ResearchJennifer Neville on What AI Failures Reveal and How Evaluation Shapes ProgressJennifer Neville, partner research manager at Microsoft Research and Purdue professor, discusses how careful evaluation uncovers surprising failures in current AI systems and guides improvements.4 min read
ResearchExperiment: Claude Opus 5.5 and Scrimshaw Jukebox Generate Text-Defined Game MusicA user tested whether Claude Opus 5.5 can compose computer game music by pairing it with Scrimshaw Jukebox and asking for a simple text-based music format plus an artifact that plays it, including example tracks.2 min read
ResearchAI-generated mathematical results and Lean formalizations published on GitHubAn organization has published a collection of new mathematical results produced by an internal frontier model, accompanied by formal proof files and transparency materials on GitHub.2 min read
ResearchOpenAI and Ironclad train GPT‑6 Astra on real contracting tasks, report higher accuracy and speedOpenAI partnered with Ironclad to turn real contracting workflows into research tasks for training and evaluating GPT‑6 Astra.4 min read
ResearchBenchmark shows AI can perform many tasks in partly automated scientific labsC5R Corp's SciUniverse benchmark evaluates how well frontier AI models can operate largely automated laboratory equipment across 92 tasks in 17 task families, covering chemistry, biology and materials science.2 min read
ResearchGPT‑6 Astra Deciphers 217‑Year‑Old Encrypted Letter Relating Napoleon's OrdersGPT‑6 Astra, OpenAI's latest model, reportedly decrypted a 217‑year‑old encrypted letter within about six hours.3 min read
ResearchLocal DGX Spark run shows Qwen3.8-27B can add large numbers in words with high one-shot accuracyA recent re-run of an experiment originally posted by Colin Frasier tested whether large-language-model Qwen3.8-27B could compute numeric sums and return the results in words.2 min read
ResearchAnthropic's Claude helps discover CRISPR-like enzyme system in bacteriophagesResearchers have identified a previously unknown enzyme system in bacteriophages — an array-associated reverse transcriptase (ART) — with the discovery aided by Anthropic's Claude AI.2 min read
ResearchCriticism of Bridgemind AI benchmarks' lack of methodologyA user publicly criticized Bridgemind AI's published benchmark post, saying it does not detail how the tests were carried out (e.g., harness, tasks, number of runs, identification of day-to-day variance, APIs, traces).1 min read
ResearchFrom GPU kernels to recursive language models: Alex Zhang on harnesses, agent swarms and untapped capabilitiesAlex Zhang (MIT) discusses his path from GPU kernels and KernelBench to Recursive Language Models (RLMs), Prime Agent and large multi‑agent experiments.6 min read
ResearchWhy AI’s Greatest Contribution May Be Doing the Boring WorkHemanth Asirvatham and Elliott Mokski argue that as intelligence — human or artificial — advances, its greatest bottleneck may be execution: the institutional, logistical and physical work needed to realize ideas.5 min read
ResearchGoogle launches orbital TPU test satellite as part of long-term Project SuncatcherGoogle launched a prototype orbital compute satellite carrying a Tensor Processing Unit (TPU) built by Planet Labs to validate power, cooling and radiation performance in space.4 min read
