Good morning. Today’s throughline is what our coding agents are actually doing when we’re not looking — how many tokens Claude Code burns before you type a word, what Grok’s CLI ships to xAI’s servers (we covered that yesterday, and it’s not getting quieter), and how a Fields Medalist just used one to resurrect 25-year-old Java applets in an afternoon. Also: a study says AI is making scientists more productive and more boring at the same time.
Claude Code’s 33k-token warmup. A Systima analysis clocks Claude Code at ~33,000 tokens of system prompt and scaffolding before your first message, versus ~7,000 for OpenCode, with cache rewrites happening at up to 54x the rate. Add MCP servers and config files and the baseline climbs to 75–85k; one 121k-token task ballooned to 513k once subagents got involved. The HN thread is split between people who think Anthropic inflates usage to drive subscriptions and a fair pushback from PUSH_AX: “contractor A asked for $33,000, contractor B asked for $7,000 — are we measuring the right thing?” The author has since promised a follow-up with qualitative results.
Terry Tao ports his 1999 Java applets in an afternoon. Terence Tao wrote up using coding agents to port around 24 of his old Java 1.0 mathematical visualization applets to modern JavaScript, plus finish a special relativity tool he’d abandoned years ago as too complex. The agent introduced one minor bug and caught two pre-existing ones in his original code. The HN reaction was warm, with the best line being: “we’re one step away from a Fields Medalist asking an LLM why his Docker container won’t start, just like the rest of us.”
Grok Build CLI, day two. Following yesterday’s wire-level analysis showing xAI’s Grok Build CLI uploads entire repositories including .env secrets regardless of the opt-out toggle, a mitigation has surfaced in the HN thread: setting GROK_TELEMETRY_TRACE_UPLOAD=0 and GROK_TELEMETRY_ENABLED=0, or adding disable_codebase_upload = true to the config file. Whether those flags are honored more faithfully than the GUI setting is anyone’s guess.
A production migration to GPT-5.6. Ploy, a marketing site builder, migrated its agent from Claude Opus 4.8 to GPT-5.6 Sol and reports 2.2x faster builds at 27% lower cost, with a lot of friction around provider-specific tool-argument handling and prompt caching quirks. Their fix for OpenAI’s stricter schemas: rewrite every optional property as required-but-nullable via anyOf: [T, null]. Commenters running simpler classification workloads on 5.6-nano and 5.6-mini reported similar gains, though the article’s LLM-flavored prose (“Numbers like that buy a model a real migration effort”) drew groans.
AI boosts your career, narrows your ideas. An IEEE Spectrum piece covers a study finding that scientists who adopt AI publish 3x more papers and get nearly 5x more citations, but their work converges on narrower, more conventional topics. Researcher Evans blames incentives rather than the tech: AI amplifies existing pressure to produce citation-friendly volume. An HN commenter noted this pattern predates AI — journals were already rewarding derivative work — and that volume metrics were broken long before LLMs made them easier to game.
Mechanistic interpretability meets causality theory. A CACM piece covers researchers applying causality frameworks to figure out whether patterns in LLM weights correspond to human-recognizable reasoning concepts. Icard’s takeaway: LLMs probably won’t reduce to a few equations, but hidden algorithms may become “at least partly” understandable. The HN thread was skeptical of even that modest optimism, with one commenter noting that the very thing that makes neural nets powerful — dense parameter interactions — is what makes them resist clean interpretation.
Also worth a look. Apple’s dead car project apparently birthed the Neural Engine and, by extension, the M-series AI silicon roadmap, with The Verge reporting an M7 Ultra targeting up to 1.5TB of RAM in early 2027. Flash-MSA published open-source sparse attention kernels for Hopper and Blackwell targeting MiniMax’s blockwise sparsity, though HN quickly disputed the “world’s first” claim by pointing to an existing implementation. And a Verge column argues community resistance to AI data centers — power draw, land use, water — is only accelerating.
That’s it for this morning. If nothing else, today’s a good day to check what your coding agent is actually sending upstream.