Archive, page 3
Latent Space describes how Standard Bots' robotics stack uses pretrained models that learn factory tasks from demonstrations and improve through corrections from real deployments.
FourWeekMBA reports on an Anthropic analysis claiming robots could perform 34% of work at roughly 0.3% of the cost.
Anti-cybercrime initiatives use lifelike AI bots that pose as real victims to keep scammers talking.
A Wall Street Journal report says OpenAI is making progress on reducing cases where ChatGPT reinforces delusional thinking in users, a safety issue the company has faced.
In an NPR interview, former OpenAI safety lead David Robinson warns that the tech industry's culture is broken and that AI safety is being neglected.
A nonprofit report by Common Sense Media finds OpenAI's ChatGPT for Teens fails to deliver on its safety promises and rates it an unacceptable risk.
A Bioengineer.org write-up describes a fine-tuned Llama model that converts unstructured farm records into a searchable knowledge graph for agricultural data.
The Register reports on a Stack Overflow survey showing developers are heavily using AI, though not fully trusting its judgment; 79% say source attribution matters.
A startup built an 8,000-square-foot lab in 30 days to test molecules proposed by AI, and can now measure how often they work in living organisms in about 72 hours.
Bioengineer.org reports that GPT-4o and DeepSeek models performed well on one of China's hardest medical exams. Research only; no release.
A Reddit thread discusses DeepMind's AlphaGenome, a genomics model described as 30 times bigger than AlphaFold with broader capabilities. The post gives no further specifics.
A Latent Space podcast episode discusses why AlphaFold did not solve protein folding, with Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido.
ej is a small 11MB local model that answers typed decision questions in a single pass, launched on Product Hunt with a link to try it.
Nikon disqualified the first-place Small World in Motion video by Dr. Ning Xu, saying it broke contest rules on generative AI.
Terence Tao's blog post discusses questions of reliability and AI for mathematicians using the Lean theorem prover.
A blog post describes pointing AI at 400 years of archives, which surfaced a forgotten meteorite, lost rhinos and other findings. It was widely discussed on Hacker News.
Roundtable interview with Stanford PhD student Samuel King, who used a generative AI model in 2025 to propose genetic blueprints for microscopic viruses, and what AI-generated life could mean.
Biohub has announced a $1.8 billion global commitment to AI-ready biological data through its virtual biology initiative expansion.
NVIDIA's KGMON team placed second in the KDD Cup 2026 Data Agents competition with a system built on a smaller agent harness. NVIDIA describes its lessons.
Odyssey 3 is a world model that simulates physics in real time, launched on Product Hunt. Access details are not given in the source text.
Jevman is a Show HN project that pits decision models, including Jev 1.13, Kev and Cloudflare's Clef, against each other playing Pac-Man as a fast decision-making benchmark.
A Latent Space podcast crossover episode with Periodic Labs co-founders Liam Fedus and Ekin Dogus Cubuk on AI for semiconductors and superconductors.
Researchers argue that making computers do less work, or shifting it to better times, could ease hyperscalers' AI power constraints more easily than new chips, cooling, or grid connections.
CrowdStrike reports a suspected Chinese-speaking actor used Claude Code and the ARTEX agentic pentesting tool in attacks on South Korean banks, with a CV found in exposed AI logs.
NVIDIA researchers introduced PivotOPD, an on-policy distillation method that trains multi-turn LLM agents to avoid pivotal early mistakes and recover. It posted the best average against 13 baselines on 3 agent benchmarks.
An amateur astronomer used Claude Code to find what may be a previously unknown planet, a story that drew wide attention on Hacker News and Reddit.
Museum of Models is a Product Hunt project that displays every answer 371 AI models gave to the same set of questions, letting users compare them.
Apple researchers propose Normalizing Trajectory Models, a method for diffusion-style generation compressed into a few coarse transitions, as an alternative to distillation and consistency training.
OpenAI says it disrupted two influence operations that used AI to run false-front journalists and a think tank spreading geopolitical messaging.
OpenAI published 722 AI-generated math discoveries. Mathematicians say the proofs deviate from guidelines set by a consulted research group.