For years now, I've been looking for signs of nontrivial zero-shot transfer across seen embodiments. When I saw the Alohas unhang tools from a wall used only on our Frankas I knew we had it! Gemini Robotics 1.5 is the first VLA to achieve such transfer!!

— Konstantinos Bousmalis, Research Scientist at DeepMind, working on robotic learning

Links for 2025-10-09 [Part 1]

AI

1. MIT researchers model human behavior as scripts rather than complex reasoning MIT researchers developed ROTE, which treats everyday human actions like predictable code snippets—”wait for green light, then go”—instead of assuming complex decision-making, beating existing methods by 50%. https://arxiv.org/abs/2510.01272

2. Scientific Algorithm Discovery by Augmenting AlphaEvolve with Deep Research https://arxiv.org/abs/2510.06056

3. MoEs Are Stronger than You Think: Hyper-Parallel Inference Scaling with RoE https://arxiv.org/abs/2509.17238

4. Reverse-engineering a model that successfully learns multiplication via implicit chain-of-thought https://www.arxiv.org/abs/2510.00184

5. Selective Underfitting in Diffusion Models https://selective-underfitting.github.io/

6. Gemini 2.5 Computer Use model https://blog.google/technology/google-deepmind/gemini-computer-use-model/

7. AI breakthrough helps astronomers spot cosmic events with just a handful of examples https://www.ox.ac.uk/news/2025-10-08-ai-breakthrough-helps-astronomers-spot-cosmic-events-just-handful-examples

8. AI will become the operating system https://www.youtube.com/watch?v=LCEmiRjPEtQ

9. JPMorgan CEO Jamie Dimon said the money the bank has poured into AI has already begun paying for itself. The company invests about $2 billion a year in AI, but has reduced headcount and improved efficiency to save at least the same amount, Dimon told Bloomberg. https://www.bloomberg.com/news/articles/2025-10-07/jpmorgan-s-dimon-says-ai-cost-savings-now-matching-money-spent [no paywall: https://archive.is/J5CVb]

10. Lufthansa to shed 4,000 jobs with help from AI https://www.dw.com/en/lufthansa-to-shed-4000-jobs-with-help-from-ai/a-74173278

11. AI can now do math. But can it ask good questions? - Ken Ono https://epochai.substack.com/p/ai-can-now-do-math-but-can-it-ask

12. Why frontier AI can’t solve this professor’s math problem - Greta Panova https://epochai.substack.com/p/why-frontier-ai-cant-solve-this-professors

13. The future of AI is already written https://www.mechanize.work/blog/technological-determinism/

14. Greg Brockman on Move 37: “I do think next year we will have some incredible models. The milestone I am most excited about is having models that can solve hard problems. And the the analogy I like to make is think back to AlphaGo, you know, Move 37, that changes people’s understanding of the game. Imagine that in coding, imagine that in material science, imagine that in medicine. Having real breakthroughs that are either potentially the AI by itself, or the AI assisted with top humans. I think we’ll start seeing that.” https://www.youtube.com/watch?app=desktop&v=5yA4o9fSJek

15. Defying Transformers: Searching for “Fixed Points” of Pretrained LLMs https://liujch1998.github.io/2025/10/07/fixed-point.html

16. “For years now, I’ve been looking for signs of nontrivial zero-shot transfer across seen embodiments. When I saw the Alohas unhang tools from a wall used only on our Frankas I knew we had it! Gemini Robotics 1.5 is the first VLA to achieve such transfer!!” — Konstantinos Bousmalis, Research Scientist at DeepMind, working on robotic learning https://x.com/bousmalis/status/1973359357447512508

AI alignment

1. Spooky Collusion at a Distance with Superrational AI https://www.lesswrong.com/posts/JEtAWvp2sAe8nqpfy/spooky-collusion-at-a-distance-with-superrational-ai

2. Can Large Language Models Develop Gambling Addiction? https://arxiv.org/abs/2509.22818

3. Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment https://www.lesswrong.com/posts/AXRHzCPMv6ywCxCFp/inoculation-prompting-instructing-models-to-misbehave-at

4. AI-Enabled Influence Operation Against Iran https://www.schneier.com/blog/archives/2025/10/ai-enabled-influence-operation-against-iran.html

Links for 2025-10-09 [Part 2]

Understanding the SEI formation in lithium-ion batteries using machine learning

1. Evidence for significant multi-Li+ clustering in common lithium-ion battery electrolytes https://chemrxiv.org/engage/chemrxiv/article-details/68c08e353e708a7649010b9a

2. Scalable and accurate simulation of electrolyte solutions with quantum chemical accuracy https://iopscience.iop.org/article/10.1088/2632-2153/adaf76

3. Exploring Lithium Diffusion in LiF with Machine Learning Potentials: From Point Defects to Collective Ring Diffusion https://chemrxiv.org/engage/chemrxiv/article-details/68cd701e9008f1a467a2e37b

Neuroscience

1. A Mysterious Web of Tunnels Connects Brain Cells—Like the Network of Trees in a Forest https://singularityhub.com/2025/10/07/a-mysterious-we-of-tunnels-connects-brain-cells-like-the-network-of-trees-in-a-forest/

2. Self-supervised predictive learning accounts for cortical layer-specificity https://www.nature.com/articles/s41467-025-61399-5

3. Large-scale cortical functional networks are organized in structured cycles https://www.nature.com/articles/s41593-025-02052-8

4. A glimpse at what NeuroAI brain models might enable: a topographic vision model predicts stimulation patterns that steer complex object recognition behavior. This could be a key ‘software’ component for visual prosthetic hardware. https://arxiv.org/abs/2510.03684

Science

1. Genetically engineered human mesenchymal progenitor (stem) cells, termed senescence-resistant cells (SRCs), can not only halt aging but also roll back its clinical and functional negative effects. https://www.cell.com/cell/abstract/S0092-8674(25)00571-9

2. MIT researchers discovered a hidden atomic order that persists in metals even after extreme processing. https://news.mit.edu/2025/uncovering-new-physics-metals-manufacturing-1008

Technology

1. Europe’s first production plant for rare-earth magnets opens https://www.ft.com/content/6b16c87e-bdfd-45ab-ab01-b69d97541126 [no paywall: https://archive.is/GEyTZ]

2. The AI Boom Has a Copper Problem. Are Microbes the Solution? https://spectrum.ieee.org/ai-copper

3. SHIELD Activated: Researchers build a defense to protect drones from cyberattacks https://news.fiu.edu/2025/shield-activated-researchers-build-a-defense-to-protect-drones-from-cyberattacks

Miscellaneous

1 .You Should Get a Reusable Mask https://www.lesswrong.com/posts/wXwjMbtiSqALMEw2g/you-should-get-a-reusable-mask

2. Originality is bullshit because everything good is just a remix Joan Westenberg argues that chasing originality is creativity’s biggest scam—Shakespeare stole plots, Jobs remixed Xerox, and every viral TikTok is just iterations on iterations. https://www.joanwestenberg.com/p/originality-is-a-scam

AI capabilities have been steadily improving across a wide range of skills, and show no sign of slowing down in the near term: https://epoch.ai/data-insights/ai-capabilities-over-past-year

Across benchmarks covering coding, math, scientific knowledge, common sense and visual reasoning, and more, state-of-the-art models have improved by 20 to 50 percentage points in the last year.

Revenues at frontier AI companies have more than tripled in the past year.

For those who still don't get it:

- Researchers working at the frontier labs are genuinely convinced that transformative AI is coming.

- They spend their time worrying about the social/political consequences of transformative AI, not the "AI bubble".

- Recursive self-improvement is on the table, and is the subject of serious discussions.

Read more: https://www.transformernews.ai/p/were-all-behind-the-curve-ai-bubble-crash-risk

via prinz

German drone manufacturer Quantum Systems presented its Jäger Interceptor-UAV to the public. The drone has a maximum operational altitude of 5km, a speed of up to 405km/h, a range of 25km and a 0,8kg warhead.

The drone also features a solid fuel rocket booster that is meant to get the drone to its operating altitude as fast as possible.

https://quantum-systems.com/

Thread by Terence Tao about using the Pro version of GPT-5: https://mathstodon.xyz/@tao/115351400633010670

Even if AI progress halts roughly where it is now, it's almost certain that AI will become an indispensable tool for professionals.

GPT-5 Pro sets a new record on FrontierMath Tier 4.

FrontierMath Tier 4 consists of research-level math problems developed by professional mathematicians. These problems can take experts weeks to solve. The problems the models are tested on are not public.

One problem solved in both of the GPT-5 Pro runs has not been solved by any other model. The problem author had this to say about it:

GPT-5 Pro is on the right track, figuring out the main trick required. It didn't verify that the conditions for applying a particular theorem were satisfied, but that isn't the hardest part.

— Yizhe Zhu, Assistant Professor of Mathematics University of Southern California

Trump got a COVID-19 booster vaccination.

Vaccination and alcohol consumption are two things where both Trump and Putin are role models. They are both boostered and mostly avoid alcohol.

A reminder that the world’s most sophisticated technology is entirely owned by a relatively small German company, which is entirely owned by a 19th-century foundation that owns itself and is prohibited by statute from ever being sold.

https://www.youtube.com/watch?v=eHBVVHgOEsM

UC Berkley used OpenEvolve to discover algorithms across diverse domains that “outperform SOTA human designs”, achieving up to 5.0× runtime improvements or 50% cost reductions.

Conclusion:

One of the most profound implication is the potential for a virtuous cycle. We can use an ADRS to improve itself. As recent work has shown, models can learn to refine their own reasoning, debug their code, and discover more effective strategies.

Paper: https://arxiv.org/abs/2510.06189

Links for 2025-10-11

AI

1. Large Language Models Achieve Gold Medal Performance at the International Olympiad on Astronomy & Astrophysics (IOAA) https://arxiv.org/abs/2510.05016

2. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models — “The model writes, reflects, and rewrites its own prompt over and over until it becomes a self-improving system” https://www.arxiv.org/abs/2510.04618

3. Nvidia researchers boost LLMs reasoning skills by getting them to ‘think’ during pre-training https://venturebeat.com/ai/nvidia-researchers-boost-llms-reasoning-skills-by-getting-them-to-think

4. h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning https://arxiv.org/abs/2510.07312

5. Code World Models for General Game Playing https://arxiv.org/abs/2510.04542

6. Towards a Typology of Strange LLM Chains-of-Thought https://www.lesswrong.com/posts/qgvSMwRrdqoDMJJnD/towards-a-typology-of-strange-llm-chains-of-thought

7. Researchers working at the frontier labs are genuinely convinced that transformative AI is coming. https://www.transformernews.ai/p/were-all-behind-the-curve-ai-bubble-crash-risk

8. Talking about the Future of AI in Law with David Wakeling https://artificialinvestment.substack.com/p/interview-with-david-wakeling-on

9. My Favorite Streamer is an LLM: Discovering, Bonding, and Co-Creating in AI VTuber Fandom https://arxiv.org/abs/2509.10427

10. Moloch’s Bargain: Emergent Misalignment When LLMs Compete for Audiences https://arxiv.org/abs/2510.06105

11. How a 23-year-old former OpenAI researcher turned a viral AI prophecy into profit, with a $1.5 billion hedge fund and outsize influence from Silicon Valley to D.C. https://fortune.com/2025/10/08/leopold-aschenbrenner-openai-ftx-1-5-billion-hedge-fund-situational-awareness/

12. A small number of samples can poison LLMs of any size https://www.anthropic.com/research/small-samples-poison

13. How well can large language models predict the future? https://forecastingresearch.substack.com/p/ai-llm-forecasting-model-forecastbench-benchmark

Science and Engineering

1. UMass Engineers Create First Artificial Neurons That Could Directly Communicate With Living Cells https://www.umass.edu/news/article/umass-engineers-create-first-artificial-neurons-could-directly-communicate-living

2. A cGAS-mediated mechanism in naked mole-rats potentiates DNA repair and delays aging https://www.science.org/doi/10.1126/science.adp5056

3. Scientists Just Reversed Alzheimer’s in Mice. Could They Do It in Humans? https://gizmodo.com/scientists-just-reversed-alzheimers-in-mice-could-they-do-it-in-humans-2000669049

4. Programmable Aptamer-Embedded Circular RNAs for Targeted Antigen-Presenting Cells Immunotherapy https://www.biorxiv.org/content/10.1101/2025.09.28.679023v1

5. Why China Built 162 Square Miles of Solar Panels on the World’s Highest Plateau https://www.nytimes.com/2025/10/10/business/china-solar-tibetan-plateau.html [no paywall: https://archive.is/vuOYu]

6. The government’s own data rebuts Trump’s claims about wind and solar prices https://www.politico.com/news/2025/10/07/green-electricity-costs-cheap-trump-00594123

Total trade war?

1. “China has asserted sweeping control over the entire global semiconductor supply chain, putting export license requirements on all rare earths used to manufacture advanced chips. If enforced aggressively, this policy could mean “lights out” for the US AI boom, and likely lead to a recession/economic crisis in the US in the short term.” https://x.com/deanwball/status/1976260051351343195

2. October 9 2025 Ministry of Commerce/GAC Export Control actions https://sinocism.notion.site/October-9-2025-Ministry-of-Commerce-GAC-Export-Control-actions-28784ece41d78059ba33cb4cca154bf4

This is Charlie Gee

One of the world’s most respected and renowned stonemason

His work in performing restoration works on Cathedrals throughout Europe is some of the best

A skill set that is becoming rarer to find and master in a more modern world

He’s only 23 years old

A rock lath master at work in the 1940s USA

Chart from the Federal Reserve Bank of Dallas: https://www.dallasfed.org/research/economics/2025/0624