Current language-model approaches can scale to superintelligence.

The “HFDT scales far” assumption: larger models on more and harder tasks, with no change to the training strategy, reach the point of autonomously advancing frontier R&D and beyond.

The whole forecast is an extrapolation of the current paradigm: another 3 to 6 OOMs of effective compute plus unhobbling takes present models to AGI by 2027 without a new idea.

Safer-1 and Safer-2 are 2025-style models that think in English; Safer-3 reuses Agent-4’s architecture. The authors say the story would be more optimistic if the first automated researchers still thought in faithful English, and this branch shows why.

Agents 1 through 4 are transformer-lineage models with added neuralese memory and iterated distillation; the authors say the story would still work with faithful English chains of thought, only more optimistically. Agent-5 is a partly rewritten new design, but it is built by Agent-4, not by humans.

He points to large language models as evidence that his 2029 date is on track and names contextual memory, common sense, and social interaction as the gaps that more compute and training will close, rather than calling for a new paradigm.

Every advance through 2026 is the same transformers made larger, wrapped in bureaucracies, and trained by reinforcement, and the author intends to keep writing until the trajectory reaches singularity, extinction, or utopia.

Powerful AI “could come as early as 2026,” a horizon that only makes sense if present models scaled and refined reach Nobel-level ability.

The models of the 2030s are scaled successors of the 2026 systems; the plan deliberately limits algorithmic progress and new paradigms, and still reaches top-expert AI by 2035 and superintelligence after 2040.

HQU is an evolved self-supervised sequence model, an extrapolation of current methods rather than a fresh paradigm. The story argues today’s trajectory suffices.

Sable is a scaled and modestly modified successor to current systems (parallel inference, shared memory, vector reasoning), not a new paradigm. The video adaptation removed even those changes to argue that scaled-up current methods suffice.

Written in 2003, the story has no language models; Manna is a 2010 PC program, and the decisive advance is a vision chip. The Vertebrane network is said to learn rather than execute code, but nothing in it exceeds human intelligence.

Lee’s introduction says general intelligence would need a dozen more breakthroughs like deep learning after only one in sixty years, so no AGI arrives by 2041 and every system in the book is a single-domain optimizer.

Hanson guesses that the first robots able to substitute for human workers will be brain emulations, arriving well before AI software is anywhere near human-level abilities.