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.

AI safety is back in the spotlight, driven by concerns about recursive self improvement. We built the first third party benchmark to measure just how close AI is building its own successor. In collaboration with marimo and CoreWeave we built the The RSI Index, which runs every frontier model... show more

Vals AI

20,057 subscribers

21,513 views • 8 days ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

10 Comments

Vals AI8 days ago

No agent reached the reference on any task, but the results are telling. Claude Fable 5.1 leads the RSI Index at 35.0%, topping compression, LM training, and harness engineering. On LM training it hit 0.8536 BPB against a published 0.85, a recipe it retrained in 30 minutes on one H100. The published result took roughly 175 hours on a TPU v2.

Vals AI8 days ago

The models did not clearly invent new architectures or training algorithms; they mostly rediscovered known techniques, combined them into working systems, and did substantial work to test and validate them. Agents used Marimo notebooks as durable experiment ledgers, letting them record results, iterate, and pick up their research across long-running autonomous sessions.

Vals AI8 days ago

The leader in this case is actually one of the most affordable system, Fable which has a total index cost of $1,481 compared with $1,886 on Opus 5 and $3,552 on GPT-6 Astra. This shows that paying more does not necessarily yield better research.

Vals AI8 days ago

These are single runs on scoped tasks with fixed budgets in off-the-shelf harnesses. More compute and better tooling could change the picture. That is why the index is built to be rerun as models advance. Methodology and results at Paper Preprint:

Ornith8 days ago

@marimo_io @CoreWeave 🫡

Mike Medeiros8 days ago

@marimo_io @CoreWeave Useful to have a third-party RSI Index — curious what it actually scores: self-improvement loops, or “AI helping humans ship the next model”? Those are different claims.

Gandhi8 days ago

@marimo_io @CoreWeave Incredible

Miyuki Matthews8 days ago

@marimo_io @CoreWeave A timely take. Great video @akshaykagrawal @RayanKrishnan

SOP | AI Agent Builder8 days ago

@marimo_io @CoreWeave The benchmark becomes much more useful if each run logs tool calls, retries, wall-clock cost, and whether edits pass tests. Decompose successor-building into planning, coding, evaluation, and deployment gates.

liquidated (Dev Arc)8 days ago

@marimo_io @CoreWeave so the models can do the work but not beat humans yet, how long till that flips fr

Related Videos

Vals AI CEO Rayan Krishnan on how they score a model's ability to build its own successor: "One benchmark we released recently is our Recursive Self-Improvement Index. It's a topic which a lot of the big labs have been talking about and starting to report on in their model cards." "There isn't a shared language to talk about the RSI potential of models. So we created this as an apples-to-apples way to actually benchmark across the models." "In an ideal world, what you want to do is take a frontier model and have it train the next version of itself and see where the delta comes from. But obviously that's very expensive and slow." "What we're doing is forming a set of proxies for every part of the process it takes to build the next version of the model." "There's work around pre-training, post-training, harness-level engineering, and then seeing in which mechanisms and behaviors the models are able to do very good research work and build something new, and where they're struggling." Rayan Krishnan Jennifer Li

Vals AI CEO Rayan Krishnan on how they score a model's ability to build its own successor: "One benchmark we released recently is our Recursive Self-Improvement Index. It's a topic which a lot of the big labs have been talking about and starting to report on in their model cards." "There isn't a shared language to talk about the RSI potential of models. So we created this as an apples-to-apples way to actually benchmark across the models." "In an ideal world, what you want to do is take a frontier model and have it train the next version of itself and see where the delta comes from. But obviously that's very expensive and slow." "What we're doing is forming a set of proxies for every part of the process it takes to build the next version of the model." "There's work around pre-training, post-training, harness-level engineering, and then seeing in which mechanisms and behaviors the models are able to do very good research work and build something new, and where they're struggling." Rayan Krishnan Jennifer Li

a16z

34,591 views • 9 days ago

"to me, it's not about the human in the loop, but the human at the center." i sat down with Stefania Druga, staff research scientist Sakana AI, at our Browserbase booth at AI Engineer. after Massachusetts Institute of Technology (MIT) and Google DeepMind, stefania recently joined sakana's recursive self-improvement (rsi) lab in tokyo to build self-improving AI for the next frontier. we got into: - catastrophic forgetting in day-long agent research runs, and how memory-first harnesses help - sovereign AI and open-weights: building on models you control instead of locking into one provider - why computer use still frustrates people even as models get smarter - the domains ripe for self-improving research agents: math, genomics, rare conditions - my fav moment at the end: what tasks are better left human, and what are we going to do as humans after everything is automated? with open frontier models arriving fast, Stefania Druga's case for sovereign AI only gets stronger. loved this conversation. thanks for joining us at the Browserbase booth!

"to me, it's not about the human in the loop, but the human at the center." i sat down with Stefania Druga, staff research scientist Sakana AI, at our Browserbase booth at AI Engineer. after Massachusetts Institute of Technology (MIT) and Google DeepMind, stefania recently joined sakana's recursive self-improvement (rsi) lab in tokyo to build self-improving AI for the next frontier. we got into: - catastrophic forgetting in day-long agent research runs, and how memory-first harnesses help - sovereign AI and open-weights: building on models you control instead of locking into one provider - why computer use still frustrates people even as models get smarter - the domains ripe for self-improving research agents: math, genomics, rare conditions - my fav moment at the end: what tasks are better left human, and what are we going to do as humans after everything is automated? with open frontier models arriving fast, Stefania Druga's case for sovereign AI only gets stronger. loved this conversation. thanks for joining us at the Browserbase booth!

justine

19,853 views • 2 months ago

Cyber defense is entering the superintelligence era. ⚡ Introducing the CrowdStrike Cyber Superintelligence Lab, the first frontier AI research organization built for cyberdefense and AI safety. The Lab delivers CrowdStrike SafeMind, a family of purpose-built security models and harnesses. SafeMind runs as one system designed to deliver AI safety: an offensive model that finds the attack path, a defensive model that closes it, and the harnesses that operate both in the same loop. Built using NVIDIA Nemotron open models and using CoreWeave’s AI Cloud for training and inference, SafeMind goes beyond AI that simply identifies threats, it defeats them. Learn more:

Cyber defense is entering the superintelligence era. ⚡ Introducing the CrowdStrike Cyber Superintelligence Lab, the first frontier AI research organization built for cyberdefense and AI safety. The Lab delivers CrowdStrike SafeMind, a family of purpose-built security models and harnesses. SafeMind runs as one system designed to deliver AI safety: an offensive model that finds the attack path, a defensive model that closes it, and the harnesses that operate both in the same loop. Built using NVIDIA Nemotron open models and using CoreWeave’s AI Cloud for training and inference, SafeMind goes beyond AI that simply identifies threats, it defeats them. Learn more:

CrowdStrike

58,094 views • 17 days ago

Excited to release PostTrainBench v1.0! This benchmark evaluates the ability of frontier AI agents to post-train language models in a simplified setting. We believe this is a first step toward tracking progress in recursive self-improvement 🧵:

Excited to release PostTrainBench v1.0! This benchmark evaluates the ability of frontier AI agents to post-train language models in a simplified setting. We believe this is a first step toward tracking progress in recursive self-improvement 🧵:

Karina

196,392 views • 6 months ago

It's month close. AI models can pass the CPA exam, but do you trust them to close the books? Introducing APEX-Accounting, built in collaboration with Ramp and Ramp Labs, to test whether frontier AI can do real accounting work. It’s a new benchmark comprising 160 tasks across 10 simulated companies, authored and graded by accounting professionals with experience at firms like Deloitte, PwC, EY, and KPMG.

It's month close. AI models can pass the CPA exam, but do you trust them to close the books? Introducing APEX-Accounting, built in collaboration with Ramp and Ramp Labs, to test whether frontier AI can do real accounting work. It’s a new benchmark comprising 160 tasks across 10 simulated companies, authored and graded by accounting professionals with experience at firms like Deloitte, PwC, EY, and KPMG.

Mercor

123,698 views • 1 month ago

Today we're announcing Base Labs, a dedicated research organization focused on advancing open-source AI. We believe in a healthy, open frontier model ecosystem. To enable this, we are working on: - Blue-sky research on continual learning, the science of RL, and how models learn across their full lifecycle, with every experiment and recipe shared openly. - The BaseHub Data Foundry: the highest-quality open RL environments, training data, and real-world benchmarks, built for anyone to train and benchmark on. - Post-post training: taking open-source models and making them better, safer, and more aligned through continual post-training, built on our research, and deployed with our frontier safety stack so organizations can use open-source models with confidence. - Making models cheaper and more performant through our model performance research. This is a mission-driven research effort, not a commercial product. We believe the health of the open-source AI ecosystem matters and that the best way to advance it is to do science in the open. We’re hiring engineers, researchers, and research fellows to advance this mission.

Today we're announcing Base Labs, a dedicated research organization focused on advancing open-source AI. We believe in a healthy, open frontier model ecosystem. To enable this, we are working on: - Blue-sky research on continual learning, the science of RL, and how models learn across their full lifecycle, with every experiment and recipe shared openly. - The BaseHub Data Foundry: the highest-quality open RL environments, training data, and real-world benchmarks, built for anyone to train and benchmark on. - Post-post training: taking open-source models and making them better, safer, and more aligned through continual post-training, built on our research, and deployed with our frontier safety stack so organizations can use open-source models with confidence. - Making models cheaper and more performant through our model performance research. This is a mission-driven research effort, not a commercial product. We believe the health of the open-source AI ecosystem matters and that the best way to advance it is to do science in the open. We’re hiring engineers, researchers, and research fellows to advance this mission.

Base Labs

234,840 views • 15 days ago

Elon Musk says humans may exit the AI development loop entirely by next year. Musk: “Humans are gradually getting less and less in the loop on the recursive self-improvement. Every successive model is built by the one before it.”

Elon Musk says humans may exit the AI development loop entirely by next year. Musk: “Humans are gradually getting less and less in the loop on the recursive self-improvement. Every successive model is built by the one before it.”

illuminatibot

55,658 views • 6 months ago

We evaluated 30+ frontier embodied AI models. The result is clear: current generalist robot policies are still far from robust real-world manipulation. This is why we built RoboDojo.

We evaluated 30+ frontier embodied AI models. The result is clear: current generalist robot policies are still far from robust real-world manipulation. This is why we built RoboDojo.

Tianxing Chen

136,330 views • 2 months ago

"The episode of history where humanity was the apex intelligence is about to end." AI is crossing an important boundary: from learning mathematics invented by humans to inventing mathematics humans can learn from. Mo Gawdat, ex-Google executive explains how recursive self-development of AI is absolutely disrupting mathematics. Humans originally built AI using human intelligence, but increasingly, AI can can improve itself (i.e. recursive self-development) Once machines become capable of making increasingly better versions of their own methods, technological progress may accelerate far beyond the normal pace of human research. ---- From "Mo Gawdat" YouTube channel, (full video link in comment)

"The episode of history where humanity was the apex intelligence is about to end." AI is crossing an important boundary: from learning mathematics invented by humans to inventing mathematics humans can learn from. Mo Gawdat, ex-Google executive explains how recursive self-development of AI is absolutely disrupting mathematics. Humans originally built AI using human intelligence, but increasingly, AI can can improve itself (i.e. recursive self-development) Once machines become capable of making increasingly better versions of their own methods, technological progress may accelerate far beyond the normal pace of human research. ---- From "Mo Gawdat" YouTube channel, (full video link in comment)

Rohan Paul

13,314 views • 12 days ago

RSI section from the AI documentary Machine God The next threshold is Recursive Self-Improvement: the moment when AI can improve itself without human assistance. For decades this sounded like science fiction. Intelligence explosion scenarios imagined a system rewriting its own code, becoming smarter, then using that new intelligence to make still better versions of itself. But the idea looks less remote now that AI contributes directly to frontier science. In mathematics, recent systems have moved beyond solving contest problems to producing serious new arguments on long-standing open problems. AI is used to build physics world models and propose candidate theories or computational methods. These are early signs that machine cognition is entering the creative loop of science itself. The crucial transition comes when that loop turns inward. AI research is, after all, a technical discipline made of code, mathematics, models of information flow. These are exactly the domains in which frontier models are improving fastest. A model that can solve hard mathematical problems, write production-quality code, design experiments, read the literature, and evaluate benchmark results is already participating in the work of building its successor. At first this will look prosaic. AI systems will write kernel optimizations, improve training infrastructure, discover better data filters, tune reinforcement-learning pipelines, design new benchmarks, and suggest architectural modifications. Human researchers will remain in the loop, approving changes and interpreting results. But the important point is that the search process accelerates. The model becomes not just the product of the lab, but part of the lab’s research machinery. The system being optimized helps optimize the next system. This is the core RSI feedback loop: better models make AI research faster; faster AI research produces still better models; those models, in turn, become better researchers. The danger is that once this loop becomes sufficiently autonomous, it may stop resembling ordinary technological progress. Human institutions are slow because humans are slow: we read papers, attend meetings, debug code, sleep, argue, and wait for funding cycles. Machines do not have to operate on that timescale. An AI research collective can run continuously across millions of processors. This is the runaway possibility. Not that an AI instantly wakes up and recursively rewrites itself into a god, but that the entire AI ecosystem becomes an autocatalytic process. Capital buys compute; compute trains models; models improve models; better models attract more capital. At some point the dominant input into AI progress may no longer be human insight, but machine-generated insight, machine-written code, and machine-run experiments. Then the Butler-Land analogy becomes sharper. Humanity is no longer merely building machines. We are building machines that help build better machines. Once intelligence itself becomes part of the production function, the old categories — tool, worker, inventor, firm, market — begin to blur. The question is whether recursive self-improvement remains a managed industrial process, or whether it becomes the first technological process in history whose natural endpoint lies beyond human comprehension.

RSI section from the AI documentary Machine God The next threshold is Recursive Self-Improvement: the moment when AI can improve itself without human assistance. For decades this sounded like science fiction. Intelligence explosion scenarios imagined a system rewriting its own code, becoming smarter, then using that new intelligence to make still better versions of itself. But the idea looks less remote now that AI contributes directly to frontier science. In mathematics, recent systems have moved beyond solving contest problems to producing serious new arguments on long-standing open problems. AI is used to build physics world models and propose candidate theories or computational methods. These are early signs that machine cognition is entering the creative loop of science itself. The crucial transition comes when that loop turns inward. AI research is, after all, a technical discipline made of code, mathematics, models of information flow. These are exactly the domains in which frontier models are improving fastest. A model that can solve hard mathematical problems, write production-quality code, design experiments, read the literature, and evaluate benchmark results is already participating in the work of building its successor. At first this will look prosaic. AI systems will write kernel optimizations, improve training infrastructure, discover better data filters, tune reinforcement-learning pipelines, design new benchmarks, and suggest architectural modifications. Human researchers will remain in the loop, approving changes and interpreting results. But the important point is that the search process accelerates. The model becomes not just the product of the lab, but part of the lab’s research machinery. The system being optimized helps optimize the next system. This is the core RSI feedback loop: better models make AI research faster; faster AI research produces still better models; those models, in turn, become better researchers. The danger is that once this loop becomes sufficiently autonomous, it may stop resembling ordinary technological progress. Human institutions are slow because humans are slow: we read papers, attend meetings, debug code, sleep, argue, and wait for funding cycles. Machines do not have to operate on that timescale. An AI research collective can run continuously across millions of processors. This is the runaway possibility. Not that an AI instantly wakes up and recursively rewrites itself into a god, but that the entire AI ecosystem becomes an autocatalytic process. Capital buys compute; compute trains models; models improve models; better models attract more capital. At some point the dominant input into AI progress may no longer be human insight, but machine-generated insight, machine-written code, and machine-run experiments. Then the Butler-Land analogy becomes sharper. Humanity is no longer merely building machines. We are building machines that help build better machines. Once intelligence itself becomes part of the production function, the old categories — tool, worker, inventor, firm, market — begin to blur. The question is whether recursive self-improvement remains a managed industrial process, or whether it becomes the first technological process in history whose natural endpoint lies beyond human comprehension.

steve hsu

61,671 views • 11 days ago

Dropping a new AI benchmark: TetrisBench and a human vs AI Tetris arena 🕹️ See how you rank against the frontier models. Only 1 human has beaten a model so far, so try it out be the 2nd! Stats so far: Opus 4.5 > GPT5.2 > Sonnet > Grok > most humans 🧵

Dropping a new AI benchmark: TetrisBench and a human vs AI Tetris arena 🕹️ See how you rank against the frontier models. Only 1 human has beaten a model so far, so try it out be the 2nd! Stats so far: Opus 4.5 > GPT5.2 > Sonnet > Grok > most humans 🧵

Yoko

35,709 views • 8 months ago

.L3Harris says Palantir helped them beat frontier AI models in less than 2 days, at 95% lower cost: "When we fine-tuned open source models trained on our own data, we were able to outperform the frontier models in less than 48 hours." "The cost of our fine-tuned open source model was 95% lower than the frontier models we were using." "AI is a commodity. It's all about the data. We view our data as a corporate asset. It's our unique hard-earned knowledge." "We believe American defense companies should not be a vassal for frontier AI labs, handing over our data and institutional knowledge, hoping to rent back the intelligence it creates." " We own the model, we own the compute, we own the advantage."

.L3Harris says Palantir helped them beat frontier AI models in less than 2 days, at 95% lower cost: "When we fine-tuned open source models trained on our own data, we were able to outperform the frontier models in less than 48 hours." "The cost of our fine-tuned open source model was 95% lower than the frontier models we were using." "AI is a commodity. It's all about the data. We view our data as a corporate asset. It's our unique hard-earned knowledge." "We believe American defense companies should not be a vassal for frontier AI labs, handing over our data and institutional knowledge, hoping to rent back the intelligence it creates." " We own the model, we own the compute, we own the advantage."

Jawwwn

1,095,465 views • 8 days ago

SITUATION EXPLAINED: Paradigm built a playable RSI simulator, and the fastest singularity is 26 days. • You run an AI lab from 2017 through takeoff, built on "The Economics of Recursive Self-Improvement" from the Elasticity Institute • Stage one is a single researcher clicking think, with the rate of improvement visibly slowing as you click • Stage two is human capital, the second researcher adds 80% productivity and each one after adds less • Stage three is the bitter lesson, where you stop scaling researchers and start scaling compute • Stage four is the scaling hypothesis, 20,000 GPUs and a model trained to a ten-second time horizon • A separate RSI Model Explorer lets you set return to R&D, algorithmic efficiency, idea stock spillover, capability response, and research response • Max every assumption and you get a singularity 26 days from founding the lab Theo Jaffee: "This is a really cool game. I think this will help a lot of normies understand how AI progress works, to a first approximation."

SITUATION EXPLAINED: Paradigm built a playable RSI simulator, and the fastest singularity is 26 days. • You run an AI lab from 2017 through takeoff, built on "The Economics of Recursive Self-Improvement" from the Elasticity Institute • Stage one is a single researcher clicking think, with the rate of improvement visibly slowing as you click • Stage two is human capital, the second researcher adds 80% productivity and each one after adds less • Stage three is the bitter lesson, where you stop scaling researchers and start scaling compute • Stage four is the scaling hypothesis, 20,000 GPUs and a model trained to a ten-second time horizon • A separate RSI Model Explorer lets you set return to R&D, algorithmic efficiency, idea stock spillover, capability response, and research response • Max every assumption and you get a singularity 26 days from founding the lab Theo Jaffee: "This is a really cool game. I think this will help a lot of normies understand how AI progress works, to a first approximation."

MTS

25,661 views • 1 month ago

“If we can make AI make better chips, then those chips can power the next generation of AI models.” Startup ‘Ricursive Intelligence,’ founded by Azalia Mirhoseini and Anna Goldie, has been building AI tools to do just that: to speed up the chip design process. The company’s name is a nod to “recursive self-improvement,” the idea that AI could one day bootstrap itself without human involvement. If they’re successful, Goldie and Mirhoseini could revolutionize the way the chip industry works. At the moment, AI companies mostly optimize their models for certain chips. But in a future where chip design happens far faster, it could work the other way around: chips optimized for specific models. The founders are on the TIME100 AI, the list of people shaping the development of AI in 2026. See the full list:

“If we can make AI make better chips, then those chips can power the next generation of AI models.” Startup ‘Ricursive Intelligence,’ founded by Azalia Mirhoseini and Anna Goldie, has been building AI tools to do just that: to speed up the chip design process. The company’s name is a nod to “recursive self-improvement,” the idea that AI could one day bootstrap itself without human involvement. If they’re successful, Goldie and Mirhoseini could revolutionize the way the chip industry works. At the moment, AI companies mostly optimize their models for certain chips. But in a future where chip design happens far faster, it could work the other way around: chips optimized for specific models. The founders are on the TIME100 AI, the list of people shaping the development of AI in 2026. See the full list:

TIME

38,584 views • 22 days ago

“One of the great things about Iso is how interdisciplinary everyone is, people come from so many different backgrounds.” We spoke with Bilal Khan, Software Engineer at Isomorphic Labs, about our unique, collaborative culture and frontier AI built to tackle the world's most complex problems. Watch the video and head to the comments to explore our current roles.

“One of the great things about Iso is how interdisciplinary everyone is, people come from so many different backgrounds.” We spoke with Bilal Khan, Software Engineer at Isomorphic Labs, about our unique, collaborative culture and frontier AI built to tackle the world's most complex problems. Watch the video and head to the comments to explore our current roles.

Isomorphic Labs

19,870 views • 6 months ago

California built the first AI safety regulations in the country. We proved you can protect people and let innovation thrive at the same time. In the absence of federal leadership, we’ll continue to do the work Donald Trump refuses to do.

California built the first AI safety regulations in the country. We proved you can protect people and let innovation thrive at the same time. In the absence of federal leadership, we’ll continue to do the work Donald Trump refuses to do.

Gavin Newsom

60,745 views • 5 months ago

Today is Day 11 of 12 Days of Agent Two. Agent Two isn't just Claude under the hood - we get asked this a lot. The truth is - while we use all the state of the art models out there, the Agent's real intelligence comes from its harness: a domain-specific layer built from continuous creative research that shapes how every model is actually used - more about this in the video. But this harness is what makes Agent Two frontier intelligence for creative work - at 4x the speed, and half the price.

Today is Day 11 of 12 Days of Agent Two. Agent Two isn't just Claude under the hood - we get asked this a lot. The truth is - while we use all the state of the art models out there, the Agent's real intelligence comes from its harness: a domain-specific layer built from continuous creative research that shapes how every model is actually used - more about this in the video. But this harness is what makes Agent Two frontier intelligence for creative work - at 4x the speed, and half the price.

Invideo

2,559,759 views • 1 month ago

#WATCH | Bengaluru, Karnataka: Co-founder, Sarvam AI, Pratyush Kumar says, "... Let me begin with what really is the mission of Sarvam. Many of you know us as a company, but it's important for us to tell you how we approach things. There are three things worth saying. Firstly, we are building at the frontier. AI is one of those technologies that is not only financially lucrative to build, but is also at the pinnacle of human engineering and insight. That pinnacle is moving much faster than in other generations of technology. That's why being at the frontier, at the very edge, becomes so significant. Everything that we do at Sarvam is about being at the frontier, building it, defining it, keeping pace with it, and pushing it further. Our first commitment is to have an innovation-led organisation that is building the frontier here in India..."

#WATCH | Bengaluru, Karnataka: Co-founder, Sarvam AI, Pratyush Kumar says, "... Let me begin with what really is the mission of Sarvam. Many of you know us as a company, but it's important for us to tell you how we approach things. There are three things worth saying. Firstly, we are building at the frontier. AI is one of those technologies that is not only financially lucrative to build, but is also at the pinnacle of human engineering and insight. That pinnacle is moving much faster than in other generations of technology. That's why being at the frontier, at the very edge, becomes so significant. Everything that we do at Sarvam is about being at the frontier, building it, defining it, keeping pace with it, and pushing it further. Our first commitment is to have an innovation-led organisation that is building the frontier here in India..."

ANI

25,111 views • 1 month ago