A chronological archive of technical articles, notes, and case analyses.
Less than a year ago, GPT-5 looked like the frontier. Today GPT-5.6 Sol crushes it on the benchmarks that still allow comparison, models are solving decades-old mathematical problems, and last year’s level of capability keeps getting cheaper. I am trying to understand where this pace is taking us.
Read articleWhy a neural network can make mistakes in simple calculations and why it needs external tools
Read articleWithin months, SpaceX combined with xAI, monetized Colossus, went public, and agreed to buy Cursor. Here is my reading of Musk’s strategy ahead of a possible AI-market correction.
Read articleOpenRouter restricts Russian access to models whose providers prohibit use in certain regions. Other models still work, but dependence on someone else’s policy is becoming harder to ignore.
Read articleWhile American flagships remain closed services, Chinese competition is making powerful open-weight models more accessible to developers.
Read articleTalkie‑1930 was trained only on text published before 1931 and later taught to tackle modern programming tasks. What does the experiment tell us about generalization?
Read articleA Claude Code npm package accidentally included a source map exposing the client’s source code. It was not a model leak, but it is another reason to discuss centralization risk.
Read articleGoogle demonstrated major KV-cache compression, while the market heard “memory is no longer needed.” Here is why that conclusion is convenient — and far too simplistic.
Read articleFear of AI has become a convenient marketing tool: it sells subscriptions, supports market expectations, and helps blur responsibility.
Read article