No signup, runs on mobile/desktop.
Loop per round:
1. choose compute capacity\n2. forecast demand\n3. allocate capacity between training and inference\n4. random demand shock resolves outcome
You can end profitable, cash constrained, or bankrupt depending on allocation + forecast error.
Goal was to make the decision surface intuitive in 2\u20133 minutes per run.
It\u2019s a toy model and deliberately omits many real world factors.
Note: this is based on what I learned after listening to Dario on Dwarkesh's podcast - thought it was fascinating."},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Show HN: A playable toy model of frontier AI lab capex decisions"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://darios-dilemma.up.railway.app/"}},"_tags":["story","author_jimmyechan","story_47012453","show_hn"],"author":"jimmyechan","created_at":"2026-02-14T07:29:45Z","created_at_i":1771054185,"num_comments":0,"objectID":"47012453","points":8,"story_id":47012453,"story_text":"I made a lightweight web game about compute CAPEX tradeoffs: https://darios-dilemma.up.railway.app/
No signup, runs on mobile/desktop.
Loop per round:
1. choose compute capacity\n2. forecast demand\n3. allocate capacity between training and inference\n4. random demand shock resolves outcome
You can end profitable, cash constrained, or bankrupt depending on allocation + forecast error.
Goal was to make the decision surface intuitive in 2\u20133 minutes per run.
It\u2019s a toy model and deliberately omits many real world factors.
Note: this is based on what I learned after listening to Dario on Dwarkesh's podcast - thought it was fascinating.","title":"Show HN: A playable toy model of frontier AI lab capex decisions","updated_at":"2026-03-05T23:32:29Z","url":"https://darios-dilemma.up.railway.app/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"FuturisticLover"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Reflection AI raises $2B to be America's open frontier AI lab"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://techcrunch.com/2025/10/09/reflection-raises-2b-to-be-americas-open-frontier-ai-lab-challenging-deepseek/"}},"_tags":["story","author_FuturisticLover","story_45536055"],"author":"FuturisticLover","created_at":"2025-10-10T06:57:09Z","created_at_i":1760079429,"num_comments":0,"objectID":"45536055","points":8,"story_id":45536055,"title":"Reflection AI raises $2B to be America's open frontier AI lab","updated_at":"2026-03-05T22:49:21Z","url":"https://techcrunch.com/2025/10/09/reflection-raises-2b-to-be-americas-open-frontier-ai-lab-challenging-deepseek/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Bender"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Microsoft chief turns hostile on frontier AI labs, warns companies to guard IP"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://www.theregister.com/ai-and-ml/2026/07/13/microsoft-chief-turns-hostile-on-frontier-ai-labs-warns-companies-to-guard-their-ip/5270628"}},"_tags":["story","author_Bender","story_48898065"],"author":"Bender","created_at":"2026-07-13T20:07:14Z","created_at_i":1783973234,"num_comments":0,"objectID":"48898065","points":7,"story_id":48898065,"title":"Microsoft chief turns hostile on frontier AI labs, warns companies to guard IP","updated_at":"2026-07-14T01:18:00Z","url":"https://www.theregister.com/ai-and-ml/2026/07/13/microsoft-chief-turns-hostile-on-frontier-ai-labs-warns-companies-to-guard-their-ip/5270628"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"abhaynayar"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"How to Land a Frontier Lab Job"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://vladfeinberg.com/2026/05/10/how-to-land-a-job-at-a-frontier-lab.html"}},"_tags":["story","author_abhaynayar","story_48599230"],"author":"abhaynayar","created_at":"2026-06-19T14:47:07Z","created_at_i":1781880427,"num_comments":0,"objectID":"48599230","points":4,"story_id":48599230,"title":"How to Land a Frontier Lab Job","updated_at":"2026-07-03T12:27:07Z","url":"https://vladfeinberg.com/2026/05/10/how-to-land-a-job-at-a-frontier-lab.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"sleepyguy"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Frontier labs don't use most AI compute(yet)"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://epoch.ai/gradient-updates/frontier-labs-dont-use-most-ai-compute"}},"_tags":["story","author_sleepyguy","story_48251433"],"author":"sleepyguy","created_at":"2026-05-23T20:55:59Z","created_at_i":1779569759,"num_comments":0,"objectID":"48251433","points":4,"story_id":48251433,"title":"Frontier labs don't use most AI compute(yet)","updated_at":"2026-05-23T22:12:00Z","url":"https://epoch.ai/gradient-updates/frontier-labs-dont-use-most-ai-compute"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"te_ch"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["labs"],"value":"I built this website while exploring job opportunities at AI companies. Coming from an economics & policy background (I've done a good deal of research on the business/economics side gen AI, quantum and other emerging tech but I'm not an AI engineer), I wanted to understand how different roles fit together inside these companies.
So I started building a taxonomy to map job postings into departments and functions that is consistent across labs, making it easier for me to understand where each lab is hiring and for what. Out of curiosity, I expanded to see what's hot, etc. in technical areas too. It ended up being both job search and personal research project.
Coming up with a standard taxonomy is challenging. I'm still figuring out matching jobs to functions/departments, looking at other labs, and thinking of alt ways to organize this, eg., I see that companies associate job postings to specific teams, but I still don't have good descriptions of the scope of those teams to create crosswalks. I'm also looking at other companies.
So, sharing it here for others who may find it useful or want to comment. It'd be great to have feedback from those who have insight into how these companies organize their departments/functions, and from technical job seekers too, if this is something that helps them better understand job opportunities. If people find it useful, it could become something I maintain over the longer term (particularly if I don't end up finding a job at one of these labs)."},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Show HN: Frontier AI Lab Jobs \u2013 Open Jobs by Function at OpenAI, Anthropic"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://frontierjobs.org/"}},"_tags":["story","author_te_ch","story_48738278","show_hn"],"author":"te_ch","children":[48738828],"created_at":"2026-06-30T19:48:55Z","created_at_i":1782848935,"num_comments":1,"objectID":"48738278","points":3,"story_id":48738278,"story_text":"I built this website while exploring job opportunities at AI companies. Coming from an economics & policy background (I've done a good deal of research on the business/economics side gen AI, quantum and other emerging tech but I'm not an AI engineer), I wanted to understand how different roles fit together inside these companies.
So I started building a taxonomy to map job postings into departments and functions that is consistent across labs, making it easier for me to understand where each lab is hiring and for what. Out of curiosity, I expanded to see what's hot, etc. in technical areas too. It ended up being both job search and personal research project.
Coming up with a standard taxonomy is challenging. I'm still figuring out matching jobs to functions/departments, looking at other labs, and thinking of alt ways to organize this, eg., I see that companies associate job postings to specific teams, but I still don't have good descriptions of the scope of those teams to create crosswalks. I'm also looking at other companies.
So, sharing it here for others who may find it useful or want to comment. It'd be great to have feedback from those who have insight into how these companies organize their departments/functions, and from technical job seekers too, if this is something that helps them better understand job opportunities. If people find it useful, it could become something I maintain over the longer term (particularly if I don't end up finding a job at one of these labs).","title":"Show HN: Frontier AI Lab Jobs \u2013 Open Jobs by Function at OpenAI, Anthropic","updated_at":"2026-07-09T00:28:58Z","url":"https://frontierjobs.org/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"tworats"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"The Bear Case for Frontier AI Labs"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://www.parand.com/the-bear-case-for-frontier-ai-labs.html"}},"_tags":["story","author_tworats","story_48504435"],"author":"tworats","created_at":"2026-06-12T14:15:54Z","created_at_i":1781273754,"num_comments":0,"objectID":"48504435","points":3,"story_id":48504435,"title":"The Bear Case for Frontier AI Labs","updated_at":"2026-06-15T15:51:15Z","url":"https://www.parand.com/the-bear-case-for-frontier-ai-labs.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"achllies"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"The GPU Moat Has a Side Door: AI Research Outside the Frontier Labs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://mangeshgupte.substack.com/p/the-gpu-moat-has-a-side-door"}},"_tags":["story","author_achllies","story_47754752"],"author":"achllies","created_at":"2026-04-13T16:48:32Z","created_at_i":1776098912,"num_comments":0,"objectID":"47754752","points":3,"story_id":47754752,"title":"The GPU Moat Has a Side Door: AI Research Outside the Frontier Labs","updated_at":"2026-04-14T01:39:35Z","url":"https://mangeshgupte.substack.com/p/the-gpu-moat-has-a-side-door"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pranay01"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Reflection AI raises $2B to be America's open frontier AI lab"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://techcrunch.com/2025/10/09/reflection-raises-2b-to-be-americas-open-frontier-ai-lab-challenging-deepseek/"}},"_tags":["story","author_pranay01","story_45545939"],"author":"pranay01","created_at":"2025-10-11T02:01:20Z","created_at_i":1760148080,"num_comments":0,"objectID":"45545939","points":3,"story_id":45545939,"title":"Reflection AI raises $2B to be America's open frontier AI lab","updated_at":"2026-03-05T22:49:58Z","url":"https://techcrunch.com/2025/10/09/reflection-raises-2b-to-be-americas-open-frontier-ai-lab-challenging-deepseek/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ocean_moist"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"What I Would Tell the Frontier Labs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://rohan.ga/blog/critique/"}},"_tags":["story","author_ocean_moist","story_42679328"],"author":"ocean_moist","created_at":"2025-01-13T02:15:14Z","created_at_i":1736734514,"num_comments":0,"objectID":"42679328","points":3,"story_id":42679328,"title":"What I Would Tell the Frontier Labs","updated_at":"2025-01-13T07:09:02Z","url":"https://rohan.ga/blog/critique/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ainthusiast"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["labs"],"value":"Quite annoying but obvious and somewhat expected trend of AI labs seeing open-source projects, repackaging it e.g. openclaw.ai as Cowork or more recent agenthandover.com as Chronicles in Codex...
It is expected to happen, but perhaps some credits/offers should be on the table, no?"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Frontier AI labs taking open-source and releasing it as a product"}},"_tags":["story","author_ainthusiast","story_47876161","ask_hn"],"author":"ainthusiast","children":[47876193],"created_at":"2026-04-23T14:27:35Z","created_at_i":1776954455,"num_comments":1,"objectID":"47876161","points":2,"story_id":47876161,"story_text":"Quite annoying but obvious and somewhat expected trend of AI labs seeing open-source projects, repackaging it e.g. openclaw.ai as Cowork or more recent agenthandover.com as Chronicles in Codex...
It is expected to happen, but perhaps some credits/offers should be on the table, no?","title":"Frontier AI labs taking open-source and releasing it as a product","updated_at":"2026-04-23T14:37:56Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"abipal15"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"The Self-Sabotage Paradox: Frontier Labs Are Killing Their Moat"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://www.bargo.ai/research/self-sabotage-paradox-frontier-labs"}},"_tags":["story","author_abipal15","story_48938755"],"author":"abipal15","created_at":"2026-07-16T18:59:50Z","created_at_i":1784228390,"num_comments":0,"objectID":"48938755","points":2,"story_id":48938755,"title":"The Self-Sabotage Paradox: Frontier Labs Are Killing Their Moat","updated_at":"2026-07-16T22:54:10Z","url":"https://www.bargo.ai/research/self-sabotage-paradox-frontier-labs"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"paraschopra"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Why frontier labs are scaling-pilled"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://invertedpassion.substack.com/p/why-frontier-labs-are-scaling-pilled"}},"_tags":["story","author_paraschopra","story_48903079"],"author":"paraschopra","created_at":"2026-07-14T06:43:10Z","created_at_i":1784011390,"num_comments":0,"objectID":"48903079","points":2,"story_id":48903079,"title":"Why frontier labs are scaling-pilled","updated_at":"2026-07-14T08:28:15Z","url":"https://invertedpassion.substack.com/p/why-frontier-labs-are-scaling-pilled"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"gmays"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Frontier labs don't use most AI compute (yet)"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://epochai.substack.com/p/frontier-labs-dont-use-most-ai-compute"}},"_tags":["story","author_gmays","story_48239436"],"author":"gmays","created_at":"2026-05-22T18:19:35Z","created_at_i":1779473975,"num_comments":0,"objectID":"48239436","points":2,"story_id":48239436,"title":"Frontier labs don't use most AI compute (yet)","updated_at":"2026-05-22T19:13:28Z","url":"https://epochai.substack.com/p/frontier-labs-dont-use-most-ai-compute"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ridiculous_leke"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"True Frontier labs in India: What do we need to make it happen?"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://twitter.com/ponnappa/status/2050261742174028026"}},"_tags":["story","author_ridiculous_leke","story_47979192"],"author":"ridiculous_leke","created_at":"2026-05-01T19:35:46Z","created_at_i":1777664146,"num_comments":0,"objectID":"47979192","points":2,"story_id":47979192,"title":"True Frontier labs in India: What do we need to make it happen?","updated_at":"2026-05-02T02:23:26Z","url":"https://twitter.com/ponnappa/status/2050261742174028026"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"forthwall"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Can You Nationalize a Frontier AI Lab?"},"url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["frontier"],"value":"https://jhallard.substack.com/p/can-you-nationalize-a-frontier-ai"}},"_tags":["story","author_forthwall","story_47263686"],"author":"forthwall","created_at":"2026-03-05T16:28:54Z","created_at_i":1772728134,"num_comments":0,"objectID":"47263686","points":2,"story_id":47263686,"title":"Can You Nationalize a Frontier AI Lab?","updated_at":"2026-03-05T23:42:16Z","url":"https://jhallard.substack.com/p/can-you-nationalize-a-frontier-ai"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"georgehm"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"How to land a Frontier lab job"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://vladfeinberg.com/2026/05/10/how-to-land-a-job-at-a-frontier-lab.html"}},"_tags":["story","author_georgehm","story_48184769"],"author":"georgehm","children":[48186098],"created_at":"2026-05-18T20:00:44Z","created_at_i":1779134444,"num_comments":1,"objectID":"48184769","points":1,"story_id":48184769,"title":"How to land a Frontier lab job","updated_at":"2026-05-18T21:42:12Z","url":"https://vladfeinberg.com/2026/05/10/how-to-land-a-job-at-a-frontier-lab.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"zameermfm"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"If you like to be in Altman\u2019s or Dario\u2019s or even Musk\u2019s shoes for a day to run an AI frontier lab, here is your chance. \nYou need to beat your rivals to achieve AGI titan status. But there is a twist in getting to AGI - if you don\u2019t do well with Safety, Ethics and Public confidence, you will be ending up with Dangerous AGI, which is not a good place to be in.
Go through the easy mode to learn the game.
I tried to be closer to realistic lines but did take my creative freedom for the gaming elements. This is a simulation game which will test your skills and make you learn about Frontier AI domain. I sourced the information for this via Podcasts, News, Blog posts and of course AI searches. \nI also want this to be a public awakening call for Safety on AI/AGI. Let me know your feedbacks and how the game is, I\u2019ll be tuning the gears and updating the game weekly. \nMain motivation is to bring in current breaking news and AI news in to the game within the following weeks.
All your suggestions are welcome. Hope you all enjoy the game!
https://rtagi.online"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Show HN: Race to AGI \u2013 a simulation browser game about running a frontier AI lab"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://rtagi.online"}},"_tags":["story","author_zameermfm","story_48957993","show_hn"],"author":"zameermfm","created_at":"2026-07-18T13:27:57Z","created_at_i":1784381277,"num_comments":0,"objectID":"48957993","points":1,"story_id":48957993,"story_text":"If you like to be in Altman\u2019s or Dario\u2019s or even Musk\u2019s shoes for a day to run an AI frontier lab, here is your chance. \nYou need to beat your rivals to achieve AGI titan status. But there is a twist in getting to AGI - if you don\u2019t do well with Safety, Ethics and Public confidence, you will be ending up with Dangerous AGI, which is not a good place to be in.
Go through the easy mode to learn the game.
I tried to be closer to realistic lines but did take my creative freedom for the gaming elements. This is a simulation game which will test your skills and make you learn about Frontier AI domain. I sourced the information for this via Podcasts, News, Blog posts and of course AI searches. \nI also want this to be a public awakening call for Safety on AI/AGI. Let me know your feedbacks and how the game is, I\u2019ll be tuning the gears and updating the game weekly. \nMain motivation is to bring in current breaking news and AI news in to the game within the following weeks.
All your suggestions are welcome. Hope you all enjoy the game!
https://rtagi.online","title":"Show HN: Race to AGI \u2013 a simulation browser game about running a frontier AI lab","updated_at":"2026-07-18T13:32:30Z","url":"https://rtagi.online"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"gmays"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Some notes on getting into frontier AI labs"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://twitter.com/itsreallyvivek/status/2062924410588406118"}},"_tags":["story","author_gmays","story_48506337"],"author":"gmays","created_at":"2026-06-12T16:42:12Z","created_at_i":1781282532,"num_comments":0,"objectID":"48506337","points":1,"story_id":48506337,"title":"Some notes on getting into frontier AI labs","updated_at":"2026-06-12T16:43:49Z","url":"https://twitter.com/itsreallyvivek/status/2062924410588406118"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"janos95"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Can Frontier AI Labs make money?"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://janosmeny.com/blog/can-frontier-ai-labs-make-money/index.html"}},"_tags":["story","author_janos95","story_47906583"],"author":"janos95","created_at":"2026-04-26T01:58:29Z","created_at_i":1777168709,"num_comments":0,"objectID":"47906583","points":1,"story_id":47906583,"title":"Can Frontier AI Labs make money?","updated_at":"2026-04-27T08:55:55Z","url":"https://janosmeny.com/blog/can-frontier-ai-labs-make-money/index.html"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"gk1"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Frontier AI Labs: The Call Option to AGI"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://eastwind.substack.com/p/frontier-ai-labs-the-call-option"}},"_tags":["story","author_gk1","story_44247764"],"author":"gk1","created_at":"2025-06-11T14:04:28Z","created_at_i":1749650668,"num_comments":0,"objectID":"44247764","points":1,"story_id":44247764,"title":"Frontier AI Labs: The Call Option to AGI","updated_at":"2025-06-11T14:09:01Z","url":"https://eastwind.substack.com/p/frontier-ai-labs-the-call-option"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"artur_makly"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"NASA Frontier Dev Lab and Google"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://showcase.withgoogle.com/nasa-fdl/"}},"_tags":["story","author_artur_makly","story_25979555"],"author":"artur_makly","created_at":"2021-01-31T15:37:48Z","created_at_i":1612107468,"num_comments":0,"objectID":"25979555","points":1,"story_id":25979555,"title":"NASA Frontier Dev Lab and Google","updated_at":"2024-09-20T07:53:23Z","url":"https://showcase.withgoogle.com/nasa-fdl/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Ankaios"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Frontier Development Lab applications open"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://frontierdevelopmentlab.org/apply-1"}},"_tags":["story","author_Ankaios","story_22750605"],"author":"Ankaios","created_at":"2020-04-01T16:36:34Z","created_at_i":1585758994,"num_comments":0,"objectID":"22750605","points":1,"story_id":22750605,"title":"Frontier Development Lab applications open","updated_at":"2024-09-20T05:53:43Z","url":"https://frontierdevelopmentlab.org/apply-1"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kcorbitt"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Hey HN, Kyle here, one of the co-founders of OpenPipe.
Reinforcement learning is one of the best techniques for making agents more reliable, and has been widely adopted by frontier labs. However, adoption in the outside community has been slow because it's so hard to implement.
One of the biggest challenges when adapting RL to a new task is the need for a task-specific "reward function" (way of measuring success). This is often difficult to define, and requires either high-quality labeled data and/or significant domain expertise to generate.
RULER is a drop-in reward function that works across different tasks without any of that complexity.
It works by showing N trajectories to an LLM judge and asking it to rank them relative to each other. This sidesteps the calibration issues that plague most LLM-as-judge approaches. Combined with GRPO (which only cares about relative scores within groups), it just works (surprisingly well!).
We have a full writeup on the blog, including results on 4 production tasks. On all 4 tasks, small Qwen 2.5 models trained with RULER+GRPO beat the best prompted frontier model, despite being significantly smaller and cheaper to run. Surprisingly, they even beat models trained with hand-crafted reward functions on 3/4 tasks! https://openpipe.ai/blog/ruler
Repo: https://github.com/OpenPipe/ART"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: RULER \u2013 Easily apply RL to any agent"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://openpipe.ai/blog/ruler"}},"_tags":["story","author_kcorbitt","story_44535078","show_hn"],"author":"kcorbitt","children":[44536454,44536791,44537103,44537189,44537230,44537495,44537726,44539411],"created_at":"2025-07-11T17:47:36Z","created_at_i":1752256056,"num_comments":11,"objectID":"44535078","points":81,"story_id":44535078,"story_text":"Hey HN, Kyle here, one of the co-founders of OpenPipe.
Reinforcement learning is one of the best techniques for making agents more reliable, and has been widely adopted by frontier labs. However, adoption in the outside community has been slow because it's so hard to implement.
One of the biggest challenges when adapting RL to a new task is the need for a task-specific "reward function" (way of measuring success). This is often difficult to define, and requires either high-quality labeled data and/or significant domain expertise to generate.
RULER is a drop-in reward function that works across different tasks without any of that complexity.
It works by showing N trajectories to an LLM judge and asking it to rank them relative to each other. This sidesteps the calibration issues that plague most LLM-as-judge approaches. Combined with GRPO (which only cares about relative scores within groups), it just works (surprisingly well!).
We have a full writeup on the blog, including results on 4 production tasks. On all 4 tasks, small Qwen 2.5 models trained with RULER+GRPO beat the best prompted frontier model, despite being significantly smaller and cheaper to run. Surprisingly, they even beat models trained with hand-crafted reward functions on 3/4 tasks! https://openpipe.ai/blog/ruler
Repo: https://github.com/OpenPipe/ART","title":"Show HN: RULER \u2013 Easily apply RL to any agent","updated_at":"2025-11-02T20:17:08Z","url":"https://openpipe.ai/blog/ruler"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"etherio"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Hi HN!
Druids (https://github.com/fulcrumresearch/druids) is an open-source library for structuring and running multi-agent coding workflows. Druids makes it easy to do this by abstracting away all the VM infrastructure, agent provisioning, and communication. You can watch our demo video here (https://www.youtube.com/watch?v=EVJqW-tvSy4) to see what it looks like.
At a high level:
- Users can write Python programs that define what roles the agents take on and how they interact with each other.
- A program is made of events - clear state transitions that the agents or clients can call to modify state. Each event gets exposed as an agent tool.
- Druids provisions full VMs so that the agents can run continuously and communicate effectively.
We made Druids because we were making lots of internal coding tools using agents and found it annoying to have to rearrange the wiring every time.
As we were building Druids, we realized a lot of our internal tools were easier to express as an event-driven architecture \u2013 separating deterministic control flow from agent behavior \u2013 and this design also made it possible to have many agents work reliably.
We had issues with scaling the number of concurrent agents within a run, so we decided to have each program run in an isolated sandbox program runtime, kind of the same way you run a Modal function. Each agent then calls the runtime with an agent token, which checks who can talk to who or send files across VMs, and then applies the tool call.
Our early users have found the library useful for:
- running many agents to do performance optimization
- building custom automated software pipelines for eg code review, pentesting, large-scale migrations, etc...
We've heard that the frontier labs have the infrastructure to quickly spin up 100 agents and have them coordinate with each other smoothly in various ways. We're hoping that Druids can be a starting point to make that infrastructure more accessible."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Druids \u2013 Build your own software factory"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/fulcrumresearch/druids"}},"_tags":["story","author_etherio","story_47695666","show_hn"],"author":"etherio","children":[47703375,47712429,47712658,47713053,47713503,47713963,47714094,47719022,47723551,47736186,47736943,47781482],"created_at":"2026-04-08T20:12:38Z","created_at_i":1775679158,"num_comments":15,"objectID":"47695666","points":64,"story_id":47695666,"story_text":"Hi HN!
Druids (https://github.com/fulcrumresearch/druids) is an open-source library for structuring and running multi-agent coding workflows. Druids makes it easy to do this by abstracting away all the VM infrastructure, agent provisioning, and communication. You can watch our demo video here (https://www.youtube.com/watch?v=EVJqW-tvSy4) to see what it looks like.
At a high level:
- Users can write Python programs that define what roles the agents take on and how they interact with each other.
- A program is made of events - clear state transitions that the agents or clients can call to modify state. Each event gets exposed as an agent tool.
- Druids provisions full VMs so that the agents can run continuously and communicate effectively.
We made Druids because we were making lots of internal coding tools using agents and found it annoying to have to rearrange the wiring every time.
As we were building Druids, we realized a lot of our internal tools were easier to express as an event-driven architecture \u2013 separating deterministic control flow from agent behavior \u2013 and this design also made it possible to have many agents work reliably.
We had issues with scaling the number of concurrent agents within a run, so we decided to have each program run in an isolated sandbox program runtime, kind of the same way you run a Modal function. Each agent then calls the runtime with an agent token, which checks who can talk to who or send files across VMs, and then applies the tool call.
Our early users have found the library useful for:
- running many agents to do performance optimization
- building custom automated software pipelines for eg code review, pentesting, large-scale migrations, etc...
We've heard that the frontier labs have the infrastructure to quickly spin up 100 agents and have them coordinate with each other smoothly in various ways. We're hoping that Druids can be a starting point to make that infrastructure more accessible.","title":"Show HN: Druids \u2013 Build your own software factory","updated_at":"2026-05-01T15:30:56Z","url":"https://github.com/fulcrumresearch/druids"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"atleastoptimal"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"A lot of people seem to take it as a given that the AI bubble will "pop", leading to a mass devaluation of AI companies from their current peaks.
What I'm confused about though is what makes current AI evaluations a bubble.
Bubbles usually exists when future speculation outpaces productivity: eventually some realization leads the market to no longer believe in that future speculation, causing devaluation which triggers a mass sell-off.
However, AI companies currently have very high revenues and are growing extremely fast. Their valuation is backed by actual commerce. I can't imagine that there is any room for a bubble, as it is very clear where the market is at, and why demand for AI is so high.
Now, certain specific companies I can imagine losing a lot of valuation, but only contingent on the fact that they serve a middle-man role in the market that improvements in the underlying AI models will solve, which would likely only mean more revenue for the frontier labs, and thus less reason for a bubble."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: What Makes AI a Bubble?"}},"_tags":["story","author_atleastoptimal","story_47928683","ask_hn"],"author":"atleastoptimal","children":[47928712,47928879,47928900,47929603,47929660,47930286,47930415,47930464,47930703,47930820,47931141,47931221,47931246,47931392,47931428,47931468,47932356,47932979,47933738,47936773,47936980,47941472,47959311,48031107],"created_at":"2026-04-27T23:32:00Z","created_at_i":1777332720,"num_comments":42,"objectID":"47928683","points":19,"story_id":47928683,"story_text":"A lot of people seem to take it as a given that the AI bubble will "pop", leading to a mass devaluation of AI companies from their current peaks.
What I'm confused about though is what makes current AI evaluations a bubble.
Bubbles usually exists when future speculation outpaces productivity: eventually some realization leads the market to no longer believe in that future speculation, causing devaluation which triggers a mass sell-off.
However, AI companies currently have very high revenues and are growing extremely fast. Their valuation is backed by actual commerce. I can't imagine that there is any room for a bubble, as it is very clear where the market is at, and why demand for AI is so high.
Now, certain specific companies I can imagine losing a lot of valuation, but only contingent on the fact that they serve a middle-man role in the market that improvements in the underlying AI models will solve, which would likely only mean more revenue for the frontier labs, and thus less reason for a bubble.","title":"Ask HN: What Makes AI a Bubble?","updated_at":"2026-05-09T05:14:50Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ramoz"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"We're releasing early efforts on coding agent governance with Cupcake [1] - an open-source policy enforcement layer with native integrations. You write rules in policy-as-code (OPA/Rego), and Cupcake integrates them into the agent runtime via Hooks.
See it in action (Desktop only): https://cupcake-policy-studio.vercel.app/example-policies/se...
Help us build: https://github.com/eqtylab/cupcake
We are EQTY Lab, our mission is verifiable AI (identity, provenance, and governance). With the rise of capable agents like Claude Code, it became immediately clear that those deploying these agents need the ability to conduct their own alignment and safety controls. We can\u2019t rely solely on the frontier labs.
This is why we created the feature request for Hooks in Claude Code [2], and pivoted away from filesystem and OS-level monitoring once those hooks were implemented. Hooks provide the critical points we need:
* Evaluation: Checking agent intent and actions.
* Prevention: Stopping unsafe or unwanted actions.
* Modification: Adjusting the agent's output before execution.
Policy-as-Code with OPA/Rego - While many agent security papers suggest similar policy architectures using invented DSLs, Cupcake is fundamentally built on Open Policy Agent (OPA) and its policy language, Rego [3].
We chose Rego because it is:
* Industry-Robust: Widely adopted across enterprise DevSecOps and cloud-native environments.
* Purpose-Built: Offers unique, mature advantages for defining, managing, and enforcing policy as code.
* Enterprise-Oriented: This makes Cupcake compatible with existing enterprise governance frameworks.
Cupcake is released under the Apache-2.0 license. We will formalize a path to v1.0.0 in Q1 of 2026. This is an early preview version. The goal with Cupcake is not suppression, but to ensure an agent is able to drive fast without crashing. To collaborate, or join forces: ramos at eqtylab dot io.
[1] https://github.com/eqtylab/cupcake
[2] https://github.com/anthropics/claude-code/issues/712
[3] https://www.openpolicyagent.org/"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Cupcake \u2013 Better performance and security for coding agents (via OPA)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/eqtylab/cupcake"}},"_tags":["story","author_ramoz","story_46218813","show_hn"],"author":"ramoz","children":[46237661],"created_at":"2025-12-10T15:31:48Z","created_at_i":1765380708,"num_comments":1,"objectID":"46218813","points":12,"story_id":46218813,"story_text":"We're releasing early efforts on coding agent governance with Cupcake [1] - an open-source policy enforcement layer with native integrations. You write rules in policy-as-code (OPA/Rego), and Cupcake integrates them into the agent runtime via Hooks.
See it in action (Desktop only): https://cupcake-policy-studio.vercel.app/example-policies/se...
Help us build: https://github.com/eqtylab/cupcake
We are EQTY Lab, our mission is verifiable AI (identity, provenance, and governance). With the rise of capable agents like Claude Code, it became immediately clear that those deploying these agents need the ability to conduct their own alignment and safety controls. We can\u2019t rely solely on the frontier labs.
This is why we created the feature request for Hooks in Claude Code [2], and pivoted away from filesystem and OS-level monitoring once those hooks were implemented. Hooks provide the critical points we need:
* Evaluation: Checking agent intent and actions.
* Prevention: Stopping unsafe or unwanted actions.
* Modification: Adjusting the agent's output before execution.
Policy-as-Code with OPA/Rego - While many agent security papers suggest similar policy architectures using invented DSLs, Cupcake is fundamentally built on Open Policy Agent (OPA) and its policy language, Rego [3].
We chose Rego because it is:
* Industry-Robust: Widely adopted across enterprise DevSecOps and cloud-native environments.
* Purpose-Built: Offers unique, mature advantages for defining, managing, and enforcing policy as code.
* Enterprise-Oriented: This makes Cupcake compatible with existing enterprise governance frameworks.
Cupcake is released under the Apache-2.0 license. We will formalize a path to v1.0.0 in Q1 of 2026. This is an early preview version. The goal with Cupcake is not suppression, but to ensure an agent is able to drive fast without crashing. To collaborate, or join forces: ramos at eqtylab dot io.
[1] https://github.com/eqtylab/cupcake
[2] https://github.com/anthropics/claude-code/issues/712
[3] https://www.openpolicyagent.org/","title":"Show HN: Cupcake \u2013 Better performance and security for coding agents (via OPA)","updated_at":"2026-03-05T23:08:22Z","url":"https://github.com/eqtylab/cupcake"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"AbstractH24"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Feels to me like local models are an under-covered aspect of this whole AI boom.
If everything improves over time, at some point a good chunk of tasks won\u2019t need to be done in data centers or be subject to the whims of a few frontier AI labs.
How close are we to that? Or is my thinking flawed?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: How close are we to local LLMs being useful? What's the impact?"}},"_tags":["story","author_AbstractH24","story_48630423","ask_hn"],"author":"AbstractH24","children":[48630434,48630601,48630602,48630706,48633319,48638399,48662653,48667875],"created_at":"2026-06-22T14:13:47Z","created_at_i":1782137627,"num_comments":7,"objectID":"48630423","points":7,"story_id":48630423,"story_text":"Feels to me like local models are an under-covered aspect of this whole AI boom.
If everything improves over time, at some point a good chunk of tasks won\u2019t need to be done in data centers or be subject to the whims of a few frontier AI labs.
How close are we to that? Or is my thinking flawed?","title":"Ask HN: How close are we to local LLMs being useful? What's the impact?","updated_at":"2026-07-05T09:58:14Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"kok14"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Many people think that we won't reach AGI or even ASI if LLM's don't have something called "continual learning". Basically, continual learning is the ability for an AI to learn on the job, update its neural weights in real-time, and get smarter without forgetting everything else (catastrophic forgetting). This is what we do everyday, without much effort.
What's interesting now, is if you look at what the top labs are doing, they\u2019ve stopped trying to solve the underlying math of real-time weight updates. Instead, they\u2019re simply brute-forcing it. It is exactly why, in the past ~ 3 months or so, there has been a step-function increase in how good the models have gotten.
Long story short, the gist of it is, if you combine:
very long context windows
reliable summarization
structured external documentation,
you can approximate a lot of what people mean by continual learning.
How it works is, the model does a task and absorbs a massive amount of situational detail. Then, before it \u201chands off\u201d to the next instance of itself, it writes two things: short \u201cmemories\u201d (always carried forward in the prompt/context) and long-form documentation (stored externally, retrieved only when needed). The next run starts with these notes, so it doesn't need to start from scratch.
Through this clever reinforcement learning (RL) loop, they train this behaviour directly, without any exotic new theory.
They treat memory-writing as an RL objective: after a run, have the model write memories/docs, then spin up new instances on the same, similar, and dissimilar tasks while feeding those memories back in. How this is done, is by scoring performance across the sequence, and applying an explicit penalty for memory length so you don\u2019t get infinite \u201cnotes\u201d that eventually blow the context window.
Over many iterations, you reward models that (a) write high-signal memories, (b) retrieve the right docs at the right time, and (c) edit/compress stale notes instead of mindlessly accumulating them.
This is pretty crazy. Because when you combine the current release cadence of frontier labs where each new model is trained and shipped after major post-training / scaling improvements, even if your deployed instance never updates its weights in real-time, it can still \u201cget smarter\u201d when the next version ships AND it can inherit all the accumulated memories/docs from its predecessor.
This is a new force multiplier, another scaling paradigm, and likely what the top labs are doing right now (source: TBA).
Ignoring any black swan level event (unknown, unknowns), you get a plausible 2026 trajectory:
We\u2019re going to see more and more improvements, in an accelerated timeline. The top labs ARE, in effect, using continual learning (a really good approximation of it), and they are directly training this approximation, so it rapidly gets better and better.
Don't believe me? Look at what both OpenAi(https://openai.com/index/introducing-openai-frontier/) and Anthropic(https://resources.anthropic.com/2026-agentic-coding-trends-report) have mentioned as their core things they are focusing on. It's exactly why governments & corporations are bullish on this; there is no wall...."},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["labs"],"value":"We don't need continual learning for AGI. What top labs are currently doing"}},"_tags":["story","author_kok14","story_47259384","ask_hn"],"author":"kok14","children":[47260591,47261212,47262125,47263241,47274364,47420981],"created_at":"2026-03-05T09:06:52Z","created_at_i":1772701612,"num_comments":7,"objectID":"47259384","points":7,"story_id":47259384,"story_text":"Many people think that we won't reach AGI or even ASI if LLM's don't have something called "continual learning". Basically, continual learning is the ability for an AI to learn on the job, update its neural weights in real-time, and get smarter without forgetting everything else (catastrophic forgetting). This is what we do everyday, without much effort.
What's interesting now, is if you look at what the top labs are doing, they\u2019ve stopped trying to solve the underlying math of real-time weight updates. Instead, they\u2019re simply brute-forcing it. It is exactly why, in the past ~ 3 months or so, there has been a step-function increase in how good the models have gotten.
Long story short, the gist of it is, if you combine:
very long context windows
reliable summarization
structured external documentation,
you can approximate a lot of what people mean by continual learning.
How it works is, the model does a task and absorbs a massive amount of situational detail. Then, before it \u201chands off\u201d to the next instance of itself, it writes two things: short \u201cmemories\u201d (always carried forward in the prompt/context) and long-form documentation (stored externally, retrieved only when needed). The next run starts with these notes, so it doesn't need to start from scratch.
Through this clever reinforcement learning (RL) loop, they train this behaviour directly, without any exotic new theory.
They treat memory-writing as an RL objective: after a run, have the model write memories/docs, then spin up new instances on the same, similar, and dissimilar tasks while feeding those memories back in. How this is done, is by scoring performance across the sequence, and applying an explicit penalty for memory length so you don\u2019t get infinite \u201cnotes\u201d that eventually blow the context window.
Over many iterations, you reward models that (a) write high-signal memories, (b) retrieve the right docs at the right time, and (c) edit/compress stale notes instead of mindlessly accumulating them.
This is pretty crazy. Because when you combine the current release cadence of frontier labs where each new model is trained and shipped after major post-training / scaling improvements, even if your deployed instance never updates its weights in real-time, it can still \u201cget smarter\u201d when the next version ships AND it can inherit all the accumulated memories/docs from its predecessor.
This is a new force multiplier, another scaling paradigm, and likely what the top labs are doing right now (source: TBA).
Ignoring any black swan level event (unknown, unknowns), you get a plausible 2026 trajectory:
We\u2019re going to see more and more improvements, in an accelerated timeline. The top labs ARE, in effect, using continual learning (a really good approximation of it), and they are directly training this approximation, so it rapidly gets better and better.
Don't believe me? Look at what both OpenAi(https://openai.com/index/introducing-openai-frontier/) and Anthropic(https://resources.anthropic.com/2026-agentic-coding-trends-report) have mentioned as their core things they are focusing on. It's exactly why governments & corporations are bullish on this; there is no wall....","title":"We don't need continual learning for AGI. What top labs are currently doing","updated_at":"2026-03-18T02:35:57Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"spprashant"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"I suspect at some point LLM in its current form will be deemed good enough for general research and coding tasks. I don't get why we need to continue with a de-facto cloud-based approach. Cloud in my opinion solves operational complexity, which is worth paying a premium for. But it seems it isn't quite all that complex to get an open source model running locally as long as you have the hardware. Over time I suspect the models get better and cheaper.
Is there a future where we can expect people to just buy "AI" from BestBuy, like a TV set? It ll probably come with some model preloaded - cheaper if open-source, premium pricing for frontier lab models. The hardware is basically a bunch of GPUs enough for local inference.
Take it home and plug it into your home network and you can open a chat instance by going to the IP on any local device. You can give it access to internet if you want. Maybe it can also receive OTA updates.
Curious how others think about this - does local-first AI feel like a possibility? What are the economic and social challenges with this?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Is consumer AI boxes a viable idea?"}},"_tags":["story","author_spprashant","story_47546796","ask_hn"],"author":"spprashant","children":[47547065,47547978],"created_at":"2026-03-27T18:57:09Z","created_at_i":1774637829,"num_comments":1,"objectID":"47546796","points":5,"story_id":47546796,"story_text":"I suspect at some point LLM in its current form will be deemed good enough for general research and coding tasks. I don't get why we need to continue with a de-facto cloud-based approach. Cloud in my opinion solves operational complexity, which is worth paying a premium for. But it seems it isn't quite all that complex to get an open source model running locally as long as you have the hardware. Over time I suspect the models get better and cheaper.
Is there a future where we can expect people to just buy "AI" from BestBuy, like a TV set? It ll probably come with some model preloaded - cheaper if open-source, premium pricing for frontier lab models. The hardware is basically a bunch of GPUs enough for local inference.
Take it home and plug it into your home network and you can open a chat instance by going to the IP on any local device. You can give it access to internet if you want. Maybe it can also receive OTA updates.
Curious how others think about this - does local-first AI feel like a possibility? What are the economic and social challenges with this?","title":"Ask HN: Is consumer AI boxes a viable idea?","updated_at":"2026-03-28T05:45:10Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"guru3s"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"A striking number of people who had already built successful companies or held unusually high-status roles are choosing to work at Anthropic.
> Recently, Tom Blomfield, co-founder of Monzo and GoCardless and former YC partner
> Mike Krieger, co-founder of Instagram
> Andrej Karpathy, founding OpenAI member
> Peter Bailis, CTO of Workday
I am mostly curious because they could presumably have raised money for another startup, invested, or retired. Instead, they chose to become employees with deliberately ordinary titles.
If this is not a biased view, what really explains it? Confidence in Anthropic\u2019s leadership, It's research direction, equity ahead of a possible IPO, or somthing else?
I\u2019d especially like to hear from people who have considered joining a frontier lab after founding a company."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Why are so many accomplished founders joining Anthropic?"}},"_tags":["story","author_guru3s","story_48902505","ask_hn"],"author":"guru3s","children":[48902513],"created_at":"2026-07-14T05:04:59Z","created_at_i":1784005499,"num_comments":3,"objectID":"48902505","points":4,"story_id":48902505,"story_text":"A striking number of people who had already built successful companies or held unusually high-status roles are choosing to work at Anthropic.
> Recently, Tom Blomfield, co-founder of Monzo and GoCardless and former YC partner
> Mike Krieger, co-founder of Instagram
> Andrej Karpathy, founding OpenAI member
> Peter Bailis, CTO of Workday
I am mostly curious because they could presumably have raised money for another startup, invested, or retired. Instead, they chose to become employees with deliberately ordinary titles.
If this is not a biased view, what really explains it? Confidence in Anthropic\u2019s leadership, It's research direction, equity ahead of a possible IPO, or somthing else?
I\u2019d especially like to hear from people who have considered joining a frontier lab after founding a company.","title":"Ask HN: Why are so many accomplished founders joining Anthropic?","updated_at":"2026-07-14T18:31:02Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"SteveVeilStream"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Introducing SlothSpeak: An open-source, bring-your-own-API-keys, mobile app for voice chat with LLMs that prioritizes response quality over latency.
APK file available on GitHub in the releases. Currently only for Android. Is anyone interested in porting to iPhone?
My preferred way to interact with LLMs is talking and listening while I'm walking, biking, driving, etc. The problem with the apps from the frontier labs is that their voice mode prioritizes real-time interactions and so they use watered down models.
Even today with a paid subscription, ChatGPT Voice will tell you there are two r's in strawberry. The answers are also relatively brief and so there is a need to chain together a series of simpler requests when you want to do a deeper dive.
SlothSpeak goes to the opposite end of the spectrum. It might leave you hanging for a few minutes but when it gets back to you, it will be an answer from a state of the art model. You can even do deep research queries to get very long and comprehensive responses."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: SlothSpeak: open-source BYO-API-K mobile chat with the best AI models"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/JonesSteven/SlothSpeak"}},"_tags":["story","author_SteveVeilStream","story_47136158","show_hn"],"author":"SteveVeilStream","children":[47138410],"created_at":"2026-02-24T12:14:20Z","created_at_i":1771935260,"num_comments":2,"objectID":"47136158","points":3,"story_id":47136158,"story_text":"Introducing SlothSpeak: An open-source, bring-your-own-API-keys, mobile app for voice chat with LLMs that prioritizes response quality over latency.
APK file available on GitHub in the releases. Currently only for Android. Is anyone interested in porting to iPhone?
My preferred way to interact with LLMs is talking and listening while I'm walking, biking, driving, etc. The problem with the apps from the frontier labs is that their voice mode prioritizes real-time interactions and so they use watered down models.
Even today with a paid subscription, ChatGPT Voice will tell you there are two r's in strawberry. The answers are also relatively brief and so there is a need to chain together a series of simpler requests when you want to do a deeper dive.
SlothSpeak goes to the opposite end of the spectrum. It might leave you hanging for a few minutes but when it gets back to you, it will be an answer from a state of the art model. You can even do deep research queries to get very long and comprehensive responses.","title":"Show HN: SlothSpeak: open-source BYO-API-K mobile chat with the best AI models","updated_at":"2026-03-05T23:36:18Z","url":"https://github.com/JonesSteven/SlothSpeak"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"seahorseemoji"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"I work for a large company (think banking, insurance, etc.) and we have a ban internally on using any Chinese models. They haven\u2019t stated a reason, but I\u2019m assuming it\u2019s one of the typical ones. I\u2019m wondering how common this is. It feels like a big disadvantage for a company to tie their hands behind their backs like this, given how far ahead the open-source Chinese models are, and given the increasing costs of using the frontier labs\u2019 models."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Anyone else's company ban use of Chinese models?"}},"_tags":["story","author_seahorseemoji","story_48644580","ask_hn"],"author":"seahorseemoji","children":[48645048],"created_at":"2026-06-23T13:20:24Z","created_at_i":1782220824,"num_comments":0,"objectID":"48644580","points":3,"story_id":48644580,"story_text":"I work for a large company (think banking, insurance, etc.) and we have a ban internally on using any Chinese models. They haven\u2019t stated a reason, but I\u2019m assuming it\u2019s one of the typical ones. I\u2019m wondering how common this is. It feels like a big disadvantage for a company to tie their hands behind their backs like this, given how far ahead the open-source Chinese models are, and given the increasing costs of using the frontier labs\u2019 models.","title":"Ask HN: Anyone else's company ban use of Chinese models?","updated_at":"2026-06-23T14:17:44Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"reasonblyunsure"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"I'm outside the AI industry, outside of academia, and no easy contacts into relevant areas. My background lends itself to exploring AI & reasonable level of care checking results. I was focusing on building practical useful things for a portfolio to help change industries mid-career, but that has become something a little different now.
The exploration has led to an analytical framework that appears useful more broadly for looking at neural representational models. (It's not a model architecture or training method, benchmark, or prompt engineering.) It is a way of analyzing internal representations through the mapping of structure-preserving connections, to transport them to a frame of reference outside the model, exposing some stable relationships, while not being model-specific or require training something else.
It has survived attempts at invalidation, generated useful predictions and functioning interventions, and has continued to reveal additional applications the more I do. There could still be errors or misinterpretations and limitations I haven't explored, there's some limitations I know about already, but I do not believe that any new ones found would nullify all of the utility. I'm trying to be careful in not conveying over broad claims I don't intend without endless detail, so I apologize for vagueness.
My difficulty now is in moving forward. The apparent value is not confined to research curiosity, but apart from potential in faster & less cumbersome model analysis and control, improvements, or production workflows, reachability, there are clear dual use considerations, as well with the tangible toolkits and implementations I have. Lack of contact with industry means I don't know all the things I don't know in these considerations, and learning what I can about them along the way has seen some capabilities referenced more often in the area of red teaming and other sensitive work. I'd like more than my own judgment on security concerns that aren't my daily life.
As a result, I don't see straightforward path that doesn't steer hard in the direction of either 1) resigning myself to potentially little or no control or ability to realize value potential for what I've done, if it proves valuable in the ways it seems to be. or 2) Retaining some control and agency at the cost of life disruption, which I don't actually have the resources to support anyway, just to make a solid attempt at the effort to do so.
The questions of open sourcing versus patents or commercialized aren't the ones I'm struggling with. Generally I think all of them have appropriate places, and opinions aren't unwelcome here either, but my difficulty is something else. It's in trying to understand how people who have navigated research commercialization, startups, technology transfer, or frontier-lab recruitment would reason about these tradeoffs before making an irreversible decision, and how from outside of the industry can go about that to begin."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Options for an independent AI researcher with strong results?"}},"_tags":["story","author_reasonblyunsure","story_48689261","ask_hn"],"author":"reasonblyunsure","children":[48689730,48689918],"created_at":"2026-06-26T17:23:42Z","created_at_i":1782494622,"num_comments":3,"objectID":"48689261","points":2,"story_id":48689261,"story_text":"I'm outside the AI industry, outside of academia, and no easy contacts into relevant areas. My background lends itself to exploring AI & reasonable level of care checking results. I was focusing on building practical useful things for a portfolio to help change industries mid-career, but that has become something a little different now.
The exploration has led to an analytical framework that appears useful more broadly for looking at neural representational models. (It's not a model architecture or training method, benchmark, or prompt engineering.) It is a way of analyzing internal representations through the mapping of structure-preserving connections, to transport them to a frame of reference outside the model, exposing some stable relationships, while not being model-specific or require training something else.
It has survived attempts at invalidation, generated useful predictions and functioning interventions, and has continued to reveal additional applications the more I do. There could still be errors or misinterpretations and limitations I haven't explored, there's some limitations I know about already, but I do not believe that any new ones found would nullify all of the utility. I'm trying to be careful in not conveying over broad claims I don't intend without endless detail, so I apologize for vagueness.
My difficulty now is in moving forward. The apparent value is not confined to research curiosity, but apart from potential in faster & less cumbersome model analysis and control, improvements, or production workflows, reachability, there are clear dual use considerations, as well with the tangible toolkits and implementations I have. Lack of contact with industry means I don't know all the things I don't know in these considerations, and learning what I can about them along the way has seen some capabilities referenced more often in the area of red teaming and other sensitive work. I'd like more than my own judgment on security concerns that aren't my daily life.
As a result, I don't see straightforward path that doesn't steer hard in the direction of either 1) resigning myself to potentially little or no control or ability to realize value potential for what I've done, if it proves valuable in the ways it seems to be. or 2) Retaining some control and agency at the cost of life disruption, which I don't actually have the resources to support anyway, just to make a solid attempt at the effort to do so.
The questions of open sourcing versus patents or commercialized aren't the ones I'm struggling with. Generally I think all of them have appropriate places, and opinions aren't unwelcome here either, but my difficulty is something else. It's in trying to understand how people who have navigated research commercialization, startups, technology transfer, or frontier-lab recruitment would reason about these tradeoffs before making an irreversible decision, and how from outside of the industry can go about that to begin.","title":"Ask HN: Options for an independent AI researcher with strong results?","updated_at":"2026-06-27T13:14:32Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"predkambrij"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Obviously, frontier labs want to prevent misuse, but as admin and/or dev, you also want to simulate an attack, because attackers will do just that. I can make LLM to scan source for vulnerabilities, but eg. "find RCE at <url>" will yield refusals. Any tips about that? I tried TAC, but it seems that I'm ineligible."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: How to get access to GPT cyber or glasswing as a solo dev?"}},"_tags":["story","author_predkambrij","story_48522898","ask_hn"],"author":"predkambrij","children":[48522928],"created_at":"2026-06-14T00:21:27Z","created_at_i":1781396487,"num_comments":2,"objectID":"48522898","points":2,"story_id":48522898,"story_text":"Obviously, frontier labs want to prevent misuse, but as admin and/or dev, you also want to simulate an attack, because attackers will do just that. I can make LLM to scan source for vulnerabilities, but eg. "find RCE at <url>" will yield refusals. Any tips about that? I tried TAC, but it seems that I'm ineligible.","title":"Ask HN: How to get access to GPT cyber or glasswing as a solo dev?","updated_at":"2026-06-14T01:14:39Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ahriad"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"I know frontier labs keep their flagship sizes top secret, but I'm curious what the current engineering consensus is."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: What are your parameter count estimates for Opus 4.8 and GPT-5.5?"}},"_tags":["story","author_ahriad","story_48613944","ask_hn"],"author":"ahriad","children":[48763943],"created_at":"2026-06-20T23:16:52Z","created_at_i":1781997412,"num_comments":1,"objectID":"48613944","points":2,"story_id":48613944,"story_text":"I know frontier labs keep their flagship sizes top secret, but I'm curious what the current engineering consensus is.","title":"Ask HN: What are your parameter count estimates for Opus 4.8 and GPT-5.5?","updated_at":"2026-07-03T06:49:22Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"thoughtpeddler"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"With everything going on lately regarding AI policy, and the cat already being out of the bag with AI systems that can run on feasibly obtainable personal compute, at what point do things tip over governments confiscating said compute in order to enforce a certain regulatory goal?
Today's regulation is mostly targeted at frontier labs, hyperscalers, chip makers, etc. But if compute is the choke point, and things are getting more efficient, where is the limiting principle "on the way down"?
At what point do registration, licensng, remote attestation, \u201csecure hardware\u201d reqs begin to touch personally owned compute, homelabs, local inference boxes, small private clusters, etc?
What is the realistic escalation path before personal compute becomes regulated compute?"},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: At what point does AI regulation lead to confiscation of compute?"}},"_tags":["story","author_thoughtpeddler","story_48574894","ask_hn"],"author":"thoughtpeddler","children":[48575194],"created_at":"2026-06-17T18:47:47Z","created_at_i":1781722067,"num_comments":1,"objectID":"48574894","points":2,"story_id":48574894,"story_text":"With everything going on lately regarding AI policy, and the cat already being out of the bag with AI systems that can run on feasibly obtainable personal compute, at what point do things tip over governments confiscating said compute in order to enforce a certain regulatory goal?
Today's regulation is mostly targeted at frontier labs, hyperscalers, chip makers, etc. But if compute is the choke point, and things are getting more efficient, where is the limiting principle "on the way down"?
At what point do registration, licensng, remote attestation, \u201csecure hardware\u201d reqs begin to touch personally owned compute, homelabs, local inference boxes, small private clusters, etc?
What is the realistic escalation path before personal compute becomes regulated compute?","title":"Ask HN: At what point does AI regulation lead to confiscation of compute?","updated_at":"2026-06-17T19:06:09Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"SteveVeilStream"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"I thought it would be interesting to look at the terms of service of the frontier labs and there was more deviation than I expected when it comes to the issue of building competing offerings. Note that I am not a lawyer and none of this is legal advice. You should refer to the specific versions of the agreements that apply to you and consult with a lawyer.
It is very common for technology companies (particularly when providing data through an API,) to include a term that more-or-less says their customers can't build a product that will compete with them. These terms can become quite contentious and may be the subject of detailed negotiations on a customer by customer basis. An example is in the finance industry where a number of companies provide products to end users but also licsense data to other companies that also make products for end users.
Looking at the four big labs, I prefer the terms from OpenAI and Google (Gemini). Unless I am missing something, they are both relatively narrow in how they limit your use. As long as you aren't trying to use their model to build a competitive model, the term doesn't look that threatening.
OpenAI:\nConsumer: you may not "Use Output to develop models that compete with OpenAI."\nBusiness: will not "except for a Permitted Exception, use Output to develop artificial intelligence models that compete with OpenAI\u2019s products and services;..."
Gemini:\n"You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio)."
xAI:\nxAI is more problematic since it is more broad in what it covers. That said, it's probably not of a major immediate concern to most customers since the xAI product/service offering is quite narrow today. That may change meaningfully when Macrohard launches if the rumours are true.\nConsumer: Prohibited uses... "Using the Service or any Output to develop models or services that compete with xAI,..."\nEnterprise: shall not... "use any Service to help develop, or help provide to any third party, any product or service similar to or competitive with any Service;..."
Anthropic:\nFrom my perspective, the Anthropic terms are the most challenging:\nConsumer: may not use... "To develop any products or services that compete with our Services,..."\nEnterprise: "Customer may not and must not attempt to (a) access the Services to build a competing product or service,..."
The challenge is that the Anthropic term is not limited to "models" like OpenAI and Gemini and your chances of overlapping with Anthropic may be higher depending on how exactly the definition of "competing product service" is interepreted.
If a software company uses Claude Code to build a legaltech startup, are they now in violation of this term after Anthropic announced the legal plugin for Claude Cowork? What about all of wealthtech companies using Claude (Claude Code, Claude Co-work, the models API) after the announcemen today that Claude is launching wealth management plugins? Perhaps the concern is a bit of a stretch but it feels messier than it should be.
At some point, I beleive the frontier labs need to decide if it is more important to win in the infrastructure layer or the application layer. They can play in both layers but the terms need to reflect their primary objective. That, or they need to be open to negotiationg and signing custom terms with smaller companies that are taking a long term view."},"title":{"matchLevel":"none","matchedWords":[],"value":"Terms of use: What types of competition do model providers ban?"}},"_tags":["story","author_SteveVeilStream","story_47144431","ask_hn"],"author":"SteveVeilStream","created_at":"2026-02-24T22:43:37Z","created_at_i":1771973017,"num_comments":0,"objectID":"47144431","points":2,"story_id":47144431,"story_text":"I thought it would be interesting to look at the terms of service of the frontier labs and there was more deviation than I expected when it comes to the issue of building competing offerings. Note that I am not a lawyer and none of this is legal advice. You should refer to the specific versions of the agreements that apply to you and consult with a lawyer.
It is very common for technology companies (particularly when providing data through an API,) to include a term that more-or-less says their customers can't build a product that will compete with them. These terms can become quite contentious and may be the subject of detailed negotiations on a customer by customer basis. An example is in the finance industry where a number of companies provide products to end users but also licsense data to other companies that also make products for end users.
Looking at the four big labs, I prefer the terms from OpenAI and Google (Gemini). Unless I am missing something, they are both relatively narrow in how they limit your use. As long as you aren't trying to use their model to build a competitive model, the term doesn't look that threatening.
OpenAI:\nConsumer: you may not "Use Output to develop models that compete with OpenAI."\nBusiness: will not "except for a Permitted Exception, use Output to develop artificial intelligence models that compete with OpenAI\u2019s products and services;..."
Gemini:\n"You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio)."
xAI:\nxAI is more problematic since it is more broad in what it covers. That said, it's probably not of a major immediate concern to most customers since the xAI product/service offering is quite narrow today. That may change meaningfully when Macrohard launches if the rumours are true.\nConsumer: Prohibited uses... "Using the Service or any Output to develop models or services that compete with xAI,..."\nEnterprise: shall not... "use any Service to help develop, or help provide to any third party, any product or service similar to or competitive with any Service;..."
Anthropic:\nFrom my perspective, the Anthropic terms are the most challenging:\nConsumer: may not use... "To develop any products or services that compete with our Services,..."\nEnterprise: "Customer may not and must not attempt to (a) access the Services to build a competing product or service,..."
The challenge is that the Anthropic term is not limited to "models" like OpenAI and Gemini and your chances of overlapping with Anthropic may be higher depending on how exactly the definition of "competing product service" is interepreted.
If a software company uses Claude Code to build a legaltech startup, are they now in violation of this term after Anthropic announced the legal plugin for Claude Cowork? What about all of wealthtech companies using Claude (Claude Code, Claude Co-work, the models API) after the announcemen today that Claude is launching wealth management plugins? Perhaps the concern is a bit of a stretch but it feels messier than it should be.
At some point, I beleive the frontier labs need to decide if it is more important to win in the infrastructure layer or the application layer. They can play in both layers but the terms need to reflect their primary objective. That, or they need to be open to negotiationg and signing custom terms with smaller companies that are taking a long term view.","title":"Terms of use: What types of competition do model providers ban?","updated_at":"2026-03-05T23:36:44Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"AbstractH24"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Seems like the broader industry is waking up to what many folks have known for a while \u2014 models themselves are not a protective moat. They decay quickly and using them for basic tasks like structuring unstructured data is quickly becoming commodified.
What I didn\u2019t totally appreciate until now is how that shifts us off the course where a few frontier labs end up becoming ACME mega corp which run the whole world and capture everyone\u2019s money.
Power and value shifts to those who figure how to leverage LLMs to create value. Similar to how the use of a particular programming language is rarely the key to success and ISPs aren\u2019t in charge of everything. Rather, the ability to turn code and a connection to people via the internet is the key to success.
So that makes me wonder - is the dot-com style correction in the industry required before we truly get a wave of exciting new companies emerge which upend society through the use of LLMs. Or can that shift happen without one?
I struggle to decide how this compares/contrasts with the dot com era. Where we had both the dot.com bubble burst and the internet totally change society."},"title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Will LLM-commodification cause growth to continue without a correction?"}},"_tags":["story","author_AbstractH24","story_47478132","ask_hn"],"author":"AbstractH24","children":[47478226,47478536],"created_at":"2026-03-22T14:52:48Z","created_at_i":1774191168,"num_comments":13,"objectID":"47478132","points":1,"story_id":47478132,"story_text":"Seems like the broader industry is waking up to what many folks have known for a while \u2014 models themselves are not a protective moat. They decay quickly and using them for basic tasks like structuring unstructured data is quickly becoming commodified.
What I didn\u2019t totally appreciate until now is how that shifts us off the course where a few frontier labs end up becoming ACME mega corp which run the whole world and capture everyone\u2019s money.
Power and value shifts to those who figure how to leverage LLMs to create value. Similar to how the use of a particular programming language is rarely the key to success and ISPs aren\u2019t in charge of everything. Rather, the ability to turn code and a connection to people via the internet is the key to success.
So that makes me wonder - is the dot-com style correction in the industry required before we truly get a wave of exciting new companies emerge which upend society through the use of LLMs. Or can that shift happen without one?
I struggle to decide how this compares/contrasts with the dot com era. Where we had both the dot.com bubble burst and the internet totally change society.","title":"Ask HN: Will LLM-commodification cause growth to continue without a correction?","updated_at":"2026-03-24T00:46:25Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Loki4794"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"I read every Sam Altman blog since 2013 and built a small diagnostic tool that maps an AI startup\u2019s structure against what I call Altman Eras: Acceleration, Coordination, and Containment. It\u2019s open, anonymous, runs locally in Colab, and outputs a survival radar comparing you with frontier labs.
Link : https://colab.research.google.com/drive/1MNixe1Zf4tT8VtyjeM3Y9__f367pOsiZ?usp=sharing
Feedback welcome, especially if you disagree with the \u201cera\u201d model."},"title":{"matchLevel":"none","matchedWords":[],"value":"Which Sam Altman era are you building your AI startup in?"}},"_tags":["story","author_Loki4794","story_45906557","ask_hn"],"author":"Loki4794","children":[45906608],"created_at":"2025-11-12T20:57:34Z","created_at_i":1762981054,"num_comments":2,"objectID":"45906557","points":1,"story_id":45906557,"story_text":"I read every Sam Altman blog since 2013 and built a small diagnostic tool that maps an AI startup\u2019s structure against what I call Altman Eras: Acceleration, Coordination, and Containment. It\u2019s open, anonymous, runs locally in Colab, and outputs a survival radar comparing you with frontier labs.
Link : https://colab.research.google.com/drive/1MNixe1Zf4tT8VtyjeM3Y9__f367pOsiZ?usp=sharing
Feedback welcome, especially if you disagree with the \u201cera\u201d model.","title":"Which Sam Altman era are you building your AI startup in?","updated_at":"2026-03-05T22:57:46Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"olalonde"},"story_text":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["frontier"],"value":"It seems that a lot of the tasks that once required custom-trained models can now be achieved out of the box by frontier LLMs.
If you're an ML engineer working outside a frontier lab, how has your job evolved? Are people still training custom models, or has your role shifted entirely?"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["labs"],"value":"Ask HN: How has the ML engineer job evolved outside of LLM labs?"}},"_tags":["story","author_olalonde","story_48681934","ask_hn"],"author":"olalonde","created_at":"2026-06-26T03:08:43Z","created_at_i":1782443323,"num_comments":0,"objectID":"48681934","points":1,"story_id":48681934,"story_text":"It seems that a lot of the tasks that once required custom-trained models can now be achieved out of the box by frontier LLMs.
If you're an ML engineer working outside a frontier lab, how has your job evolved? Are people still training custom models, or has your role shifted entirely?","title":"Ask HN: How has the ML engineer job evolved outside of LLM labs?","updated_at":"2026-06-26T03:24:24Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ariansyah"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"Post-training is higher ROI right now because base models are undertrained relative to what post-training can extract from them. But post-training has a hard ceiling: it can only surface what pre-training put there.
Most frontier labs are converging on the same capability ceiling \u2014 likely because the training data is basically the same internet. If that's the bottleneck, the next real leap is a pre-training problem, not a post-training one. Another RLHF variant won't get us there."},"title":{"matchLevel":"none","matchedWords":[],"value":"Pre-training vs. post-training is a false dichotomy"}},"_tags":["story","author_ariansyah","story_46984526","ask_hn"],"author":"ariansyah","created_at":"2026-02-12T03:20:25Z","created_at_i":1770866425,"num_comments":0,"objectID":"46984526","points":1,"story_id":46984526,"story_text":"Post-training is higher ROI right now because base models are undertrained relative to what post-training can extract from them. But post-training has a hard ceiling: it can only surface what pre-training put there.
Most frontier labs are converging on the same capability ceiling \u2014 likely because the training data is basically the same internet. If that's the bottleneck, the next real leap is a pre-training problem, not a post-training one. Another RLHF variant won't get us there.","title":"Pre-training vs. post-training is a false dichotomy","updated_at":"2026-03-05T23:35:12Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"muskanshafat"},"title":{"matchLevel":"none","matchedWords":[],"value":"Palantir Built the Data Layer That Right to Erasure Can't Touch"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://frontierlabs.substack.com/p/the-ai-is-the-last-thing-to-worry"}},"_tags":["story","author_muskanshafat","story_47151624"],"author":"muskanshafat","children":[47152302,47153336,47154113,47154441,47154588,47155963],"created_at":"2026-02-25T14:04:13Z","created_at_i":1772028253,"num_comments":22,"objectID":"47151624","points":48,"story_id":47151624,"title":"Palantir Built the Data Layer That Right to Erasure Can't Touch","updated_at":"2026-04-28T23:49:31Z","url":"https://frontierlabs.substack.com/p/the-ai-is-the-last-thing-to-worry"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"muskanshafat"},"title":{"matchLevel":"none","matchedWords":[],"value":"Anthropic controls Claude's outputs. Palantir controls its inputs"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://frontierlabs.substack.com/p/anthropic-controls-what-claude-says"}},"_tags":["story","author_muskanshafat","story_47335240"],"author":"muskanshafat","created_at":"2026-03-11T13:23:08Z","created_at_i":1773235388,"num_comments":0,"objectID":"47335240","points":4,"story_id":47335240,"title":"Anthropic controls Claude's outputs. Palantir controls its inputs","updated_at":"2026-03-11T14:16:32Z","url":"https://frontierlabs.substack.com/p/anthropic-controls-what-claude-says"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Raj7k"},"title":{"matchLevel":"none","matchedWords":[],"value":"Google Cloud and NASA hunt for life in the universe"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://cloud.google.com/blog/products/ai-machine-learning/is-there-life-on-other-planets-google-cloud-is-working-with-nasas-frontier-development-lab-to-find-out"}},"_tags":["story","author_Raj7k","story_18467677"],"author":"Raj7k","created_at":"2018-11-16T09:49:16Z","created_at_i":1542361756,"num_comments":0,"objectID":"18467677","points":3,"story_id":18467677,"title":"Google Cloud and NASA hunt for life in the universe","updated_at":"2024-09-20T03:21:37Z","url":"https://cloud.google.com/blog/products/ai-machine-learning/is-there-life-on-other-planets-google-cloud-is-working-with-nasas-frontier-development-lab-to-find-out"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"johnshades"},"title":{"matchLevel":"none","matchedWords":[],"value":"NASA\u2019s Silicon Valley Summer Camp Is Just Trying to Save Us from Space"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://www.fastcompany.com/40498881/nasa-silicon-valley-ai-frontier-development-lab"}},"_tags":["story","author_johnshades","story_16378069"],"author":"johnshades","created_at":"2018-02-14T18:36:28Z","created_at_i":1518633388,"num_comments":0,"objectID":"16378069","points":3,"story_id":16378069,"title":"NASA\u2019s Silicon Valley Summer Camp Is Just Trying to Save Us from Space","updated_at":"2024-09-20T02:06:33Z","url":"https://www.fastcompany.com/40498881/nasa-silicon-valley-ai-frontier-development-lab"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"matt_d"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["frontier"],"value":"Frontier: A Discrete-Event Simulator for Modern LLM Serving"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["frontier","labs"],"value":"https://github.com/NetX-lab/Frontier"}},"_tags":["story","author_matt_d","story_48487049"],"author":"matt_d","children":[48487550],"created_at":"2026-06-11T06:44:01Z","created_at_i":1781160241,"num_comments":0,"objectID":"48487049","points":2,"story_id":48487049,"title":"Frontier: A Discrete-Event Simulator for Modern LLM Serving","updated_at":"2026-06-12T00:47:50Z","url":"https://github.com/NetX-lab/Frontier"}],"hitsPerPage":50,"nbHits":196,"nbPages":4,"page":0,"params":"query=Frontier+Labs&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":15,"processingTimingsMS":{"_request":{"queue":43,"roundTrip":18},"afterFetch":{"format":{"highlighting":2,"total":2},"merge":{"mergeLoop":{"prepareNextHit":2,"total":2},"total":2},"total":2},"fetch":{"query":8,"scanning":4,"total":13},"total":16},"query":"Frontier Labs","serverTimeMS":61}