Links for 2024-12-14
AI:
1. Microsoft Phi-4 is a 14B parameter LM trained heavily on synthetic data, with very strong performance, even exceeding GPT-4o on GPQA and MATH benchmarks.
https://techcommunity.microsoft.com/blog/aiplatformblog/introducing-phi-4-microsoft%E2%80%99s-newest-small-language-model-specializing-in-comple/43570902. 10,000x Faster: AI Discovers New Microscopy Techniques in Record Time
https://www.nature.com/articles/s41467-024-54696-y3. AutoReason: Automatic Few-Shot Reasoning Decomposition
https://arxiv.org/abs/2412.069754. Bootstrapping Language-Guided Navigation Learning with Self-Refining Data Flywheel
https://www.arxiv.org/abs/2412.084675. Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
https://arxiv.org/abs/2412.026746. “The convergence of neuroscience, artificial intelligence and computing has created an unprecedented opportunity to understand intelligence itself.”
https://www.thetransmitter.org/neuroai/solving-intelligence-requires-new-research-and-funding-models/7. Neural networks are often trained on data generated by agents pursuing long-term plans (whether human data, or e.g. distilling MCTS as in AlphaZero). So it's natural that networks might learn long-term planning. This would have big implications for generalization and safety.
https://axrp.net/episode/2024/12/12/episode-38_3-erik-jenner-learned-look-ahead.html8. “Ever notice how some people pace when they’re deep in thought? Surprisingly, neural networks do something similar—and it boosts their performance! We made this discovery while exploring the planning behavior of a recurrent neural network (RNN) trained to play the complex puzzle game Sokoban.”
https://far.ai/post/2024-07-learned-planners/9. DeepNose: An Equivariant Convolutional Neural Network Predictive Of Human Olfactory Percepts
https://arxiv.org/abs/2412.0874710. Frontier LLMs have shrunk dramatically: GPT-4 had ~1.8T params, while GPT-4o likely has ~200B and Claude 3.5 Sonnet ~400B. The next generation of models, corresponding to GPT-5 and Claude 4 (or Claude 3.5 Opus) will probably return to or slightly exceed the size of the original GPT-4.
https://epoch.ai/gradient-updates/frontier-language-models-have-become-much-smaller11. Ilya Sutskever at NeurIPS 2024 speaks about the forthcoming arrival of superintelligence
https://youtu.be/1yvBqasHLZs?si=ya0LXSZLnX0b6hX412. Microsoft CEO Satya Nadella says OpenAI has a 2-year lead in the AI race and this gives them an "escape velocity" advantage
https://youtu.be/9NtsnzRFJ_o?si=7qQLfEsonCvCW_cF&t=249713. ARK Invest's Chief Futurist Brett Winton explains why AI foundation models will be worth $15-20 trillion by 2030 and OpenAI could grab the lion's share of that market
https://youtu.be/SImm15uF_3Q?si=R0c2WeilV2Bp2ZlB&t=8314. Europe jumps into ‘incredibly costly’ AI supercomputing race
https://www.politico.eu/article/europe-costly-artificial-intelligence-race-supercomputer-startups/15. Elon Musk wanted an OpenAI for-profit — “You can’t sue your way to AGI.”
https://openai.com/index/elon-musk-wanted-an-openai-for-profit/Miscellaneous:
1. Craig Mundie says the nuclear fusion company backed by Sam Altman will surprise the world by showing fusion electrical generation next year, becoming the basis for a "radical transformation of the energy system" due to safe, cheap power
https://www.youtube.com/live/Z246nuPpeOQ?si=ZWGqa2FDnn_aFkb5&t=38462. Ferroelectric Devices Could Make IoT Data Unhackable — FeFET array enables homomorphic encryption in battery-powered devices
https://spectrum.ieee.org/unhackable-phone3. "Natural selection... has been acting on us for the past 3,000 years, right up to the modern day, new research suggests. And it seems to be acting in surprising ways on complex traits encoded by multiple genes, such as those tied to intelligence..." [published in 2021]
https://www.livescience.com/natural-selection-human-genes4. How this cancer drug could make radiation a slam dunk therapy
https://www.sciencedaily.com/releases/2024/12/241210115059.htm