Links for 2025-01-22
AI:
Anthropic CEO: "I am more confident than I have ever been at any previous time...What I’ve seen inside Anthropic over the last few months led me to believe that AI will surpass almost all humans at almost all tasks in 2-3 years." https://www.youtube.com/watch?v=snkOMOjiVOk
Announcing the Stargate Project, a $500 billion AI infrastructure investment in the US, Sam Altman and President Trump say this is about building AGI, with one of the most important applications being the cure of diseases at an unprecedented rate https://youtu.be/IYUoANr3cMo?si=-5dNbyFZZinTQzr9&t=544
“The first data centers are actually under construction in Texas already, and we'll be turning them over to Sam” https://youtu.be/m9HJu26Oh_k?si=tnybgvJmiyBCPuCd&t=260
OpenAI's Sarah Friar: we internally feel that we have a path to AGI and the rapid iteration in the o-series of reasoning models shows that the pace of innovation is accelerating https://www.youtube.com/watch?v=oyFniEgH0jI
Another pieces of evidence for increasingly sophisticated out-of-context reasoning capabilities of LLMs: “We train LLMs on a particular behavior, e.g. always choosing risky options in economic decisions. They can *describe* their new behavior, despite no explicit mentions in the training data. So LLMs have a form of intuitive self-awareness.” https://www.lesswrong.com/posts/xrv2fNJtqabN3h6Aj/tell-me-about-yourself-llms-are-aware-of-their-implicit
“…we report a demonstration of a form of Out-of-Context Reasoning where training on documents which discuss (but don’t demonstrate) Claude’s tendency to reward hack can lead to an increase or decrease in reward hacking behavior.” https://www.lesswrong.com/posts/qXYLvjGL9QvD3aFSW/training-on-documents-about-reward-hacking-induces-reward
Google presents Learn-by-interact: A Data-Centric Framework for Self-Adaptive Agents in Realistic Environments https://arxiv.org/abs/2501.10893
Agent-R: Training Language Model Agents to Reflect via Iterative Self-Training https://arxiv.org/abs/2501.11425
Alibaba presents Mobile-Agent-E: Self-Evolving Mobile Assistant for Complex Tasks https://x-plug.github.io/MobileAgent/
Bytedance's LLM for browser operations achieves high-accuracy processing even with a 2B-sized model. With this level of efficiency, the potential use cases for agents will likely expand significantly. https://arxiv.org/abs/2501.12326
Google released an updated reasoning model. The latest version also includes code execution, and a 1M token content window. You can try it right now for free here: https://aistudio.google.com/prompts/new_chat?model=gemini-2.0-flash-thinking-exp-01-21
Physics of Skill Learning https://arxiv.org/abs/2501.12391
Optical Interposers Could Start Speeding Up AI in 2025 https://spectrum.ieee.org/optical-interposers
When you have short timelines, it's pretty dishonest to say they don't have the money. They'll raise the money pretty soon when investors see what's coming out of these labs and what they have internally.
You can disagree that AI will be big enough to raise $500 billion over the next 4 years. But Musk has short timelines.
Those who liken artificial intelligence to Web 3.0 or Bitcoin overlook its profound distinction. Intelligence is not merely a tool—it is the universe’s most potent force, the catalyst behind every discovery and innovation. Artificial General Intelligence (AGI) represents humanity’s final necessary invention, for it transcends being a singular breakthrough; it is the engine that will autonomously unlock all future advancements.
To replicate general intelligence is to democratize labor itself, transforming it into a limitless resource akin to sunlight: infinitely available, universally accessible, and unrestrained by scarcity. AGI is not just a technological leap but a self-propagating seed, capable of cultivating an entire ecosystem of progress. It is the automation of innovation—a system that will exponentially amplify its own capabilities through recursive self-improvement, redesigning its architecture faster than any human effort ever could.
If aligned with human values, AGI could elevate civilization to unprecedented heights. Imagine a world where every individual, irrespective of innate talent, enjoys prosperity surpassing today’s wealthiest elites. Even humanity’s grandest ambitions—like reshaping galaxies via self-replicating von Neumann probes—would transition from science fiction to feasible engineering. AGI is not merely transformative; it is the key to unlocking a post-scarcity cosmos, where the boundaries of possibility dissolve.
Miscellaneous:
People randomly assigned to farm rice are more collectivistic than those randomly assigned to farm wheat https://www.stevestewartwilliams.com/p/sowing-the-seeds-of-culture
How to turn Nuclear Waste into Nuclear Fuel https://www.youtube.com/watch?v=o2prJNZLuoM
