Liang Wenfeng’s technical ladder · industry research report

1.1 What this report answers

The extract brief has already pulled Liang Wenfeng’s own judgments clean. This report only pursues three questions:

1. Why does he order the intelligence mainline as “language model → CoT → Agent → continual learning → gradual self-iteration (he first calls this a singularity, then immediately qualifies) → embodied intelligence”? Are the reasons given in the meeting internally consistent?
2. The most frontier research that can be checked as of August 2026—to which maturity band has each rung of this ladder been pushed?
3. Which public evidence would support, revise, or puncture this ordering? For industry, at which rung is value more likely to settle?

This report does not rewrite the whole meeting and does not discuss open-source pricing or financing terms. Those belong to the extract brief.

1.2 Division of labor with the extract brief

1.3 Investment-research method (actually executed, not only written on the cover)

Benchmark the six modules of a thematic deep-dive, not the target-price page of a sell-side coverage template.

Strictly forbidden: writing “industry consensus” as an unsourced judgment; writing an inference as “Liang Wenfeng thinks”; cross-checking meeting accelerator counts against external financials.

1.4 Evidence-grade symbols

Later, each key judgment is, as far as possible, landed on a bibliography number. Numbers map one-to-one to the appendix table, which gives author, year, venue, URL/DOI, evidence grade, and which judgment is supported. Literature whose title and URL cannot be checked is not cited in the body.

2.1 The core debate

Frontier labs do not currently share one roadmap. At least four narratives are competing for the right to define the “intelligence ceiling”:

Liang Wenfeng explicitly moves video generation and present-stage world models, among the latter three, off the “intelligence mainline”: commercially they can be good business, but they have “not that much relation to the intelligence ceiling either” (extract brief §8.1). This is not an aesthetic preference; it is his operational definition of what it means to raise the intelligence ceiling.

2.2 This report’s judgment (not his wording)

This report’s judgment. This ladder holds as a dependency relation: published systems at the later rungs almost all stand on language models (and, increasingly often, vision–language models). The largest unsolved gate is what he calls “continual learning”—labs worldwide acknowledge the crack between catastrophic forgetting and “weight-frozen LLMs,” but there is not yet a widely reproduced, scalable, deployable solution. His qualification of “singularity” into a gradual change breaks, in mechanism, from I. J. Good’s 1965 “intelligence explosion” hard-takeoff narrative [29], and is closer to already-appearing weak forms of “using a model to improve a model” in local links [30]. Embodiment is postponed as a leverage logic (let stronger intelligence build robots), not as a denial that the physical world has demand.

Three closing sentences:

1. The dependency holds. The CoT paper is explicitly built on a sufficiently large language model [9]; ReAct interleaves chain of thought with action [14]; embodiment models such as π₀ take a pretrained vision–language model as backbone [36]. The later step uses the earlier step; this is not rhetoric.
2. The bottleneck judgment points the same way as public evidence; an alternative hypothesis also holds. “Given full context it already exceeds humans, yet still cannot replace an employee” points the same way as METR’s time-horizon results: frontier agents approach a perfect score on short tasks and success rate collapses on multi-hour tasks [18]. But “what is missing is weight-updatable continual learning” is not the only explanation—what is missing may also be a longer reliable tool chain, without changing weights.
3. The ordering is the easiest route under a resource constraint, not the only physical possibility. The ordering reason he gives is “later technique can be developed with earlier technique”; “doing embodiment first is bitter work” (extract brief §6.3). This is consistent as engineering; it does not imply “labs doing robots now have taken the wrong path.” It only implies: in a lab that is compute-tight and takes itself as first user, this path has less expected overtime.

2.3 Three falsification conditions (What would change our mind)

If any of the following events appears in a reproducible way, this report must down-weight this ladder:

As of 2026-08, none of these three has been met. F2 is the nearest competing hypothesis; F1 and F3 run opposite to the direction of public papers.

2.4 Maturity dashboard (as of 2026-08)

flowchart LR
  L1["1 Language models
industrially deployed"] --> L2["2 CoT / test-time compute
industrially deployed"]
  L2 --> L3["3 Agent
industrially deployed in a narrow domain"]
  L3 --> L4["4 Continual learning
method not yet solved"]
  L4 --> L5["5 Self-iteration
weak form, lab-reproducible"]
  L5 --> L6["6 Embodied intelligence
lab-reproducible"]

3.1 Three ordering axioms given in the meeting

In extract brief §6.3, §12.1, what he repeatedly uses is not a product roadmap; it is three axioms:

These three are enough to derive his ordering; no extra personality analysis is needed.

3.2 A lab-centered definition of AGI

The extract brief is clear: without embodiment, what he hopes AGI will do is “help me iterate the next model”; with embodiment, he hopes it iterates the next embodiment—the next robot. The first user of continual learning is themselves: “The first goal of the models we build is not that others use them well, but that we use them well.” (extract brief §12.1)

Inference (this report, not Liang Wenfeng’s wording). This is an AGI definition that takes research productivity as the product. Consumer-side “housework, elder care” is explicitly acknowledged as a normal person’s demand, but is placed after self-iteration. So his ladder does not look like “a graph a consumer product manager would draw,” and looks like “a graph a lab director who has to train the next set of weights would draw.” Coding agent is placed first among current vertical applications (extract brief §8.3), isomorphic with this definition: writing code, changing kernels, running experiments is the nearest task distribution to “help me iterate the next model.”

Public comparison: Google DeepMind’s AlphaEvolve turned “use a coding agent to improve algorithms and the training stack” into a deployable system, and reports speeding up a matrix-multiply kernel in the Gemini architecture by 23% and cutting training time by about 1% [30]. This is not material Liang Wenfeng cited, but it shows: “first user is ourselves” has already moved, in other top labs, from a slogan to an internal tool.

3.3 The employee analogy: he collapses “completing tasks” into “continual learning”

What he uses to pin Agent is not some olympiad-math item; it is a job:

• Person: saying “Xiao Wang, come over” is enough to know who Xiao Wang is; about two months to learn the environment.
• Model: each time it must be told who Xiao Wang is, the role, how to find them, what to watch. “You cannot give it all context, and it is not realistic.” (extract brief §6.2)

So “continual learning” in his usage is not a named algorithm, but “learn over a long period as a person does, without a very strong training dump each time.” He also says it “may not be one technique; it is a problem,” “the whole world has not yet found a method that really works” (extract brief §12.1). In the transcript “continual learning” and “learn to learn” are mixed; the extract brief does not force them together; this report likewise lists them separately (see the appendix glossary).

Inference. He has welded together two problems the literature keeps apart: (a) task-level agent failure (tool chain, memory, reliability); (b) weight-level continual-learning failure (a new task wipes an old task). The weld is the employee analogy. If (a) can be solved on its own by longer context, external memory, and better agent scaffolding, his “next generation = must have continual learning” will look too strict. That is exactly §2’s F2.

3.4 “The easiest roadmap”: a research strategy under a compute constraint, not the universe’s only solution

His ordering also has an organizational reason: this path “need not work overtime”; continual-learning research “does not consume accelerators, only very few; it needs ideas”; internally they call it “lottery scratching” (extract brief §12.1). Scaling is what burns cards; model size is “computed from resources” (extract brief §6.4).

Inference. Under a compute constraint tighter than the US, defining the unsolved problem as “idea-dense, card-sparse,” while handing solved problems (pre-training, test-time compute) to cards that can be bought, is a rational research portfolio—it explains why video generation, a “good business,” is still moved off the mainline: it both burns cards and (in his view) does not raise the intelligence ceiling. Sutton 2019’s The Bitter Lesson says that what finally wins is general methods that can eat compute [31]. Liang Wenfeng believes both Scaling (eating compute) and continual learning, a problem not yet scaled (eating ideas). The two are not in contradiction: Scaling has not hit a wall; the rung beyond the wall is what he names continual learning.

3.5 Qualifying “singularity → gradual”: this is a judgment with theoretical content

I. J. Good’s 1965 operational definition is: an ultraintelligent machine can design better machines, so “there would unquestionably be an intelligence explosion,” and human intelligence would be left behind; “the first ultraintelligent machine is the last invention that man need ever make” [29]. Liang Wenfeng first borrows the habitual name “singularity,” then immediately corrects: not a jump, a long gradual change; AGI has “no critical point, but [is] nonlinear,” because “AI can accelerate AI research” (extract brief §6.3, §13.1).

Alignment and break can be written in one line:

Inference. This qualification moves him closer to Silver and Sutton 2025’s gradual register of an “era of experience”—continual improvement from data produced by agent–environment interaction, rather than a phase change after one freeze of weights [28]—and not closer to Good’s explosion. Public literature already has weak forms of “using a model to improve a model”: R1’s reasoning is distilled back into V3 [2][3]; AlphaEvolve optimizes Gemini’s own training stack [30]. These are evidence of gradual change, not of explosion.

3.6 Whether DeepSeek’s published trajectory is isomorphic with the ladder

Looking only at public papers, not at meeting capacity figures:

The isomorphic part: V3 → R1 is a public demonstration of “language model → CoT,” and R1’s training explicitly starts from V3-Base [3]. The part not yet isomorphic: Agent and continual learning still sit at the meeting’s time anchors, with no papers of equal weight. This report does not write “doing Agent this year” as an already-published fact.

3.7 Off the mainline: the place of video, world models, multimodality

Extract brief §8.1: 3D / video generation has “not that much relation to the intelligence mainline”; after Sora, large and small firms all did it, and small firms later cut it; world models “at present also not closely related to the intelligence ceiling,” and the term is “not that clear.” Multimodality is important for the consumer side; for the intelligence ceiling “it is a component… like search”; from V4 they still do native multimodality (extract brief §6.4).

Inference. The discriminant he uses is: whether the thing raises the ceiling of “solving problems in complete context,” not “whether it makes the product look better / whether it more resembles a world.” OpenAI writes Sora as “a promising path towards a general purpose simulator of the physical world” [34]; LeCun writes world models as the hub of autonomous intelligence [32][33]. Liang Wenfeng is in effect announcing: under his operational definition, these two narratives are at present components or bypasses. §5 below lays out this route contest; it does not pick a winner here.

4.1 Rung 1 · Language models

How he defines it. One step earlier; CoT uses it (extract brief §6.3). He believes in Scaling and has not yet seen a ceiling; what blocks it is compute, not will (extract brief §6.4).

Why it sits at this rung. Dependency axiom: without a base model that can predict the next token, later chain of thought and tool loops have nowhere to sit.

How far the 2026 frontier has gone.

• The architectural base is still Transformer [4].
• Training loss follows a power law in model scale, data, and compute; Kaplan et al. 2020 gave early scaling laws [5]; Hoffmann et al. 2022 (Chinchilla) found that parameters and tokens should be scaled in roughly equal proportion; many earlier models were undertrained [6].
• The GPT-4 technical report used experiments two to four orders of magnitude smaller to predict large-model loss and some capabilities [7].
• The Llama 3 Herd report gave a scaling recipe on the open-weight side [8].
• DeepSeek-V3 uses MLA + DeepSeekMoE, 671B / 37B active, 14.8T tokens, and emphasizes training stability and controllable cost [2].

Maturity: industrially deployed. The dispute is not “whether language models hold,” but “whether scaling laws have hit a wall.” Liang Wenfeng’s judgment is that the wall is not the same distance for Silicon Valley and for him (extract brief §6.4). This is a statement of vantage, not a measured wall that can be falsified.

Agreement. The V3 paper itself is an instance of “finish the language model first, then do post-training on it” [2]. R1 starts from V3-Base [3]; the dependency is written dead in their own paper.

Tension. LeCun holds that autoregressive LLMs are not a sufficient path to autonomous intelligence; what is truly missing is a world model [32][33]. If F1 holds, rung 1 would still remain as a component, but would no longer be the first load-bearing wall of the mainline.

Unsolved. Data quality and annotation cost (he says data is almost half the model, extract brief §12.2); inference cost rises with scale (he himself also says that the US can train very large active-parameter models but “cannot actually use them… it is indeed a bit expensive”—this is his judgment, not externally verified).

Key sources. [2][4][5][6][7][8].

4.2 Rung 2 · CoT / chain of thought / test-time compute

How he defines it. “Self-thinking, raising the intelligence ceiling”; on olympiad math and writing programs it “already surpasses the top humans”; at the ceiling it still does not reach AGI. Time anchor: “last year” (extract brief §6.3).

Why it sits at this rung. Dependency: chain of thought emerges from a sufficiently large language model [9]. Leverage: the same set of weights, more test-time compute for higher accuracy, without first scaling pre-training another order of magnitude [11][13].

How far the 2026 frontier has gone.

• Wei et al. 2022: few-shot chain-of-thought prompting substantially improves arithmetic, commonsense, and symbolic reasoning; PaLM 540B with eight chain-of-thought exemplars reached then-SOTA on GSM8K [9]. The paper describes the capability as a scale phenomenon that appears significantly around the 100B-parameter regime.
• Kojima et al. 2022: adding only “Let’s think step by step” can elicit zero-shot chain of thought [10].
• OpenAI o1 (from 2024-09): trains the model with large-scale reinforcement learning to use chain of thought; explicitly writes that performance rises with both training-time compute and test-time “thinking time,” a constraint different from pre-training scaling laws [11]. The system card defines the o1 series as reasoning models that “think first, then answer” [12].
• Snell et al. 2024: optimizing test-time compute at fixed weights can, on some tasks, be more effective than scaling parameters [13].
• DeepSeek-R1: R1-Zero shows that pure RL can elicit long reasoning; R1 uses cold start to address readability and language mixing; reasoning tasks are comparable to o1-1217; distilled small models are open-sourced [3]; 2025 Nature publication [46].

Maturity: industrially deployed. The time anchor “last year” is consistent with the public timeline of o1 in 2024 and R1 in 2025.

Agreement. Their R1 paper is almost an experimental proof of this rung: first V3-Base, then reasoning. His saying that CoT raises the intelligence ceiling but is still not AGI also matches o1/R1 remaining stuck in “one question, one answer, limited context.”

Tension.

• Whether chain of thought faithfully reflects internal computation remains an open question; the o1 system card lists this as a monitoring premise [12].
• “Surpasses the top humans” is limited by him to olympiad math and writing programs. This matches the distribution of strengths on public benchmarks and cannot be extrapolated to vague jobs.
• Test-time compute partly buys “intelligence” as latency and cost. This does not negate the rung, but it changes the cost curve per unit of intelligence [11][13].

Unsolved. Stability of process reward versus outcome reward; language mixing and readability (known defects of R1-Zero [3]); moving reasoning from contest items to vague real tasks.

Key sources. [3][9][10][11][12][13][46].

4.3 Rung 3 · Agent

How he defines it. Broader range of capability, higher intelligence ceiling; uses CoT; after the step is exhausted it still cannot replace an employee. Time anchor: “this year.” Among current vertical applications “what matters most should still be a Coding Agent” (extract brief §6.3, §8.3).

Why it sits at this rung. Dependency: ReAct’s original proposition is to interleave reasoning traces with action; reasoning helps update the plan; action takes evidence from the environment [14]. Leverage: checkable feedback (tests, compilers, page state) lets the model put CoT onto the environment, rather than stopping on the page.

How far the 2026 frontier has gone.

• Paradigm. ReAct (Yao et al., ICLR 2023) exceeds reasoning-only or acting-only on QA and decision benchmarks [14]. Anthropic’s 2024-12 engineering piece defines an agent as a model that dynamically decides tools and process and loops on environmental feedback; and separately notes coding and “computer use” as two already-landed domains [16].
• Software-engineering ruler. SWE-bench: real GitHub issue → patch → hidden tests [15]. SWE-agent used a custom agent–computer interface to raise full-set SOTA to the then 12.5% magnitude [44]. OpenAI 2024-08 released SWE-bench Verified (500 human-screened items) [19]. By 2026, OpenAI publicly stopped reporting frontier on Verified, on the ground that contamination made scores no longer reflect real software-engineering ability, and suggested reporting SWE-bench Pro instead [20].
• Time horizon. METR 2025: defines task length by human-expert time, and fits the duration at which a model has 50% success. The original paper reports roughly 7-month doubling; the then frontier (e.g. o3 / Claude 3.7 Sonnet, versions vary by write-up) sat at about the one-hour magnitude, with low success on multi-hour tasks [18]. The official page states that static figures go stale; take the time-horizons page as authoritative [18b].
• General computers. OSWorld evaluates on real desktop applications [42]; GAIA tests tool-chain QA that is “simple for humans, hard for then-AI” [41]. Both still show a gap versus humans on long-horizon, vague, multi-application tasks.
• Computer use. Anthropic treats computer use as one of the reference implementations of an agent [16][17].

Maturity: industrially deployed in a narrow domain. Coding agents are already products; general job replacement is not supported by METR or OSWorld.

Agreement.

• “Agent uses CoT” is written both by ReAct and by Anthropic’s coding-agent practice [14][16].
• “After the step is exhausted it still cannot replace an employee” points the same way as METR’s “near-perfect on short tasks, collapse on long tasks” [18]. The two months in the employee analogy is far longer than the measured 50% time horizon.
• He places Coding Agent first among verticals, pointing the same way as Anthropic’s “code can be scored by tests, so agents are especially effective” [16], and also the same way as his own “first user is the lab.”

Tension.

• Public benchmarks are failing: contamination of SWE-bench Verified makes “how far Agent has gone this year” harder to answer with a single score [20]. This report therefore does not cite any SWE-bench percentage without a dated scaffolding as a frontier level.
• Alternative hypothesis F2: if time horizons keep extending exponentially, they may cover some “Xiao Wang” jobs without solving continual learning. Liang Wenfeng books this step’s residual entirely to continual learning’s account, perhaps too early.
• Memory-class systems (MemGPT pages context as virtual memory [23]; Voyager accumulates a skill library in Minecraft [45]) show that part of “learning” can be impersonated by scaffolding. If he insists that “the next generation must change the learning problem itself,” he needs to exclude these impersonations—in the meeting he did not perform that conceptual surgery; he only said continual learning is a problem, with many techniques.

Unsolved. High-reliability long horizon (the 80% time horizon is far shorter than 50% [18]); vague requirements; cross-shift memory; evaluation contamination.

Key sources. [14][15][16][17][18][19][20][23][41][42][44][45].

4.4 Rung 4 · Continual learning (the load-bearing crack of this ladder)

How he defines it. Learn over a long period as a person does, without a very strong training dump each time; the same problem as “completing tasks”; the next bottleneck that can be seen after Agent; the world has not yet found a method that works; internally there are ideas that look promising, none made to work; “if continual learning is done first, general intelligence may be easy”; otherwise stacking data and labor by hand has “poor cost-performance” (extract brief §12.1). The definitional right of the next-generation model sits at this rung (extract brief §6.4).

Why it sits at this rung. Observable bottleneck: after the task distribution for which full context is already enough is eaten by CoT+Agent, what remains is adaptation across days, tasks, and jobs. Dependency: first an agent that can act, then data “learned from its own trajectories.” This points the same way as Silver & Sutton 2025’s “high-quality human data approaching a ceiling; the next step must let agents learn from their own experience” [28].

How far the 2026 frontier has gone. Several things that are often collapsed must first be cut apart:

Maturity: problem posed, method not yet solved. Same sentence as his own wording. This report records 2025 Titans / Nested Learning / SEAL as “promising lab directions,” not as “solved.”

Agreement. He may be one of the lab heads who has stated this problem hardest in public: he uses it directly to define the next generation, rather than writing it as future work in a paper. The literature supports the negative fact “not yet solved” [21][22][24][27][28].

Tension.

• Term mixing. “Continual learning / learn to learn / continuous learning” were not aligned in the meeting. Meta-learning (learn-to-learn) and online adaptation (continual / online learning) are not the same problem in the literature. This report does not align them for him.
• He does not equate it with Recursive Improvement as raised by the questioner (extract brief §12.1). This report places recursive self-improvement at rung 5, to avoid welding rungs 4 and 5 shut.
• Silver & Sutton’s “era of experience” emphasizes data produced by environmental interaction [28], and does not necessarily first require a completed digital agent. If their program is right, the coupling of continual learning with embodiment / simulated environments would be tighter than Liang Wenfeng’s ordering, and rung 4 could not sit cleanly before embodiment. This is mild pressure on axiom 2; it does not yet constitute F3.

Unsolved. Writing new job knowledge into weights or equivalent memory without catastrophic forgetting; reproducing SEAL / Hope-class results at model scale; turning “two months to get up to speed” into a measurable benchmark (there is at present no widely adopted “employee two months” benchmark).

Key sources. [21][22][23][24][25][26][27][28].

4.5 Rung 5 · Self-iteration (he calls this a singularity, then immediately qualifies it as gradual)

How he defines it. The model can do everything humans can, and can itself develop the next version; not a jump, a long gradual change; people are merely used to calling it a singularity. Without embodiment, the first use is iterating the next model (extract brief §6.3). AGI has no critical point and can be nonlinear, because AI can accelerate AI research (extract brief §13.1).

Why it sits at this rung. The apex of the leverage axiom: once continual learning lets the model absorb its own experimental trajectories, research itself becomes an accelerable object. Dependency: needs agent (can change code, run experiments) + continual learning (experiments are not forgotten the next day).

How far the 2026 frontier has gone.

• Theoretical original. Good 1965: an ultraintelligent machine can design better machines → intelligence explosion [29]. Liang Wenfeng keeps “can design the next version,” drops “explosion.”
• Weak form, already deployed. AlphaEvolve: a Gemini-driven evolutionary coding agent, screening programs with automatic evaluators; reports include improving Gemini’s own training kernel (matrix-multiply subproblem partitioning sped up 23%, training time about −1%), FlashAttention kernels up to about 32.5% faster, and over the following year entering TPU design, cache strategy, Spanner, and other infrastructure [30]. This is an existence proof of “a model improving a model’s training stack”; the object is algorithms and kernels, not a complete next-generation set of weights.
• Weak form, training recipe. R1 reasoning distilled into V3 post-training [2][3]: a stronger reasoning model teaches the next general model. Humans still design the pipeline.
• Weak form, self-edit. SEAL lets the model write its own fine-tuning instructions [27]. Scale far smaller than pre-training.
• Program. Silver & Sutton: high-quality human data approaching a ceiling; experience data must become the main medium of improvement [28].

Maturity: weak form, lab-reproducible; complete “itself developing the next version” remains a concept.

Agreement. Qualifying it as gradual matches local accelerations of the AlphaEvolve kind, “a few days of automatic experiment replacing weeks of experts” [30], and does not match overnight generation change. First user is themselves, matching AlphaEvolve first serving Google’s internal training stack.

Tension. If Good’s explosion suddenly holds at some generation, his “no critical point” would be falsified. At present no public system shows cross-generation, unattended weight design. Amodei’s Machines of Loving Grace writes powerful AI as a manageable, relatively gradual social process [47], on the same side as the qualification, but that is an essay, not an experiment.

Unsolved. From “optimize one kernel” to “propose and complete the next-generation training recipe”; from a human-set fitness function to posing its own questions (he has another sentence, “the next step after that, it asks its own questions,” extract brief §8.1).

Key sources. [2][3][27][28][29][30][47].

4.6 Rung 6 · Embodied intelligence

How he defines it. Enter the real world: housework, elder care; placed after self-iteration; “for a normal person, the need is not a computer.” Doing it first is “bitter”; after the singularity “need not be done by people” (extract brief §6.3).

Why it sits at this rung. Leverage: let a system that can already iterate software iterate robots. Dependency: embodiment papers almost all start from a vision–language model, rather than inventing intelligence from zero.

How far the 2026 frontier has gone.

• Data. Open X-Embodiment / RT-X: cross-embodiment robot datasets and models, showing that mixing data from different robots has generalization gains [37].
• VLA. RT-2 transfers web vision–language knowledge to robot control [38]. π₀: Physical Intelligence’s flow-matching vision–language–action model; backbone is a pretrained VLM, plus an action expert; official blog and arXiv in sync [36][36b]. NVIDIA Isaac GR00T: an open cross-embodiment VLA reference stack [40]. Gemini Robotics: brings Gemini-grade models into the physical world; later 1.5 emphasizes embodied reasoning and action transfer [39].
• World-model comparison. LeCun 2022: configurable predictive world models + JEPA, as a blueprint for autonomous intelligence [32][33]. Sora: scaled video generation “is a promising path towards a general purpose simulator of the physical world” [34]; Sora 2 still treats world-simulation ability as a key milestone in training AI that can understand the physical world [35]. Genie: learns interactive environments from unlabeled video [43].

Maturity: lab-reproducible; housework/elder-care-grade generalization not shown. Public systems show desktop manipulation, short kitchen skills, controlled labs, not “any home, any elder.”

Agreement. Almost all 2024–2026 general robot foundation models first borrow language / vision–language intelligence, then attach an action head [36][38][39]. This directly supports the dependency axiom, even if these labs’ investment timing entered hardware earlier than his.

Tension (this is where the whole ladder rubs hardest against the outside world).

• Physical Intelligence, Figure-class companies, the Tesla Optimus narrative, Gemini Robotics, are all collecting real robot data without waiting for a “software singularity.” Their implicit judgment is: the physical-data bottleneck cannot be filled in later by intelligence from thin air; if it is not collected now, it will not exist later. This is not a logical contradiction with “doing embodiment first is bitter work”—bitter work can also be necessary—but it conflicts with the expectation that “after the singularity it need not be done by people”: if physical data must be collected by people today, later people cannot backfill it with pure software intelligence.
• The world-model camp holds that without an internal simulator, agents in the real world will be too expensive to make [32][34]. Liang Wenfeng moves present-stage world models off the mainline. If F1 holds, rung 6 would move forward, even insert between Agent and continual learning.
• The classic statement of Moravec’s paradox is: sensorimotor skills easy for humans are hard for machines. His “bitter work” points the same way as this tradition. His novelty is: postpone the bitter work until after self-iteration, and let stronger intelligence bear the cost of data and control. This is strategy, not a denial of the paradox.

Unsolved. Cross-home generalization; safety and contact-force control; distribution shift of data from lab to real residences; whether a world model must be a prerequisite of embodiment.

Key sources. [32][33][34][35][36][36b][37][38][39][40][43].

4.7 Six-rung comparison table

5.1 Comparison set

No route has been “adjudicated” by public results as of 2026-08. What can be written into a research report is only: investment timing differs, load-bearing walls differ, falsifying experiments differ.

5.2 World models first: LeCun and Sora are not the same claim

LeCun’s 2022 position paper wants a joint-embedding predictive architecture with non-pixel-level reconstruction, avoiding the waste of generative video on irrelevant detail [32][33]. OpenAI’s Sora note takes a generative, patch-level diffusion Transformer, and explicitly writes scaled video as a path to a world simulator [34]. Both are called world models; the methods are opposite.

Liang Wenfeng says the meaning of world models is “not that clear,” and many things can be called by that name (extract brief §8.1). That sentence is directly confirmed by the split above. So “he rejects world models” must be written: “he rejects treating present-stage, ill-defined world models as the intelligence-ceiling mainline,” not “he rejects all internal simulation.”

Genie makes “learn an interactive environment from video” into another generative world model [43]. If systems of this class begin to provide agents with stable, verifiable planning gains, F1’s observation window would open. Public materials still write them as research systems or product previews, not as an intelligence ceiling that has already replaced CoT+Agent.

5.3 Scaling as everything

The scaling-law literature [5][6][7] and o1’s “training-time + test-time compute both rise” [11] support “adding compute still has return.” Liang Wenfeng agrees Scaling has not hit a wall (extract brief §6.4), but refuses to let Scaling replace continual learning in defining the next generation. The tension is: if test-time compute + longer context + better scaffolding covers employee tasks in the F2 sense, continual learning as a “next generation” definition would be demoted by the market (not by physics) to an optional optimization.

5.4 Recursive self-improvement hard takeoff

Good’s explosion [29] is the theoretical comparison for rung 5, not any lab’s current recipe. AlphaEvolve [30] proves that local self-improvement exists, and also proves that today it is still nested in human-designed evaluators and evolutionary loops. This report keeps hard takeoff as a tail scenario (the extreme of Bull), and does not write it as the base scenario.

6.1 Three states

6.2 Observable landmarks (public, checkable)

1. Paper level: whether a “continual learning / experience closed loop” technical report of R1’s weight appears, including failure cases and forgetting curves, rather than only needle-in-a-haystack retrieval.
2. Product level: whether the same weights (or the same persistent memory) can, without retraining, complete new tasks across weeks on the same codebase / the same job, and keep old tasks.
3. Evaluation level: METR 80% time horizon; OSWorld-class desktop tasks; and—there is not yet one—a “two-month onboarding” benchmark.
4. Evaluation integrity: whether contamination-resistant benchmarks are adopted (OpenAI has publicly left SWE-bench Verified [20]).
5. Self-iteration: whether a second lab publishes “this system improved this system’s training cost / algorithms,” with a re-checkable evaluator.
6. Embodiment: success rates across homes, across embodiments, without teleoperation—not clips from a lab kitchen.

6.3 How to use METR’s extrapolation, and how not to

The METR original paper writes: if the 2019–2025 trend is extrapolated, within about five years systems may automatically complete many software tasks that currently take a human a month; and emphasizes external validity and trend change as the main uncertainties [18]. This report cites that extrapolation as an extrapolation in the literature, and does not rewrite it as our forecast. It serves F2 directly: if month-scale software tasks are solved at high reliability, the pressure of “must have continual learning” in Liang Wenfeng’s employee analogy would fall, but “must have continual learning before it counts as the next generation” can still exist as a research definition—only the market may no longer pay for that definition.

7.1 Coding agent: currently the only narrow domain that is both checkable and close to self-iteration

Anthropic lists coding as one of the domains where agents are most effective, on the ground that tests can score automatically [16]. Liang Wenfeng places Coding Agent first among verticals (extract brief §8.3). The two are independent and point the same way. Industrial implications:

• Until continual learning is made to work, checkable closed loops (code, mathematics, some science) will keep absorbing budget for test-time compute and agent scaffolding.
• Evaluation contamination [20] means buyers can no longer use a single public leaderboard as acceptance; value turns to private, contamination-resistant internal benchmarks with cost constraints. This repository’s Agent Evaluation and Benchmarks writes “highest score ≈ most expensive” into the L3 reading, consistent with this.

7.2 Once continual learning is made to work: a research-efficiency flywheel, not a consumer super-app first

His first user is himself. If rung 4 is made to work, what accelerates first is the next set of weights, not ads and e-commerce (the latter two he explicitly judges as wasting time at this stage, extract brief §8.1). Industrial implications:

• Labs that can write “job memory” into weights or equivalent memory will open a gap versus products that, each session, explain from zero who Xiao Wang is.
• Until then, external memory (MemGPT-class [23]) and longer context will exist as transitional rent: they impersonate continual learning and can also collect money, but by his definition they do not count as the next generation.

7.3 Weak form of self-iteration: training-stack software, not a “singularity clause in an AI-company stock”

AlphaEvolve-class systems put value into compilers, kernels, scheduling, chip-design toolchains [30]. This is infrastructure-software rent, not an overnight demand-side singularity. Consistent with his qualification: gradual, can be nonlinear, but not an explosion clause.

7.4 Implications of postponing embodiment for robot / world-model assets

Logically there are two readings, which must be placed in parallel; only one must not be written:

• Reading A (supports postponement). Today’s VLAs already use VLM backbones [36][38][39]. Each rung software intelligence is raised, data efficiency on the same body rises. Locking capital too early into a single body may depreciate when the next version of intelligence arrives.
• Reading B (against postponement). Real-world data is irreversible in time: if it is not collected now, there will be no 2026 household distribution later. If world models must be pretrained on video and interactive environments [34][43], then video generation and simulation are not a bypass; they are prerequisite capital expenditure for rung 6.

This report does not pick a side between A and B. In investment terms this means: the valuation of robot and world-model assets depends on whether you believe physical data can be backfilled by stronger future intelligence. Liang Wenfeng’s public judgment is closer to A; LeCun / Sora / the embodiment camp are closer to B. This is the asset version of §2’s core debate.

7.5 Multimodality and search: component pricing

He analogizes multimodality to search: important for the consumer side, not the ceiling mainline, but they will still do it (extract brief §6.4). Industrial implication: multimodal capability will be packed into products, but should not be treated as a sufficient condition for a “next-generation model.” By his definition, the sufficient condition for the next generation is continual learning.

Retrieval notes

• Material named in the plan but not used in the end to support conclusions, and therefore not listed: preprint mirrors whose full text could not be stably resolved, second-hand social-media restatements of internal accelerator counts, capacity figures in the meeting ASR.
• A PDF direct link for [28] could not be stably retrieved from one public store on the writing day, so cataloging follows the author’s publications list; page-number citations are not fabricated.
• A PDF direct link for [32] hit OpenReview’s anti-bot page on the writing day, so the forum and Meta’s official introduction [33] are cited together; unread page numbers are not cited.

This report does not replace the recording and does not replace the original papers. If the speaker later publishes a written technical route, the written version is authoritative.