Summary
A quiet day on curated retweets (zero from @bayesiansapien across both captured slots), with signal concentrated in two clusters and a lot of off-topic noise around it. The strongest single thread is the agentic-coding tension: George Hotz's "Eternal Sloptember" blog argues coding agents fundamentally cannot program (morning), while evening brings two independent data points pushing the other direction — xAI's Grok V9-Medium (1.5T parameters) reportedly finishing training and dropping in two weeks, and Grok Build 0.1 showing up alongside Stepfun, Gemini 3.5 Flash, and Qwen 3.7 Max as a serious option on the Kilo coding gateway. The second thread is Hassabis on both sides of the AGI line in the same week: "nowhere near AGI" (morning) sitting against his "foothills of the singularity" framing covered in today's digest. A one-line @magicsilicon take on Huawei's SoC-to-SiP framing pairs cleanly with today's BitCPM-CANN result on Ascend NPUs. Everything else is French politics, LA mayoral content, Tesla FSD promo, and Nicki Minaj keyword-filter false positives.
Posts
- Geohot: coding agents cannot actually program (@ns123abc · blog) [morning]. Hotz spent six months trying agents on tinygrad work and on reverse-engineering a USB-to-PCIe chip, and concluded he could have done both faster and better by hand. The agents are accurate statistical mimics, which is exactly why their bugs are getting harder to detect. Strongest single dissent the wiki has captured this month against the everything-is-fine framing on agentic coding.
- Grok V9-Medium (1.5T) finished training, drops in 2 weeks (@ns123abc) [evening]. Quoted internal text says a 3x scale jump over the 0.5T V8-small that currently serves all Grok production traffic, with emphasis on difficult coding tasks. If real, the most consequential xAI signal of the week.
- Kilo Gateway usage: three challengers to Stepfun (@kilocode) [evening]. Stepfun still 30%+ of usage, but Grok Build 0.1, Gemini 3.5 Flash, and Qwen 3.7 Max are now in the mix on the same workflows. Real market-share data on what coders actually pick when routed across frontier models. Ties to Qwen 3.7 Max autonomous coding.
- Grok Build "right there with Claude" on real coding (@brivael, repost of @scion_x_) [evening]. Hands-on developer report after a full agentic session: model sharp, agentic flow holds, some UX rough edges. Single anecdote, but paired with the Kilo numbers it is the first time Grok Build has registered as a serious Claude competitor in the wiki.
- Hassabis: today's systems are nowhere near AGI (@ns123abc) [morning]. "Doesn't matter how many Erdos problems you solve... far from what a true invention or someone like a Ramanujan would have been able to do." Same Hassabis who said "foothills of the singularity" the same week (see today's digest Industry Pulse). Both can hold, but the juxtaposition is the story.
- @ns123abc flips to the LeCun/Marcus camp on LLMs (@ns123abc) [morning]. "Real programming agents will need world models, not some RLVR shit." Posted as direct reaction to the Geohot piece. The skeptical camp is broadening, not narrowing.
- Huawei: Moore's Law shifts from SoC to SiP (@magicsilicon) [morning]. One-line take on Huawei's positioning that monolithic die scaling is being replaced by chiplet packaging. The substrate story under today's BitCPM-CANN result on Ascend NPU 1.58-bit training.
- Pope Leo XIV: Vatican and Anthropic to work together on "a path for humanity in the AI era" (@brivael, repost of French source) [evening]. AI must be "disarmed" like nuclear energy. Video only, no primary English source. Fits the pattern of Anthropic doing institutional alignment work outside standard tech-policy channels.
- Ronak Malde leaves DeepMind, hints at Cursor/Codeium news (@ellev3n11) [morning]. "Some cool stuff soon" after the $2.4B Codeium acquisition. Watch for product news in 30-60 days.
- Ilya Sutskever posted something (@ns123abc) [morning]. Tweet body is just "Ilya Sutskever just posted this" with no quoted content captured. Click through to read.
- SF founder-residency directory (cluster of 3) (@Scobleizer, hf0 demo day, piano AI musing) [morning]. 22-link list of SF AI hacker-houses and residencies. Useful as a directory, no research content. Skip.
- Robotics in-home data collection (@Scobleizer) [morning]. Argument that paid in-home camera setups will scale because users will accept data collection in exchange for chores. Industry positioning, no research substance. Skip.
- Tesla FSD with Grok and Hey Grok geological guide (cluster of 2) (@Tesla, Hey Grok) [morning]. Marketing for Tesla FSD Supervised plus Grok-as-voice-tour-guide. Skip.
- @brivael French politics and lifestyle (cluster of 30 across both slots) (morning, evening) [morning + evening]. Social class theory, Macron, Palantir's Karp on his PhD, French ENA, XChat homescreen, Attal video, Elon-as-French-president fiction, business school as ponzi, Yalc.ai promo, coffee machine, Musk wealth math, UAE growth. Not AI-research-relevant. Skip.
- @spencerpratt LA mayoral campaign (cluster of 6 across both slots) (morning, evening) [morning + evening]. Karen Bass clips, NBC News attention-vs-votes, LA Times crime-rate framing, Drew Carey. Caught by keyword filter. Skip.
- Nicki Minaj off-topic (cluster of 2) (@NICKIMINAJ) [morning]. Keyword-filter false positives. Skip.
- @AustinJustice Travis County storage burglar (@AustinJustice) [evening]. Local Austin crime story. Skip.