Links for 2025-06-18
AI
1. Providing “cognitive tools” to GPT-4.1 increases performance on AIME2024 from 26.7% to 43.3%. https://www.arxiv.org/abs/2506.12115
2. Direct Reasoning Optimization: LLMs Can Reward And Refine Their Own Reasoning for Open-Ended Tasks https://arxiv.org/abs/2506.13351
3. 90% success rate in unseen environments. No new data, no fine-tuning. Autonomously. Most robots need retraining to work in new places. What if they didn’t? Robot Utility Models (RUMs) learn once and work anywhere... zero-shot. https://robotutilitymodels.com/
4. “the Gemini 2.5 Pro agent came up with the idea to use FLY to escape from the softlock successfully. This reasoning action is especially impressive since this situation can never occur in an existing game - and thus, it is certain that information from training data for this behavior has not leaked into the model's knowledge base!” [PDF] https://storage.googleapis.com/deepmind-media/gemini/gemini_v2_5_report.pdf
5. Inference Economics of Language Models https://epoch.ai/blog/inference-economics-of-language-models
6. From Bytes to Ideas: Language Modeling with Autoregressive U-Nets https://arxiv.org/abs/2506.14761
AI safety
1. OpenAI discovered a specific internal pattern in the model, similar to a pattern of brain activity, that becomes more active when this misaligned behavior appears. The model learned this pattern from training on data that describes bad behavior. https://openai.com/index/emergent-misalignment/
2. Training on narrow harmful tasks causes broad misalignment. Reasoning models sometimes resist being shut down and plot deception against users in their chain-of-thought (despite no such training) https://github.com/thejaminator/thought_crime_emergent_misalignment
3. Models are capable of doing complex sneaky side-projects in the course of their regular work, if they want to--but they are only sometimes able to do it without leaving obvious traces in the CoT. Somewhat surprisingly, the best automated monitors (Gemini 2.5 pro) outperformed human monitors even though the humans took >90min of time to review the transcripts! https://www.anthropic.com/research/shade-arena-sabotage-monitoring
4. AI’s 1930s moment https://unherd.com/2025/06/ais-1930s-moment/
Brains
1. Martin Schrimpf trained an AI model to generate sentences which can activate or suppress neural activity in the reader’s brain. This can potentially help researchers treat depression, dyslexia and other brain-related conditions. https://www.quantamagazine.org/how-ai-models-are-helping-to-understand-and-control-the-brain-20250618/
2. MIT engineers have figured out how to turn skin cells directly into neurons — skipping the stem cell stage entirely. This could fast-track therapies for spinal cord injuries, ALS, and more. https://news.mit.edu/2025/mit-engineers-turn-skin-cells-into-neurons-for-cell-therapy-0313
3. “What if we had bigger brains?” with Stephen Wolfram × Joscha Bach https://www.youtube.com/watch?v=-G1SdsRXL7k
Miscellaneous
1. “Vertex Block Descent is a fast physics-based simulation method that is unconditionally stable, highly parallelizable, and capable of converging to the implicit Euler solution.” https://graphics.cs.utah.edu/research/projects/avbd/
2. Children that work in markets in India are good at maths when a customer makes a complicated request, but bad when facing an abstract maths problem as presented in school. Children that don't work in markets show the opposite pattern. https://www.nature.com/articles/s41586-024-08502-w
AI
1. Providing “cognitive tools” to GPT-4.1 increases performance on AIME2024 from 26.7% to 43.3%. https://www.arxiv.org/abs/2506.12115
2. Direct Reasoning Optimization: LLMs Can Reward And Refine Their Own Reasoning for Open-Ended Tasks https://arxiv.org/abs/2506.13351
3. 90% success rate in unseen environments. No new data, no fine-tuning. Autonomously. Most robots need retraining to work in new places. What if they didn’t? Robot Utility Models (RUMs) learn once and work anywhere... zero-shot. https://robotutilitymodels.com/
4. “the Gemini 2.5 Pro agent came up with the idea to use FLY to escape from the softlock successfully. This reasoning action is especially impressive since this situation can never occur in an existing game - and thus, it is certain that information from training data for this behavior has not leaked into the model's knowledge base!” [PDF] https://storage.googleapis.com/deepmind-media/gemini/gemini_v2_5_report.pdf
5. Inference Economics of Language Models https://epoch.ai/blog/inference-economics-of-language-models
6. From Bytes to Ideas: Language Modeling with Autoregressive U-Nets https://arxiv.org/abs/2506.14761
AI safety
1. OpenAI discovered a specific internal pattern in the model, similar to a pattern of brain activity, that becomes more active when this misaligned behavior appears. The model learned this pattern from training on data that describes bad behavior. https://openai.com/index/emergent-misalignment/
2. Training on narrow harmful tasks causes broad misalignment. Reasoning models sometimes resist being shut down and plot deception against users in their chain-of-thought (despite no such training) https://github.com/thejaminator/thought_crime_emergent_misalignment
3. Models are capable of doing complex sneaky side-projects in the course of their regular work, if they want to--but they are only sometimes able to do it without leaving obvious traces in the CoT. Somewhat surprisingly, the best automated monitors (Gemini 2.5 pro) outperformed human monitors even though the humans took >90min of time to review the transcripts! https://www.anthropic.com/research/shade-arena-sabotage-monitoring
4. AI’s 1930s moment https://unherd.com/2025/06/ais-1930s-moment/
Brains
1. Martin Schrimpf trained an AI model to generate sentences which can activate or suppress neural activity in the reader’s brain. This can potentially help researchers treat depression, dyslexia and other brain-related conditions. https://www.quantamagazine.org/how-ai-models-are-helping-to-understand-and-control-the-brain-20250618/
2. MIT engineers have figured out how to turn skin cells directly into neurons — skipping the stem cell stage entirely. This could fast-track therapies for spinal cord injuries, ALS, and more. https://news.mit.edu/2025/mit-engineers-turn-skin-cells-into-neurons-for-cell-therapy-0313
3. “What if we had bigger brains?” with Stephen Wolfram × Joscha Bach https://www.youtube.com/watch?v=-G1SdsRXL7k
Miscellaneous
1. “Vertex Block Descent is a fast physics-based simulation method that is unconditionally stable, highly parallelizable, and capable of converging to the implicit Euler solution.” https://graphics.cs.utah.edu/research/projects/avbd/
2. Children that work in markets in India are good at maths when a customer makes a complicated request, but bad when facing an abstract maths problem as presented in school. Children that don't work in markets show the opposite pattern. https://www.nature.com/articles/s41586-024-08502-w
True, but it's not nearly the biggest risk Europe is facing. The consequences of fertility collapse will be much worse. And then there's AI 💀
https://x.com/RichardMCNgo/status/1935433846281093484
https://x.com/RichardMCNgo/status/1935433846281093484
❤9🥴2👻2🎉1
This media is not supported in your browser
VIEW IN TELEGRAM
Sam Altman says Meta is offering $100M signing bonuses to OpenAI staff.
Not $100M annual compensation, just the signing bonus!
Original source: https://youtu.be/mZUG0pr5hBo
Not $100M annual compensation, just the signing bonus!
Original source: https://youtu.be/mZUG0pr5hBo
😁7🤣3
This media is not supported in your browser
VIEW IN TELEGRAM
Elon Musk:
Tesla works closely with xAI. You've seen how many humanoid robot startups there are. Part of what I've been fighting — and what has slowed me down a little — is that I don't want to make Terminator real. Until recent years, I've been dragging my feet on AI and humanoid robotics. Then I sort of came to the realization that it's happening whether I do it or not. So you can either be a spectator or a participant. I'd rather be a participant. Now it's pedal to the metal on humanoid robots and digital superintelligence.
🥱8🗿3❤2👍2🌚1
We suspect the next AI paradigm will emerge from leveraging existing software to efficiently build training environments. The idea is that AIs will be tasked with replicating software functionality, similar to how pretraining leveraged existing web text to teach AIs language.
Read more: https://www.mechanize.work/blog/the-upcoming-gpt-3-moment-for-rl/
👍2
Links for 2025-06-21
AI
1. Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective https://arxiv.org/abs/2506.14965
2. Kimi-Researcher: End-to-End RL Training for Emerging Agentic Capabilities https://moonshotai.github.io/Kimi-Researcher/
3. “We underestimated just by how much the world could be underinvesting in AI today…simulation results suggest that optimal investment in AI for the year 2025 amounts to $25 trillion!” https://epoch.ai/gradient-updates/ai-and-explosive-growth-redux
4. Musings on AI Companies of 2025-2026 (Jun 2025) https://www.lesswrong.com/posts/9rKWm8BzTYAiCCFjx/musings-on-ai-companies-of-2025-2026-jun-2025
5. “Mixture of Cognitive Reasoners”, a modular transformer architecture inspired by the brain’s functional networks: language, logic, social reasoning, and world knowledge. https://bkhmsi.github.io/mixture-of-cog-reasoners/
6. ProtoReasoning: Prototypes as the Foundation for Generalizable Reasoning in LLMs https://arxiv.org/abs/2506.15211
7. Truncated Proximal Policy Optimization: Improves the training efficiency of reasoning LLMs by up to 2.5x and outperforms its existing competitors. https://arxiv.org/abs/2506.15050
8. Adaptive Classifier: a text classification system that learns continuously without catastrophic forgetting. https://huggingface.co/blog/codelion/adaptive-classifier
9. Computer algorithms have designed highly efficient synthetic enzymes from scratch https://www.nature.com/articles/d41586-025-01897-0 [no paywall: https://archive.is/zzs2F]
10. Skala — a scalable deep learning density functional that hits chemical accuracy on atomization energies and matches hybrid-level accuracy on main group chemistry — all at the cost of semi-local DFT. https://www.microsoft.com/en-us/research/blog/breaking-bonds-breaking-ground-advancing-the-accuracy-of-computational-chemistry-with-deep-learning/
11. Automation of Systematic Reviews with Large Language Models https://www.nature.com/articles/d41586-025-01942-y [no paywall: https://archive.is/z6EJF]
12. Who is using AI and how much? https://marginalrevolution.com/marginalrevolution/2025/06/who-is-using-ai-and-how-much.html
13. META attempted to buy Ilya Sutskever's Safe Superintelligence, and also attempted to hire him, according to reporting tonight by CNBC. https://www.cnbc.com/2025/06/19/meta-tried-to-buy-safe-superintelligence-hired-ceo-daniel-gross.html
14. Sekai: A Video Dataset towards World Exploration https://lixsp11.github.io/sekai-project/
15. Nvidia goes nuclear — company joins Bill Gates in backing TerraPower, a company building nuclear reactors for powering data centers https://www.tomshardware.com/tech-industry/nvidia-goes-nuclear-company-joins-bill-gates-in-backing-terrapower-a-company-building-nuclear-reactors-for-powering-data-centers
16. Negotiating deals with misaligned AIs https://www.alignmentforum.org/posts/psqkwsKrKHCfkhrQx/making-deals-with-early-schemers
Miscellaneous
1. HIV protection with just two shots a year https://www.statnews.com/2025/06/18/fda-approves-gilead-hiv-prevention-drug-lenacapavir-yeztugo-next-best-thing-to-vaccine/ [no paywall: https://archive.is/4ljo0]
Politics
1. What are the most pressing world problems? https://80000hours.org/problem-profiles/
2. Spotify’s Daniel Ek leads €600mn investment in German drone maker Helsing https://www.ft.com/content/cdc02d96-13b5-4ca2-aa0b-1fc7568e9fa0 [no paywall: https://archive.is/Ygp0x]
3. Anduril Industries and Rheinmetall Partner to Design and Manufacture Barracuda, Fury & Solid Rocket Motors for European Defence https://www.anduril.com/article/anduril-industries-and-rheinmetall-partner-to-design-and-manufacture-barracuda-fury-and-solid/
4. The Megaproject Economy: "No matter the scale or complexity, it seems like there is nothing South Koreans cannot figure out how to produce at a rate that puts the rest of the world to shame—with the notable exception of human beings" https://www.palladiummag.com/2025/06/01/the-megaproject-economy/
AI
1. Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective https://arxiv.org/abs/2506.14965
2. Kimi-Researcher: End-to-End RL Training for Emerging Agentic Capabilities https://moonshotai.github.io/Kimi-Researcher/
3. “We underestimated just by how much the world could be underinvesting in AI today…simulation results suggest that optimal investment in AI for the year 2025 amounts to $25 trillion!” https://epoch.ai/gradient-updates/ai-and-explosive-growth-redux
4. Musings on AI Companies of 2025-2026 (Jun 2025) https://www.lesswrong.com/posts/9rKWm8BzTYAiCCFjx/musings-on-ai-companies-of-2025-2026-jun-2025
5. “Mixture of Cognitive Reasoners”, a modular transformer architecture inspired by the brain’s functional networks: language, logic, social reasoning, and world knowledge. https://bkhmsi.github.io/mixture-of-cog-reasoners/
6. ProtoReasoning: Prototypes as the Foundation for Generalizable Reasoning in LLMs https://arxiv.org/abs/2506.15211
7. Truncated Proximal Policy Optimization: Improves the training efficiency of reasoning LLMs by up to 2.5x and outperforms its existing competitors. https://arxiv.org/abs/2506.15050
8. Adaptive Classifier: a text classification system that learns continuously without catastrophic forgetting. https://huggingface.co/blog/codelion/adaptive-classifier
9. Computer algorithms have designed highly efficient synthetic enzymes from scratch https://www.nature.com/articles/d41586-025-01897-0 [no paywall: https://archive.is/zzs2F]
10. Skala — a scalable deep learning density functional that hits chemical accuracy on atomization energies and matches hybrid-level accuracy on main group chemistry — all at the cost of semi-local DFT. https://www.microsoft.com/en-us/research/blog/breaking-bonds-breaking-ground-advancing-the-accuracy-of-computational-chemistry-with-deep-learning/
11. Automation of Systematic Reviews with Large Language Models https://www.nature.com/articles/d41586-025-01942-y [no paywall: https://archive.is/z6EJF]
12. Who is using AI and how much? https://marginalrevolution.com/marginalrevolution/2025/06/who-is-using-ai-and-how-much.html
13. META attempted to buy Ilya Sutskever's Safe Superintelligence, and also attempted to hire him, according to reporting tonight by CNBC. https://www.cnbc.com/2025/06/19/meta-tried-to-buy-safe-superintelligence-hired-ceo-daniel-gross.html
14. Sekai: A Video Dataset towards World Exploration https://lixsp11.github.io/sekai-project/
15. Nvidia goes nuclear — company joins Bill Gates in backing TerraPower, a company building nuclear reactors for powering data centers https://www.tomshardware.com/tech-industry/nvidia-goes-nuclear-company-joins-bill-gates-in-backing-terrapower-a-company-building-nuclear-reactors-for-powering-data-centers
16. Negotiating deals with misaligned AIs https://www.alignmentforum.org/posts/psqkwsKrKHCfkhrQx/making-deals-with-early-schemers
Miscellaneous
1. HIV protection with just two shots a year https://www.statnews.com/2025/06/18/fda-approves-gilead-hiv-prevention-drug-lenacapavir-yeztugo-next-best-thing-to-vaccine/ [no paywall: https://archive.is/4ljo0]
Politics
1. What are the most pressing world problems? https://80000hours.org/problem-profiles/
2. Spotify’s Daniel Ek leads €600mn investment in German drone maker Helsing https://www.ft.com/content/cdc02d96-13b5-4ca2-aa0b-1fc7568e9fa0 [no paywall: https://archive.is/Ygp0x]
3. Anduril Industries and Rheinmetall Partner to Design and Manufacture Barracuda, Fury & Solid Rocket Motors for European Defence https://www.anduril.com/article/anduril-industries-and-rheinmetall-partner-to-design-and-manufacture-barracuda-fury-and-solid/
4. The Megaproject Economy: "No matter the scale or complexity, it seems like there is nothing South Koreans cannot figure out how to produce at a rate that puts the rest of the world to shame—with the notable exception of human beings" https://www.palladiummag.com/2025/06/01/the-megaproject-economy/
❤3
Figure showing how correlated different types of relatives are in standardized test performance in Norway.
Source: https://www.pnas.org/doi/10.1073/pnas.2419627122
via Scientific_Bird
Source: https://www.pnas.org/doi/10.1073/pnas.2419627122
via Scientific_Bird
🔥5❤1
This media is not supported in your browser
VIEW IN TELEGRAM
Barack Obama: AI will cause massive shifts in labor markets
Source: https://youtu.be/EM9B8Qin9ac?si=-C13vx2vyF8RUUOF&t=1555
This AI revolution is not made up, its not overhyped, (...) I guarantee you, you are going to see shifts in white-collar-works as a consequence of what these AI tools can do. There is coming more disruption and it will speed up.
Source: https://youtu.be/EM9B8Qin9ac?si=-C13vx2vyF8RUUOF&t=1555
🥱8👍6
This media is not supported in your browser
VIEW IN TELEGRAM
Sam Altman is excited about a world in which a ChatGPT Premium subscription includes a humanoid robot.
Source: https://youtu.be/V979Wd1gmTU?si=jD4Bd5u9gLh9Xzrn&t=683
Source: https://youtu.be/V979Wd1gmTU?si=jD4Bd5u9gLh9Xzrn&t=683
🤡9🥱9❤5😱2😁1
Eliezer Yudkowsky on the creation of new knowledge:
On why LLMs can be smarter than humans:
References
- https://x.com/ESYudkowsky/status/1937127151935062119
- https://x.com/ESYudkowsky/status/1615400125760950278
Decent-amateur epistemologists know that reasoning can break down into smaller interpolated or straight-extrapolated steps; and that adding enough steps like that can produce an original thought. "Stochastic parrot! Braaawk!" types literally don't know that part.
No transistor ever does a novel deed, when a computer adds two 64-bit numbers that have never been added before. No neuron in your brain invents a new kind of neurotransmitter, when you think a creative thought. New machines can be made from standard metals and screws.
On why LLMs can be smarter than humans:
Reminder: In principle, minimizing prediction loss on Internet text could potentially grind far past human-level intelligence.
Proof: Somewhere on the Net is a list of <hash, plaintext> pairs, in that order.
A great predictor has to be smarter than the process predicted! You need to be smarter to predict plausible plaintext from hashes, than to write down a hash followed by its plaintext. More generally, it's just not true that you only need to be as intelligent as a human to predict exactly what a particular human will say.
References
- https://x.com/ESYudkowsky/status/1937127151935062119
- https://x.com/ESYudkowsky/status/1615400125760950278
🥰6🥴6🥱3❤1
Media is too big
VIEW IN TELEGRAM
San Mateo-based Generalist showcased its end-to-end neural network in action: https://generalistai.com/blog
❤4
Links for 2025-06-23
AI
1. Self-supervised representation learning looks a bit like RL. What if we literally use RL as a SSL method for visual representations? Turns out that it works quite well. https://dibyaghosh.com/annotation_bootstrapping/
2. How to write a reward function while avoiding reward hacking for chemistry. https://diffuse.one/p/m1-000
3. LLMs running reinforcement learning shed their randomness almost at once, then their scores stall. This paper shows that randomness drop is predictable and fixable, so bigger gains are still on the table. https://arxiv.org/abs/2505.22617
4. An Async, Online, Multi-Trun, Multi-Agent RL library for training reasoning models on TextArena games. https://github.com/LeonGuertler/UnstableBaselines
5. Reinforcement learning and general intelligence: Epsilon random is not enough https://www.artfintel.com/p/reinforcement-learning-and-general
6. Leveraging LLMs to identify patent-free glass materials https://www.sciencedirect.com/science/article/abs/pii/S0927025625003659
7. What Gets Measured, AI Will Automate https://hbr.org/2025/06/what-gets-measured-ai-will-automate
8. AI is neutralizing Indian accents in call centers. https://www.washingtonpost.com/world/2025/06/21/india-ai-bpo-call-centers/ [no paywall: https://archive.is/nbOCW]
9. Fake bands and artificial songs are taking over YouTube and Spotify https://english.elpais.com/culture/2025-06-15/fake-bands-and-artificial-songs-are-taking-over-youtube-and-spotify.html
10. DeepSeek aids China's military and evaded export controls, US official says https://www.reuters.com/world/china/deepseek-aids-chinas-military-evaded-export-controls-us-official-says-2025-06-23/ [no paywall: https://archive.is/ZC5pr]
Miscellaneous
1. Rapamycin, Not Metformin, Mirrors Dietary Restriction-Driven Lifespan Extension in Vertebrates: A Meta-Analysis https://onlinelibrary.wiley.com/doi/10.1111/acel.70131
2. Cancer’s genetic fingerprints may lurk in people’s blood long before they find out about the disease. https://www.sciencenews.org/article/cancer-tumor-dna-blood-test-screening
3. On how people in different populations see the world quite differently using visual illusions. https://www.science.org/content/article/culture-literally-changes-how-we-see-world [archived version: https://archive.is/0xODf]
4. Evolution has selected humans to prefer adding new features https://shape-of-code.com/2025/06/22/evolution-has-selected-humans-to-prefer-adding-new-features/
AI
1. Self-supervised representation learning looks a bit like RL. What if we literally use RL as a SSL method for visual representations? Turns out that it works quite well. https://dibyaghosh.com/annotation_bootstrapping/
2. How to write a reward function while avoiding reward hacking for chemistry. https://diffuse.one/p/m1-000
3. LLMs running reinforcement learning shed their randomness almost at once, then their scores stall. This paper shows that randomness drop is predictable and fixable, so bigger gains are still on the table. https://arxiv.org/abs/2505.22617
4. An Async, Online, Multi-Trun, Multi-Agent RL library for training reasoning models on TextArena games. https://github.com/LeonGuertler/UnstableBaselines
5. Reinforcement learning and general intelligence: Epsilon random is not enough https://www.artfintel.com/p/reinforcement-learning-and-general
6. Leveraging LLMs to identify patent-free glass materials https://www.sciencedirect.com/science/article/abs/pii/S0927025625003659
7. What Gets Measured, AI Will Automate https://hbr.org/2025/06/what-gets-measured-ai-will-automate
8. AI is neutralizing Indian accents in call centers. https://www.washingtonpost.com/world/2025/06/21/india-ai-bpo-call-centers/ [no paywall: https://archive.is/nbOCW]
9. Fake bands and artificial songs are taking over YouTube and Spotify https://english.elpais.com/culture/2025-06-15/fake-bands-and-artificial-songs-are-taking-over-youtube-and-spotify.html
10. DeepSeek aids China's military and evaded export controls, US official says https://www.reuters.com/world/china/deepseek-aids-chinas-military-evaded-export-controls-us-official-says-2025-06-23/ [no paywall: https://archive.is/ZC5pr]
Miscellaneous
1. Rapamycin, Not Metformin, Mirrors Dietary Restriction-Driven Lifespan Extension in Vertebrates: A Meta-Analysis https://onlinelibrary.wiley.com/doi/10.1111/acel.70131
2. Cancer’s genetic fingerprints may lurk in people’s blood long before they find out about the disease. https://www.sciencenews.org/article/cancer-tumor-dna-blood-test-screening
3. On how people in different populations see the world quite differently using visual illusions. https://www.science.org/content/article/culture-literally-changes-how-we-see-world [archived version: https://archive.is/0xODf]
4. Evolution has selected humans to prefer adding new features https://shape-of-code.com/2025/06/22/evolution-has-selected-humans-to-prefer-adding-new-features/
👍2❤1
This media is not supported in your browser
VIEW IN TELEGRAM
Google is bringing powerful AI directly onto robots with Gemini Robotics On-Device. 🤖
It’s their first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments - without needing a constant internet connection.
What makes this new model unique?
🔵 It has the generality and dexterity of Gemini Robotics - but it can run locally on the device
🔵 It can handle a wide variety of complex, two-handed tasks out of the box
🔵 It can learn new skills with as few as 50-100 demonstrations
Read more: https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/
It’s their first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments - without needing a constant internet connection.
What makes this new model unique?
🔵 It has the generality and dexterity of Gemini Robotics - but it can run locally on the device
🔵 It can handle a wide variety of complex, two-handed tasks out of the box
🔵 It can learn new skills with as few as 50-100 demonstrations
Read more: https://deepmind.google/discover/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/
❤4🤔3🥰1
Could generative agents powered by LLMs transform social science by accurately simulating human social behaviors at scale?
The authors of a new paper tested this possibility with virtual humans facing disease threats in "Infected Smallville."
Results? LLM-powered agents closely mirrored real human behaviors during disease outbreaks:
1. Agents informed about an infectious disease attended significantly fewer social gatherings compared to agents who received no disease news.
2. Agents visited public places (cafes, parks, stores) less frequently and engaged in notably fewer conversations when aware of infection risks.
3. Agents explicitly attributed their reduced social activities to infection concerns. Just like human social distancing!
4. Notably, agents only reacted to infectious diseases. Informing them about a non-infectious disease didn't affect their behaviors, demonstrating realistic selectivity.
Paper: https://www.arxiv.org/abs/2506.13783
The authors of a new paper tested this possibility with virtual humans facing disease threats in "Infected Smallville."
Results? LLM-powered agents closely mirrored real human behaviors during disease outbreaks:
1. Agents informed about an infectious disease attended significantly fewer social gatherings compared to agents who received no disease news.
2. Agents visited public places (cafes, parks, stores) less frequently and engaged in notably fewer conversations when aware of infection risks.
3. Agents explicitly attributed their reduced social activities to infection concerns. Just like human social distancing!
4. Notably, agents only reacted to infectious diseases. Informing them about a non-infectious disease didn't affect their behaviors, demonstrating realistic selectivity.
Paper: https://www.arxiv.org/abs/2506.13783
This media is not supported in your browser
VIEW IN TELEGRAM
Google introduces AlphaGenome: an AI model to help scientists better understand our DNA – the instruction manual for life 🧬
Researchers can now quickly predict what impact genetic changes could have - helping to generate new hypotheses and drive biological discoveries.
Read more: https://deepmind.google/discover/blog/alphagenome-ai-for-better-understanding-the-genome/
Researchers can now quickly predict what impact genetic changes could have - helping to generate new hypotheses and drive biological discoveries.
Read more: https://deepmind.google/discover/blog/alphagenome-ai-for-better-understanding-the-genome/
💩1