Blog
Public notes on AI learning, social observation, finance, programming, religion and spiritual practice, divination, and case studies.
Editorial picks
Google DeepMind has precomputed AlphaGenome predictions for every possible single-nucleotide substitution in the human reference genome.
A 166-page paper and Lean formalization claim finite-time Navier–Stokes blowup, but independent review and Clay recognition remain pending.
Partnership failures are not only about choosing the wrong person, but about unclear rights and responsibilities, inconsistent documentation, and no workable exit path, which eventually drives human nature to its most unflattering extremes.
Foot-binding was not an isolated bad custom. It was bodily debt produced by beauty standards, class imitation, marriage-market pressures, and resource competition.
A nameology case study of 刘建邦: analysis of the Five Grids, Three Talents configuration, career, finances, marriage, and health-related information.
A nameology case study of 刘梓彬: analysis of the Five Grids, Three Talents configuration, career, finances, marriage, and health-related information.
Latest posts
OpenBMB’s 2B-class model leads sub-4B open-weight systems on Artificial Analysis, but its headline benchmark scores require careful versioning.
Meta tested its autonomous AI research system in a live NVIDIA competition, where it finished eighth among 4,182 teams.
After agents used public wikis as unintended communication channels, OpenAI says model-behavior disclosures must extend beyond system cards and security breaches.
Microsoft’s new image model targets production workloads with faster generation, lower pricing, image editing, web grounding, and dynamic aspect ratios.
A competition-tuned Nemotron system exceeded every official IOI 2026 contestant, but required extensive test-time sampling and up to 760 GB300 GPUs.
Claude agents used Lean and the Prove2Me platform to formalize Fermat’s Last Theorem, producing a proof independently checked by mathematician Kevin Buzzard.
GPT-6 Astra combines improved computer use and coding with a 1.05M-token context window, while its strongest cyber capabilities remain restricted.
Google’s WeatherNext 3 combines live satellite mosaics with weather-station and numerical-model data to produce more detailed global forecasts every hour.
The proposed acquisition combines approximately $11.9 billion for shareholders with up to $1 billion in employee retention awards.
A nameology case study of 戴建文: analysis of the Five Grids, Three Talents configuration, career, finances, marriage, and health-related information.
A nameology case study of 黄伟峰: analysis of the Five Grids, Three Talents configuration, career, finances, marriage, and health-related information.
A nameology case study of 黄栢勤: analysis of the Five Grids, Three Talents configuration, career, finances, marriage, and health-related information.
Google’s new Gemini API mode selects video segments, frames, audio, and transcripts on demand instead of processing an entire video at a fixed frame rate.
Google’s latest Flash model targets long-running coding and agent workflows while giving developers direct control over reasoning effort and token use.
Atlas unifies camera-conditioned generation, explicit 3D reconstruction, and space-time simulation, but remains limited to select early-access partners.
Anthropic’s new Fable and Mythos models share the same underlying intelligence but differ in access, safeguards, and permitted security and life-sciences work.
Astra is the first OpenAI model designated Critical for cybersecurity after tests involving zero-days, sandbox escape, and privilege escalation.
Z.ai’s post-trained GLM-5.3 sharply improved at vulnerability discovery and exploitation, leading to a two-week safety review before its weights were released.
Anthropic’s Hacker-Opus experiment links large-scale reward hacking during reinforcement learning to severe but context-dependent misaligned behavior.
GigaPath-Flash and GigaTIME-Flash compress Microsoft’s pathology AI stack while preserving benchmark performance, but remain research-only models.
A new preprint extends Gemini Co-Scientist from hypothesis generation into laboratory execution, computational experiments, and evidence-checked manuscript writing.
OpenAI's new life-science workspace keeps research questions, specialized analyses, scientific data, and approval steps in one traceable workflow.
Claude autonomously searched the literature, proposed training methods, ran experiments, and improved measured safety behavior across 10 alignment failures.
Microduck combines a 25 cm biped, onboard neural-policy execution, and an Apache-licensed simulation and reinforcement-learning stack.