General
- The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis. ChatGPT, when used appropriately, can improve learning outcomes.
- Analysis of AI usage: Large Language Models, Small Labor Market Effects. They find evidence of significant uptake/usage, but relatively minor effects on economic indicators.
Research Insights
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example. Training over and over on a single example can yield gains.
- Interesting to compare/contrast to RL on unlimited synthetic data: Absolute Zero: Reinforced Self-play Reasoning with Zero Data.
- Sakana AI introduces Continuous Thought Machines (interactive report, preprint, code), a new neural approach where the neurons are synchronized and allowed to move their attention around over time. This allows a crude approximation of temporal “thinking” as the neurons modify their focus and state.
LLM
- OpenAI add o4-mini to their reinforcement fine-tuning API.
- ByteDance releases SeedCoder 8B.
- OpenAI adds GPT-4.1 to the ChatGPT web product.
- OpenAI release HealthBench. In addition to providing a useful way to track progress on LLMs for healthcare applications, the current results demonstrate just how effective existing LLMs can be in this application space.
Agents
- Google describes: AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms (paper). An evolutionary approach is used to optimize algorithms for tasks including math, chip design, datacenter scheduling, and fine-tuning LLMs.
- Among many other things, it was able to improve on various optimal packing math problems.
Safety
- OpenAI launch safety evaluations hub.
Audio
Video
- Google announces that the latest improvements to Gemini 2.5 (Pro and Flash) greatly improve video understanding.
- Tencent HunyuanCustom is an open-source video model (preprint, code).
- New lipsync model: KeySync: A Robust Approach for Leakage-free Lip Synchronization in High Resolution (preprint, code, demo).
- LTXV 13B Distilled claims rendering 5× faster than base model.
World Synthesis
- Enigma Labs claims they have made the first multiplayer AI-generative video game (a multiplayer car racing game). They say they will open-source the work eventually. Although the gameplay video shows crude graphics, it is further evidence that generative environments are a key part of future entertainment.
Science
- AI-Engineered DNA Turns Genes On and Off in Blood Cells.
- Publication in Cell: Design principles of cell-state-specific enhancers in hematopoiesis.
- Inference-time compute methods being applied to AI for bioscience: Test-Time Scaling Unlocks a Leap Forward in De Novo Antibody Design.
- Behind the Noise (github). Self-supervised denoising can learn meaningful representations for science images.
- Meta report: Sharing new breakthroughs and artifacts supporting molecular property prediction, language processing, and neuroscience.
- Open Molecules 2025 (OMol25) and Meta’s Universal Model for Atoms (UMA): Revolutionizing design at the atomic scale
- Adjoint Sampling: A breakthrough in highly scalable, reward-driven generative modeling
- Unlocking how the human brain develops language
Hardware
- Google announces: Gemini smarts are coming to more Android devices. E.g. Gemini will be available from one’s smart watch. A Samsung headset with Gemini integration will supposedly launch later this year.
Robots
- Tesla shows a video of Optimus robot dancing. Fluid motion like this tests the limit of hardware and software (latency, real-time compensation, etc.).