questions:
1. What educational strategy / syllabus do you have for your children?
2. At what age do you introduce smartphones / tablets for their independent usage? ()
1. What do you think of the WHO guidance [4]?\n\n 2. How much screen time do you think is beneficial / detrimental?\n4. How do you teach technological foundations (STEM) without technology?
5. What activities do you perform outside / in nature? (e.g., Usage of animal guides / plant guides)
6. My daughter sees me working a lot with tablet / laptop and wants to do the same. (She's even 'scheduling meetings'). How would you deal with this?
() I am familiar with a related question from 2014 (Ask HN: How much screen time do you let your kids have?[1]).
---
[1]: https://news.ycombinator.com/item?id=8746045
[2]: https://www.amazon.com/dp/019536936X
[3]: https://www.dk.com/us/book/9781465447562-5-language-visual-dictionary/
[4]: https://www.who.int/news/item/24-04-2019-to-grow-up-healthy-children-need-to-sit-less-and-play-more"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: What is your education and technology strategy for your children?"}},"_tags":["story","author_haensi","story_29105410","ask_hn"],"author":"haensi","children":[29111175],"created_at":"2021-11-04T10:38:54Z","created_at_i":1636022334,"num_comments":1,"objectID":"29105410","points":6,"story_id":29105410,"story_text":"My wife (staying at home) and I have a 2yo girl and a 7mo girl and raise them trilingually (English, Spanish, German).\nTo our surprise, the 2yo has a better pronunciation than monolingual peers.\nShe also has a remarkable memory and can remember names of places and people.\nMoreover, she discovers letters on signs and posters. Her favorite book is a visual dictionary. [3]\nIn general, our daughters can have access to all tools after a short introduction and under supervision.\nScissors? Hold em like this. Kitchen knife? Handle with care. China plates? Careful.
Most people in our neighborhood do not have a strategy on how to raise the children.\nThey rather trust in day care to take care of that or follow their intuition.
I feel a strong need to strategically work out an education.\nSo far, I have been studying the Montessori work, namely 'Montessori, the Science Behind the Genius' [2].
I am interested in the following questions:
1. What educational strategy / syllabus do you have for your children?
2. At what age do you introduce smartphones / tablets for their independent usage? ()
1. What do you think of the WHO guidance [4]?\n\n 2. How much screen time do you think is beneficial / detrimental?\n4. How do you teach technological foundations (STEM) without technology?
5. What activities do you perform outside / in nature? (e.g., Usage of animal guides / plant guides)
6. My daughter sees me working a lot with tablet / laptop and wants to do the same. (She's even 'scheduling meetings'). How would you deal with this?
() I am familiar with a related question from 2014 (Ask HN: How much screen time do you let your kids have?[1]).
---
[1]: https://news.ycombinator.com/item?id=8746045
[2]: https://www.amazon.com/dp/019536936X
[3]: https://www.dk.com/us/book/9781465447562-5-language-visual-dictionary/
[4]: https://www.who.int/news/item/24-04-2019-to-grow-up-healthy-children-need-to-sit-less-and-play-more","title":"Ask HN: What is your education and technology strategy for your children?","updated_at":"2024-09-20T09:43:32Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jacobedawson"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"I'm trying to process a disconnect between the unemployment crisis brought on by the virus, an incoming housing market implosion and a decreasing Index of Consumer Sentiment with the fact that the S&P 500 is up 500 points in the past month - it's on track for the best month since 1974.
I don't understand monetary policy or macro-economics in general, but I'm struggling to grasp the gap between my intuition and reality.
What's the deal?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Why is the S&P 500 rallying in a crisis?"}},"_tags":["story","author_jacobedawson","story_23027824","ask_hn"],"author":"jacobedawson","children":[23028141,23028208,23028393,23030716,23040707],"created_at":"2020-04-30T05:50:15Z","created_at_i":1588225815,"num_comments":7,"objectID":"23027824","points":5,"story_id":23027824,"story_text":"I'm trying to process a disconnect between the unemployment crisis brought on by the virus, an incoming housing market implosion and a decreasing Index of Consumer Sentiment with the fact that the S&P 500 is up 500 points in the past month - it's on track for the best month since 1974.
I don't understand monetary policy or macro-economics in general, but I'm struggling to grasp the gap between my intuition and reality.
What's the deal?","title":"Ask HN: Why is the S&P 500 rallying in a crisis?","updated_at":"2024-09-20T06:02:39Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"boa00"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"I have no idea what to make of IQ in general, and I try to figure it out. Majority of the books/videos I found which at least try to rely on some scientific methods (e.g. research) are predominantly in favor of IQ and its reliability/importance.
Could you recommend any books/articles/videos which try criticize IQ tests from the scientific standpoint? Like tests being biased/unrepresentative, most experiments are not properly constructed and anything along those lines. Most of the arguments against IQ I found though relied mostly on personal experience/intuition (e.g. it\u2019s too simplistic, intelligence is much more complex than that, etc), but I\u2019d like to hear more \u2018measurable\u2019 criticism.
P.S. I\u2019m not trying to prove or disprove anything; I just want to look at this from different points of view"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Any Evidence Against IQ?"}},"_tags":["story","author_boa00","story_27026778","ask_hn"],"author":"boa00","children":[27026805,27027184,27027390,27034246],"created_at":"2021-05-03T16:04:11Z","created_at_i":1620057851,"num_comments":4,"objectID":"27026778","points":4,"story_id":27026778,"story_text":"I have no idea what to make of IQ in general, and I try to figure it out. Majority of the books/videos I found which at least try to rely on some scientific methods (e.g. research) are predominantly in favor of IQ and its reliability/importance.
Could you recommend any books/articles/videos which try criticize IQ tests from the scientific standpoint? Like tests being biased/unrepresentative, most experiments are not properly constructed and anything along those lines. Most of the arguments against IQ I found though relied mostly on personal experience/intuition (e.g. it\u2019s too simplistic, intelligence is much more complex than that, etc), but I\u2019d like to hear more \u2018measurable\u2019 criticism.
P.S. I\u2019m not trying to prove or disprove anything; I just want to look at this from different points of view","title":"Ask HN: Any Evidence Against IQ?","updated_at":"2024-09-20T08:29:31Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mpmpmp"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hello. Currently in the process of looking for new employment, I am curious about the perception in the development community of working at Microsoft versus places like Google, a similarly well-liked software company, or some other fresh, hyped startup.
I understand this is a somewhat shallow question, but employment is closely related to identity and future employment options; and, while I'm experienced in software development, I don't have much experience in the way of socializing in the development community, so I am unfamiliar with the general perceptions on these types of things within much of the SW world.
Let's just assume for a moment one applied to and received offers from Microsoft and Google. My intuition and personal perception is that Microsoft is something of a software dinosaur, bureaucratic and rigid, a bit out of touch with the leading edge of development, whereas places like Google, smaller startups and such are more agile, developer-friendly and forward-thinking.
I have friends who work at these and other similar locations and I reactively tend to esteem the Google employees job more so than those at Microsoft. I'm aware of this personal tendency and it bothers me a bit when considering my own future employment options.
Is this a common feeling, and does it really matter for future career moves?
What is the general opinion in the development world, when meeting someone who works at Microsoft, what's the perception of such a person based on where they work? What's the stereotype? What about Microsoft Research, how does that change the opinion? How does that compare to someone who works at Google?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Working at MS vs. Google: Community opinion/Effect on resume perception?"},"url":{"matchLevel":"none","matchedWords":[],"value":""}},"_tags":["story","author_mpmpmp","story_1884343","ask_hn"],"author":"mpmpmp","children":[1884366],"created_at":"2010-11-08T22:37:59Z","created_at_i":1289255879,"num_comments":4,"objectID":"1884343","points":4,"story_id":1884343,"story_text":"Hello. Currently in the process of looking for new employment, I am curious about the perception in the development community of working at Microsoft versus places like Google, a similarly well-liked software company, or some other fresh, hyped startup.
I understand this is a somewhat shallow question, but employment is closely related to identity and future employment options; and, while I'm experienced in software development, I don't have much experience in the way of socializing in the development community, so I am unfamiliar with the general perceptions on these types of things within much of the SW world.
Let's just assume for a moment one applied to and received offers from Microsoft and Google. My intuition and personal perception is that Microsoft is something of a software dinosaur, bureaucratic and rigid, a bit out of touch with the leading edge of development, whereas places like Google, smaller startups and such are more agile, developer-friendly and forward-thinking.
I have friends who work at these and other similar locations and I reactively tend to esteem the Google employees job more so than those at Microsoft. I'm aware of this personal tendency and it bothers me a bit when considering my own future employment options.
Is this a common feeling, and does it really matter for future career moves?
What is the general opinion in the development world, when meeting someone who works at Microsoft, what's the perception of such a person based on where they work? What's the stereotype? What about Microsoft Research, how does that change the opinion? How does that compare to someone who works at Google?","title":"Ask HN: Working at MS vs. Google: Community opinion/Effect on resume perception?","updated_at":"2024-09-19T17:23:48Z","url":""},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"glenstein"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"http://en.wikipedia.org/wiki/What_Computers_Can't_Do
Full disclosure- my intuition is a strong reaction in the negative but I've only just begun reading his book, and I find it particularly staggering that the introduction frames the debate as open-and-shut, with Dreyfus having won.
Dreyfus's introduction for example spends a lot of time going over the brick walls hit by AI in the mid 20th century. That's compelling for what its worth as it leads to a deeper appreciation for the problems such researchers were trying to solve, but as an argument it's similar to saying that a lack of a man on mars counts as proof that it's impossible for men to get to mars.
Also, as a general note, I find the approach of supplying strictly philosophical arguments suspicious- there is lots of room for vagueness, and for characterizing AI researchers as taking certain philosophical positions without meaningfully engaging with the fruits of their research.
I apologize if that sounds unnecessarily harsh. I do think that I have an obligation to read and seriously mull over these arguments given that I want to disbelieve them so strongly and I'd like to think I'm open to being convinced. So I'm hoping that there is someone in this community who approached this book with a similar mindset who's willing to share their experience with Dreyfus's ideas."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Have you read Dreyfus's What Computers Can't Do? How did you react?"},"url":{"matchLevel":"none","matchedWords":[],"value":""}},"_tags":["story","author_glenstein","story_471802","ask_hn"],"author":"glenstein","created_at":"2009-02-08T04:03:44Z","created_at_i":1234065824,"num_comments":0,"objectID":"471802","points":4,"story_id":471802,"story_text":"http://en.wikipedia.org/wiki/What_Computers_Can't_Do
Full disclosure- my intuition is a strong reaction in the negative but I've only just begun reading his book, and I find it particularly staggering that the introduction frames the debate as open-and-shut, with Dreyfus having won.
Dreyfus's introduction for example spends a lot of time going over the brick walls hit by AI in the mid 20th century. That's compelling for what its worth as it leads to a deeper appreciation for the problems such researchers were trying to solve, but as an argument it's similar to saying that a lack of a man on mars counts as proof that it's impossible for men to get to mars.
Also, as a general note, I find the approach of supplying strictly philosophical arguments suspicious- there is lots of room for vagueness, and for characterizing AI researchers as taking certain philosophical positions without meaningfully engaging with the fruits of their research.
I apologize if that sounds unnecessarily harsh. I do think that I have an obligation to read and seriously mull over these arguments given that I want to disbelieve them so strongly and I'd like to think I'm open to being convinced. So I'm hoping that there is someone in this community who approached this book with a similar mindset who's willing to share their experience with Dreyfus's ideas.","title":"Ask HN: Have you read Dreyfus's What Computers Can't Do? How did you react?","updated_at":"2024-09-19T16:33:26Z","url":""},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"SuKefan"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"I've googled around and haven't been able to find good comparisons of using FastAPI vs a gRPC service in Python. My intuition is that assuming long lived clients gRPC should be faster due to binary serialization of data (since they are both using HTTP2) but I'm surprised that nobody has done a comprehensive comparison post.
I'm interested in comparisons for both performance under heavy load as well as general usability/maintainability thoughts."},"title":{"matchLevel":"none","matchedWords":[],"value":"FastAPI vs. gRPC for high volume Python services"}},"_tags":["story","author_SuKefan","story_23911389","ask_hn"],"author":"SuKefan","created_at":"2020-07-21T21:52:46Z","created_at_i":1595368366,"num_comments":0,"objectID":"23911389","points":4,"story_id":23911389,"story_text":"I've googled around and haven't been able to find good comparisons of using FastAPI vs a gRPC service in Python. My intuition is that assuming long lived clients gRPC should be faster due to binary serialization of data (since they are both using HTTP2) but I'm surprised that nobody has done a comprehensive comparison post.
I'm interested in comparisons for both performance under heavy load as well as general usability/maintainability thoughts.","title":"FastAPI vs. gRPC for high volume Python services","updated_at":"2024-09-20T06:33:33Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kovezd"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hi HN! BI-LLM (https://github.com/amVizion/BI-LLM) automates data analysis workflows on semi-structured data. The assumption is that many decisions could be led by data, but instead are made by intuition because doing analysis would be too slow or expensive. The sweet-spot is decisions based on a few hundreds to a couple thousand items: too many to analyze manually, too few to dedicate an analyst to it. I believe that powerful insights are derived from a large number of data-driven decisions.
There are 3 things special about BI-LLM:
1. Web-native: the code is fully developed on TypeScript. A native integration with web, avoids the need for specialized software, or licenses. More importantly, I hope it contributes opening the field of AI to a wider group of developers.
2. Runs-locally: BI-LLM uses a combination of statistical, machine-learning, and GenAI methods to automate the analysis. Embedding models run in TensorflowJS, and LLMs in Ollama. Running locally not only helps keeping costs down, it also contributes to data privacy, and security.
3. No data-background required: all the technical details when pre-processing data are hidden on the implementation. The analysis only requires a simple JSON file, and after a couple CLI commands, it outputs a full report with the conclusions in a language that directly leads to making decisions.
The data analysis flow starts by turning texts into embeddings, and, using LLMs to label, and score the texts. Then, texts are grouped by similarity and the scores used to write an analysis describing each cluster, and explaining the correlation between labels, and attributes in the final report. You can see in YouTube a 2-minute demo (https://youtu.be/RoLU_REypyY) of a sample analysis. In GitHub, the demo directory also has screenshots from a sample analysis. You can also find diagrams with more detailed explanations of the steps taken during analysis on the ReadME.
The project was partially inspired by feedback from the community. A few months ago I posted an analysis that stirred some controversy about whether it was written by AI. There were some encouraging comments including:
> This is the perfect use case for LLMs and Data Analysis in general. I'd legitimately pay money for a similar productivity tool.
The project took shape after speaking with an investor that had a perfect use case for it. The key insight was to move from doing unsupervised analysis (as the original article) to supervised analysis: attempting to predict a defined outcome. The final straw was when exploring some leads data, I clustered the data and so differences in the rate of conversion between 20% to 2%. The analysis required almost no work, and provided useful insights on how to identify the best prospects. I figure it out this could be a great tool for startups, and small teams that cannot afford a data analyst.
I would love to hear your feedback. Thank you!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: BI-LLM democratize data analysis"}},"_tags":["story","author_kovezd","story_41285543","show_hn"],"author":"kovezd","created_at":"2024-08-18T21:11:39Z","created_at_i":1724015499,"num_comments":0,"objectID":"41285543","points":3,"story_id":41285543,"story_text":"Hi HN! BI-LLM (https://github.com/amVizion/BI-LLM) automates data analysis workflows on semi-structured data. The assumption is that many decisions could be led by data, but instead are made by intuition because doing analysis would be too slow or expensive. The sweet-spot is decisions based on a few hundreds to a couple thousand items: too many to analyze manually, too few to dedicate an analyst to it. I believe that powerful insights are derived from a large number of data-driven decisions.
There are 3 things special about BI-LLM:
1. Web-native: the code is fully developed on TypeScript. A native integration with web, avoids the need for specialized software, or licenses. More importantly, I hope it contributes opening the field of AI to a wider group of developers.
2. Runs-locally: BI-LLM uses a combination of statistical, machine-learning, and GenAI methods to automate the analysis. Embedding models run in TensorflowJS, and LLMs in Ollama. Running locally not only helps keeping costs down, it also contributes to data privacy, and security.
3. No data-background required: all the technical details when pre-processing data are hidden on the implementation. The analysis only requires a simple JSON file, and after a couple CLI commands, it outputs a full report with the conclusions in a language that directly leads to making decisions.
The data analysis flow starts by turning texts into embeddings, and, using LLMs to label, and score the texts. Then, texts are grouped by similarity and the scores used to write an analysis describing each cluster, and explaining the correlation between labels, and attributes in the final report. You can see in YouTube a 2-minute demo (https://youtu.be/RoLU_REypyY) of a sample analysis. In GitHub, the demo directory also has screenshots from a sample analysis. You can also find diagrams with more detailed explanations of the steps taken during analysis on the ReadME.
The project was partially inspired by feedback from the community. A few months ago I posted an analysis that stirred some controversy about whether it was written by AI. There were some encouraging comments including:
> This is the perfect use case for LLMs and Data Analysis in general. I'd legitimately pay money for a similar productivity tool.
The project took shape after speaking with an investor that had a perfect use case for it. The key insight was to move from doing unsupervised analysis (as the original article) to supervised analysis: attempting to predict a defined outcome. The final straw was when exploring some leads data, I clustered the data and so differences in the rate of conversion between 20% to 2%. The analysis required almost no work, and provided useful insights on how to identify the best prospects. I figure it out this could be a great tool for startups, and small teams that cannot afford a data analyst.
I would love to hear your feedback. Thank you!","title":"Show HN: BI-LLM democratize data analysis","updated_at":"2024-09-20T17:42:21Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"brianllamar"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Last week I sat in the LLM Paper Club (hosted by the latent.space podcast). We shared a paper from DeepMind on "Discovering Multiagent Learning Algorithms with Large Language Models". Linked in the github repo.
This was my first time sitting in one of these and reading an LLM paper. I thought it was interesting that the AlphaEvolve agent in the paper essentially writes code, runs it, scores the output, and then writes better code, over and over. It's not discovering strategies through play. It's discovering algorithms through code mutation. The LLM proposes changes to how regret is accumulated or how policies are derived, a fitness function scores the result, and the best variants survive to the next generation.
The two algorithms it found (VAD-CFR and SHOR-PSRO) use mechanisms the authors describe as "non-intuitive," things like volatility-sensitive discounting and hard warm-start schedules that a human designer probably wouldn't have tried. That's the interesting part: the LLM isn't constrained by the same design intuitions we are.
To make it concrete for myself, I built a small version of this loop for a Pokemon game agent. The setup is simple: define a fitness function (turns survived, maps visited, stuck events), parameterize the strategy space (door cooldown, stuck threshold, skip distance), and let an LLM propose variants that get evaluated in parallel. Ten agents race through the game, the best parameters survive. I used tapes.dev to collect session telemetry and feed observational memory back into the fitness scoring.
The first run already surfaced something useful: shorter door cooldowns (4 vs 8) reduce stuck events from 16 to 9. Not a breakthrough, but the point is the system found it without me guessing. That's the same dynamic as the paper, just at toy scale.
What I took away from the paper club: the bottleneck in algorithm design isn't computation, it's the search process itself. If you can express your problem as "parameterized code + fitness function," an LLM evolution loop can explore the space faster than manual iteration. The paper proves it works for game theory. The link is in my startup's repo, I want to explore applying this technique for improving future general purpose and coding agent sessions.
Just pointing out Pokemon aren't the only thing evolving in that repo."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: AlphaEvolve inspired evolution harness for Pokemon"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/papercomputeco/pokemon"}},"_tags":["story","author_brianllamar","story_47327720","show_hn"],"author":"brianllamar","created_at":"2026-03-10T19:24:34Z","created_at_i":1773170674,"num_comments":0,"objectID":"47327720","points":2,"story_id":47327720,"story_text":"Last week I sat in the LLM Paper Club (hosted by the latent.space podcast). We shared a paper from DeepMind on "Discovering Multiagent Learning Algorithms with Large Language Models". Linked in the github repo.
This was my first time sitting in one of these and reading an LLM paper. I thought it was interesting that the AlphaEvolve agent in the paper essentially writes code, runs it, scores the output, and then writes better code, over and over. It's not discovering strategies through play. It's discovering algorithms through code mutation. The LLM proposes changes to how regret is accumulated or how policies are derived, a fitness function scores the result, and the best variants survive to the next generation.
The two algorithms it found (VAD-CFR and SHOR-PSRO) use mechanisms the authors describe as "non-intuitive," things like volatility-sensitive discounting and hard warm-start schedules that a human designer probably wouldn't have tried. That's the interesting part: the LLM isn't constrained by the same design intuitions we are.
To make it concrete for myself, I built a small version of this loop for a Pokemon game agent. The setup is simple: define a fitness function (turns survived, maps visited, stuck events), parameterize the strategy space (door cooldown, stuck threshold, skip distance), and let an LLM propose variants that get evaluated in parallel. Ten agents race through the game, the best parameters survive. I used tapes.dev to collect session telemetry and feed observational memory back into the fitness scoring.
The first run already surfaced something useful: shorter door cooldowns (4 vs 8) reduce stuck events from 16 to 9. Not a breakthrough, but the point is the system found it without me guessing. That's the same dynamic as the paper, just at toy scale.
What I took away from the paper club: the bottleneck in algorithm design isn't computation, it's the search process itself. If you can express your problem as "parameterized code + fitness function," an LLM evolution loop can explore the space faster than manual iteration. The paper proves it works for game theory. The link is in my startup's repo, I want to explore applying this technique for improving future general purpose and coding agent sessions.
Just pointing out Pokemon aren't the only thing evolving in that repo.","title":"Show HN: AlphaEvolve inspired evolution harness for Pokemon","updated_at":"2026-03-10T19:28:29Z","url":"https://github.com/papercomputeco/pokemon"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Venky1729"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hey Folks, I've been building an MCP server that generates beautiful Excalidraw architecture diagrams fully auto-laid-out, no manual positioning. It's open source, works with any AI IDE (Cursor, Windsurf, etc.).I wanted to share it early because the core layout engine is solid and already producing clean diagrams.
What Problem Am I Solving?\nAI IDEs generate architecture diagrams as Mermaid or ASCII art. When they attempt Excalidraw, they hallucinate coordinates and boxes overlap, arrows cross through nodes, and you spend more time fixing the diagram than drawing it yourself. LLMs understand what a system looks like, but they have zero intuition for where things go on a 2D canvas.
How Does This MCP Solve It?\nYou describe the components and connections. The MCP runs a Sugiyama hierarchical layout algorithm to compute positions deterministically -- the AI never touches coordinates. It auto-styles 50+ technologies (say "Kafka" and get a stream-styled node), stretches hub nodes like API Gateways to span their services, and routes arrows around obstacles. The result is a clean .excalidraw file in seconds, with stateful editing so you can say "add a Redis cache" without starting over.
Why Not the Official excalidraw-mcp?\nThe official MCP gives the AI a raw canvas and lets it place elements freely -- great for general sketching, but architecture diagrams still end up with overlapping boxes. excalidraw-architect-mcp takes a structured graph as input and runs a proper layout algorithm, so every diagram is overlap-free by design. It also runs fully offline in your IDE with no API keys, outputs version-controllable .excalidraw files, and includes architecture-aware styling that the general-purpose tool doesn't have.
What's the Future?\nDiagram-from-code (point at a codebase, get an architecture diagram automatically), live sync so diagrams update as code evolves, richer layouts (radial, swimlane, grid), and an expanded component library covering cloud-provider icons and CI/CD tools. The goal: architecture diagrams as a byproduct of building software, not a separate task."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Excalidraw Architect MCP for AI Based IDEs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/BV-Venky/excalidraw-architect-mcp"}},"_tags":["story","author_Venky1729","story_47289785","show_hn"],"author":"Venky1729","created_at":"2026-03-07T17:47:49Z","created_at_i":1772905669,"num_comments":0,"objectID":"47289785","points":2,"story_id":47289785,"story_text":"Hey Folks, I've been building an MCP server that generates beautiful Excalidraw architecture diagrams fully auto-laid-out, no manual positioning. It's open source, works with any AI IDE (Cursor, Windsurf, etc.).I wanted to share it early because the core layout engine is solid and already producing clean diagrams.
What Problem Am I Solving?\nAI IDEs generate architecture diagrams as Mermaid or ASCII art. When they attempt Excalidraw, they hallucinate coordinates and boxes overlap, arrows cross through nodes, and you spend more time fixing the diagram than drawing it yourself. LLMs understand what a system looks like, but they have zero intuition for where things go on a 2D canvas.
How Does This MCP Solve It?\nYou describe the components and connections. The MCP runs a Sugiyama hierarchical layout algorithm to compute positions deterministically -- the AI never touches coordinates. It auto-styles 50+ technologies (say "Kafka" and get a stream-styled node), stretches hub nodes like API Gateways to span their services, and routes arrows around obstacles. The result is a clean .excalidraw file in seconds, with stateful editing so you can say "add a Redis cache" without starting over.
Why Not the Official excalidraw-mcp?\nThe official MCP gives the AI a raw canvas and lets it place elements freely -- great for general sketching, but architecture diagrams still end up with overlapping boxes. excalidraw-architect-mcp takes a structured graph as input and runs a proper layout algorithm, so every diagram is overlap-free by design. It also runs fully offline in your IDE with no API keys, outputs version-controllable .excalidraw files, and includes architecture-aware styling that the general-purpose tool doesn't have.
What's the Future?\nDiagram-from-code (point at a codebase, get an architecture diagram automatically), live sync so diagrams update as code evolves, richer layouts (radial, swimlane, grid), and an expanded component library covering cloud-provider icons and CI/CD tools. The goal: architecture diagrams as a byproduct of building software, not a separate task.","title":"Show HN: Excalidraw Architect MCP for AI Based IDEs","updated_at":"2026-03-07T17:57:03Z","url":"https://github.com/BV-Venky/excalidraw-architect-mcp"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"anifow"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"The Aakkozzll app is a substitute for throwing dice. Players can see the probability distribution on each throw (unlike dice which have a similar pattern, but you just can't see it).
The app is great for high-school to college level stats and data-management classes, and simple enough to serve as a primer for elementary-aged students. It gives them a physical / visual way to appreciate the normal curve. Can be used in a classroom activity of Craps or even monopoly, using the Aakkozzll instead of dice. Games like Craps will take on a whole new meaning when students can actually see the chances of getting a 7.
Will give students an intuition for the binomial distribution, frequency graphs, and the normal curve. As the balls fall down, they randomly go left or right (its about 50/50) depending on how they land. Its like having a series of 50/50 coin tosses which determines the final position of the balls. Only the red ball actually determines your \"roll\" (it blinks red).
The original inspiration came from the Galton board, a physical device that was built 150 years ago to demonstrate the central tendency. We've recreated the physics of that device into an easy to use iPhone app so that it's insights can be shared with the world. History of the Bean Machine: (http://en.wikipedia.org/wiki/Bean_machine)
Why the funny name? The Aakkozzll (pronounced \"acausal\") demonstrates a paradox that although the result of any one individual cannot be predicted, the general pattern can be. This extends to the real world. The fact that a person falls into poverty is not entirely predictable, but the fact that a society will have a certain distribution of rich and poor (regardless of their particular paths towards riches and poverty) is quite constant, especially when you have a large population.
Educators, we encourage you to share this with your peers!"},"title":{"matchLevel":"none","matchedWords":[],"value":"New iPhone App encourages Classroom Activity for Stats/Math"},"url":{"matchLevel":"none","matchedWords":[],"value":"http://itunes.apple.com/us/app/aakkozzll/id561069750?mt=8 "}},"_tags":["story","author_anifow","story_4571946"],"author":"anifow","created_at":"2012-09-25T18:42:37Z","created_at_i":1348598557,"num_comments":0,"objectID":"4571946","points":2,"story_id":4571946,"story_text":"The Aakkozzll app is a substitute for throwing dice. Players can see the probability distribution on each throw (unlike dice which have a similar pattern, but you just can't see it).
The app is great for high-school to college level stats and data-management classes, and simple enough to serve as a primer for elementary-aged students. It gives them a physical / visual way to appreciate the normal curve. Can be used in a classroom activity of Craps or even monopoly, using the Aakkozzll instead of dice. Games like Craps will take on a whole new meaning when students can actually see the chances of getting a 7.
Will give students an intuition for the binomial distribution, frequency graphs, and the normal curve. As the balls fall down, they randomly go left or right (its about 50/50) depending on how they land. Its like having a series of 50/50 coin tosses which determines the final position of the balls. Only the red ball actually determines your \"roll\" (it blinks red).
The original inspiration came from the Galton board, a physical device that was built 150 years ago to demonstrate the central tendency. We've recreated the physics of that device into an easy to use iPhone app so that it's insights can be shared with the world. History of the Bean Machine: (http://en.wikipedia.org/wiki/Bean_machine)
Why the funny name? The Aakkozzll (pronounced \"acausal\") demonstrates a paradox that although the result of any one individual cannot be predicted, the general pattern can be. This extends to the real world. The fact that a person falls into poverty is not entirely predictable, but the fact that a society will have a certain distribution of rich and poor (regardless of their particular paths towards riches and poverty) is quite constant, especially when you have a large population.
Educators, we encourage you to share this with your peers!","title":"New iPhone App encourages Classroom Activity for Stats/Math","updated_at":"2024-09-19T18:58:38Z","url":"http://itunes.apple.com/us/app/aakkozzll/id561069750?mt=8 "},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"zerojames"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Evaluating visual capabilities of language models is hard.
On the one end of the evaluation spectrum, we have vibe checks which, while useful for building intuition, are time-consuming to run across a dozen or more models. On the other end, we have large benchmarks which are so large that they are intractable to most users.
Vision AI Checkup is a new tool for evaluating VLMs. The site is made up of hand-crafted prompts focused on real-world problems: defect detection, understanding how the position of one object relates to another, colour understanding, and more.
Our prompts are especially focused on industrial tasks -- serial number reading, assembly line understanding, and more -- although we're excited to add more general prompts.
The tool lets you see how models do across categories of prompts, and how different models do on a single prompt.
We have open sourced the codebase, with instructions on how to add a prompt to the assessment: https://github.com/roboflow/vision-ai-checkup. You can also add new models.
We'd love feedback and, also, ideas for areas where VLMs struggle that you'd like to see assessed!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Vision AI Checkup, an Optometrist for VLMs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://visioncheckup.com"}},"_tags":["story","author_zerojames","story_43971984","show_hn"],"author":"zerojames","created_at":"2025-05-13T12:08:02Z","created_at_i":1747138082,"num_comments":0,"objectID":"43971984","points":2,"story_id":43971984,"story_text":"Evaluating visual capabilities of language models is hard.
On the one end of the evaluation spectrum, we have vibe checks which, while useful for building intuition, are time-consuming to run across a dozen or more models. On the other end, we have large benchmarks which are so large that they are intractable to most users.
Vision AI Checkup is a new tool for evaluating VLMs. The site is made up of hand-crafted prompts focused on real-world problems: defect detection, understanding how the position of one object relates to another, colour understanding, and more.
Our prompts are especially focused on industrial tasks -- serial number reading, assembly line understanding, and more -- although we're excited to add more general prompts.
The tool lets you see how models do across categories of prompts, and how different models do on a single prompt.
We have open sourced the codebase, with instructions on how to add a prompt to the assessment: https://github.com/roboflow/vision-ai-checkup. You can also add new models.
We'd love feedback and, also, ideas for areas where VLMs struggle that you'd like to see assessed!","title":"Show HN: Vision AI Checkup, an Optometrist for VLMs","updated_at":"2025-05-14T12:49:11Z","url":"https://visioncheckup.com"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Nwt"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Problem:
We will probably create AGI before we have a formal model of human ethics, which is not an urgent problem but a potentially catastrophic one.
Proposed solution:
We assume we want that AGI (Artificial general intelligence) to be aligned with all life and we want it to produce maximum "good". We then define "good" as whatever any unit of life defines as such. (Interactions between units will produce additional emergent definitions.) Then instead of optimizing for any of those definitions we set AGI's utility function as maximization of the definitions, that is: maximization of the quantity of both units of life and interactions between units, and then to make it all safe we forbid it from interacting with either. Maximizing the quantity of life without interacting with it means both shielding life from cataclysmic events and creating space for life's expansion into places where no life exists (outer space).
More details: https://pni.ai/future
My intuition is that this can be expressed formally, simulated and proven/disproven and if proven may lead to a future with potential for growth that is hard to imagine, yet without any risk to life and without making humans obsolete. There is also no benefit in being the first to create such AGI and all the reasons to cooperate as it produces universal "good".
Can you guys poke holes in this?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Possible solution to AGI alignment and safety?"}},"_tags":["story","author_Nwt","story_23249504","ask_hn"],"author":"Nwt","created_at":"2020-05-20T17:31:38Z","created_at_i":1589995898,"num_comments":0,"objectID":"23249504","points":2,"story_id":23249504,"story_text":"Problem:
We will probably create AGI before we have a formal model of human ethics, which is not an urgent problem but a potentially catastrophic one.
Proposed solution:
We assume we want that AGI (Artificial general intelligence) to be aligned with all life and we want it to produce maximum "good". We then define "good" as whatever any unit of life defines as such. (Interactions between units will produce additional emergent definitions.) Then instead of optimizing for any of those definitions we set AGI's utility function as maximization of the definitions, that is: maximization of the quantity of both units of life and interactions between units, and then to make it all safe we forbid it from interacting with either. Maximizing the quantity of life without interacting with it means both shielding life from cataclysmic events and creating space for life's expansion into places where no life exists (outer space).
More details: https://pni.ai/future
My intuition is that this can be expressed formally, simulated and proven/disproven and if proven may lead to a future with potential for growth that is hard to imagine, yet without any risk to life and without making humans obsolete. There is also no benefit in being the first to create such AGI and all the reasons to cooperate as it produces universal "good".
Can you guys poke holes in this?","title":"Ask HN: Possible solution to AGI alignment and safety?","updated_at":"2024-09-20T06:14:23Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vdupras"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"First of all: this post isn't about the meaning of "Elon is Snowball". This is an example. No politics underneath.
So, a few minutes ago, I google this phrase to see if someone else out there has the same intuition. Then, I'm surprised by Google's AI overview:
> The statement "Elon is snowball" is a misunderstanding. Elon is likely referring to Elon Musk [...]. The term "snowball" typically refers to a phenomenon, often used metaphorically to describe something growing rapidly in size or importance. [...]
So, the AI despite all its knowledge, can't draw the parallel at all. The question isn't whether the parallel is valid or not, but I think it's obvious to any human that has read a little bit that one could see a parallel.
I don't want to "spoil" the example by being too blunt about what the parallel is, to give AI experts the time to query their favorite models before they index this post and have it too easy, so I intentionally remain vague.
I'm not a AI expert and I frankly don't spend much time thinking about AI in general, but given how intensely AI capabilities are praised all around, I'm rather surprised by this result.
So, my question to HN: Why isn't Google's AI capable of even seeing the link? Is it just their model, or is it something fundamental to LLMs? How long until LLMs can make this link all by themselves?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Can current AI models grok \"Elon is Snowball\"?"}},"_tags":["story","author_vdupras","story_44193801","ask_hn"],"author":"vdupras","children":[44193845,44193861,44193967,44194071,44194083,44194114,44194493,44201379],"created_at":"2025-06-05T17:28:52Z","created_at_i":1749144532,"num_comments":16,"objectID":"44193801","points":1,"story_id":44193801,"story_text":"First of all: this post isn't about the meaning of "Elon is Snowball". This is an example. No politics underneath.
So, a few minutes ago, I google this phrase to see if someone else out there has the same intuition. Then, I'm surprised by Google's AI overview:
> The statement "Elon is snowball" is a misunderstanding. Elon is likely referring to Elon Musk [...]. The term "snowball" typically refers to a phenomenon, often used metaphorically to describe something growing rapidly in size or importance. [...]
So, the AI despite all its knowledge, can't draw the parallel at all. The question isn't whether the parallel is valid or not, but I think it's obvious to any human that has read a little bit that one could see a parallel.
I don't want to "spoil" the example by being too blunt about what the parallel is, to give AI experts the time to query their favorite models before they index this post and have it too easy, so I intentionally remain vague.
I'm not a AI expert and I frankly don't spend much time thinking about AI in general, but given how intensely AI capabilities are praised all around, I'm rather surprised by this result.
So, my question to HN: Why isn't Google's AI capable of even seeing the link? Is it just their model, or is it something fundamental to LLMs? How long until LLMs can make this link all by themselves?","title":"Ask HN: Can current AI models grok \"Elon is Snowball\"?","updated_at":"2025-06-07T21:16:31Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"SeanAnderson"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"I'm going to be taking some time off from work here at the end of the year to pursue some creative software development.
Although I have almost 15 years of industry experience I've never built a game. Up until about six months ago, I didn't think I was /that/ type of software developer, but now, here I am, interest piqued, reading up on all sorts of technology stacks and having my mind blown by the size of the knowledge space. I am especially excited to try and express logic using an ECS/DOTS approach!
I know I am building for desktop web. Leaning towards Rust + WebGPU and hoping to time an MVP around WebGPU general availability. I know I shouldn't be in this space for the money. I know to expect game programming to be significantly more challenging than whatever my intuition believes and that I should continually and aggressively prune away at scope creep. I know there's a LOT of math and that performance is frequently an issue.
What are some joys, or challenges, you unexpectedly encountered when developing your first game which you wish you'd known before breaking ground?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: What do you wish you knew before you developed your first video game?"}},"_tags":["story","author_SeanAnderson","story_33801959","ask_hn"],"author":"SeanAnderson","children":[33801974,33802077],"created_at":"2022-11-30T15:23:18Z","created_at_i":1669821798,"num_comments":4,"objectID":"33801959","points":1,"story_id":33801959,"story_text":"I'm going to be taking some time off from work here at the end of the year to pursue some creative software development.
Although I have almost 15 years of industry experience I've never built a game. Up until about six months ago, I didn't think I was /that/ type of software developer, but now, here I am, interest piqued, reading up on all sorts of technology stacks and having my mind blown by the size of the knowledge space. I am especially excited to try and express logic using an ECS/DOTS approach!
I know I am building for desktop web. Leaning towards Rust + WebGPU and hoping to time an MVP around WebGPU general availability. I know I shouldn't be in this space for the money. I know to expect game programming to be significantly more challenging than whatever my intuition believes and that I should continually and aggressively prune away at scope creep. I know there's a LOT of math and that performance is frequently an issue.
What are some joys, or challenges, you unexpectedly encountered when developing your first game which you wish you'd known before breaking ground?","title":"Ask HN: What do you wish you knew before you developed your first video game?","updated_at":"2024-09-20T12:34:18Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nDot_io"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hi HN,
I\u2019m working on a \u201chacker science\u201d experiment called Ai_home.\nIt\u2019s a cognitive architecture prototype that I designed to explore the current limits of LLMs in terms of persistent identity, long-term memory, and autonomy.
The system is not just a simple chatbot loop, but a multi-threaded architecture:
1. Worker: Handles user interactions and tool use.\n2. Monologue: A background \u201csubconscious\u201d thread that analyzes context and logs intuitions/tips for the Worker.\n3. Memory: Manages vector-based long-term memory (Postgres + pgvector) with emotional weighting.\n4. Mind: This layer is responsible for deeper interpretation of messages and for exploring creative alternatives.
Because of this, it\u2019s not a synchronous question\u2013answer chatbot.\nThe model and the user (the Helper) can communicate in parallel, and the Worker processes this asynchronously.
Technical details:
- Hybrid Multi-LLM: I combine multiple models (Gemini, GPT-4, Groq). I use different models for creative idea generation (\u201ccreative\u201d) and for logical processing (\u201cinterpreter\u201d).
- Modes: I don\u2019t use a single context window. Depending on the operating mode (General, Developer, Analyst), I partition messages into separate contexts. I\u2019ve introduced a transition process between mode switches to ensure that the essential information is preserved across contexts.
- Dynamic Prompt: Based on memories and accumulated experience, I dynamically modify the prompt on every API call so that each conversation can gain a fresh contextual interpretation.
- Incubator: The system has an experimental environment where it can attempt to refactor its own code. The results are mixed so far, but it\u2019s fascinating to watch a model interpret its own code.
- Identity and Laws: For building identity, the system has a \u201cconstitution\u201d (fundamental laws) and tools for modifying them. The content and structure of this are still an active area of experimentation.
Disclaimer:\nThis is an architectural experiment to investigate whether functional patterns of consciousness (global workspace, recurrence) can be mimicked with LLMs in order to create more reliable agents.
I explicitly do not claim that the system is sentient, nor that this is a formal academic research project (We don\u2019t have the personnel or infrastructure for that).
I\u2019m looking for collaborators not only for coding, but also to help define a development methodology for this open, collaborative experimental project.
All feedback is very welcome!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Cognitive AI architecture prototype with identity, memory, initiative"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://ivanhonis.github.io/ai_home/"}},"_tags":["story","author_nDot_io","story_46097698","show_hn"],"author":"nDot_io","children":[46098146],"created_at":"2025-11-30T15:59:42Z","created_at_i":1764518382,"num_comments":2,"objectID":"46097698","points":1,"story_id":46097698,"story_text":"Hi HN,
I\u2019m working on a \u201chacker science\u201d experiment called Ai_home.\nIt\u2019s a cognitive architecture prototype that I designed to explore the current limits of LLMs in terms of persistent identity, long-term memory, and autonomy.
The system is not just a simple chatbot loop, but a multi-threaded architecture:
1. Worker: Handles user interactions and tool use.\n2. Monologue: A background \u201csubconscious\u201d thread that analyzes context and logs intuitions/tips for the Worker.\n3. Memory: Manages vector-based long-term memory (Postgres + pgvector) with emotional weighting.\n4. Mind: This layer is responsible for deeper interpretation of messages and for exploring creative alternatives.
Because of this, it\u2019s not a synchronous question\u2013answer chatbot.\nThe model and the user (the Helper) can communicate in parallel, and the Worker processes this asynchronously.
Technical details:
- Hybrid Multi-LLM: I combine multiple models (Gemini, GPT-4, Groq). I use different models for creative idea generation (\u201ccreative\u201d) and for logical processing (\u201cinterpreter\u201d).
- Modes: I don\u2019t use a single context window. Depending on the operating mode (General, Developer, Analyst), I partition messages into separate contexts. I\u2019ve introduced a transition process between mode switches to ensure that the essential information is preserved across contexts.
- Dynamic Prompt: Based on memories and accumulated experience, I dynamically modify the prompt on every API call so that each conversation can gain a fresh contextual interpretation.
- Incubator: The system has an experimental environment where it can attempt to refactor its own code. The results are mixed so far, but it\u2019s fascinating to watch a model interpret its own code.
- Identity and Laws: For building identity, the system has a \u201cconstitution\u201d (fundamental laws) and tools for modifying them. The content and structure of this are still an active area of experimentation.
Disclaimer:\nThis is an architectural experiment to investigate whether functional patterns of consciousness (global workspace, recurrence) can be mimicked with LLMs in order to create more reliable agents.
I explicitly do not claim that the system is sentient, nor that this is a formal academic research project (We don\u2019t have the personnel or infrastructure for that).
I\u2019m looking for collaborators not only for coding, but also to help define a development methodology for this open, collaborative experimental project.
All feedback is very welcome!","title":"Show HN: Cognitive AI architecture prototype with identity, memory, initiative","updated_at":"2026-03-05T23:10:49Z","url":"https://ivanhonis.github.io/ai_home/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"anchority"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"DECONSTRUCTING THE MYTH ABOUT VENTURE RISK: A WAY OUT OF DILUTION
VCs are just great. They are there waiting at the finish-line to see startups through the race. Angels are so graceful. They are the proverbial friend who saw the pulloff from the beginning, and even went down with the buddy to jail saying that was darn good! Whatever\u2019s wrong with the gap between money and ideas, it never should play us all dead about the natural forces of the market and what exactly we stand for in the marketplace.
Since Joe left his fine job to join Dick, fresh out of college and so stubborn for not taking that big offer from General Everything, to turn a question into a startup; someone reading their story later in time, like their families and friends now, should see that they were making or breaking it. Stupid or brilliant, only time can heal what reason can\u2019t.
They met an angel through another friend. He falls in love and tears his check leaves. The prototype goes live and the market is made. That wonderful thingumajig is sold for $3 apiece. Where no taxes apply because of tax holiday for startups in the state they live in, all the revenues amount to $450, 000. Figuring out the total cost of production of $65, 000, they are left with $385, 000 in profit.
If only they had paid back in 1 \u00bd years after company launch the $80,000 lent by the angel, they would not have had to convert the loan to 20% shareholding. According to their agreement, a share unit is worth two dollars at pre-money valuation of $800, 000. In fact, the work-out posits that the effort of each of Joe and Dick is worth twice the amount invested by the angel. Therefore, each of them has invested $160, 000 without handing in any money save for all-important brains and brawns.
$160, 000 of Joe + $160, 000 of Dick + $80, 000 of the angel = $400, 000 total investment in the startup. In the same proportion, the goodwill accruing to each of them amounts to an equal $400, 000. Because there\u2019s no established custom or popularity, the goodwill should not count now, but they figure that VCs might come calling once the promising product is launched.
Besides, they feel that they should be really compensated for risking spending and exerting so much means and pain on proof of concept delivered to two such clients without any remuneration and contractual obligation to buy. One of them was provided by the angel.
Transport, everyday office rituals, sleepless nights, involuntary fasting, costly PR and shoestring marketing went with the whole gamut. They played deep for a great amount of work effort and thus, the longer time before they could avoid the 1 \u00bd year loan clause. Just why goodwill came a fair deal.
Press rave is here about the product. Pop culture and the TV people are dying for attention from Joe and Dick. They are reminded by the angel that they could let out some steam motion, yet they should not abandon substance for vanity. They are always back at work. Giving attention to events and the people that matter at every point in time.
Looking at the future before them, sound reasoning points in the direction of more important product features, valuable employees, necessary Intellectual Property tying-here-and-there, more solid collaboration, research and development, due diligence, etc.
It all goes from making the market to meeting the demands of the exploding market. The VCs are getting in touch with one another about this new high-tech star. The shepherd and the sheep are running to graze the bounty grass. By now, everyone in town knows how much more Joe and Dick figure the growing company needs in the way of money.
Before sales, if the share price was $2, the share price after sales takes after this picture:
$385, 000 (profit after tax: recall no taxes apply in this instance)/ $800, 000 worth of existing shares in the company = $0.48125, what is called the market quotient.
Add the pre-sales share price of $2 to the market quotient of $0.48125. And you get $2.48125 for the new share price. As simple as that. This is the market price, and it represents the least rational price any new investor should be paying to own a piece of the company.
In that wise, because the quotient could have been negative as a result of loss instead of profit, such a market price should serve as the lowest extreme of the share price band. If it was so a loss, the same formula applies and the market price should possibly be the highest extreme. And in both scenarios, whatever might count for premium and added value between buyers and sellers of the company\u2019s shares should be taken into consideration in each unique trading case.
Already, some readers would say that the allegory is demonstrating the reality on stock market floors (IPO or trading publicly quoted stocks), or common everyday buying and selling of anything founded on value. In every form of enterprise, startup investing, shipbuilding, fashion house, electric car manufacturing or trivia gamemaking, it indispensably counts to let reward and risk take their natural place in the scheme of things.
We never should attempt to trick the market, if we don\u2019t want to see another market crash or an unfortunate founder suicide. The market should decide how much reward goes to any risk-taker. From above, if any VC would invest in the hypothetical startup, he should be paying at least $2.48125 and nothing less. Doing otherwise is tricking the market, which is the prevalent culture of Series A funding in Silicon Valley.
His later involvement with the company should reward the angel and the founders for all the risks they\u2019ve taken before now: the VC should stop tearing apart previously existing term sheets and creating new ones for the sake of importunely favoring himself and the fund management people he works for. His attitude is mostly driven by irrational fear.
There is no power in arbitrariness. Even if it exists anywhere, such power function never lasts. That\u2019s easily evident in Third World dictatorships and command line communist states. Being in need of more money to run a startup shouldn\u2019t make VCs twist and wriggle angels and founders because he wants to protect investments from risk.
The only optimal and efficient way of protecting investments from risk is identifying its intrinsic value before time, and taking the well-angled dive into the startup just early enough. That\u2019s what separates the best from the worst in managing funds for startup investing.
That viewpoint already puts the angels at vantage position. It explains why nowadays the species of superangels are evolving in the ecosystem of startup investing. Some VCs are realizing that \u201cearly enough means most profitably enough\u201d. Jump on it before time asks for a greater price.
Such VCs know that the absolute power the VC community wield now would not last long. Not in the face of a whole lot of market-determined eventualities: far cheaper servers, ubiquitous open-source tools and libertarian programming languages, to mention just a sliver. Dead seriously, running a startup to profit is increasingly a function of how far along determination gets and an efficient combination of these consequences.
Merely observing the market rules and staying fair to the pricing process, is way enough to autocorrect the market in a way that the levers of demand and supply are made to naturally belong to the right sides of the marketplace. The good ones would remain, and the bad ones would be gone.
Matter-of-factly, to have more of the company, using the vast warchest in his possession, the VC should just buy more units of the shareholding at the market price. It\u2019s the price he pays for coming late to the jamboree, and might as well be his deep goblet when General Everything starts an acquisition party or the stock market invites the startup to the IPO frenzy.
If startup founders would only concentrate on creating a product the market is dying to have, the cost of funding would increasingly lower out; since such an important market factor would signal the early value investors to invest when it\u2019s so cheap to do so.
As a result, should the law of large numbers prevail; as bringing about a fairly good count of value-product startups backed by a broad population of early value investors; stock dilution caused by irrational fear would be eliminated.
That\u2019s because more would-be investors would see that it\u2019s just the way to go. Of course, it would establish the pre-eminence of deciding the value of any startup (any kind of company for that matter) on the benchmark of market price.
It should even give convenient rise to founders building from day one a company meant to make them a lifetime living. It would not matter then if things turn out to be acquisition or IPO for an exit strategy on the part of market forces.
Lean startup, blind resolve to intuition, cost-bursting technology, ready candor to mend mistakes and restiveness for innovating right are raw qualities that count for guiding lights to the typical founder that will build the value-product company of the future. Together they should materialize into traction space and undiminishing liquidity. All for one reason: any semblance of that $2.48125."},"title":{"matchLevel":"none","matchedWords":[],"value":"DECONSTRUCTING THE MYTH ABOUT VENTURE RISK: A WAY OUT OF DILUTION"},"url":{"matchLevel":"none","matchedWords":[],"value":""}},"_tags":["story","author_anchority","story_1937172","ask_hn"],"author":"anchority","children":[1937186],"created_at":"2010-11-24T12:45:40Z","created_at_i":1290602740,"num_comments":1,"objectID":"1937172","points":1,"story_id":1937172,"story_text":"DECONSTRUCTING THE MYTH ABOUT VENTURE RISK: A WAY OUT OF DILUTION
VCs are just great. They are there waiting at the finish-line to see startups through the race. Angels are so graceful. They are the proverbial friend who saw the pulloff from the beginning, and even went down with the buddy to jail saying that was darn good! Whatever\u2019s wrong with the gap between money and ideas, it never should play us all dead about the natural forces of the market and what exactly we stand for in the marketplace.
Since Joe left his fine job to join Dick, fresh out of college and so stubborn for not taking that big offer from General Everything, to turn a question into a startup; someone reading their story later in time, like their families and friends now, should see that they were making or breaking it. Stupid or brilliant, only time can heal what reason can\u2019t.
They met an angel through another friend. He falls in love and tears his check leaves. The prototype goes live and the market is made. That wonderful thingumajig is sold for $3 apiece. Where no taxes apply because of tax holiday for startups in the state they live in, all the revenues amount to $450, 000. Figuring out the total cost of production of $65, 000, they are left with $385, 000 in profit.
If only they had paid back in 1 \u00bd years after company launch the $80,000 lent by the angel, they would not have had to convert the loan to 20% shareholding. According to their agreement, a share unit is worth two dollars at pre-money valuation of $800, 000. In fact, the work-out posits that the effort of each of Joe and Dick is worth twice the amount invested by the angel. Therefore, each of them has invested $160, 000 without handing in any money save for all-important brains and brawns.
$160, 000 of Joe + $160, 000 of Dick + $80, 000 of the angel = $400, 000 total investment in the startup. In the same proportion, the goodwill accruing to each of them amounts to an equal $400, 000. Because there\u2019s no established custom or popularity, the goodwill should not count now, but they figure that VCs might come calling once the promising product is launched.
Besides, they feel that they should be really compensated for risking spending and exerting so much means and pain on proof of concept delivered to two such clients without any remuneration and contractual obligation to buy. One of them was provided by the angel.
Transport, everyday office rituals, sleepless nights, involuntary fasting, costly PR and shoestring marketing went with the whole gamut. They played deep for a great amount of work effort and thus, the longer time before they could avoid the 1 \u00bd year loan clause. Just why goodwill came a fair deal.
Press rave is here about the product. Pop culture and the TV people are dying for attention from Joe and Dick. They are reminded by the angel that they could let out some steam motion, yet they should not abandon substance for vanity. They are always back at work. Giving attention to events and the people that matter at every point in time.
Looking at the future before them, sound reasoning points in the direction of more important product features, valuable employees, necessary Intellectual Property tying-here-and-there, more solid collaboration, research and development, due diligence, etc.
It all goes from making the market to meeting the demands of the exploding market. The VCs are getting in touch with one another about this new high-tech star. The shepherd and the sheep are running to graze the bounty grass. By now, everyone in town knows how much more Joe and Dick figure the growing company needs in the way of money.
Before sales, if the share price was $2, the share price after sales takes after this picture:
$385, 000 (profit after tax: recall no taxes apply in this instance)/ $800, 000 worth of existing shares in the company = $0.48125, what is called the market quotient.
Add the pre-sales share price of $2 to the market quotient of $0.48125. And you get $2.48125 for the new share price. As simple as that. This is the market price, and it represents the least rational price any new investor should be paying to own a piece of the company.
In that wise, because the quotient could have been negative as a result of loss instead of profit, such a market price should serve as the lowest extreme of the share price band. If it was so a loss, the same formula applies and the market price should possibly be the highest extreme. And in both scenarios, whatever might count for premium and added value between buyers and sellers of the company\u2019s shares should be taken into consideration in each unique trading case.
Already, some readers would say that the allegory is demonstrating the reality on stock market floors (IPO or trading publicly quoted stocks), or common everyday buying and selling of anything founded on value. In every form of enterprise, startup investing, shipbuilding, fashion house, electric car manufacturing or trivia gamemaking, it indispensably counts to let reward and risk take their natural place in the scheme of things.
We never should attempt to trick the market, if we don\u2019t want to see another market crash or an unfortunate founder suicide. The market should decide how much reward goes to any risk-taker. From above, if any VC would invest in the hypothetical startup, he should be paying at least $2.48125 and nothing less. Doing otherwise is tricking the market, which is the prevalent culture of Series A funding in Silicon Valley.
His later involvement with the company should reward the angel and the founders for all the risks they\u2019ve taken before now: the VC should stop tearing apart previously existing term sheets and creating new ones for the sake of importunely favoring himself and the fund management people he works for. His attitude is mostly driven by irrational fear.
There is no power in arbitrariness. Even if it exists anywhere, such power function never lasts. That\u2019s easily evident in Third World dictatorships and command line communist states. Being in need of more money to run a startup shouldn\u2019t make VCs twist and wriggle angels and founders because he wants to protect investments from risk.
The only optimal and efficient way of protecting investments from risk is identifying its intrinsic value before time, and taking the well-angled dive into the startup just early enough. That\u2019s what separates the best from the worst in managing funds for startup investing.
That viewpoint already puts the angels at vantage position. It explains why nowadays the species of superangels are evolving in the ecosystem of startup investing. Some VCs are realizing that \u201cearly enough means most profitably enough\u201d. Jump on it before time asks for a greater price.
Such VCs know that the absolute power the VC community wield now would not last long. Not in the face of a whole lot of market-determined eventualities: far cheaper servers, ubiquitous open-source tools and libertarian programming languages, to mention just a sliver. Dead seriously, running a startup to profit is increasingly a function of how far along determination gets and an efficient combination of these consequences.
Merely observing the market rules and staying fair to the pricing process, is way enough to autocorrect the market in a way that the levers of demand and supply are made to naturally belong to the right sides of the marketplace. The good ones would remain, and the bad ones would be gone.
Matter-of-factly, to have more of the company, using the vast warchest in his possession, the VC should just buy more units of the shareholding at the market price. It\u2019s the price he pays for coming late to the jamboree, and might as well be his deep goblet when General Everything starts an acquisition party or the stock market invites the startup to the IPO frenzy.
If startup founders would only concentrate on creating a product the market is dying to have, the cost of funding would increasingly lower out; since such an important market factor would signal the early value investors to invest when it\u2019s so cheap to do so.
As a result, should the law of large numbers prevail; as bringing about a fairly good count of value-product startups backed by a broad population of early value investors; stock dilution caused by irrational fear would be eliminated.
That\u2019s because more would-be investors would see that it\u2019s just the way to go. Of course, it would establish the pre-eminence of deciding the value of any startup (any kind of company for that matter) on the benchmark of market price.
It should even give convenient rise to founders building from day one a company meant to make them a lifetime living. It would not matter then if things turn out to be acquisition or IPO for an exit strategy on the part of market forces.
Lean startup, blind resolve to intuition, cost-bursting technology, ready candor to mend mistakes and restiveness for innovating right are raw qualities that count for guiding lights to the typical founder that will build the value-product company of the future. Together they should materialize into traction space and undiminishing liquidity. All for one reason: any semblance of that $2.48125.","title":"DECONSTRUCTING THE MYTH ABOUT VENTURE RISK: A WAY OUT OF DILUTION","updated_at":"2024-09-19T17:21:11Z","url":""},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"fengjiabo2400"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["general","intuition"],"value":"Hey HN! I built this over a weekend after a frustrating experience with our wedding photographer.
We paid $5,000 and got back a few dozen photos. The colors were off in several shots \u2013 too warm, inconsistent lighting, some faces looked washed out. I wanted to fix them myself, but photo editing is incredibly complicated. Photoshop has hundreds of buttons, layers, masks, adjustment curves. I spent hours watching YouTube tutorials just to adjust colors on one photo. There had to be a better way.
So I built Photowand.
What it does:\nUpload any photo, describe what you want in plain English, and get professional-quality results in ~20 seconds. "Fix the colors and lighting" or "change background to beach sunset" or "make this look like a professional headshot." No Photoshop skills needed.
Try it without signup: photowand.ai
How it compares to existing AI tools:
vs ChatGPT:\n- Better photorealistic quality (specialized model, not general-purpose)\n- 4x faster processing (20 seconds vs 80+ seconds)\n- Bulk generation (create 50 variations simultaneously)\n- Iterative editing (make changes to existing photos, not just one-shot generation)
vs Gemini:\n- No watermarks on any images\n- Faster generation (20 seconds vs 40+ seconds)\n- Simpler interface (just describe what you want)\n- Commercial licensing included
Technical approach:\nCustom diffusion models optimized for photorealistic photography. I trained specifically on professional photo datasets (headshots, products, real estate, food) rather than general internet images.
Key features:\n- Natural language editing ("fix the lighting" instead of adjusting curves)\n- Consistency across batches (same person in multiple poses)\n- 400+ photo pack templates (LinkedIn headshots, product photos, etc.)\n- Unlimited revisions (describe changes, apply instantly)
What happened after I launched:\n- Now processing 500k+ photos daily\n- People using it globally in more than 180 countries\n- Turns out lots of people have the same problem (photographers cost $500-$5,000)
Where it still fails:\n- Wedding photography (emotional timing requires human intuition)\n- Complex hand interactions (I flag ~5% for manual review)\n- Extreme artistic vision (brand-specific creative direction)\n- Events requiring physical presence
The irony: I made this to fix my wedding photos. Now people mostly use it for business photography (LinkedIn headshots, product catalogs, real estate listings).
Tech stack: Custom diffusion models, Next.js, R2 storage, Cloudflare
Try the demo: photowand.ai (no signup required)
I'm here all day to answer questions about the architecture, training approach, or where AI photography still needs humans.
What would make this more useful for you?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Photowand \u2013 Turn selfies into professional photos with AI"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://photowand.ai"}},"_tags":["story","author_fengjiabo2400","story_45493020","show_hn"],"author":"fengjiabo2400","created_at":"2025-10-06T16:22:03Z","created_at_i":1759767723,"num_comments":0,"objectID":"45493020","points":1,"story_id":45493020,"story_text":"Hey HN! I built this over a weekend after a frustrating experience with our wedding photographer.
We paid $5,000 and got back a few dozen photos. The colors were off in several shots \u2013 too warm, inconsistent lighting, some faces looked washed out. I wanted to fix them myself, but photo editing is incredibly complicated. Photoshop has hundreds of buttons, layers, masks, adjustment curves. I spent hours watching YouTube tutorials just to adjust colors on one photo. There had to be a better way.
So I built Photowand.
What it does:\nUpload any photo, describe what you want in plain English, and get professional-quality results in ~20 seconds. "Fix the colors and lighting" or "change background to beach sunset" or "make this look like a professional headshot." No Photoshop skills needed.
Try it without signup: photowand.ai
How it compares to existing AI tools:
vs ChatGPT:\n- Better photorealistic quality (specialized model, not general-purpose)\n- 4x faster processing (20 seconds vs 80+ seconds)\n- Bulk generation (create 50 variations simultaneously)\n- Iterative editing (make changes to existing photos, not just one-shot generation)
vs Gemini:\n- No watermarks on any images\n- Faster generation (20 seconds vs 40+ seconds)\n- Simpler interface (just describe what you want)\n- Commercial licensing included
Technical approach:\nCustom diffusion models optimized for photorealistic photography. I trained specifically on professional photo datasets (headshots, products, real estate, food) rather than general internet images.
Key features:\n- Natural language editing ("fix the lighting" instead of adjusting curves)\n- Consistency across batches (same person in multiple poses)\n- 400+ photo pack templates (LinkedIn headshots, product photos, etc.)\n- Unlimited revisions (describe changes, apply instantly)
What happened after I launched:\n- Now processing 500k+ photos daily\n- People using it globally in more than 180 countries\n- Turns out lots of people have the same problem (photographers cost $500-$5,000)
Where it still fails:\n- Wedding photography (emotional timing requires human intuition)\n- Complex hand interactions (I flag ~5% for manual review)\n- Extreme artistic vision (brand-specific creative direction)\n- Events requiring physical presence
The irony: I made this to fix my wedding photos. Now people mostly use it for business photography (LinkedIn headshots, product catalogs, real estate listings).
Tech stack: Custom diffusion models, Next.js, R2 storage, Cloudflare
Try the demo: photowand.ai (no signup required)
I'm here all day to answer questions about the architecture, training approach, or where AI photography still needs humans.
What would make this more useful for you?","title":"Show HN: Photowand \u2013 Turn selfies into professional photos with AI","updated_at":"2026-03-05T22:51:26Z","url":"https://photowand.ai"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"calebchiam"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["intuition"],"value":"Cleanlab (https://github.com/cleanlab/cleanlab) is a family of algorithms for automatically finding issues in datasets. It might seem surprising that it\u2019s possible to automatically identify label errors and out-of-distribution data; Cleanlab does this using the algorithms published in https://arxiv.org/abs/1911.00068.
Cleanlab\u2019s algorithms, while clever, are actually relatively simple. To help myself (and others!) build intuition for how they work, I built Vizzy, an interactive demo that runs in the browser. Vizzy lets you experiment with an example dataset, tweak the labels, and run Cleanlab to automatically find issues like label errors and out-of-distribution data
Vizzy includes a JavaScript port of (a part of) cleanlab, along with other neat technical nuggets including ML model training in the browser (using features from a pretrained ResNet-18, performing truncated SVD, and using an SVM model for speed). If you\u2019re interested in the details of how Vizzy works, check out this blog post: https://cleanlab.ai/blog/cleanlab-vizzy/
I\u2019m happy to answer any questions related to Vizzy, cleanlab, or confident learning and data-centric AI in general!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Cleanlab Vizzy \u2013 automatically find label errors and bad data"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://playground.cleanlab.ai/"}},"_tags":["story","author_calebchiam","story_32496807","show_hn"],"author":"calebchiam","created_at":"2022-08-17T14:23:17Z","created_at_i":1660746197,"num_comments":0,"objectID":"32496807","points":1,"story_id":32496807,"story_text":"Cleanlab (https://github.com/cleanlab/cleanlab) is a family of algorithms for automatically finding issues in datasets. It might seem surprising that it\u2019s possible to automatically identify label errors and out-of-distribution data; Cleanlab does this using the algorithms published in https://arxiv.org/abs/1911.00068.
Cleanlab\u2019s algorithms, while clever, are actually relatively simple. To help myself (and others!) build intuition for how they work, I built Vizzy, an interactive demo that runs in the browser. Vizzy lets you experiment with an example dataset, tweak the labels, and run Cleanlab to automatically find issues like label errors and out-of-distribution data
Vizzy includes a JavaScript port of (a part of) cleanlab, along with other neat technical nuggets including ML model training in the browser (using features from a pretrained ResNet-18, performing truncated SVD, and using an SVM model for speed). If you\u2019re interested in the details of how Vizzy works, check out this blog post: https://cleanlab.ai/blog/cleanlab-vizzy/
I\u2019m happy to answer any questions related to Vizzy, cleanlab, or confident learning and data-centric AI in general!","title":"Show HN: Cleanlab Vizzy \u2013 automatically find label errors and bad data","updated_at":"2024-09-20T11:55:05Z","url":"https://playground.cleanlab.ai/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"frabonacci"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"Hello HN! We're Dillon and Francesco from Cua.
We were wondering how many computer use tasks actually need a full general purpose LLM (e.g. gpt-6-astra, claude-opus-5 etc.) to think through all their decisions and steps. Some tasks require thinking about a plan, exploring different paths, recovering from failure. Other tasks are a question of making local decisions, like this value should go in this box, or should I check this box, or this element should be ignored.
We wondered how far we could go with a small model trained to only make these kinds of decisions.
Our inspiration was Typesafe's Jev and its System One Model framing. This is a nod to the dichotomy between thinking quickly, automatically, and intuitively (system 1) vs. thinking slowly, analytically (system 2), as described by Daniel Kahneman.
The interesting question for us was: what happens if you give a model an interface of current context, and a set of possible choices, and you ask it to return a probability for each choice? This kind of model does not generate output token by token like most LLMs do, but rather scores the options you give it, which you can check, trust, and use to drive your app's behavior.
CUA-S1 is our answer for narrow, specialized decision models for computer use. Our first release is CUA-S1-FORMS. We built this from ideas and code in jevlike, and then trained a second model just to handle form interactions. It has 706k parameters, and the original checkpoint is 2.8 MB.
The first training iteration took less than 30 minutes on synthetic data. Given a set of structured elements and values extracted from a document, it predicts whether to use the given value, CHECK, CLICK, or SKIP for each element. It does not predict new values for text fields, and does not consider screenshots. Element decisions are scored together, and your code can order the actions, and Cua Driver will execute them one at a time.
A first evaluation of this specialist vs. hosted Jev on our form task:
- For the whole decision set: 99.7% correct vs 83.6%.
- For the subset of steps that require an action: 100% correct vs 96%.
- For the subset of steps that are just leaving already-filled fields alone: 100% correct vs 74%.
The specialist was trained specifically for this task and convention (just press skip for already filled boxes), while hosted Jev has not been fine-tuned for it, so this is an experiment in scoped specialization.
We measured 7-9 ms to score a form locally vs. 260-280 ms per call to hosted Jev including network latency, though those samples measure different things and are not end-to-end form completion times.
Our interest here is in the space between a brittle script and a general agent loop. The content and layout of form fields vary enough that scripts get unwieldy, but the set of available decisions can remain narrow and well scoped. We want to explore the possibility of a general agent encountering something novel, and passing well understood decisions over to specialists like this.
That is a direction we are looking into. The current release is for forms only. We're open sourced the synthetic data generation, training, evaluation, and Driver integration under libs/cua-s1 with an MIT license.
Comments welcome! Especially if you are building computer-use agents and have run into a recurring decision that is too variable to script but is too narrow to call another LLM for."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: CUA-S1 \u2013 A System One Model for Computer Use"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/trycua/cua"}},"_tags":["story","author_frabonacci","story_49767564","show_hn"],"author":"frabonacci","children":[49769462,49769966,49769972,49770166,49770350,49771158,49773393,49777463],"created_at":"2026-09-19T15:52:51Z","created_at_i":1789833171,"num_comments":10,"objectID":"49767564","points":89,"story_id":49767564,"story_text":"Hello HN! We're Dillon and Francesco from Cua.
We were wondering how many computer use tasks actually need a full general purpose LLM (e.g. gpt-6-astra, claude-opus-5 etc.) to think through all their decisions and steps. Some tasks require thinking about a plan, exploring different paths, recovering from failure. Other tasks are a question of making local decisions, like this value should go in this box, or should I check this box, or this element should be ignored.
We wondered how far we could go with a small model trained to only make these kinds of decisions.
Our inspiration was Typesafe's Jev and its System One Model framing. This is a nod to the dichotomy between thinking quickly, automatically, and intuitively (system 1) vs. thinking slowly, analytically (system 2), as described by Daniel Kahneman.
The interesting question for us was: what happens if you give a model an interface of current context, and a set of possible choices, and you ask it to return a probability for each choice? This kind of model does not generate output token by token like most LLMs do, but rather scores the options you give it, which you can check, trust, and use to drive your app's behavior.
CUA-S1 is our answer for narrow, specialized decision models for computer use. Our first release is CUA-S1-FORMS. We built this from ideas and code in jevlike, and then trained a second model just to handle form interactions. It has 706k parameters, and the original checkpoint is 2.8 MB.
The first training iteration took less than 30 minutes on synthetic data. Given a set of structured elements and values extracted from a document, it predicts whether to use the given value, CHECK, CLICK, or SKIP for each element. It does not predict new values for text fields, and does not consider screenshots. Element decisions are scored together, and your code can order the actions, and Cua Driver will execute them one at a time.
A first evaluation of this specialist vs. hosted Jev on our form task:
- For the whole decision set: 99.7% correct vs 83.6%.
- For the subset of steps that require an action: 100% correct vs 96%.
- For the subset of steps that are just leaving already-filled fields alone: 100% correct vs 74%.
The specialist was trained specifically for this task and convention (just press skip for already filled boxes), while hosted Jev has not been fine-tuned for it, so this is an experiment in scoped specialization.
We measured 7-9 ms to score a form locally vs. 260-280 ms per call to hosted Jev including network latency, though those samples measure different things and are not end-to-end form completion times.
Our interest here is in the space between a brittle script and a general agent loop. The content and layout of form fields vary enough that scripts get unwieldy, but the set of available decisions can remain narrow and well scoped. We want to explore the possibility of a general agent encountering something novel, and passing well understood decisions over to specialists like this.
That is a direction we are looking into. The current release is for forms only. We're open sourced the synthetic data generation, training, evaluation, and Driver integration under libs/cua-s1 with an MIT license.
Comments welcome! Especially if you are building computer-use agents and have run into a recurring decision that is too variable to script but is too narrow to call another LLM for.","title":"Show HN: CUA-S1 \u2013 A System One Model for Computer Use","updated_at":"2026-09-20T23:30:41Z","url":"https://github.com/trycua/cua"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"domhudson"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"I've been playing with this idea to build an estimation of chess cheating likelihood from an exported chess game (PGN format).
The main thing I've developed so far is a way of estimating the complexity of each position. The intuition is that matching an engine's top moves in difficult positions is much more informative than doing so when there's an obvious best move.
The detector currently combines complexity with engine agreement and centipawn loss to make a prediction. I'd like to move beyond the heuristics and train an ML model on a decent dataset of known cheating/non-cheating games.
There is an obvious limitation though: I have fewer signals available than than Chess.com/Lichess. I generally don't have a player's game history, and most PGNs don't contain time-per-move data, both of which would presumably be very useful.
Still pretty experimental, so I'd be interested in feedback; especially from anyone who's worked on chess engines, cheating detection, or related statistical problems."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Chess Cheater Detector"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://chesscheaterdetector.com/"}},"_tags":["story","author_domhudson","story_49706032","show_hn"],"author":"domhudson","children":[49729682],"created_at":"2026-09-15T00:14:36Z","created_at_i":1789431276,"num_comments":4,"objectID":"49706032","points":1,"story_id":49706032,"story_text":"I've been playing with this idea to build an estimation of chess cheating likelihood from an exported chess game (PGN format).
The main thing I've developed so far is a way of estimating the complexity of each position. The intuition is that matching an engine's top moves in difficult positions is much more informative than doing so when there's an obvious best move.
The detector currently combines complexity with engine agreement and centipawn loss to make a prediction. I'd like to move beyond the heuristics and train an ML model on a decent dataset of known cheating/non-cheating games.
There is an obvious limitation though: I have fewer signals available than than Chess.com/Lichess. I generally don't have a player's game history, and most PGNs don't contain time-per-move data, both of which would presumably be very useful.
Still pretty experimental, so I'd be interested in feedback; especially from anyone who's worked on chess engines, cheating detection, or related statistical problems.","title":"Show HN: Chess Cheater Detector","updated_at":"2026-09-20T21:30:27Z","url":"https://chesscheaterdetector.com/"}],"hitsPerPage":50,"nbHits":37,"nbPages":1,"page":0,"params":"query=General+Intuition&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":12,"processingTimingsMS":{"_request":{"queue":39,"roundTrip":14},"afterFetch":{"format":{"highlighting":3,"total":4}},"fetch":{"query":8,"scanning":2,"total":11},"total":12},"query":"General Intuition","serverTimeMS":56}