13/12/2025
Links for 2025-12-12
AI
1. DeepMind co-founder and chief AGI researcher Shane Legg has publicly held the same prediction since 2009: there’s a 50% chance we’ll see AGI by 2028. Here he discusses why he hasn’t changed his mind and how we need to prepare before we get there. https://www.youtube.com/watch?v=l3u_FAv33G0
2. What happens when you train a transformer on one of math’s most infamous unsolved puzzles—and then study how it fails? https://axiommath.ai/territory/learning-collatz-the-mother-of-all-rabbit-holes
3. New math prover release from Nous, combining open specialized models and agentic pipelines. Strong Putnam scores on only 3B active parameters. https://huggingface.co/NousResearch/nomos-1
4. Self-Improving VLM Judges Without Human Annotations https://arxiv.org/abs/2512.05145
5. Devstral 2: State-of-the-art, open-source agentic coding models and CLI agent. https://mistral.ai/news/devstral-2-vibe-cli
6. Google dropped a Gemini agent into an unseen 3D world. The model acted as the task proposer, the agent, and the reward model - autonomously learning from self-generated experience. It surpassed human performance through self-improvement iterations. https://arxiv.org/abs/2512.04797
7. The first quantitative scaling principles for agent systems, testing 180 configurations across three LLM families (OpenAI, Google, Anthropic) and four agentic benchmarks spanning financial reasoning, web navigation, game planning, and workflow ex*****on. https://arxiv.org/abs/2512.08296
8. AI makes persuasion so cheap that elites might just manufacture whatever public opinion they want https://arxiv.org/abs/2512.04047
9. “AI can obviously create new knowledge” https://andymasley.substack.com/p/ai-can-obviously-create-new-knowledge
10. ReJump: A Tree-Jump Representation for Analyzing and Improving LLM Reasoning https://www.arxiv.org/abs/2512.00831
11. Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning https://arxiv.org/abs/2512.07461
12. ProAgent: Harnessing On-Demand Sensory Contexts for Proactive LLM Agent Systems https://arxiv.org/abs/2512.06721
13. These New AI Models Are Trained on Physics, Not Words, and They’re Driving Discovery https://www.simonsfoundation.org/2025/12/09/these-new-ai-models-are-trained-on-physics-not-words-and-theyre-driving-discovery/
14. Insights into Claude Opus 4.5 from Pokémon https://www.lesswrong.com/posts/u6Lacc7wx4yYkBQ3r/insights-into-claude-opus-4-5-from-pokemon
15. Prediction: AI will make formal verification go mainstream https://martin.kleppmann.com/2025/12/08/ai-formal-verification.html
16. Selling H200s to China Is Unwise and Unpopular https://www.lesswrong.com/posts/kmEpWTjWeFyqv4tb5/selling-h200s-to-china-is-unwise-and-unpopular
17. The Architects of AI Are TIME’s 2025 Person of the Year https://time.com/7339685/person-of-the-year-2025-ai-architects/
18. U.S. Investors Are Going Big on China AI Despite Concerns in Congress https://www.wsj.com/tech/ai/u-s-investors-are-going-big-on-china-ai-despite-concerns-in-congress-cb9a71c5 [no paywall: https://archive.is/FhLSM]
19. Bezos and Musk Race to Bring Data Centers to Space https://www.wsj.com/tech/bezos-and-musk-race-to-bring-data-centers-to-space-faa486ee [no paywall: https://archive.is/lHwA8]
20. ‘Greetings, earthlings’: Nvidia-backed Starcloud trains first AI model in space as orbital data center race heats up https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html
21. The Pentagon just launched GenAI.mil using Google Gemini for military operations https://oodaloop.com/briefs/technology/pentagon-launches-military-ai-platform-powered-by-google-gemini-for-defense-operations/
22. Karpathy uses GPT to grade decade-old Hacker News predictions with hindsight https://karpathy.bearblog.dev/auto-grade-hn/
23. SAP consultants rated AI work at 95% accuracy until they knew it was AI https://venturebeat.com/ai/the-ai-that-scored-95-until-consultants-learned-it-was-ai
24. Demonstrably Safe AI For Autonomous Driving https://waymo.com/blog/2025/12/demonstrably-safe-ai-for-autonomous-driving
25. Google DeepMind robotics lab tour with Hannah Fry https://www.youtube.com/watch?v=UALxgn1MnZo
26. Gemini Deep Research agent for developers https://blog.google/technology/developers/deep-research-agent-gemini-api/
27. You should be able to provide an LLM as a job reference, just like you would a coworker, manager, or professor. It can form an opinion and represent you without revealing any private data. — John Carmack https://x.com/ID_AA_Carmack/status/1998753499002048589
Science and Technology
1. How Scientists Are Growing Computers From Human Brain Cells—and Why They Want to Keep Doing It https://singularityhub.com/2025/12/11/how-scientists-are-growing-computers-from-human-brain-cells-and-why-they-want-to-keep-doing-it/
2. Even with steady technological progress, there can be sudden and overwhelming phase transitions https://andyljones.com/posts/horses.html
3. A massive study involving nearly 80,000 children reveals that fine motor skills, particularly handwriting, are powerful predictors of academic achievement in reading, writing, and mathematics across all age groups. These findings challenge the modern educational trend of replacing manual activities with digital alternatives, suggesting that skilled hand use is intrinsically linked to cognitive processes and essential for optimal learning. https://www.sciencedirect.com/science/article/pii/S1747938X25000855
4. Researchers have discovered the earliest known instance of human-created fire, which took place in the east of England 400,000 years ago. https://www.bbc.com/news/resources/idt-b9da7a6d-165b-492a-8785-235cd10e2e8e
5. Why Is Ice Slippery? A New Hypothesis Slides Into the Chat. https://www.quantamagazine.org/why-is-ice-slippery-a-new-hypothesis-slides-into-the-chat-20251208/
6. “Lilly’s triple agonist, retatrutide, delivered weight loss of up to an average of 71.2 lbs along with substantial relief from osteoarthritis pain in first successful Phase 3 trial” https://investor.lilly.com/news-releases/news-release-details/lillys-triple-agonist-retatrutide-delivered-weight-loss-average
The arrival of AGI | Shane Legg (co-founder of DeepMind) Shane Legg, Chief AGI Scientist and co-founder of Google DeepMind, has been talking about artificial general intelligence, or AGI, for more than a decade. So...