For the past 6 months or so I have been totally working with nodejs and its eco-system. I do agree its blazing fast and has a very good eco-system. But I really don't believe that it is that much strong in web-development. I guess the only stable web framework is Express. But you have to do alot in it to achieve things. After researching I found Sails.js is gaining popularity and has many features out of the box. But after trying it for a month or so, I found that its very much buggy. The thing that works on my local environment does not work in production, however it has the same technology stack with same versions.
npm packages are outdated, dependencies not updated. I wonder why there is so much hype about nodejs and what it is really good at?
What would you say?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Why Node.js?"},"url":{"matchLevel":"none","matchedWords":[],"value":""}},"_tags":["story","author_mahadazad","story_10008803","ask_hn"],"author":"mahadazad","children":[10008833,10008859,10008861,10009501,10009719,10009848,10010590,10011797,10015449,10021591],"created_at":"2015-08-05T10:23:16Z","created_at_i":1438770196,"num_comments":14,"objectID":"10008803","points":13,"story_id":10008803,"story_text":"I am a developer with a very strong PHP background. Experience over 5-6 years in PHP and related technologies.
For the past 6 months or so I have been totally working with nodejs and its eco-system. I do agree its blazing fast and has a very good eco-system. But I really don't believe that it is that much strong in web-development. I guess the only stable web framework is Express. But you have to do alot in it to achieve things. After researching I found Sails.js is gaining popularity and has many features out of the box. But after trying it for a month or so, I found that its very much buggy. The thing that works on my local environment does not work in production, however it has the same technology stack with same versions.
npm packages are outdated, dependencies not updated. I wonder why there is so much hype about nodejs and what it is really good at?
What would you say?","title":"Why Node.js?","updated_at":"2024-09-19T22:07:10Z","url":""},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"AbhinavX"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Hi HN, we\u2019re Abhinav, Andy, and Jeremy, and we\u2019re building Lucidic AI (https://dashboard.lucidic.ai), an AI agent interpretability tool to help observe/debug AI agents.
Here is a demo: https://youtu.be/Zvoh1QUMhXQ.
Getting started is easy with just one line of code. You just call lai.init() in your agent code and log into the dashboard. You can see traces of each run, cumulative trends across sessions, built-in or custom evals, and grouped failure modes. Call lai.create_step() with any metadata you want, memory snapshots, tool outputs, stateful info, and we'll index it for debugging.
We did NLP research at Stanford AI Lab (SAIL), where we worked on creating an AI agent (w/ fine-tuned models and DSPy) to solve math olympiad problems (focusing on AIME/USAMO); and we realized debugging these agents was hard. But the last straw was when we built an e-commerce agent that could buy items online. It kept failing at checkout, and every one-line change, tweaking a prompt, switching to Llama, adjusting tool logic, meant another 10-minute rerun just to see if we hit the same checkout page.
At this point, we were all like, this sucks, so we improved agent interpretability with better debugging, monitoring, and evals.
We started by listening to users who told us traditional LLM observability platforms don't capture the complexity of agents. Agents have tools, memories, events, not just input/output pairs. So we automatically transform OTel (and/or regular) agent logs into interactive graph visualizations that cluster similar states based on memory and action patterns.\nWe heard that people wanted to test small changes even with the graphs, so we created \u201ctime traveling,\u201d where you can modify any state (memory contents, tool outputs, context), then re-simulate 30\u201340 times to see outcome distributions. We embed the responses, cluster by similarity, and show which modifications lead to stable vs. divergent behaviors.
Then we saw people running their agent 10 times on the same task, watching each run individually, and wasting hours looking at mostly repeated states. So we built trajectory clustering on similar state embeddings (like similar tools or memories) to surface behavioral patterns across mass simulations.
We then use that to create a force-directed layout that automatically groups similar paths your agent took, which displays states as nodes, actions as edges, and failure probability as color intensity. The clusters make failure patterns obvious; you see trends across hundreds of runs, not individual traces.
Finally, when people saw our observability features, they naturally wanted evaluation capabilities. So we developed a concept for people to make their own evals called "rubrics," which lets you define specific criteria, assign weights to each criterion, and set score definitions, giving you a structured way to measure agent performance against your exact requirements.
To evaluate these criteria, we used our own platform to build an investigator agent that reviews your criteria and evaluates performance much more effectively than traditional LLM-as-a-judge approaches.
To get started visit dashboard.lucidic.ai and https://docs.lucidic.ai/getting-started/quickstart. You can use it for free for 1,000 event and step creations.
Look forward to your thoughts! And don\u2019t hesitate to reach out at team@lucidic.ai"},"title":{"matchLevel":"none","matchedWords":[],"value":"Launch HN: Lucidic (YC W25) \u2013 Debug, test, and evaluate AI agents in production"}},"_tags":["story","author_AbhinavX","story_44735843","launch_hn"],"author":"AbhinavX","children":[44736020,44736243,44736389,44736461,44737009,44737110,44737185,44737435,44737824,44738655,44739338,44739456,44740539,44741816,44742393,44773750],"created_at":"2025-07-30T15:54:02Z","created_at_i":1753890842,"num_comments":39,"objectID":"44735843","points":116,"story_id":44735843,"story_text":"Hi HN, we\u2019re Abhinav, Andy, and Jeremy, and we\u2019re building Lucidic AI (https://dashboard.lucidic.ai), an AI agent interpretability tool to help observe/debug AI agents.
Here is a demo: https://youtu.be/Zvoh1QUMhXQ.
Getting started is easy with just one line of code. You just call lai.init() in your agent code and log into the dashboard. You can see traces of each run, cumulative trends across sessions, built-in or custom evals, and grouped failure modes. Call lai.create_step() with any metadata you want, memory snapshots, tool outputs, stateful info, and we'll index it for debugging.
We did NLP research at Stanford AI Lab (SAIL), where we worked on creating an AI agent (w/ fine-tuned models and DSPy) to solve math olympiad problems (focusing on AIME/USAMO); and we realized debugging these agents was hard. But the last straw was when we built an e-commerce agent that could buy items online. It kept failing at checkout, and every one-line change, tweaking a prompt, switching to Llama, adjusting tool logic, meant another 10-minute rerun just to see if we hit the same checkout page.
At this point, we were all like, this sucks, so we improved agent interpretability with better debugging, monitoring, and evals.
We started by listening to users who told us traditional LLM observability platforms don't capture the complexity of agents. Agents have tools, memories, events, not just input/output pairs. So we automatically transform OTel (and/or regular) agent logs into interactive graph visualizations that cluster similar states based on memory and action patterns.\nWe heard that people wanted to test small changes even with the graphs, so we created \u201ctime traveling,\u201d where you can modify any state (memory contents, tool outputs, context), then re-simulate 30\u201340 times to see outcome distributions. We embed the responses, cluster by similarity, and show which modifications lead to stable vs. divergent behaviors.
Then we saw people running their agent 10 times on the same task, watching each run individually, and wasting hours looking at mostly repeated states. So we built trajectory clustering on similar state embeddings (like similar tools or memories) to surface behavioral patterns across mass simulations.
We then use that to create a force-directed layout that automatically groups similar paths your agent took, which displays states as nodes, actions as edges, and failure probability as color intensity. The clusters make failure patterns obvious; you see trends across hundreds of runs, not individual traces.
Finally, when people saw our observability features, they naturally wanted evaluation capabilities. So we developed a concept for people to make their own evals called "rubrics," which lets you define specific criteria, assign weights to each criterion, and set score definitions, giving you a structured way to measure agent performance against your exact requirements.
To evaluate these criteria, we used our own platform to build an investigator agent that reviews your criteria and evaluates performance much more effectively than traditional LLM-as-a-judge approaches.
To get started visit dashboard.lucidic.ai and https://docs.lucidic.ai/getting-started/quickstart. You can use it for free for 1,000 event and step creations.
Look forward to your thoughts! And don\u2019t hesitate to reach out at team@lucidic.ai","title":"Launch HN: Lucidic (YC W25) \u2013 Debug, test, and evaluate AI agents in production","updated_at":"2025-08-08T04:22:45Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pseudolus"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["sail"],"value":"Ultrafast transfer of low-mass payloads to Mars using aerographite solar sails"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"https://www.researchgate.net/publication/373552196_Ultrafast_transfer_of_low-mass_payloads_to_Mars_and_beyond_using_aerographite_solar_sails"}},"_tags":["story","author_pseudolus","story_37771582"],"author":"pseudolus","children":[37772010],"created_at":"2023-10-04T20:50:57Z","created_at_i":1696452657,"num_comments":1,"objectID":"37771582","points":3,"story_id":37771582,"title":"Ultrafast transfer of low-mass payloads to Mars using aerographite solar sails","updated_at":"2024-09-20T15:18:18Z","url":"https://www.researchgate.net/publication/373552196_Ultrafast_transfer_of_low-mass_payloads_to_Mars_and_beyond_using_aerographite_solar_sails"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"yzh"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["research"],"value":"HloEnv: An environment based on XLA for DL compiler optimization research"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["sail"],"value":"https://github.com/sail-sg/hloenv"}},"_tags":["story","author_yzh","story_33528994"],"author":"yzh","created_at":"2022-11-09T09:10:06Z","created_at_i":1667985006,"num_comments":0,"objectID":"33528994","points":2,"story_id":33528994,"title":"HloEnv: An environment based on XLA for DL compiler optimization research","updated_at":"2024-09-20T12:25:39Z","url":"https://github.com/sail-sg/hloenv"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ColinWright"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["sail"],"value":"Oumuamua may be alien light sail: Harvard astronomers"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"https://www.news.com.au/technology/science/space/oumuamua-harvard-researchers-suggest-strange-interstellar-object-may-be-alien-light-sail/news-story/6527afbbd9e1350a77ba8618da2c1b9e"}},"_tags":["story","author_ColinWright","story_19084457"],"author":"ColinWright","children":[19084490],"created_at":"2019-02-05T11:47:16Z","created_at_i":1549367236,"num_comments":1,"objectID":"19084457","points":1,"story_id":19084457,"title":"Oumuamua may be alien light sail: Harvard astronomers","updated_at":"2024-09-20T03:45:29Z","url":"https://www.news.com.au/technology/science/space/oumuamua-harvard-researchers-suggest-strange-interstellar-object-may-be-alien-light-sail/news-story/6527afbbd9e1350a77ba8618da2c1b9e"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"King-Aaron"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["research"],"value":"Australian University researchers light up pathway to another planetary system"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["sail"],"value":"https://www.abc.net.au/news/2021-06-08/canberra-anu-scientists-laser-sail-interstellar-space-discovery/100198228"}},"_tags":["story","author_King-Aaron","story_27433767"],"author":"King-Aaron","created_at":"2021-06-08T11:37:22Z","created_at_i":1623152242,"num_comments":0,"objectID":"27433767","points":1,"story_id":27433767,"title":"Australian University researchers light up pathway to another planetary system","updated_at":"2024-09-20T08:43:06Z","url":"https://www.abc.net.au/news/2021-06-08/canberra-anu-scientists-laser-sail-interstellar-space-discovery/100198228"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"MargoRooo"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Most people think procrastination is a discipline or time management problem. But after analyzing data from 80,000+ program participants and 3,000+ coaching sessions, we discovered something unexpected.
Procrastination Is Emotional Dysregulation, Not Laziness
Key finding: People don\u2019t postpone tasks because of laziness. They postpone due to fear, perfectionism, overwhelm, or uncertainty. These are subcortical brain structures protecting from perceived risk.
When we explained to participants that their \u201cinner saboteur\u201d is an ancient part of the brain, they stopped blaming themselves for weak willpower and started working with resistance constructively.
Agile Approach to Life Works Better Than Long-term Planning
Pattern: 3 weeks of focus + 1 week of reflection give better results than trying to plan months ahead.
Participants said: \u201cFor the first time, planning doesn\u2019t cause stress.\u201d Short sprints allow experimenting and course correction, instead of \u201cpushing through\u201d toward unclear goals.
\u201cBoat Bottom\u201d Is More Important Than \u201cSail\u201d
Key mistake: 70% of participants set ambitious goals (sail) without covering basic needs (boat bottom).
We divided goals into two levels:
- Basic goals \u2014 sleep, energy, emotional state\n- Growth goals \u2014 career, self-actualization, new skills
When people started with the basic level, growth goals were achieved naturally. When they tried to \u201craise the sail\u201d on a leaky boat \u2014 they burned out quickly.
Three Levels of Working with Procrastination
From the data, a system crystallized:
1. Energy \u2014 restoring healthy dopamine cycles\n2. Goals \u2014 turning dreams into clear action plans\n3. Psychology \u2014 working with internal blocks based on psychotype
Attempts to immediately jump to level 3, bypassing 1 and 2, led to sabotage and regression.
Practical Takeaways
- Procrastination is a symptom, not the disease. You need to treat the causes\n- Short sprints (3 weeks) are more effective than long-term plans\n- Basic needs must be met BEFORE ambitious goals\n- Understanding neurobiology reduces self-flagellation and increases effectiveness
Application
Based on this data, we built an AI system that determines what level a person is at and adapts the approach. Early results are encouraging: 200 users, 300+ completed tasks, early sales exceeded $100k.
But most importantly \u2014 we now have a data-driven understanding of how procrastination works. And these principles can be applied regardless of any apps.
Looking for Early Testers
We\u2019re finalizing the full mobile app AJI (AI Journal & Insights) and need people to test the first versions. We\u2019re especially interested in participants with chronic procrastination \u2014 the approach shows best results with them.
If you\u2019re ready to try it and share feedback \u2014 write in the comments.
Data collected from productivity programs, coaching sessions, and scientific research from 2020-202"},"title":{"matchLevel":"none","matchedWords":[],"value":"We Studied Procrastination on a Scale. Here's What Works"}},"_tags":["story","author_MargoRooo","story_44267715","ask_hn"],"author":"MargoRooo","children":[44267717,44267867,44267916],"created_at":"2025-06-13T11:57:05Z","created_at_i":1749815825,"num_comments":2,"objectID":"44267715","points":5,"story_id":44267715,"story_text":"Most people think procrastination is a discipline or time management problem. But after analyzing data from 80,000+ program participants and 3,000+ coaching sessions, we discovered something unexpected.
Procrastination Is Emotional Dysregulation, Not Laziness
Key finding: People don\u2019t postpone tasks because of laziness. They postpone due to fear, perfectionism, overwhelm, or uncertainty. These are subcortical brain structures protecting from perceived risk.
When we explained to participants that their \u201cinner saboteur\u201d is an ancient part of the brain, they stopped blaming themselves for weak willpower and started working with resistance constructively.
Agile Approach to Life Works Better Than Long-term Planning
Pattern: 3 weeks of focus + 1 week of reflection give better results than trying to plan months ahead.
Participants said: \u201cFor the first time, planning doesn\u2019t cause stress.\u201d Short sprints allow experimenting and course correction, instead of \u201cpushing through\u201d toward unclear goals.
\u201cBoat Bottom\u201d Is More Important Than \u201cSail\u201d
Key mistake: 70% of participants set ambitious goals (sail) without covering basic needs (boat bottom).
We divided goals into two levels:
- Basic goals \u2014 sleep, energy, emotional state\n- Growth goals \u2014 career, self-actualization, new skills
When people started with the basic level, growth goals were achieved naturally. When they tried to \u201craise the sail\u201d on a leaky boat \u2014 they burned out quickly.
Three Levels of Working with Procrastination
From the data, a system crystallized:
1. Energy \u2014 restoring healthy dopamine cycles\n2. Goals \u2014 turning dreams into clear action plans\n3. Psychology \u2014 working with internal blocks based on psychotype
Attempts to immediately jump to level 3, bypassing 1 and 2, led to sabotage and regression.
Practical Takeaways
- Procrastination is a symptom, not the disease. You need to treat the causes\n- Short sprints (3 weeks) are more effective than long-term plans\n- Basic needs must be met BEFORE ambitious goals\n- Understanding neurobiology reduces self-flagellation and increases effectiveness
Application
Based on this data, we built an AI system that determines what level a person is at and adapts the approach. Early results are encouraging: 200 users, 300+ completed tasks, early sales exceeded $100k.
But most importantly \u2014 we now have a data-driven understanding of how procrastination works. And these principles can be applied regardless of any apps.
Looking for Early Testers
We\u2019re finalizing the full mobile app AJI (AI Journal & Insights) and need people to test the first versions. We\u2019re especially interested in participants with chronic procrastination \u2014 the approach shows best results with them.
If you\u2019re ready to try it and share feedback \u2014 write in the comments.
Data collected from productivity programs, coaching sessions, and scientific research from 2020-202","title":"We Studied Procrastination on a Scale. Here's What Works","updated_at":"2025-06-15T14:10:30Z"},{"_highlightResult":{"author":{"fullyHighlighted":true,"matchLevel":"partial","matchedWords":["sail"],"value":"sails"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["research"],"value":"Hey!
We\u2019ve built a data extraction tool to flexibly automate data and document processing.
You\u2019ve probably seen a few of these, so have we! A few of us have been varyingly stuck trying to automate the extraction of borrower financials for the past 5 years.
We think that there are a few missing features of most data extraction tools.
* They are usually too complex to quickly get up and running
* They are overly constrained in terms of what workflows and documents they support
We\u2019ve always felt like speed and flexibility were sticking points, so we went slightly orthogonal to the alternatives.
It is deliberately very simple, with three key features
1. Generate a custom schema that is the target of the data extraction
2. Logic and \u201cexpert insights\u201d can get added to field level prompts
3. Ability to share with externals via chat to get data and feedback, quickly!
The demo page is at https://go.sea.dev with an explainer and link to demo.
The demo app is constrained, but feel free to get in touch with me with details if you\u2019d like to unlock the full capabilities
Some of the features we are working on:
* Improvements on OCR
* Advanced data management
* Email or WhatsApp to accept responses
* API and embedded data collection copilot for SaaS
* Meeting recording plugin, and streaming audio
* Securely share partial submissions with 3rd party for completion
* Enrichment via website scraping
* Integrations (Hubspot, Sheets etc)
We are a very small team, with a mix of researchers and engineers. We are leaning into improvements in data extraction, conversation evals and data structuring innovation to make the product as powerful and seamless as possible.
Our background is in financial services, so most of our effort is going into that use-case (business risk assessment to be specific), we\u2019ve solved a bunch of our internal use cases as well as general purpose customer problems, so we thought to share it with a wider audience.
Would love to get your feedback and comments!"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Data Extraction with Flexible Schemas"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://go.sea.dev"}},"_tags":["story","author_sails","story_42430490","show_hn"],"author":"sails","created_at":"2024-12-16T12:42:57Z","created_at_i":1734352977,"num_comments":0,"objectID":"42430490","points":5,"story_id":42430490,"story_text":"Hey!
We\u2019ve built a data extraction tool to flexibly automate data and document processing.
You\u2019ve probably seen a few of these, so have we! A few of us have been varyingly stuck trying to automate the extraction of borrower financials for the past 5 years.
We think that there are a few missing features of most data extraction tools.
* They are usually too complex to quickly get up and running
* They are overly constrained in terms of what workflows and documents they support
We\u2019ve always felt like speed and flexibility were sticking points, so we went slightly orthogonal to the alternatives.
It is deliberately very simple, with three key features
1. Generate a custom schema that is the target of the data extraction
2. Logic and \u201cexpert insights\u201d can get added to field level prompts
3. Ability to share with externals via chat to get data and feedback, quickly!
The demo page is at https://go.sea.dev with an explainer and link to demo.
The demo app is constrained, but feel free to get in touch with me with details if you\u2019d like to unlock the full capabilities
Some of the features we are working on:
* Improvements on OCR
* Advanced data management
* Email or WhatsApp to accept responses
* API and embedded data collection copilot for SaaS
* Meeting recording plugin, and streaming audio
* Securely share partial submissions with 3rd party for completion
* Enrichment via website scraping
* Integrations (Hubspot, Sheets etc)
We are a very small team, with a mix of researchers and engineers. We are leaning into improvements in data extraction, conversation evals and data structuring innovation to make the product as powerful and seamless as possible.
Our background is in financial services, so most of our effort is going into that use-case (business risk assessment to be specific), we\u2019ve solved a bunch of our internal use cases as well as general purpose customer problems, so we thought to share it with a wider audience.
Would love to get your feedback and comments!","title":"Show HN: Data Extraction with Flexible Schemas","updated_at":"2024-12-16T15:05:32Z","url":"https://go.sea.dev"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"charlycst"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Miralis is a RISC-V firmware that virtualizes RISC-V firmware. In other words, it runs firmware in user-space (M-mode software in U-mode).
The fact that this is even possible is interesting: indeed, not all ISAs are virtualizable, and the same applies for their firmware mode. It all boils down to the virtualization requirements [1], which is a great read if you haven't come across it yet. Arm's EL3 cannot be virtualized, for instance, because some instructions, such as `cpsid`, are sensitive but do not trap (`cpsid` is a nop in user-space).
If you have a VisionFive 2 or a HiFive Premier P550, you can try it out, the instructions are in the documentation [2, 3]. Of course, it runs on QEMU too.
As Miralis is a research project, we have also been using it as a vehicle to explore other research ideas, such as automated verification of hypervisors [4]. For instance, we verified instruction emulation by comparing Miralis' implementation with the reference RISC-V executable specification [5], which we translated to Rust.
It has been fun working on Miralis, I hope you'll find it interesting too!
[1]: https://dl.acm.org/doi/pdf/10.1145/361011.361073
[2]: https://miralis-firmware.github.io/docs/platforms/visionfive...
[3]: https://miralis-firmware.github.io/docs/platforms/premierp55...
[4]: https://charlycst.github.io/papers/lightweight-hypervisor-ve...
[5]: https://github.com/riscv/sail-riscv"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Miralis \u2013 a RISC-V virtual firmware monitor"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/CharlyCst/miralis"}},"_tags":["story","author_charlycst","story_43947576","show_hn"],"author":"charlycst","created_at":"2025-05-10T17:58:11Z","created_at_i":1746899891,"num_comments":0,"objectID":"43947576","points":4,"story_id":43947576,"story_text":"Miralis is a RISC-V firmware that virtualizes RISC-V firmware. In other words, it runs firmware in user-space (M-mode software in U-mode).
The fact that this is even possible is interesting: indeed, not all ISAs are virtualizable, and the same applies for their firmware mode. It all boils down to the virtualization requirements [1], which is a great read if you haven't come across it yet. Arm's EL3 cannot be virtualized, for instance, because some instructions, such as `cpsid`, are sensitive but do not trap (`cpsid` is a nop in user-space).
If you have a VisionFive 2 or a HiFive Premier P550, you can try it out, the instructions are in the documentation [2, 3]. Of course, it runs on QEMU too.
As Miralis is a research project, we have also been using it as a vehicle to explore other research ideas, such as automated verification of hypervisors [4]. For instance, we verified instruction emulation by comparing Miralis' implementation with the reference RISC-V executable specification [5], which we translated to Rust.
It has been fun working on Miralis, I hope you'll find it interesting too!
[1]: https://dl.acm.org/doi/pdf/10.1145/361011.361073
[2]: https://miralis-firmware.github.io/docs/platforms/visionfive...
[3]: https://miralis-firmware.github.io/docs/platforms/premierp55...
[4]: https://charlycst.github.io/papers/lightweight-hypervisor-ve...
[5]: https://github.com/riscv/sail-riscv","title":"Show HN: Miralis \u2013 a RISC-V virtual firmware monitor","updated_at":"2025-05-11T18:20:42Z","url":"https://github.com/CharlyCst/miralis"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vox_mollis"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Today I got stopped out at level 5 of tptacek and patio11's excellent Jailbreak Challenge.
I am a professional developer in my dayjob, and I am mostly proud of the systems that I've engineered over the years. There have always been large challenges and tight constraints, but the path to solution has always been straightforward, given a little research. There has never been a time in my career as a developer where I've been in an "I can't see any possible way to do this" scenario.
Enter Jailbreak. I have never been more frustrated in my entire life. Nothing but wild goose chases and dead ends. While others sail through the levels with ease, I eventually ended up having to give up entirely at level 5 due to total exhaustion of all available options.
Having eaten my humble pie, my question to HN is: how do I improve at this? Are there fundamental cognitive constraints or modes of thinking that some people like me will simply never be able to push through? Why are solutions so clear and obvious during normal software development whereas these class of problems do little more than frustrate? Am I simply too dumb (or old, at 36, due to brain plasticity)?
What resources are available to improve at this, if any? Ideally I'd like to improve my skills at this and retry level 5 in whatever timeframe (a year?) is expected to become competent at this. Any help or suggestions would be appreciated. Thanks."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: How do I improve at reversing/cracking challenges?"}},"_tags":["story","author_vox_mollis","story_12098784","ask_hn"],"author":"vox_mollis","created_at":"2016-07-15T03:27:28Z","created_at_i":1468553248,"num_comments":0,"objectID":"12098784","points":3,"story_id":12098784,"story_text":"Today I got stopped out at level 5 of tptacek and patio11's excellent Jailbreak Challenge.
I am a professional developer in my dayjob, and I am mostly proud of the systems that I've engineered over the years. There have always been large challenges and tight constraints, but the path to solution has always been straightforward, given a little research. There has never been a time in my career as a developer where I've been in an "I can't see any possible way to do this" scenario.
Enter Jailbreak. I have never been more frustrated in my entire life. Nothing but wild goose chases and dead ends. While others sail through the levels with ease, I eventually ended up having to give up entirely at level 5 due to total exhaustion of all available options.
Having eaten my humble pie, my question to HN is: how do I improve at this? Are there fundamental cognitive constraints or modes of thinking that some people like me will simply never be able to push through? Why are solutions so clear and obvious during normal software development whereas these class of problems do little more than frustrate? Am I simply too dumb (or old, at 36, due to brain plasticity)?
What resources are available to improve at this, if any? Ideally I'd like to improve my skills at this and retry level 5 in whatever timeframe (a year?) is expected to become competent at this. Any help or suggestions would be appreciated. Thanks.","title":"Ask HN: How do I improve at reversing/cracking challenges?","updated_at":"2024-09-19T23:31:10Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"chlor"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"On the journey of exploring deep learning, the performance of GPU is the key to the speed of progress. I had tried multiple platforms, attempting to find the "computing power engine" that could carry my deep learning dreams. However, the reality often fell short of expectations. Stuttering and latency were like shadows, restricting the speed of model training. When dealing with large-scale datasets and complex network architectures, the computing efficiency of these platforms seemed inadequate, and I was constantly tormented by stuttering and latency. Even more troublesome was that video memory management also became a major obstacle on my way forward. When handling tasks rich in sequential data such as video action recognition, the GPU video memory of common platforms was often stretched to the limit. When processing batches of high-definition video frames in long sequences with 16GB or 24GB of video memory, video memory overflow became a common occurrence. It was not until I encountered Burncloud (https://www.burncloud.com/835.html) that I truly unlocked the door to efficient deep learning.\n Its computing efficiency is truly astonishing. When I trained a conventional convolutional neural network for image classification on a certain commonly used old cloud platform, the average time for a single iteration was 25 minutes. However, the NVIDIA A100 GPU equipped on Burncloud, with its 6,912 CUDA cores combined with an advanced architecture, reduced the iteration time for the same model and dataset to around 4 minutes. The efficiency improvement was more than 5 times, as if a high-speed motor had been installed for model training, rapidly pushing the loss function towards the optimal value.\n Video memory management is also outstanding. When handling tasks rich in sequential data such as video action recognition, the GPU of common platforms usually has 16GB or 24GB of video memory. When processing batches of high-definition video frames in long sequences, there is a more than 70% probability of encountering video memory overflow. Burncloud is different. With large-capacity video memory starting from 40GB and reaching up to 80GB, it can easily accommodate the features of massive video frames. The whole training process is as stable as a mountain, and even complex 3D convolutional long short-term memory networks can set sail smoothly, ensuring that the exploration of deep learning does not run aground.\n Multi-GPU collaboration is also a highlight of Burncloud. In distributed training for language models of the GPT scale, other platforms are limited by internal interconnection bandwidth, and the data synchronization latency between cards exceeds 50 microseconds, with a computing power utilization rate of less than 60%. Burncloud relies on the NVLink ultra-high-speed channel to reduce the latency to within 10 microseconds. The computing power of multiple cards can be expanded almost linearly, and the utilization rate exceeds 90%. It's like a group of well-trained elite teams working together to charge forward and overcome complex model problems. I sincerely recommend it to my fellow professionals. Come to Burncloud and press the accelerator button for your scientific research and projects!"},"title":{"matchLevel":"none","matchedWords":[],"value":"The Ideal \"Computing Power Engine\" for Deep Learning \u2013 Burncloud GPU Platform"}},"_tags":["story","author_chlor","story_42480075","ask_hn"],"author":"chlor","created_at":"2024-12-21T15:03:02Z","created_at_i":1734793382,"num_comments":0,"objectID":"42480075","points":2,"story_id":42480075,"story_text":"On the journey of exploring deep learning, the performance of GPU is the key to the speed of progress. I had tried multiple platforms, attempting to find the "computing power engine" that could carry my deep learning dreams. However, the reality often fell short of expectations. Stuttering and latency were like shadows, restricting the speed of model training. When dealing with large-scale datasets and complex network architectures, the computing efficiency of these platforms seemed inadequate, and I was constantly tormented by stuttering and latency. Even more troublesome was that video memory management also became a major obstacle on my way forward. When handling tasks rich in sequential data such as video action recognition, the GPU video memory of common platforms was often stretched to the limit. When processing batches of high-definition video frames in long sequences with 16GB or 24GB of video memory, video memory overflow became a common occurrence. It was not until I encountered Burncloud (https://www.burncloud.com/835.html) that I truly unlocked the door to efficient deep learning.\n Its computing efficiency is truly astonishing. When I trained a conventional convolutional neural network for image classification on a certain commonly used old cloud platform, the average time for a single iteration was 25 minutes. However, the NVIDIA A100 GPU equipped on Burncloud, with its 6,912 CUDA cores combined with an advanced architecture, reduced the iteration time for the same model and dataset to around 4 minutes. The efficiency improvement was more than 5 times, as if a high-speed motor had been installed for model training, rapidly pushing the loss function towards the optimal value.\n Video memory management is also outstanding. When handling tasks rich in sequential data such as video action recognition, the GPU of common platforms usually has 16GB or 24GB of video memory. When processing batches of high-definition video frames in long sequences, there is a more than 70% probability of encountering video memory overflow. Burncloud is different. With large-capacity video memory starting from 40GB and reaching up to 80GB, it can easily accommodate the features of massive video frames. The whole training process is as stable as a mountain, and even complex 3D convolutional long short-term memory networks can set sail smoothly, ensuring that the exploration of deep learning does not run aground.\n Multi-GPU collaboration is also a highlight of Burncloud. In distributed training for language models of the GPT scale, other platforms are limited by internal interconnection bandwidth, and the data synchronization latency between cards exceeds 50 microseconds, with a computing power utilization rate of less than 60%. Burncloud relies on the NVLink ultra-high-speed channel to reduce the latency to within 10 microseconds. The computing power of multiple cards can be expanded almost linearly, and the utilization rate exceeds 90%. It's like a group of well-trained elite teams working together to charge forward and overcome complex model problems. I sincerely recommend it to my fellow professionals. Come to Burncloud and press the accelerator button for your scientific research and projects!","title":"The Ideal \"Computing Power Engine\" for Deep Learning \u2013 Burncloud GPU Platform","updated_at":"2024-12-21T16:25:52Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"angiosperm"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Marshall Eubanks wants to send a swarm of 1000 flyweight probes to the\nnearest star, reaching it in 20 years.
https://www.researchgate.net/publication/373833951_Swarming_Proxima_Centauri_Optical_Communications_Over_Interstellar_Distances
However:
1. Sending just 1000 probes is fatally timid. We should think in terms of millions of identical probes. Produced in volume, their price-per drops to near nothing. The entire probe could have just one chip plus a couple of capacitors and coils.
2. Launching just one probe swarm would be wasteful: having built the solar-pumped launching laser/maser, we have no reason to turn it off: keep launching probes throughout the life of the mission. The cost of a million probes is small vs. the cost of the project overall. Over the next 20 years, launch 1000 probes (or more: 10,000, 50,000) each week. Having accelerated to cruising speed, probes turn to fly edge-on, minimizing collision risk.
3. With the launch apparatus operating throughout the mission, there is no need to accelerate to cruising speed immediately. This allows a much lesser beam intensity, and radically reduces G loading on the probes.
4. There is no value in having probes formed up in a plane at the destination. A dispersed cloud of a million probes flying through continuously for 20 years offers enormously greater value. The first few thousand through the system can identify interesting bodies, and subsequent swarms can tack against starlight and concentrate on those. This eliminates the "predict position of planetary bodies" problem.
5. Probes spend almost all their flight time close to zero K. This creates design opportunities denied to other projects. The sail surface may be an atomically thin wafer of superconductor, providing perfect reflection of any long-enough wavelength, stiffened by its own field.
6. Trying to transmit data home via laser would be a monumental blunder: the starlight behind the probes interferes with any optical signal. Worse, at optical frequency, phased-array transmission by clouds of probes is impossible; but at ordinary radio frequency, a cloud may transmit as a phased array with no difficulty, having triangulated their relative positions and synchronized to a pilot signal from home. Receivers throughout the asteroid belt may remain synchronized to operate as a phased array using atomic clocks (as was used to image M87).
7. There is no need for probes at the destination to transmit their data all the way home. They need only send to the next swarm coming up a light-day or light-week behind, which may relay data to the next.
8. Upon approaching the destination, the swarm may focus incident light (laser/maser/starlight) on a few, selected probes to slow them down enough to enter orbit. Subsequent swarms flying through can provide phased-array communication, and assist altering their orbits, with the occasional additional probe joining them.
9. A launch laser may be a simple laser cavity pumped with sunlight reflected from an array of monochromatic mirrors ranged around it, with no conversion loss. Monochromatic mirrors are easily produced by depositing alternating layers of transparent material with different dielectric constant but identical thickness. Given continuous operation throughout the mission, a much lower power output (10GW? 1GW?) becomes wholly adequate.
10. But consider the alternative of launching with an array of masers, instead. Masers are much more easily constructed and operated. An array of independent masers can be maintained in phase with the aid of atomic clocks.
11. With two swarms sent out in different directions, you get a phased-array radio telescope light-years across that can (e.g.) image the surface of a pulsar, using the pulsar itself to synchronize. The second swarm could get away with many fewer probes than the first, but you might better target another star with it.
Can this be improved upon?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Project \"Breakthrough Starshot\" Improvements"}},"_tags":["story","author_angiosperm","story_39191932","ask_hn"],"author":"angiosperm","created_at":"2024-01-30T16:20:08Z","created_at_i":1706631608,"num_comments":0,"objectID":"39191932","points":2,"story_id":39191932,"story_text":"Marshall Eubanks wants to send a swarm of 1000 flyweight probes to the\nnearest star, reaching it in 20 years.
https://www.researchgate.net/publication/373833951_Swarming_Proxima_Centauri_Optical_Communications_Over_Interstellar_Distances
However:
1. Sending just 1000 probes is fatally timid. We should think in terms of millions of identical probes. Produced in volume, their price-per drops to near nothing. The entire probe could have just one chip plus a couple of capacitors and coils.
2. Launching just one probe swarm would be wasteful: having built the solar-pumped launching laser/maser, we have no reason to turn it off: keep launching probes throughout the life of the mission. The cost of a million probes is small vs. the cost of the project overall. Over the next 20 years, launch 1000 probes (or more: 10,000, 50,000) each week. Having accelerated to cruising speed, probes turn to fly edge-on, minimizing collision risk.
3. With the launch apparatus operating throughout the mission, there is no need to accelerate to cruising speed immediately. This allows a much lesser beam intensity, and radically reduces G loading on the probes.
4. There is no value in having probes formed up in a plane at the destination. A dispersed cloud of a million probes flying through continuously for 20 years offers enormously greater value. The first few thousand through the system can identify interesting bodies, and subsequent swarms can tack against starlight and concentrate on those. This eliminates the "predict position of planetary bodies" problem.
5. Probes spend almost all their flight time close to zero K. This creates design opportunities denied to other projects. The sail surface may be an atomically thin wafer of superconductor, providing perfect reflection of any long-enough wavelength, stiffened by its own field.
6. Trying to transmit data home via laser would be a monumental blunder: the starlight behind the probes interferes with any optical signal. Worse, at optical frequency, phased-array transmission by clouds of probes is impossible; but at ordinary radio frequency, a cloud may transmit as a phased array with no difficulty, having triangulated their relative positions and synchronized to a pilot signal from home. Receivers throughout the asteroid belt may remain synchronized to operate as a phased array using atomic clocks (as was used to image M87).
7. There is no need for probes at the destination to transmit their data all the way home. They need only send to the next swarm coming up a light-day or light-week behind, which may relay data to the next.
8. Upon approaching the destination, the swarm may focus incident light (laser/maser/starlight) on a few, selected probes to slow them down enough to enter orbit. Subsequent swarms flying through can provide phased-array communication, and assist altering their orbits, with the occasional additional probe joining them.
9. A launch laser may be a simple laser cavity pumped with sunlight reflected from an array of monochromatic mirrors ranged around it, with no conversion loss. Monochromatic mirrors are easily produced by depositing alternating layers of transparent material with different dielectric constant but identical thickness. Given continuous operation throughout the mission, a much lower power output (10GW? 1GW?) becomes wholly adequate.
10. But consider the alternative of launching with an array of masers, instead. Masers are much more easily constructed and operated. An array of independent masers can be maintained in phase with the aid of atomic clocks.
11. With two swarms sent out in different directions, you get a phased-array radio telescope light-years across that can (e.g.) image the surface of a pulsar, using the pulsar itself to synchronize. The second swarm could get away with many fewer probes than the first, but you might better target another star with it.
Can this be improved upon?","title":"Project \"Breakthrough Starshot\" Improvements","updated_at":"2024-09-20T16:19:20Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"angiosperm"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Marshall Eubanks wants to send a swarm of 1000 flyweight probes to the nearest star, reaching it in 20 years.
https://www.researchgate.net/publication/373833951_Swarming_Proxima_Centauri_Optical_Communications_Over_Interstellar_Distances
However:
1. Sending just 1000 probes is fatally timid. We should think in terms of millions of identical probes. Produced in volume, their price-per drops to near nothing, even when more capable. The entire probe could have just one graphene chip plus a couple of capacitors and coils.
2. Launching just one probe swarm would be wasteful: having built the solar-pumped launching laser/maser, we have no reason to turn it off: keep launching probes throughout the life of the mission. The cost of a million probes is small vs. the cost of the project overall. Over the next 20 years, launch 1000 probes (or more: 10,000, 50,000) each week. Having accelerated to cruising speed, probes turn to fly edge-on, minimizing erosion.
3. With the launch apparatus operating throughout the mission, there is no need to accelerate to cruising speed immediately. This allows a much lesser beam intensity, and radically reduces G loading on the probes.
4. There is no value in having probes form up in a plane at the destination. A dispersed cloud of a million probes flying through continuously for 20 years offers enormously greater value. The first few thousand through the system can identify interesting bodies, and subsequent swarms can tack against starlight and concentrate on those. This eliminates the "predict position of planetary bodies" problem.
5. Probes spend almost all their flight time close to zero K. This creates design opportunities denied to other projects. The sail surface may be an atomically thin wafer of superconductor, providing perfect reflection of any long-enough wavelength, stiffened by its own field.
6. Trying to transmit data home via laser would be a monumental blunder: at optical wavelengths, phased-array transmission by clouds of probes is impossible; but at ordinary RF, a cloud may transmit as a phased array with no difficulty, having triangulated their relative positions and synchronized to a pilot signal from home. Receivers throughout the asteroid belt may remain synchronized to operate as a phased array using atomic clocks (as was used to image M87).
7. There is no need for probes at the destination to transmit their data all the way home. They need only send to the next swarm coming up a light-day or light-week behind, which may relay data to the next.
8. Upon approaching the destination, the swarm may focus incident light (laser/maser/starlight) on a few, selected probes to slow them down enough to enter orbit. Subsequent swarms flying through can provide phased-array communication, and assist altering their orbits, with the occasional additional probe joining them.
9. A launch laser may be a simple laser cavity pumped with sunlight reflected from an array of monochromatic mirrors ranged around it, with no conversion loss. Monochromatic mirrors are easily produced by depositing alternating layers of transparent material with different dielectric constant but identical thickness. Given continuous operation throughout the mission, a much lower power output (10GW? 1GW?) becomes wholly adequate.
10. But consider the alternative of launching with an array of masers, instead. Masers are much more easily constructed and operated. An array of independent masers can be maintained in phase with the aid of atomic clocks.
11. With two swarms sent out in different directions, you get a phased-array radio telescope light-years across that can (e.g.) image the surface of a pulsar, using the pulsar itself to synchronize. The second swarm could get away with many fewer probes than the first, but you might better target another star with it.
Can this be improved upon?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Project Starswarm"}},"_tags":["story","author_angiosperm","story_39252584","ask_hn"],"author":"angiosperm","created_at":"2024-02-04T17:58:08Z","created_at_i":1707069488,"num_comments":0,"objectID":"39252584","points":1,"story_id":39252584,"story_text":"Marshall Eubanks wants to send a swarm of 1000 flyweight probes to the nearest star, reaching it in 20 years.
https://www.researchgate.net/publication/373833951_Swarming_Proxima_Centauri_Optical_Communications_Over_Interstellar_Distances
However:
1. Sending just 1000 probes is fatally timid. We should think in terms of millions of identical probes. Produced in volume, their price-per drops to near nothing, even when more capable. The entire probe could have just one graphene chip plus a couple of capacitors and coils.
2. Launching just one probe swarm would be wasteful: having built the solar-pumped launching laser/maser, we have no reason to turn it off: keep launching probes throughout the life of the mission. The cost of a million probes is small vs. the cost of the project overall. Over the next 20 years, launch 1000 probes (or more: 10,000, 50,000) each week. Having accelerated to cruising speed, probes turn to fly edge-on, minimizing erosion.
3. With the launch apparatus operating throughout the mission, there is no need to accelerate to cruising speed immediately. This allows a much lesser beam intensity, and radically reduces G loading on the probes.
4. There is no value in having probes form up in a plane at the destination. A dispersed cloud of a million probes flying through continuously for 20 years offers enormously greater value. The first few thousand through the system can identify interesting bodies, and subsequent swarms can tack against starlight and concentrate on those. This eliminates the "predict position of planetary bodies" problem.
5. Probes spend almost all their flight time close to zero K. This creates design opportunities denied to other projects. The sail surface may be an atomically thin wafer of superconductor, providing perfect reflection of any long-enough wavelength, stiffened by its own field.
6. Trying to transmit data home via laser would be a monumental blunder: at optical wavelengths, phased-array transmission by clouds of probes is impossible; but at ordinary RF, a cloud may transmit as a phased array with no difficulty, having triangulated their relative positions and synchronized to a pilot signal from home. Receivers throughout the asteroid belt may remain synchronized to operate as a phased array using atomic clocks (as was used to image M87).
7. There is no need for probes at the destination to transmit their data all the way home. They need only send to the next swarm coming up a light-day or light-week behind, which may relay data to the next.
8. Upon approaching the destination, the swarm may focus incident light (laser/maser/starlight) on a few, selected probes to slow them down enough to enter orbit. Subsequent swarms flying through can provide phased-array communication, and assist altering their orbits, with the occasional additional probe joining them.
9. A launch laser may be a simple laser cavity pumped with sunlight reflected from an array of monochromatic mirrors ranged around it, with no conversion loss. Monochromatic mirrors are easily produced by depositing alternating layers of transparent material with different dielectric constant but identical thickness. Given continuous operation throughout the mission, a much lower power output (10GW? 1GW?) becomes wholly adequate.
10. But consider the alternative of launching with an array of masers, instead. Masers are much more easily constructed and operated. An array of independent masers can be maintained in phase with the aid of atomic clocks.
11. With two swarms sent out in different directions, you get a phased-array radio telescope light-years across that can (e.g.) image the surface of a pulsar, using the pulsar itself to synchronize. The second swarm could get away with many fewer probes than the first, but you might better target another star with it.
Can this be improved upon?","title":"Project Starswarm","updated_at":"2024-09-20T16:17:51Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vessenes"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"So, Lecun has been quite public saying that he believes LLMs will never fix hallucinations because, essentially, the token choice method at each step leads to runaway errors -- these can't be damped mathematically.
In exchange, he offers the idea that we should have something that is an 'energy minimization' architecture; as I understand it, this would have a concept of the 'energy' of an entire response, and training would try and minimize that.
Which is to say, I don't fully understand this. That said, I'm curious to hear what ML researchers think about Lecun's take, and if there's any engineering done around it. I can't find much after the release of ijepa from his group."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Any insider takes on Yann LeCun's push against current architectures?"}},"_tags":["story","author_vessenes","story_43325049","ask_hn"],"author":"vessenes","children":[43325670,43348309,43365259,43365343,43365355,43365365,43365456,43365503,43365581,43365660,43365718,43365762,43365788,43365809,43366061,43366161,43366174,43366246,43366306,43366463,43366681,43366931,43367047,43367126,43367415,43367519,43367529,43367543,43367642,43367692,43367709,43367717,43367746,43367760,43367765,43367802,43368085,43368609,43368660,43368709,43368787,43368832,43368865,43369058,43369083,43369131,43369139,43369202,43369250,43369367,43369370,43369418,43369854,43369868,43370048,43371664,43372073,43372206,43372471,43387067],"created_at":"2025-03-10T19:41:37Z","created_at_i":1741635697,"num_comments":325,"objectID":"43325049","points":385,"story_id":43325049,"story_text":"So, Lecun has been quite public saying that he believes LLMs will never fix hallucinations because, essentially, the token choice method at each step leads to runaway errors -- these can't be damped mathematically.
In exchange, he offers the idea that we should have something that is an 'energy minimization' architecture; as I understand it, this would have a concept of the 'energy' of an entire response, and training would try and minimize that.
Which is to say, I don't fully understand this. That said, I'm curious to hear what ML researchers think about Lecun's take, and if there's any engineering done around it. I can't find much after the release of ijepa from his group.","title":"Ask HN: Any insider takes on Yann LeCun's push against current architectures?","updated_at":"2026-10-08T02:53:09Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"hxii"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Hi there. My name is Paul. I\u2019m a developer, an explorer and I\u2019ve got ADHD. As of the beginning of August, I\u2019m also unemployed.
This last part, along with everything else this past year, has caused me a great deal of stress.
While our balcony zen garden project is yet to be completed, I had an idea to create a virtual one that everyone can use.
It's an idea that, unfortunately, I\u2019ve been postponing for a while now, mostly because I have no fucking idea how to do it as I don\u2019t know JavaScript, and I don't have the capacity to learn it right now.
So I let perfect be the enemy of good and, well... just kept the idea to myself.
Then I said "fuck it" and used AI to make the thing I really wanted to make.
I realized I didn't want "perfect". I wanted "good enough".
I tweaked, added, removed, drew, researched, questioned, tested... I just wasn't the one coding it.
So now, instead of occupying my brain, it now lives on the internet for others to enjoy.
Yes, there\u2019s something noble about making something entirely on your own, but what good is an idea that just sits in my head?
So here I am. I made the thing. The weird, little, quiet koi pond.
The silly project of passion. The little corner of the internet to let strangers watch fish quietly, together.
I hope this pond helps you as much as it helped me."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Koi.rest \u2013 watch some fish and regain your balance"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://koi.rest"}},"_tags":["story","author_hxii","story_49837006","show_hn"],"author":"hxii","children":[49837291,49837446,49837516,49837765,49837791,49837808,49837833,49837863,49837875,49837916,49837928,49838038,49838148,49838237,49838258,49838265,49838399,49838467,49838568,49839173,49839439,49839559,49839584,49839805,49840296,49841240,49841570,49841899,49843129,49844366,49844992,49859617,49865646],"created_at":"2026-09-24T21:33:03Z","created_at_i":1790285583,"num_comments":65,"objectID":"49837006","points":228,"story_id":49837006,"story_text":"Hi there. My name is Paul. I\u2019m a developer, an explorer and I\u2019ve got ADHD. As of the beginning of August, I\u2019m also unemployed.
This last part, along with everything else this past year, has caused me a great deal of stress.
While our balcony zen garden project is yet to be completed, I had an idea to create a virtual one that everyone can use.
It's an idea that, unfortunately, I\u2019ve been postponing for a while now, mostly because I have no fucking idea how to do it as I don\u2019t know JavaScript, and I don't have the capacity to learn it right now.
So I let perfect be the enemy of good and, well... just kept the idea to myself.
Then I said "fuck it" and used AI to make the thing I really wanted to make.
I realized I didn't want "perfect". I wanted "good enough".
I tweaked, added, removed, drew, researched, questioned, tested... I just wasn't the one coding it.
So now, instead of occupying my brain, it now lives on the internet for others to enjoy.
Yes, there\u2019s something noble about making something entirely on your own, but what good is an idea that just sits in my head?
So here I am. I made the thing. The weird, little, quiet koi pond.
The silly project of passion. The little corner of the internet to let strangers watch fish quietly, together.
I hope this pond helps you as much as it helped me.","title":"Show HN: Koi.rest \u2013 watch some fish and regain your balance","updated_at":"2026-10-08T20:24:11Z","url":"https://koi.rest"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jv22222"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["sail","research"],"value":"Hi HN, I'm Justin.
Breadcrumb records everything you do on your Mac (screen + meetings + AI transcripts + what you and your AI decided) and turns it into memory your AI can search. It's local and encrypted.
You can also teach it rules by talking to it and it makes sure the right rules turn up in the right context. Works with Claude Code / Codex / Cursor / opencode.
All of this is exposed to your AI as 30+ MCP tools (here's the definitions): https://innerloop.works/breadcrumb/mcp
I started it in June because I wanted to understand what was going on with my AI dev work. Then it just kept growing as I jammed on it 10 hours a day for 3 months.
I started by recording every screen so I could jump back to a moment and see what I'd been doing with the AI. Then I added MCP so the AI could search it. Then time tracking. Then meetings with transcripts and timestamped screenshots.
After that I started feeding the data back in. At first it was just a few relevant memories in the prompt. Then I started adding rules and rules modules.
Then I started recording the AI transcripts and all the other stuff the AI did (subagents, tools etc). I wanted a complete record so I could go back and see what the AI did and why.
I found the more data the AI had the more useful it became. One example. I got off a zoom call with the team at the day job and said:
"Review the meeting I just had, find all the bugs we discussed, and create JIRA tickets with screenshots."
It read the transcript. Found every bug we mentioned, pulled the screens from the screen timeline, spun up my local server, diagnosed each bug, reviewed the code, created 14 tickets in JIRA with diagnosis and likely fixes, attached 12 screenshots as evidence, and created a nice Slack message to send to my colleagues.
I never built a workflow for any of that. Claude had the meeting record, Breadcrumb tools, and rules I'd accumulated. It basically chained the steps together itself.
My favorite thing about the system is observability of my own work as an AI developer. I can trace a bug all the way back to the first thing I said (or the first thing the AI did) that caused it. Then I can add a rule so it doesn't happen again.
Anyway, my working system is simply Claude Desktop + Breadcrumb + Markdown. I find I don't need anything else.
It's local software for one person. Local models for transcription and diarization. Data encrypted with SQLCipher and a passphrase. You can exclude any apps, websites, words and phrases.
Breadcrumb itself only makes one network call which is a daily update check. (but of course your AI will use it and that is local or cloud depending on what you use).
It runs cool about 12% of one core on average (M series mac has 8+). You can set it to just save OCR text or to save screenshots. If just OCR it's about 12mb day (a lot of telemetry, indexing etc). If you record screenshots on meetings it's about 120mb/hour (1 every 2 secs). It can also record screenshots of everything that's heavy at about 160mb/day. It does local classification/naming with ollama with Gemma 3 4B (2.5 GB).
It's 100% free. There's no account you can just download and start using it. I'll monetize it via E2EE cloud sync and team sync. It's open beta. Just one developer working on it. Requires 16GB+ M Series Mac.
Thanks for reading!
https://innerloop.works/breadcrumb (homepage)
https://innerloop.works/breadcrumb/download (direct download)
https://innerloop.works/breadcrumb/mcp (mcp definition)"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Breadcrumb, record everything on your mac + context manager for AI"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://innerloop.works/breadcrumb"}},"_tags":["story","author_jv22222","story_49924943","show_hn"],"author":"jv22222","children":[49925986,49927447,49927457,49932154,49938466,49944236,49944353,49958242],"created_at":"2026-10-01T17:58:37Z","created_at_i":1790877517,"num_comments":9,"objectID":"49924943","points":49,"story_id":49924943,"story_text":"Hi HN, I'm Justin.
Breadcrumb records everything you do on your Mac (screen + meetings + AI transcripts + what you and your AI decided) and turns it into memory your AI can search. It's local and encrypted.
You can also teach it rules by talking to it and it makes sure the right rules turn up in the right context. Works with Claude Code / Codex / Cursor / opencode.
All of this is exposed to your AI as 30+ MCP tools (here's the definitions): https://innerloop.works/breadcrumb/mcp
I started it in June because I wanted to understand what was going on with my AI dev work. Then it just kept growing as I jammed on it 10 hours a day for 3 months.
I started by recording every screen so I could jump back to a moment and see what I'd been doing with the AI. Then I added MCP so the AI could search it. Then time tracking. Then meetings with transcripts and timestamped screenshots.
After that I started feeding the data back in. At first it was just a few relevant memories in the prompt. Then I started adding rules and rules modules.
Then I started recording the AI transcripts and all the other stuff the AI did (subagents, tools etc). I wanted a complete record so I could go back and see what the AI did and why.
I found the more data the AI had the more useful it became. One example. I got off a zoom call with the team at the day job and said:
"Review the meeting I just had, find all the bugs we discussed, and create JIRA tickets with screenshots."
It read the transcript. Found every bug we mentioned, pulled the screens from the screen timeline, spun up my local server, diagnosed each bug, reviewed the code, created 14 tickets in JIRA with diagnosis and likely fixes, attached 12 screenshots as evidence, and created a nice Slack message to send to my colleagues.
I never built a workflow for any of that. Claude had the meeting record, Breadcrumb tools, and rules I'd accumulated. It basically chained the steps together itself.
My favorite thing about the system is observability of my own work as an AI developer. I can trace a bug all the way back to the first thing I said (or the first thing the AI did) that caused it. Then I can add a rule so it doesn't happen again.
Anyway, my working system is simply Claude Desktop + Breadcrumb + Markdown. I find I don't need anything else.
It's local software for one person. Local models for transcription and diarization. Data encrypted with SQLCipher and a passphrase. You can exclude any apps, websites, words and phrases.
Breadcrumb itself only makes one network call which is a daily update check. (but of course your AI will use it and that is local or cloud depending on what you use).
It runs cool about 12% of one core on average (M series mac has 8+). You can set it to just save OCR text or to save screenshots. If just OCR it's about 12mb day (a lot of telemetry, indexing etc). If you record screenshots on meetings it's about 120mb/hour (1 every 2 secs). It can also record screenshots of everything that's heavy at about 160mb/day. It does local classification/naming with ollama with Gemma 3 4B (2.5 GB).
It's 100% free. There's no account you can just download and start using it. I'll monetize it via E2EE cloud sync and team sync. It's open beta. Just one developer working on it. Requires 16GB+ M Series Mac.
Thanks for reading!
https://innerloop.works/breadcrumb (homepage)
https://innerloop.works/breadcrumb/download (direct download)
https://innerloop.works/breadcrumb/mcp (mcp definition)","title":"Show HN: Breadcrumb, record everything on your mac + context manager for AI","updated_at":"2026-10-08T04:21:53Z","url":"https://innerloop.works/breadcrumb"}],"hitsPerPage":50,"nbHits":28,"nbPages":1,"page":0,"params":"query=Sail+Research&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":18,"processingTimingsMS":{"_request":{"queue":46,"roundTrip":15},"afterFetch":{"format":{"highlighting":2,"total":2}},"fetch":{"query":12,"scanning":3,"total":16},"total":18},"query":"Sail Research","serverTimeMS":67}