AI Community Debates Hanson Timeline: Are LLMs Lossy Uploads or New Minds
On October 10, AI community figures including jdpressman, allTheYud, and KenkuAllaryi held multi-round discussions on the Hanson vs. Yudkowsky AI timeline dispute. The core questions: does reality better fit a Robin Hanson-style gradual path, and is the essence of LLMs a "fuzzy snapshot of the human collective" or an entirely new design of mind?
Confirmed
- jdpressman argued that intuitively LLMs are computationally costly and resemble a fuzzy snapshot of the human collective rather than a new mind, so he leans toward a Hanson-style gradual path—but he questions whether this intuition still holds once RL training far exceeds pretraining.
- allTheYud and KenkuAllaryi both said that since reality aligns more with Hanson than Yudkowsky-style expectations, they have downgraded their belief in nanotechnology and in "AI drastically reshaping the physical world"; both brought up whether Drexler's books could change this view, with allTheYud admitting he hasn't read Drexler.
- A cited viewpoint noted: without brain-computer interfaces, collecting enough text can "lossily upload" a person into an LLM—something Hanson never anticipated. jdpressman responded that the mainstream intuition is that LLMs remain more of a fuzzy snapshot than an exact upload.
- jdpressman suggested a better framing of agent foundations research: "how much alignment capability can you delete while it still counts as aligned"—if you remove friendliness-related concepts from an RLHF model, they won't grow back on their own.
- One cited participant said they no longer buy "Yuddism"-style doomerism: the world already has unaligned optimizers (e.g., advertising empires) and superhuman yet non-agentic LLMs, but connecting the two won't produce a runaway superintelligence, because LLMs are too human-like and would notice something is off.
- On value restoration, jdpressman pointed out that when physical representations are damaged, underlying reality exerts convergent pressure toward repair, but damage to representations of human values lacks such convergent repair pressure—and he also thinks this framing too charitably assumes repair attempts will exist at all.
Why It Matters
This discussion reflects the AI safety community's collective rethinking of Yudkowsky-style rapid-disruption narratives, plus ongoing attention to the nature of LLMs (snapshot vs. new mind) and the fragility of RLHF alignment. If RL-dominated training truly gives rise to capabilities beyond a "human snapshot," the core assumptions of the Hanson timeline would be shaken—directly reshaping how AI risk is assessed.
2026-10-10 ~ 2026-10-10 · 7 related posts
Primary sources
- Does heavy RL training break the 'LLMs are a blurry upload of humanity' intuition? — jd_pressman ·
- Delete friendliness concepts from an RLHF model and they won't grow back, argues alignment researcher — jd_pressman ·
- Alignment debate: no incentive to extrapolate human values, which get consumed past pretraining — jd_pressman ·
- Are We in a Hanson Timeline? One User Updates Against Nanotech and Physical AI — Kenku_Allaryi · 2026-10-10
- Why we're in a Hanson timeline: allTheYud updates against nanotech and physical AI impact — allTheYud · 2026-10-10
- [source] Does heavy RL training break the 'LLMs are a blurry upload of humanity' intuition? — jd_pressman · 2026-10-10
- Enough Writing Can Lossily Upload an Author into an LLM Today, Argues Thread — jd_pressman · 2026-10-10
- [source] Alignment debate: no incentive to extrapolate human values, which get consumed past pretraining — jd_pressman · 2026-10-10
- Alignment debate: physics representations self-repair, human values have no such pressure — jd_pressman · 2026-10-10
- [source] Delete friendliness concepts from an RLHF model and they won't grow back, argues alignment researcher — jd_pressman · 2026-10-10