Why this page exists
Most of this manual leans toward taking AGI risk seriously, since most of it covers what happens if the optimists are right. That posture makes it easy to flatten skeptics into a strawman — “they just don’t understand scaling” — and move on. That’s a disservice to a substantive position held by researchers who built some of the field’s foundational tools. This page represents their actual arguments, not a version built to lose, without refereeing who is correct; see how to weigh claims about AGI and AGI timeline forecasts for how the range of opinion is normally reported.
Yann LeCun: no world model, no planning
Yann LeCun left Meta in November 2025 after twelve years, including founding its FAIR lab, and by March 2026 was Executive Chairman and co-founder of AMI Labs, a Paris-based startup that raised a $1.03 billion seed round — reportedly the largest European startup seed round on record — to build what he calls “world models” (TechCrunch). That matters: LeCun’s skepticism didn’t soften into a hedge — it hardened into a company he is betting his career on.
His technical objection is specific. Auto-regressive LLMs spend a fixed amount of computation per output token, so they cannot allocate more effort to a harder sub-problem the way a person mulling a hard question can — there’s no mechanism for genuine planning, only next-token prediction dressed up with longer chains of tokens. They have no persistent memory beyond a context window, and no grounding in physical cause and effect: LeCun’s example is a toddler who knows a dropped pen will fall but doesn’t know which way it will spin, and doesn’t bother guessing — the toddler has an internal physical model with known gaps. An LLM will confidently guess anyway, because it has no internal model to consult, only a distribution over plausible text (Newsweek). His fix — training on video, spatial, and sensorimotor data instead of text alone — bets that AGI, if it happens, needs an architecture that doesn’t yet exist at scale.
Andrew Ng: raise the evidentiary bar
Andrew Ng — founder of DeepLearning.AI, Managing General Partner at AI Fund, and adjunct professor at Stanford — argues “AGI has turned into a term of hype rather than a term with a precise meaning,” and that the industry is decades away from a system able to do any intellectual task a human can, from driving a truck to writing a doctoral thesis (PYMNTS). His concrete contribution is methodological: a proposed “Turing-AGI Test,” where a judge designs a multi-day real-work trial — actual tasks, not a curated demo — and the system passes only if it performs as well as a skilled human. The point: move goalposts off leaderboard scores and onto sustained performance, defusing what he sees as bubble-risk hype.
Gary Marcus: hallucination, reasoning collapse, and the arithmetic of compounding assumptions
Gary Marcus, Professor Emeritus of Psychology and Neural Science at NYU and founder of Geometric Intelligence and Robust.AI, has built a multi-front case. On reliability: LLMs “literally don’t know the difference between truth and falsehood,” and he cites a formal proof they will hallucinate even given perfect data and unlimited compute — borne out, he says, when GPT-5 (August 2025) failed to solve hallucination and OpenAI’s o3 hallucinated more than its predecessors (Marcus on AI). On reasoning: “we are not going to solve physics with systems that can’t play Tower of Hanoi on a tower of 8 discs” — citing Apple’s finding that LLM “reasoning” models collapse past a complexity threshold rather than degrade gracefully.
His most widely discussed intervention is methodological: a takedown of the “AI 2027” forecasting scenario as “fiction dressed as science” — borrowing narrative techniques while presenting citations and footnotes as though the result were a forecast. Marcus decomposes the scenario into roughly eight sequential “leap of faith” milestones, multiplies their implied individual probabilities, and shows the compounded probability of the full chain landing on schedule is close to zero (Substack) — backed by forecasting’s own track record: hallucinations declared “solved” in 2024, driverless cars supposedly “everywhere” by 2017. Notably, Marcus isn’t arguing “AGI is impossible” — “anyone who thinks AGI is impossible” is as wrong, he says, as anyone who thinks it’s imminent — and now puts plausible arrival by 2100, not 2027. That stance is itself evidence the skeptic camp is a spread, not a single position — a point the timeline forecasts page makes about the field as a whole.
Tim Dettmers: computation is physical
Tim Dettmers, a researcher at the Allen Institute for AI, makes an argument that doesn’t depend on model architecture — it’s about physics. GPU performance-per-dollar plateaued around 2018, he argues, and most visible gains since have come from lower-precision numerics — doing the same arithmetic more cheaply — rather than genuine compute growth. He estimates “maybe one, maybe two more years of scaling left” before gains hit memory, I/O, and power walls, where energy and cost grow exponentially for only linear capability gains (timdettmers.com). He frames “AGI” as requiring economically meaningful competence at physical tasks, not benchmark scores, and extends the logic to its endpoint: a superintelligence would face the same physical constraints improving its own inference stack. Unlike the arguments above, this one doesn’t depend on being right about LLMs specifically — it’s a claim about the substrate any digital intelligence runs on.
The shorter roster: also see
A shorter roster rounds out the case. Rodney Brooks, CTO of Robust.AI and Professor Emeritus of Robotics at MIT, runs an annual, dated, falsifiable predictions scorecard — rare discipline among AI commentators — and argues from decades of robotics experience that narrow benchmark progress keeps getting mistaken for general competence. François Chollet, creator of the ARC-AGI benchmark, points to models scoring near zero on tasks built to resist memorization — the newer ARC-AGI-3 saw 2026’s best models score below 1% against 100% for humans — as evidence intelligence is skill-acquisition efficiency, not scale. Subbarao Kambhampati of Arizona State argues LLMs cannot plan or self-verify regardless of prompting, and now questions whether chain-of-thought tokens represent real reasoning at all. Melanie Mitchell of the Santa Fe Institute argues LLM success reflects pattern-matching over training distributions rather than causal reasoning. Emily Bender, a University of Washington linguist and co-author of “Stochastic Parrots,” argues fluent output is form without grounded understanding — models trained to predict word sequences lack access to communicative intent.
The forecasting methodology itself is a target
Beyond any individual researcher, skeptics take direct aim at how AGI timelines get produced. Expert-survey methodology shows a well-documented “moving horizon” effect — respondents place AGI roughly fifteen to twenty-five years out regardless of when they’re asked — compounded by selection effects and wording sensitivity. A March 2026 RAND report catalogs these problems systematically across survey, prediction-market, and compute-extrapolation methods alike (RAND).
METR’s own “time horizon” metric — the capability curve most often cited as evidence of acceleration, and discussed on the timeline forecasts page — has drawn scrutiny including from METR itself. A January 2026 notice disclosed a 35% correction applied to a fixed-slope logistic fit, that roughly a third of task-difficulty estimates are inferred rather than measured, and that METR has “no idea” whether a model’s true time horizon actually sits inside the confidence interval it reports (METR). An independent critique published days earlier went further: the tasks are unrealistically clean, public solutions may have contaminated the data, and human baseline times came from a small, financially incentivized sample that likely inflated how long tasks “should” take — making the AI’s apparent acceleration partly an artifact of the comparison (Transformer News). Either way, a large share of the field’s timeline inferences reprice off one noisy curve fit its own authors can’t fully vouch for.
That compounds with an older problem: Goodhart’s Law applied to benchmarks. Once a benchmark becomes an optimization target, labs fine-tune toward the leaderboard rather than the construct it measures — MMLU went from a 43-point GPT-3 score against a 90-point human baseline in 2020 to top models clustering so tightly by mid-2024 that differences fall inside prompt-formatting noise. Benchmark-leakage contamination makes rising scores harder to read as genuine gain. See how to weigh claims about AGI and the capability ladder for why “beats a benchmark” and “reaches AGI” are different claims.
What this page is not claiming
None of this amounts to a case that AGI is impossible, and none of the researchers above make that claim — Marcus is explicit both extremes are wrong. What it establishes is real substance: architectural limits that don’t obviously resolve with scale, a physical compute ceiling indifferent to any roadmap, and a forecasting apparatus that, by METR’s own admission, cannot fully vouch for the curve everyone cites. Whether that adds up to “decades away” or “closer than this page suggests” is exactly what this manual declines to resolve for the reader. What it will not do is let the skeptic case be a caricature.