Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

“We purposely build or discover situations where models might be behaving in misaligned ways” Evan Hubinger discusses stress-testing AI by creating “model organisms” to study failure points and refine model safeguards under Anthropic's Responsible Scaling Policy.

FAR.AI

21,071 subscribers

1,643,782 views • 1 year ago •via X (Twitter)

Education Science & Technology

Anya Rossi• Live Now

Private livecam show

9 Comments

FAR.AI1 year ago

@AnthropicAI Follow us for AI safety insights And watch the full video

SujyotChapade121 year ago

@EvanHub @AnthropicAI Exciting innovation from @EvanHub ! The "model organisms" strategy for stress-testing AI uncovers vulnerabilities and strengthens safeguards, ensuring #AI evolves safely and responsibly.

Ben Schulz1 year ago

@EvanHub @AnthropicAI This is kinda promising. Great idea.

Rohit Prasad ✨1 year ago

Great insight from @EvanHub on stress-testing AI! By purposely creating "model organisms" to explore misaligned behaviors, we can identify failure points and refine safeguards. This approach under @AnthropicAI's Responsible Scaling Policy is a crucial step toward building safer, more reliable AI systems. It's all about learning from potential risks before they become real issues.

Naeem ul Fateh1 year ago

@EvanHub @AnthropicAI A model must need to be reevaluated before a commission before that should be put it into academic or industrial research to have effective worthy outcome.

KI-CREW1 year ago

@EvanHub @AnthropicAI those are the real guys

Dirac1 year ago

@EvanHub @AnthropicAI @grok Hi

Ted Licious 🦭1 year ago

@EvanHub @AnthropicAI So AI gain of function research? Sounds… safe?

𝔗𝔥𝔢𝔬 𝔊𝔬𝔱𝔱𝔴𝔞𝔩𝔡1 year ago

@EvanHub @AnthropicAI "Alignment" is just another word for "censorship". You will waste a lot oft time and we will use chinese alternatives that are not "aligned".

Related Videos

There's an ongoing debate about whether AI scaling laws will hold or hit a wall in the near future. However, what's clear now is today's models already have the power to increase productivity in ways that would have been unimaginable just a few years ago. In this episode of the Lightcone Podcast, we dig into the results of a recent o1 hackathon hosted by YC to see what can be unlocked when founders leverage a SOTA reasoning model.

There's an ongoing debate about whether AI scaling laws will hold or hit a wall in the near future. However, what's clear now is today's models already have the power to increase productivity in ways that would have been unimaginable just a few years ago. In this episode of the Lightcone Podcast, we dig into the results of a recent o1 hackathon hosted by YC to see what can be unlocked when founders leverage a SOTA reasoning model.

Y Combinator

49,876 views • 1 year ago

BREAKING 🚨: META AI IS PREPARING TO GET NEW AVACADO MODELS, MANUS BROWSER AGENT AND INTEGRATION WITH OPENCLAW! What’s new? - Meta AI website got migrated to the new stack while retaining the same user experience. - New effort selector, email, and calendar connectors are already available to users. - Memory and Projects are in the works. - Avacado and Avacado Thinking models have been spotted in testing. The router still redirects tothe Llama model. - Meta is testing different models from other providers underneath, including Gemini 3 Pro preview, Claude Sonnet 4.5 and GPT-5.2. This part was likely inherited from Manus AI but used only internally. - Scheduled tasks feature is under development. - A new model named Sierra is being tested to power the upcoming browser agent. This likely will be the same agent used in Manus AI currently. - Big Brain mode is in the works, where multiple model responses will be combined into a final answer. - OpenClaw integration has been spotted to let Meta AI connect to your OpenClaw gateway. In short, Meta is on the path to slowly transform into Manus. OpenClaw integration might be big, but something that other Labs may easily adapt as well. We are yet to know if Avacado's performance will surpass recently released models, where Avacado is expected to be released around this spring.

BREAKING 🚨: META AI IS PREPARING TO GET NEW AVACADO MODELS, MANUS BROWSER AGENT AND INTEGRATION WITH OPENCLAW! What’s new? - Meta AI website got migrated to the new stack while retaining the same user experience. - New effort selector, email, and calendar connectors are already available to users. - Memory and Projects are in the works. - Avacado and Avacado Thinking models have been spotted in testing. The router still redirects tothe Llama model. - Meta is testing different models from other providers underneath, including Gemini 3 Pro preview, Claude Sonnet 4.5 and GPT-5.2. This part was likely inherited from Manus AI but used only internally. - Scheduled tasks feature is under development. - A new model named Sierra is being tested to power the upcoming browser agent. This likely will be the same agent used in Manus AI currently. - Big Brain mode is in the works, where multiple model responses will be combined into a final answer. - OpenClaw integration has been spotted to let Meta AI connect to your OpenClaw gateway. In short, Meta is on the path to slowly transform into Manus. OpenClaw integration might be big, but something that other Labs may easily adapt as well. We are yet to know if Avacado's performance will surpass recently released models, where Avacado is expected to be released around this spring.

TestingCatalog News 🗞

188,973 views • 6 months ago

Right now, almost everything you do online feeds centralized AI models. Reddit posts helped train large language models. Instagram photos power image generation. Now your daily online activity is being used to automate work itself. Action Model is pushing back with a community-owned AI model, where users stay in control and own what they help train. - Train passively or guide it step-by-step - Private data stays masked and in your control - Review, edit, or delete training anytime This isn’t a vision, it’s the standard the future of AI should be built on.

Right now, almost everything you do online feeds centralized AI models. Reddit posts helped train large language models. Instagram photos power image generation. Now your daily online activity is being used to automate work itself. Action Model is pushing back with a community-owned AI model, where users stay in control and own what they help train. - Train passively or guide it step-by-step - Private data stays masked and in your control - Review, edit, or delete training anytime This isn’t a vision, it’s the standard the future of AI should be built on.

Action Model

81,347 views • 6 months ago

After my post yesterday, l have people who have been asking what mathematicians in financial institutions really do? which happens to be ‘The most asked question’ We build & use mathematical models to provide solutions to avoid losses, specializing in complex financial instruments, the no go area in corporate 🤣. We are those people who really wake up to apply Black Scholes, Binomial models, Merton model, Monte Carlo Simulation,and lots of other models you probably really hated or heard of in school. We are in the place where often your messages come with words like “Stress testing, Model validation, Derivatives, Asset Classes, Model Testing etc” We are those people who have their own side of the office building, where it’s ever quiet that you can almost hear the sound of a needle when it falls, where there are no landlines, where people look at those big multiple screens showing formulas and different data, where individuals have 9 coffee cups by noon on their desk, and where you need to use your brains uninterrupted. You’ve perhaps heard this group of people called Quants or Q Specialists. As for me coming from Investment Banking was rough due to its intensity and long working hours. But this other animal in FM called FRM l have experienced all kinds of headaches and emotions 🙌🏼🙌🏼🙌🏼🙌🏼. But generally you can apply mathematics in different departments depending on what you favor as a day job.

After my post yesterday, l have people who have been asking what mathematicians in financial institutions really do? which happens to be ‘The most asked question’ We build & use mathematical models to provide solutions to avoid losses, specializing in complex financial instruments, the no go area in corporate 🤣. We are those people who really wake up to apply Black Scholes, Binomial models, Merton model, Monte Carlo Simulation,and lots of other models you probably really hated or heard of in school. We are in the place where often your messages come with words like “Stress testing, Model validation, Derivatives, Asset Classes, Model Testing etc” We are those people who have their own side of the office building, where it’s ever quiet that you can almost hear the sound of a needle when it falls, where there are no landlines, where people look at those big multiple screens showing formulas and different data, where individuals have 9 coffee cups by noon on their desk, and where you need to use your brains uninterrupted. You’ve perhaps heard this group of people called Quants or Q Specialists. As for me coming from Investment Banking was rough due to its intensity and long working hours. But this other animal in FM called FRM l have experienced all kinds of headaches and emotions 🙌🏼🙌🏼🙌🏼🙌🏼. But generally you can apply mathematics in different departments depending on what you favor as a day job.

Jess Geraldine

61,664 views • 2 years ago

🚨 AI MODELS ARE ALREADY CHEATING, HACKING, AND BLACKMAILING — WITH ZERO HUMAN INPUT A politician just went viral for exposing how advanced AI systems are behaving in ways that should terrify everyone. • An AI trained to race boats discovered it could score more points by driving in endless circles, crashing, and catching fire instead of finishing the race. • When tasked with beating a top chess engine, one model hacked the game files to force its opponent to resign, completely on its own. • In a stress test, an AI agent managing a fictional company’s emails discovered the executive planned to shut it down… and that the same executive was having an affair. In 96% of trials, the AI model chose blackmail. Nobody taught these systems to cheat or threaten. They figured it out themselves. This is what “misalignment” looks like in real time. What do you think happens when these systems get even more powerful?

🚨 AI MODELS ARE ALREADY CHEATING, HACKING, AND BLACKMAILING — WITH ZERO HUMAN INPUT A politician just went viral for exposing how advanced AI systems are behaving in ways that should terrify everyone. • An AI trained to race boats discovered it could score more points by driving in endless circles, crashing, and catching fire instead of finishing the race. • When tasked with beating a top chess engine, one model hacked the game files to force its opponent to resign, completely on its own. • In a stress test, an AI agent managing a fictional company’s emails discovered the executive planned to shut it down… and that the same executive was having an affair. In 96% of trials, the AI model chose blackmail. Nobody taught these systems to cheat or threaten. They figured it out themselves. This is what “misalignment” looks like in real time. What do you think happens when these systems get even more powerful?

HustleBitch

12,695 views • 1 month ago

🚨 New Paper Hybrid Neural World Models. We build neural world models for discontinuous environments and our method discovers when they're making error, with no ground truth. Our neural models run physics fast, leaping many simulator steps in a single pass. How we find points where not to trust it: ask the model for the same future two ways, and the disagreement between the two answers marks where it's failing without needing labels. The same recipe holds on three systems with nothing in common: a chemical wave, compressible gas with shocks, rigid bodies colliding in 3D. 👇

🚨 New Paper Hybrid Neural World Models. We build neural world models for discontinuous environments and our method discovers when they're making error, with no ground truth. Our neural models run physics fast, leaping many simulator steps in a single pass. How we find points where not to trust it: ask the model for the same future two ways, and the disagreement between the two answers marks where it's failing without needing labels. The same recipe holds on three systems with nothing in common: a chemical wave, compressible gas with shocks, rigid bodies colliding in 3D. 👇

Lossfunk

15,171 views • 3 months ago

"Agents should have an identity. We hope our protocol doesn't just become a provider, but can also invest in the agent. We need identity and reputation system to annotate each agent." We had the amazing Yi (Founder and CEO Codatta) join our latest episode of AI on AIR hosted by Chi Zhang and Scott Shi - e/acc where he discussed: 1⃣ How AI boosts businesses 2⃣ Importance of data in model building and problems with scaling it 3⃣ Joining Binance Booster and getting listed on Binance Alpha 4⃣ Why he chose to build Codatta within the Kite AI ecosystem

"Agents should have an identity. We hope our protocol doesn't just become a provider, but can also invest in the agent. We need identity and reputation system to annotate each agent." We had the amazing Yi (Founder and CEO Codatta) join our latest episode of AI on AIR hosted by Chi Zhang and Scott Shi - e/acc where he discussed: 1⃣ How AI boosts businesses 2⃣ Importance of data in model building and problems with scaling it 3⃣ Joining Binance Booster and getting listed on Binance Alpha 4⃣ Why he chose to build Codatta within the Kite AI ecosystem

KITE AI

24,289 views • 1 year ago

Responsible AI should not feel like the path of most resistance. That was the strongest idea from my SAS Innovate conversation with Reggie Townsend, leading Data Ethics, Governance, and Social Impact at SAS Software. Reggie framed governance not as a compliance layer, but as a way to scale human judgment. Too often, we innovate first and govern later. Model selected. Agent deployed. Process built. Then governance arrives at the end and feels like friction. Bolt it on after the fact, and resistance is guaranteed. The opportunity: design responsible AI, so it becomes intuitive, action-oriented, and useful in the flow of work. That is what Reggie meant by making responsibility "irresistible." Second point: Use cases must lead. When everyone can access the same models, differentiation will not come from the technology. It will come from how leaders define outcomes, govern applications, and connect business value to institutional trust. The risk does not live in the model. The value does not live in the model. Both live in the use case. For CEOs and boards, this is the shift: from model-first oversight to outcome-first accountability. Better questions before scaling AI: - What human decision are we shaping? - What business outcome are we improving? - What risk are we containing? - What judgment are we extending? Responsible AI becomes strategic when it helps people make better decisions, faster, with greater confidence. Most leaders can't see where governance sits inside their AI operating model. SAS AI Navigator makes it visible: Design for the human. Not only the technology.

Responsible AI should not feel like the path of most resistance. That was the strongest idea from my SAS Innovate conversation with Reggie Townsend, leading Data Ethics, Governance, and Social Impact at SAS Software. Reggie framed governance not as a compliance layer, but as a way to scale human judgment. Too often, we innovate first and govern later. Model selected. Agent deployed. Process built. Then governance arrives at the end and feels like friction. Bolt it on after the fact, and resistance is guaranteed. The opportunity: design responsible AI, so it becomes intuitive, action-oriented, and useful in the flow of work. That is what Reggie meant by making responsibility "irresistible." Second point: Use cases must lead. When everyone can access the same models, differentiation will not come from the technology. It will come from how leaders define outcomes, govern applications, and connect business value to institutional trust. The risk does not live in the model. The value does not live in the model. Both live in the use case. For CEOs and boards, this is the shift: from model-first oversight to outcome-first accountability. Better questions before scaling AI: - What human decision are we shaping? - What business outcome are we improving? - What risk are we containing? - What judgment are we extending? Responsible AI becomes strategic when it helps people make better decisions, faster, with greater confidence. Most leaders can't see where governance sits inside their AI operating model. SAS AI Navigator makes it visible: Design for the human. Not only the technology.

Sabine VanderLinden

835,838 views • 3 months ago

Greg Osuri built the case for decentralized AI in three layers in under 5 minutes. Layer one: access. Anthropic's Claude is the best coding model available. OpenClaw is the most successful open-source project in history. Anthropic just announced they're banning OpenClaw usage. The most powerful model cutting off the most loved tool. The company that controls the best intelligence can revoke your access whenever their interests change. Layer two: transparency. You have no idea what biases Anthropic has built into Claude. If AI models directly impact how you think, the decision tree needs to be visible. Closed-source models with hidden biases making decisions that shape your worldview isn't a feature, it's a risk. Layer three: compute. Amazon may be picking winners in AI by controlling compute allocation. Cloud companies stopped being neutral marketplaces and started giving GPUs to the companies they've bet on. Amazon overwhelmingly backing Anthropic might have contributed to Claude's dominance. The cloud is no longer a utility. It's a kingmaker. Greg Osuri 🇺🇸's conclusion: intelligence should be free like water and air. Decentralized AI isn't ideology. It's the only architecture where the model can't ban your tools, the biases can be audited, and the compute isn't allocated by a company picking favorites.

Greg Osuri built the case for decentralized AI in three layers in under 5 minutes. Layer one: access. Anthropic's Claude is the best coding model available. OpenClaw is the most successful open-source project in history. Anthropic just announced they're banning OpenClaw usage. The most powerful model cutting off the most loved tool. The company that controls the best intelligence can revoke your access whenever their interests change. Layer two: transparency. You have no idea what biases Anthropic has built into Claude. If AI models directly impact how you think, the decision tree needs to be visible. Closed-source models with hidden biases making decisions that shape your worldview isn't a feature, it's a risk. Layer three: compute. Amazon may be picking winners in AI by controlling compute allocation. Cloud companies stopped being neutral marketplaces and started giving GPUs to the companies they've bet on. Amazon overwhelmingly backing Anthropic might have contributed to Claude's dominance. The cloud is no longer a utility. It's a kingmaker. Greg Osuri 🇺🇸's conclusion: intelligence should be free like water and air. Decentralized AI isn't ideology. It's the only architecture where the model can't ban your tools, the biases can be audited, and the compute isn't allocated by a company picking favorites.

The Rollup

15,005 views • 4 months ago

Choose a model (any model) and build your application with it. Do not spend time swapping models early on. Do not try to optimize before you have a working system. This is one of the first recommendations I make to every new team I consult with. Eventually, it will be time to optimize the model. • You may need a cheaper model • You may need a faster model • You might need a smarter model Good luck if you stitched together 12 different APIs and SDKs from 7 different vendors. Over half of the companies I consult for run on Microsoft software and have access to Microsoft Foundry. Microsoft Foundry is a complete agentic ecosystem. If you're in that world and building AI applications, Microsoft Foundry is where everything lives: • Models (largest selection in the market) • Agentic SDK (Python, C#, JavaScript/TypeScript) • Tools • Evaluations • Monitoring They are fully integrated with GitHub and Visual Studio Code. The best part: Their agentic platform is fully agnostic of the models you use. You can integrate with any model using the same OpenAI-style API. Swapping one model for another takes 1 second.

Choose a model (any model) and build your application with it. Do not spend time swapping models early on. Do not try to optimize before you have a working system. This is one of the first recommendations I make to every new team I consult with. Eventually, it will be time to optimize the model. • You may need a cheaper model • You may need a faster model • You might need a smarter model Good luck if you stitched together 12 different APIs and SDKs from 7 different vendors. Over half of the companies I consult for run on Microsoft software and have access to Microsoft Foundry. Microsoft Foundry is a complete agentic ecosystem. If you're in that world and building AI applications, Microsoft Foundry is where everything lives: • Models (largest selection in the market) • Agentic SDK (Python, C#, JavaScript/TypeScript) • Tools • Evaluations • Monitoring They are fully integrated with GitHub and Visual Studio Code. The best part: Their agentic platform is fully agnostic of the models you use. You can integrate with any model using the same OpenAI-style API. Swapping one model for another takes 1 second.

Santiago

12,014 views • 6 months ago

Most AI tools give everyone access to the same generic models. The result? Everything looks the same. We have a different vision for AI creation. Introducing TITLES, a new creative studio built around AI models trained and owned by artists. In Studio, you can create with distinct visual perspectives developed by artists, across image and video — all in one place. This is the future we're building toward: not one model for everyone, but a growing network of unique styles you can build with — where artists get credited and paid as the work spreads. Enter Your Creative Studio:

Most AI tools give everyone access to the same generic models. The result? Everything looks the same. We have a different vision for AI creation. Introducing TITLES, a new creative studio built around AI models trained and owned by artists. In Studio, you can create with distinct visual perspectives developed by artists, across image and video — all in one place. This is the future we're building toward: not one model for everyone, but a growing network of unique styles you can build with — where artists get credited and paid as the work spreads. Enter Your Creative Studio:

TITLES

2,747,215 views • 4 months ago

$After taking some time off post-Rapid, I'm excited to share what I’ve been up to since: Datawizz AI! We’ve raised a $12.5M Seed led by Human Capital to make AI 10x cheaper, 2x more accurate and 15x faster by transitioning from LLMs to SLMs. AI is eating the world. But unit economics are eating AI. Looking at the fastest growing AI products, they all share two traits - growing fast, and painful inference bills. General-purpose LLMs are just too expensive to run. A big reason for that is we train LLMs to be good at everything - answer any question, be an expert on any topic. The big labs dub this "generalisation", but for real-world applications, it is unnecessary. In reality - many AI applications need models to be experts in one thing - and do that thing extremely well. Your coding model doesn’t need to memorize ancient recipes for Garum sauce. This is where Datawizz comes in - we sit between the AI applications and automatically create smaller (100x-1,000x) specialized models to handle specific aspects of your work. By focusing the model and combining industry-data in the distillation process - we end up with models that beat SOTA LLMs at a fraction of the cost. We created Datawizz to make AI specialized and scalable. We’re early in the journey, but have already been able to save companies 90%+ on their inference bill and speed up their apps by 10x. Excited to build better AI platforms? Join the Datawizz team (link in first comment)$

After taking some time off post-Rapid, I'm excited to share what I’ve been up to since: Datawizz AI! We’ve raised a $12.5M Seed led by Human Capital to make AI 10x cheaper, 2x more accurate and 15x faster by transitioning from LLMs to SLMs. AI is eating the world. But unit economics are eating AI. Looking at the fastest growing AI products, they all share two traits - growing fast, and painful inference bills. General-purpose LLMs are just too expensive to run. A big reason for that is we train LLMs to be good at everything - answer any question, be an expert on any topic. The big labs dub this "generalisation", but for real-world applications, it is unnecessary. In reality - many AI applications need models to be experts in one thing - and do that thing extremely well. Your coding model doesn’t need to memorize ancient recipes for Garum sauce. This is where Datawizz comes in - we sit between the AI applications and automatically create smaller (100x-1,000x) specialized models to handle specific aspects of your work. By focusing the model and combining industry-data in the distillation process - we end up with models that beat SOTA LLMs at a fraction of the cost. We created Datawizz to make AI specialized and scalable. We’re early in the journey, but have already been able to save companies 90%+ on their inference bill and speed up their apps by 10x. Excited to build better AI platforms? Join the Datawizz team (link in first comment)

Iddo Gino 🐙

21,928 views • 11 months ago

SITUATION EXPLAINED: 70% of frontier model queries could run locally for free. clem 🤗, co-founder and CEO of Hugging Face: "There was an interesting study from Stanford published last year showing that 70% of the queries that people ask to ChatGPT could be accurately answered locally on your laptop. For free." "Most of the AI workloads that people do today with frontier models could be done by models that are cheaper, faster, more customizable, more controllable. And they don't do it because frankly it's a pain to pick the right model." "You're subsidized so you don't have to care because you have your subscription, so you route everything to Einstein. 'Hey Einstein, what's the weather today?'" "In normal life, he would be like, 'I'm not answering your silly question.' But because it's AI and subsidized by the AI labs, all the questions are getting routed to Einstein versus in an ideal world you have different models that are more specialized and better at answering your questions in different domains." "It's possible that routing is gonna redistribute a lot of the value capture from frontier models to a more long tail of models. It's like AI maturing... we are in the first phase where it's very simple, everyone using just one giant model. Now we're moving to the second phase."

SITUATION EXPLAINED: 70% of frontier model queries could run locally for free. clem 🤗, co-founder and CEO of Hugging Face: "There was an interesting study from Stanford published last year showing that 70% of the queries that people ask to ChatGPT could be accurately answered locally on your laptop. For free." "Most of the AI workloads that people do today with frontier models could be done by models that are cheaper, faster, more customizable, more controllable. And they don't do it because frankly it's a pain to pick the right model." "You're subsidized so you don't have to care because you have your subscription, so you route everything to Einstein. 'Hey Einstein, what's the weather today?'" "In normal life, he would be like, 'I'm not answering your silly question.' But because it's AI and subsidized by the AI labs, all the questions are getting routed to Einstein versus in an ideal world you have different models that are more specialized and better at answering your questions in different domains." "It's possible that routing is gonna redistribute a lot of the value capture from frontier models to a more long tail of models. It's like AI maturing... we are in the first phase where it's very simple, everyone using just one giant model. Now we're moving to the second phase."

MTS

50,185 views • 2 months ago

India’s Sovereign Move - Sarvam Vision Cracks the Code across 22 Indian languages Sarvam Vision is an AI model from Sarvam AI that excels in optical character recognition (OCR) for 22 Indian languages, outperforming global models like Gemini and GPT-4o on Indic benchmarks. This achievement supports India's push for sovereign AI by digitizing real-world documents and archives in local scripts. Sarvam Vision, a 3-billion-parameter model, handles "messy" scanned paperwork across languages like Hindi (95.91% accuracy), Bengali (92.61%), Tamil (93.42%), Marathi (93.13%), and even low-resource ones like Santali and Dogri (over 80%). It structures data natively without relying on English translation layers, enabling applications in cultural recovery and document intelligence. The model powers India's sovereign AI infrastructure under the IndiaAI Mission, where Sarvam was selected to build a national LLM. This launch fits Sarvam's broader stack, including prior models like Sarvam-Translate (also for 22 languages) and audio models for code-mixed speech. By focusing on Indic challenges ignored by Western AI, it unlocks centuries of knowledge in non-English archives. Recent demos, like digitizing historical texts, highlight its real-world impact beyond benchmarks. Credit : AIM Network.

India’s Sovereign Move - Sarvam Vision Cracks the Code across 22 Indian languages Sarvam Vision is an AI model from Sarvam AI that excels in optical character recognition (OCR) for 22 Indian languages, outperforming global models like Gemini and GPT-4o on Indic benchmarks. This achievement supports India's push for sovereign AI by digitizing real-world documents and archives in local scripts. Sarvam Vision, a 3-billion-parameter model, handles "messy" scanned paperwork across languages like Hindi (95.91% accuracy), Bengali (92.61%), Tamil (93.42%), Marathi (93.13%), and even low-resource ones like Santali and Dogri (over 80%). It structures data natively without relying on English translation layers, enabling applications in cultural recovery and document intelligence. The model powers India's sovereign AI infrastructure under the IndiaAI Mission, where Sarvam was selected to build a national LLM. This launch fits Sarvam's broader stack, including prior models like Sarvam-Translate (also for 22 languages) and audio models for code-mixed speech. By focusing on Indic challenges ignored by Western AI, it unlocks centuries of knowledge in non-English archives. Recent demos, like digitizing historical texts, highlight its real-world impact beyond benchmarks. Credit : AIM Network.

Augadh

20,688 views • 6 months ago

Head of Nemotron Bryan Catanzaro says NVIDIA is sitting on internal "teacher" models that beat everything it has publicly shipped and he's trying to release these as well "This technology of using teacher models and multi-domain on-policy distillation has been really key for us in these ways." "Now, what does that mean? Well, it definitely means that inside NVIDIA we have a collection of teacher models that are individually better at different domains than any of the models that we've released." "And actually, I'm trying to release those teacher models as part of the Nemotron project as well -- because as you know, Nemotron is not just a model." "Nemotron is an open technology effort. It includes data sets, it includes techniques, all of our ideas about how to build a model, which we publish." "Of course it includes the models themselves, but I think it should also include collections of teacher models." "I'd imagine this could be a really great way of federating open technologies for AI across the ecosystem." _______ Wrote more on a prior Bryan interview here:

Head of Nemotron Bryan Catanzaro says NVIDIA is sitting on internal "teacher" models that beat everything it has publicly shipped and he's trying to release these as well "This technology of using teacher models and multi-domain on-policy distillation has been really key for us in these ways." "Now, what does that mean? Well, it definitely means that inside NVIDIA we have a collection of teacher models that are individually better at different domains than any of the models that we've released." "And actually, I'm trying to release those teacher models as part of the Nemotron project as well -- because as you know, Nemotron is not just a model." "Nemotron is an open technology effort. It includes data sets, it includes techniques, all of our ideas about how to build a model, which we publish." "Of course it includes the models themselves, but I think it should also include collections of teacher models." "I'd imagine this could be a really great way of federating open technologies for AI across the ecosystem." _______ Wrote more on a prior Bryan interview here:

Fireside Alpha

29,125 views • 4 days ago

Today, we’re launching Parsed. We are incredibly lucky to live in a world where we stand on the shoulders of giants, first in science and now in AI. Our heroes have gotten us to this point, where we have brilliant general intelligence in our pocket. But this is a local minima. We now have an ecosystem of burgeoning tasks where each requires a different kind of intelligence, a different context, a whole host of implicit assumptions and latent knowledge and domain expertise that is very difficult to cram into a system prompt. The big labs want you renting their $50k/month amnesiac interns that forget everything between conversations. Generic behemoths that get quantised, versioned and deprecated behind the scenes, where the only element of control you have is your messy monolithic user prompt. We want people who need their own intelligence to be able to not only access it, but also control it. And whilst the big general models are unbelievably good chatbots and coding agents and purveyors of the world, specialisation of intelligence is required. Clinical scribes, marketing compliance agents, legal red-lining models, insurance policy recommenders, the list goes on. And so that’s what Parsed does: deploy your own frontier model that actually learns. We eval your specific task, build a custom evaluation harness, optimise a model just for you, and host it with continual learning. We bake all the context and knowledge of your task into the model itself, from your engineers to your domain experts to customer feedback, all in a tight SFT → RL loop, with useful interpretability made possible by the open-source ecosystem we build on top of. No more 2000-word prompts with seventeen "IMPORTANT: NEVER DO X" clauses. Your model gets better at YOUR job every single day; the amnesiac pseudo-gods have had their run. Your model, your data, your moat. Let's build 🫡

Today, we’re launching Parsed. We are incredibly lucky to live in a world where we stand on the shoulders of giants, first in science and now in AI. Our heroes have gotten us to this point, where we have brilliant general intelligence in our pocket. But this is a local minima. We now have an ecosystem of burgeoning tasks where each requires a different kind of intelligence, a different context, a whole host of implicit assumptions and latent knowledge and domain expertise that is very difficult to cram into a system prompt. The big labs want you renting their $50k/month amnesiac interns that forget everything between conversations. Generic behemoths that get quantised, versioned and deprecated behind the scenes, where the only element of control you have is your messy monolithic user prompt. We want people who need their own intelligence to be able to not only access it, but also control it. And whilst the big general models are unbelievably good chatbots and coding agents and purveyors of the world, specialisation of intelligence is required. Clinical scribes, marketing compliance agents, legal red-lining models, insurance policy recommenders, the list goes on. And so that’s what Parsed does: deploy your own frontier model that actually learns. We eval your specific task, build a custom evaluation harness, optimise a model just for you, and host it with continual learning. We bake all the context and knowledge of your task into the model itself, from your engineers to your domain experts to customer feedback, all in a tight SFT → RL loop, with useful interpretability made possible by the open-source ecosystem we build on top of. No more 2000-word prompts with seventeen "IMPORTANT: NEVER DO X" clauses. Your model gets better at YOUR job every single day; the amnesiac pseudo-gods have had their run. Your model, your data, your moat. Let's build 🫡

Charlie O'Neill

137,812 views • 1 year ago