Nenad Tomasev
@nenadtomasev.bsky.social
Developing AI responsibly.
Senior Staff Research Scientist at Google DeepMind. Opinions are my own.
Senior Staff Research Scientist at Google DeepMind. Opinions are my own.
Reposted by Nenad Tomasev
Great work from some of my colleagues, check this out!
Such a creative application of generative AI! ♟️
www.chess.com/news/view/ai...
Such a creative application of generative AI! ♟️
www.chess.com/news/view/ai...
DeepMind's AI Learns To Create Original Chess Puzzles, Praised By GMs
In a new study, researchers from Google DeepMind have created an AI system that is capable of generating creative chess puzzles, some of which impressed experts in chess compositions.
www.chess.com
October 29, 2025 at 8:54 PM
Great work from some of my colleagues, check this out!
Such a creative application of generative AI! ♟️
www.chess.com/news/view/ai...
Such a creative application of generative AI! ♟️
www.chess.com/news/view/ai...
Reposted by Nenad Tomasev
Take a sneak peak at The Paradox of the Organism at Google books books.google.com/books?hl=sv&...
Don’t forget to pre-order!
Don’t forget to pre-order!
September 14, 2025 at 10:35 PM
Take a sneak peak at The Paradox of the Organism at Google books books.google.com/books?hl=sv&...
Don’t forget to pre-order!
Don’t forget to pre-order!
Another concept that we explore is that of Mission Economies realized through steerable AI agent markets.
Happy to share a new preprint:
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
Virtual Agent Economies
The rapid adoption of autonomous AI agents is giving rise to a new economic layer where agents transact and coordinate at scales and speeds beyond direct human oversight. We propose the "sandbox econo...
arxiv.org
September 15, 2025 at 1:14 PM
Another concept that we explore is that of Mission Economies realized through steerable AI agent markets.
One of the concepts that we discuss is that of a hypothetical High-Frequency Negotiation (HFN) framework between Personal AI Assistants, on behalf of their users, and in line with individual preferences.
Happy to share a new preprint:
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
Virtual Agent Economies
The rapid adoption of autonomous AI agents is giving rise to a new economic layer where agents transact and coordinate at scales and speeds beyond direct human oversight. We propose the "sandbox econo...
arxiv.org
September 15, 2025 at 1:13 PM
One of the concepts that we discuss is that of a hypothetical High-Frequency Negotiation (HFN) framework between Personal AI Assistants, on behalf of their users, and in line with individual preferences.
Happy to share a new preprint:
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
Virtual Agent Economies
The rapid adoption of autonomous AI agents is giving rise to a new economic layer where agents transact and coordinate at scales and speeds beyond direct human oversight. We propose the "sandbox econo...
arxiv.org
September 15, 2025 at 1:12 PM
Happy to share a new preprint:
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
Virtual Agent Economies
arxiv.org/abs/2509.10147
where we discuss a number of possible frameworks for establishing steerable agent markets.
For those with interest in mental health and AI, and in particular on how potentially sensitive data gets collected and used there - there is now a Delphi survey that you can fill out to inform how this gets done, as a part of PARQAIR-MH project (www.parqair.org/home), via:
redcap.link/2l91l043
redcap.link/2l91l043
www.parqair.org
May 9, 2025 at 11:53 AM
For those with interest in mental health and AI, and in particular on how potentially sensitive data gets collected and used there - there is now a Delphi survey that you can fill out to inform how this gets done, as a part of PARQAIR-MH project (www.parqair.org/home), via:
redcap.link/2l91l043
redcap.link/2l91l043
Our work on concept discovery towards bridging the human-AI knowledge gap in AlphaZero has now been published in PNAS.
As future AI systems become even more capable, we should be thinking of ways of utilizing them not only to perform tasks, but also to further our own knowledge and understanding.
As future AI systems become even more capable, we should be thinking of ways of utilizing them not only to perform tasks, but also to further our own knowledge and understanding.
April 1, 2025 at 2:13 PM
Our work on concept discovery towards bridging the human-AI knowledge gap in AlphaZero has now been published in PNAS.
As future AI systems become even more capable, we should be thinking of ways of utilizing them not only to perform tasks, but also to further our own knowledge and understanding.
As future AI systems become even more capable, we should be thinking of ways of utilizing them not only to perform tasks, but also to further our own knowledge and understanding.
Reposted by Nenad Tomasev
This was cool: "Gemini 2.5, create a sim that is cross between Jujujajaki networks and cellular automata"
Gemini: "What's a Jujujajaki network?" I paste in a paper.
Gemini: "Got it, a dynamic network with local search & exploration." Worked in one shot.
Me: "Make it nicer"
Some cleverness here
Gemini: "What's a Jujujajaki network?" I paste in a paper.
Gemini: "Got it, a dynamic network with local search & exploration." Worked in one shot.
Me: "Make it nicer"
Some cleverness here
April 1, 2025 at 1:48 AM
This was cool: "Gemini 2.5, create a sim that is cross between Jujujajaki networks and cellular automata"
Gemini: "What's a Jujujajaki network?" I paste in a paper.
Gemini: "Got it, a dynamic network with local search & exploration." Worked in one shot.
Me: "Make it nicer"
Some cleverness here
Gemini: "What's a Jujujajaki network?" I paste in a paper.
Gemini: "Got it, a dynamic network with local search & exploration." Worked in one shot.
Me: "Make it nicer"
Some cleverness here
An interesting read on self-organizing systems, and how order and disorder may manifest differently at different scales: www.nature.com/articles/s44...
Self-organizing systems: what, how, and why? - npj Complexity
npj Complexity - Self-organizing systems: what, how, and why?
www.nature.com
March 26, 2025 at 10:54 AM
An interesting read on self-organizing systems, and how order and disorder may manifest differently at different scales: www.nature.com/articles/s44...
Reposted by Nenad Tomasev
Dataset Distillation (2018/2020)
They show that it is possible to compress 60,000 MNIST training images into just 10 synthetic distilled images (one per class) and achieve close to original performance with only a few gradient descent steps, given a fixed network initialization.
They show that it is possible to compress 60,000 MNIST training images into just 10 synthetic distilled images (one per class) and achieve close to original performance with only a few gradient descent steps, given a fixed network initialization.
March 5, 2025 at 12:23 AM
Dataset Distillation (2018/2020)
They show that it is possible to compress 60,000 MNIST training images into just 10 synthetic distilled images (one per class) and achieve close to original performance with only a few gradient descent steps, given a fixed network initialization.
They show that it is possible to compress 60,000 MNIST training images into just 10 synthetic distilled images (one per class) and achieve close to original performance with only a few gradient descent steps, given a fixed network initialization.
I'm happy to advertise an upcoming Student Researcher position on my Agent Frontiers team here at the Google DeepMind Foundational Research Unit, aimed for a start date early in the summer (currently listed as late June, but obviously somewhat flexible).
March 3, 2025 at 2:02 PM
I'm happy to advertise an upcoming Student Researcher position on my Agent Frontiers team here at the Google DeepMind Foundational Research Unit, aimed for a start date early in the summer (currently listed as late June, but obviously somewhat flexible).
Reposted by Nenad Tomasev
Constrained roads to complex brains -
Neural development and brain circuit evolution converged in birds and mammals www.science.org/doi/10.1126/...
Neural development and brain circuit evolution converged in birds and mammals www.science.org/doi/10.1126/...
Constrained roads to complex brains
Neural development and brain circuit evolution converged in birds and mammals
www.science.org
February 14, 2025 at 6:20 PM
Constrained roads to complex brains -
Neural development and brain circuit evolution converged in birds and mammals www.science.org/doi/10.1126/...
Neural development and brain circuit evolution converged in birds and mammals www.science.org/doi/10.1126/...
Reposted by Nenad Tomasev
LLMs Mastering Board Games: ZurichNLP Meetup - Feb 20th!
Excited to share insights from my student research at Google DeepMind at the upcoming ZurichNLP meetup! I'll present how we achieved high-level play in board games using LLMs with a search budget comparable to human chess grandmasters.
Excited to share insights from my student research at Google DeepMind at the upcoming ZurichNLP meetup! I'll present how we achieved high-level play in board games using LLMs with a search budget comparable to human chess grandmasters.
February 14, 2025 at 5:07 PM
LLMs Mastering Board Games: ZurichNLP Meetup - Feb 20th!
Excited to share insights from my student research at Google DeepMind at the upcoming ZurichNLP meetup! I'll present how we achieved high-level play in board games using LLMs with a search budget comparable to human chess grandmasters.
Excited to share insights from my student research at Google DeepMind at the upcoming ZurichNLP meetup! I'll present how we achieved high-level play in board games using LLMs with a search budget comparable to human chess grandmasters.
Reposted by Nenad Tomasev
Model-free deep RL algorithms like NFSP, PSRO, ESCHER, & R-NaD are tailor-made for games with hidden information (e.g. poker).
We performed the largest-ever comparison of these algorithms.
We find that they do not outperform generic policy gradient methods, such as PPO.
arxiv.org/abs/2502.08938
1/N
We performed the largest-ever comparison of these algorithms.
We find that they do not outperform generic policy gradient methods, such as PPO.
arxiv.org/abs/2502.08938
1/N
February 14, 2025 at 6:41 PM
Model-free deep RL algorithms like NFSP, PSRO, ESCHER, & R-NaD are tailor-made for games with hidden information (e.g. poker).
We performed the largest-ever comparison of these algorithms.
We find that they do not outperform generic policy gradient methods, such as PPO.
arxiv.org/abs/2502.08938
1/N
We performed the largest-ever comparison of these algorithms.
We find that they do not outperform generic policy gradient methods, such as PPO.
arxiv.org/abs/2502.08938
1/N
Reposted by Nenad Tomasev
Can LLMs be used to discover interpretable models of human and animal behavior?🤔
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
February 10, 2025 at 12:21 PM
Can LLMs be used to discover interpretable models of human and animal behavior?🤔
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
Turns out: yes!
Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior.
1/12
Reposted by Nenad Tomasev
🙏🙏🙏
See also full set of video lectures at m.youtube.com/playlist?lis...
See also full set of video lectures at m.youtube.com/playlist?lis...
February 9, 2025 at 4:09 PM
🙏🙏🙏
See also full set of video lectures at m.youtube.com/playlist?lis...
See also full set of video lectures at m.youtube.com/playlist?lis...
Reposted by Nenad Tomasev
No company in the self-driving industry has invested in self-play at scale. I've been dying for someone to finally do this—many others have felt the same. This landmark work finally shows the potential in this approach.
This is a challenge to the industry.
This is a challenge to the industry.
We've built a simulated driving agent that we trained on 1.6 billion km of driving with no human data.
It is SOTA on every planning benchmark we tried.
In self-play, it goes 20 years between collisions.
It is SOTA on every planning benchmark we tried.
In self-play, it goes 20 years between collisions.
February 8, 2025 at 2:06 AM
No company in the self-driving industry has invested in self-play at scale. I've been dying for someone to finally do this—many others have felt the same. This landmark work finally shows the potential in this approach.
This is a challenge to the industry.
This is a challenge to the industry.
One of the things that I am most excited about is how we can use AI (and language models in particular) to help us generate new scientific hypotheses and accelerate scientific discovery.
February 7, 2025 at 6:54 AM
One of the things that I am most excited about is how we can use AI (and language models in particular) to help us generate new scientific hypotheses and accelerate scientific discovery.
Reposted by Nenad Tomasev
We launched a bunch of Gemini 2.0 models today. Compared to the 1.5 series models, each of the 2.0 models is generally better than the "one size up" model in the 1.5 series.
2.0 Flash & Flash-Lite set new standards in the quality/cost Pareto frontier.
More details:
blog.google/technology/g...
2.0 Flash & Flash-Lite set new standards in the quality/cost Pareto frontier.
More details:
blog.google/technology/g...
February 5, 2025 at 11:45 PM
We launched a bunch of Gemini 2.0 models today. Compared to the 1.5 series models, each of the 2.0 models is generally better than the "one size up" model in the 1.5 series.
2.0 Flash & Flash-Lite set new standards in the quality/cost Pareto frontier.
More details:
blog.google/technology/g...
2.0 Flash & Flash-Lite set new standards in the quality/cost Pareto frontier.
More details:
blog.google/technology/g...
Reposted by Nenad Tomasev
In December, I posted about our new paper on mastering board games using internal + external planning. 👇
Here's a talk now on Youtube about it given by my awesome colleague John Schultz!
www.youtube.com/watch?v=JyxE...
Here's a talk now on Youtube about it given by my awesome colleague John Schultz!
www.youtube.com/watch?v=JyxE...
January 17, 2025 at 5:26 PM
In December, I posted about our new paper on mastering board games using internal + external planning. 👇
Here's a talk now on Youtube about it given by my awesome colleague John Schultz!
www.youtube.com/watch?v=JyxE...
Here's a talk now on Youtube about it given by my awesome colleague John Schultz!
www.youtube.com/watch?v=JyxE...
Reposted by Nenad Tomasev
John's talk is now available online!
www.youtube.com/watch?v=JyxE...
www.youtube.com/watch?v=JyxE...
January 17, 2025 at 3:27 PM
John's talk is now available online!
www.youtube.com/watch?v=JyxE...
www.youtube.com/watch?v=JyxE...
Reposted by Nenad Tomasev
Check out the Autonomous Agents for Social Good (AASG) Workshop taking place at #AAMAS 2025!
Are you doing research on using autonomous agents or multi-agent systems to address social challenges to make the world a better place?
Submission deadline: Feb. 4th.
panosd.eu/aasg2025/
Are you doing research on using autonomous agents or multi-agent systems to address social challenges to make the world a better place?
Submission deadline: Feb. 4th.
panosd.eu/aasg2025/
January 16, 2025 at 2:15 PM
Check out the Autonomous Agents for Social Good (AASG) Workshop taking place at #AAMAS 2025!
Are you doing research on using autonomous agents or multi-agent systems to address social challenges to make the world a better place?
Submission deadline: Feb. 4th.
panosd.eu/aasg2025/
Are you doing research on using autonomous agents or multi-agent systems to address social challenges to make the world a better place?
Submission deadline: Feb. 4th.
panosd.eu/aasg2025/
For those interested in AI for improving mental health, registrations are still open for the next MEXA hackathon.
www.linkedin.com/posts/mexaco...
www.linkedin.com/posts/mexaco...
MEXA on LinkedIn: #mexa #hackathon #mentalhealth #ai #innovation
Have you registered for MEXA's next Virtual Hackathon, 𝙈𝙚𝙣𝙩𝙖𝙡 𝙃𝙚𝙖𝙡𝙩𝙝 𝙖𝙣𝙙 𝘼𝙄: 𝙄𝙣𝙩𝙚𝙧𝙫𝙚𝙣𝙩𝙞𝙤𝙣? Now is the time!⏰
Join us in…
www.linkedin.com
January 16, 2025 at 8:14 AM
For those interested in AI for improving mental health, registrations are still open for the next MEXA hackathon.
www.linkedin.com/posts/mexaco...
www.linkedin.com/posts/mexaco...
Reposted by Nenad Tomasev
Join John's talk to get insights on our paper on mastering board games with language models!
January 13, 2025 at 1:35 PM
Join John's talk to get insights on our paper on mastering board games with language models!
Reposted by Nenad Tomasev
A new paper dropped from DeepMind: Deliberation in Latent Space via Differentiable Cache Augmentation
The trouble is, it's not very readable. I tried making a thread here, but it got far too long, so it's a blog now:
timkellogg.me/blog/2024/12...
The trouble is, it's not very readable. I tried making a thread here, but it got far too long, so it's a blog now:
timkellogg.me/blog/2024/12...
Explainer: Latent Space Experts - Tim Kellogg
timkellogg.me
December 24, 2024 at 2:26 PM
A new paper dropped from DeepMind: Deliberation in Latent Space via Differentiable Cache Augmentation
The trouble is, it's not very readable. I tried making a thread here, but it got far too long, so it's a blog now:
timkellogg.me/blog/2024/12...
The trouble is, it's not very readable. I tried making a thread here, but it got far too long, so it's a blog now:
timkellogg.me/blog/2024/12...