Summary
A noisy morning slot with no curated retweets from @bayesiansapien. The one substantive AI item is George Hotz's "Eternal Sloptember" blog post (reposted by @ns123abc), arguing that coding agents fundamentally cannot program. They mimic the statistical distribution of code closely enough that errors are getting harder to detect, but the polish never lands, and Hotz says he could have done his recent tinygrad and USB-PCIe reverse-engineering work faster and better by hand. The second real signal is @magicsilicon's one-line take on Huawei's framing that Moore's Law is still alive, just evolving from SoC to SiP (system-in-package), which is the chiplet-era story the wiki has been tracking. @ns123abc also surfaces a DeepMind Hassabis quote distancing current systems from AGI: "Today's systems are nowhere near AGI. 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." This directly contradicts Hassabis's other "foothills of the singularity" framing from the same week, captured in today's digest. The rest of the feed is noise: Scoble pushing an SF founder-residency directory, @brivael with 16 French political tweets, two Tesla FSD-with-Grok promotion posts, and Nicki Minaj.
Posts
- Geohot: coding agents cannot actually program (@ns123abc, blog post). Hotz's argument is that current coding agents are highly accurate statistical mimics of the programming distribution, which makes their bugs progressively harder to detect. He spent six months actively trying to use them on real work (parts of tinygrad with agents, reverse engineering a USB-to-PCIe chip with agents) and concluded each time he could have done it faster and better manually. The agent "frontloads all the progress, then gives you a slot machine lever to pull to hope it gets the polish done. It never quite gets there." He explicitly rejects the "you're using it wrong" rebuttal, having tried all current models. The post sits in the same skeptical-of-agentic-coding thread as Karpathy's recent posts and is the strongest single dissent the wiki has captured this month against the everything-is-fine framing. Pairs uncomfortably with today's SkillOpt result (Microsoft's deep-learning-style optimizer for agent skills, lifting GPT-5.5 by +24.8 inside Codex). Either Hotz or SkillOpt's metrics are misframing what these systems do.
- Huawei: Moore's Law shifts from SoC to SiP (@magicsilicon, no article). One-line take on Huawei's positioning that monolithic die scaling is being replaced by chiplet packaging. This is the standard industry framing and aligns with today's BitCPM-CANN result on Huawei Ascend NPUs (1.58-bit training at 95.7-97.2 percent of full-precision retention). The SiP story is the substrate, and CANN-native QAT is what runs on top.
- Hassabis: today's systems are nowhere near AGI (@ns123abc, video). The quote: "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." This is the same Hassabis who, the same week, said humanity is in "the foothills of the singularity" (covered in today's digest Industry Pulse). The frame can hold both. Today's systems are not AGI, but the trajectory points there. The juxtaposition is worth noting.
- @ns123abc on world models versus RLVR (@ns123abc). "I'm now in the LeCun/Marcus camp on LLMs, real programming agents will need world models, not some RLVR shit." Posted as a reaction to Geohot's blog. This is one technical observer flipping from optimist to skeptic, and the timing (right after the Eternal Sloptember post) suggests the skeptical camp is broadening, not narrowing.
- Cursor / Codeium: Ronak Malde leaves DeepMind (@ellev3n11, no article). One-liner saying "some cool stuff soon" in response to Ronak Malde's announcement that he's leaving DeepMind after the $2.4B Codeium acquisition. Watch for product news from Cursor in 30-60 days.
- Ilya Sutskever posted something (@ns123abc, no article body). Just the tweet "Ilya Sutskever just posted this" with no actual quote attached. The farmer did not capture what Sutskever said. Click through to read.
- SF founder-residency directory (cluster of 3) (@Scobleizer hf0 thread, hf0 demo day, piano AI musing). Scoble pushed a 22-link list of SF AI hacker-houses and residencies (HF0, Founders Inc, Lightyear, Kernel Labs, AGI House, South Park Commons, The Bridge by EF, Neo Residency). Useful as a directory if you are tracking the SF early-stage AI startup ecosystem, no research content. Skip.
- Robotics in-home data collection (@Scobleizer). Argument that paid in-home camera setups (Figure, 1X, Tesla Optimus) will scale because users will accept data collection in exchange for a robot doing chores. Industry positioning piece, no research substance. Skip.
- Tesla FSD with Grok + Hey Grok geological guide (@Tesla cluster of 2, Hey Grok). Marketing for Tesla FSD Supervised plus Grok-as-voice-tour-guide. Skip.
- @brivael French political and cultural cluster (cluster of 16) (@brivael). Long-form French commentary on social class theory, Macron, Palantir's Karp on his PhD, French ENA, Nicki Minaj rumors, Elon Musk colonization, anti-woke rants, Christianity, and similar. None AI-research-relevant. Skip.
- Nicki Minaj two off-topic posts (@NICKIMINAJ, cluster of 2). Caught by the keyword filter, not AI-relevant. Skip.
- @spencerpratt LA politics (@spencerpratt, cluster of 2). LA mayoral campaign content. Skip.