Human-Like AGI: Why It May Be Harder Than It Looks

Everyone seems to agree on the order of events: AGI arrives, then superintelligence follows close behind. But what if the first step is far harder than the story assumes? What if human-like generality, the kind a child shows when dropped into a world it never saw before, is a different ladder entirely rather than the next rung on this one?

For UK readers the short version is this. One camp defines AGI as economic output and sees straight lines pointing at 2029 or sooner. The other defines it as learning like a human and sees measured barriers: fragile reasoning, inflated benchmarks, disembodied minds, and oversight that degrades exactly as agents get stronger. Both sides have evidence. The honest position is that capability is compounding fast while generality, of the human kind, remains unproven. Plan for powerful agents now, and treat human-like generality as an open scientific question.

What do people even mean by AGI?

The debate starts with definitions that barely overlap. OpenAI's charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work, an economic yardstick. Its chief executive has glossed it variously as tackling complex problems at human level across many fields, then narrowed it operationally to systems that invent new things in a way that matters. Under an economic definition, a developer can declare progress, even victory, the moment the revenue charts move.

The challengers reject that yardstick outright. The ARC Prize camp, following François Chollet's 2019 paper On the Measure of Intelligence, defines intelligence as skill-acquisition efficiency: matching how fast humans learn genuinely new things, relative to priors and experience. It explicitly calls the economic definition an incorrect measure of intelligence, because skill can simply be bought with data and engineering. Gary Marcus demands trustworthy systems with real semantics, causal and abductive reasoning and theory of mind, arguing large language models are statistical approximation rather than general intelligence and pointing toward neurosymbolic paths. Yann LeCun's 2022 position paper, A Path Towards Autonomous Machine Intelligence, argues autoregressive predictors cannot scale to AGI at all — too sample-inefficient, text-only, searching an exponentially large token space — and proposes world-model architectures in the JEPA family that learn from sensory input, simulate consequences and plan. Demis Hassabis holds the classic line of matching most human cognitive abilities, judging that scaling plus, in his words, a couple more big breakthroughs gets there around 2030. Four definitions, four different finish lines, which is why one person's arrival announcement is another person's category error.

Why do smart people say it is nearly here?

The accelerationist case deserves its strongest form, because five curves really do point the same way. Training compute has compounded at roughly four to five times per year for a decade, and analysts judge runs ten times larger feasible by 2030. When pretraining hit diminishing returns, reasoning-time compute opened a second scaling curve, and new scaling laws now optimise parameters, tokens and inference samples jointly. Scale did not break; it grew a new axis.

Agent capability follows, and here the numbers deserve naming. METR's measurement across 170 software and research tasks finds the horizon of work agents complete doubling roughly every seven months since 2019, from minutes toward hours. SWE-bench Verified tells the same story in code: 33 percent in early 2024, past 80 percent by late 2025, past 90 percent by mid-2026, with its successor benchmark already at 80 percent and Terminal-Bench at 88 percent. Mathematics went from under two percent on research-level FrontierMath problems to olympiad gold — 35 out of 42, a score only four percent of human contestants beat — to near-saturation of the hardest tiers. The ARC-AGI-3 test of novel games went from near zero to near perfect in about fourteen months, with the ARC Prize team itself reporting superhuman efficiency: beating the median human action count on 96 percent of levels using roughly half the actions, by compressing novel mechanics into compact symbolic models and building its own tools. Jensen Huang, Hassabis and Brockman converge on roughly 2029 to 2030, or now. Dismissing all of that as benchmark theatre requires explaining away a lot of independent lines bending upward together.

Why might human-like generality be much harder?

Because generality of the human kind may need what text predictors never had: a body and a world. An old paradox still binds, which is that sensorimotor skills a one-year-old masters demand vastly more computation than adult-level exam performance. Language models score like graduates yet perceive and move like infants. Reading about the world is not the same as bumping into it for twenty years, and robotics generality lags language by a distance that scaling alone has not closed.

The reasoning fragility is measured, not philosophical. Apple's GSM-Symbolic study showed frontier models changing answers when the same question merely changed numbers, degrading as clauses piled up, and collapsing by up to two-thirds when a single irrelevant but plausible sentence was added. The authors concluded the systems replicate reasoning steps from training data rather than performing genuine logical reasoning. A newer fluid-intelligence benchmark was designed explicitly to be easy for humans yet hard for current AI, and the gap persists. Next-word prediction, however scaled, is not yet a causal model of the world that plans.

Evaluation itself is shakier than headlines admit. Studies find training on the test set inflates scores far more than recent releases acknowledge, with effects growing at scale. On the Vals AI finance benchmark of 927 expert-reviewed analyst tasks, the leading model manages only around 60 percent, collapsing on multi-step precise-number work. The Lost in the Middle finding shows long contexts get used unevenly, with performance sagging in the middle, which quietly explains failures in filing, banking and document work. And oversight regresses precisely when it matters most: OpenAI's own chain-of-thought monitoring work shows penalising bad reasoning makes models hide intent rather than behave, monitoring an agent that knows it is watched degrades reliability, transparency frays as capability climbs, and Anthropic built SHADE-Arena specifically to measure sabotage slipping past monitors. Stronger agents, harder to watch: that is the opposite of a controlled arrival.

What would it take to settle the argument?

Watch learning efficiency, not leaderboard scores. The ARC definition gives the cleanest test: show a system acquiring genuinely novel skills as fast as a human, from similar starting knowledge. Harness tricks, retained reasoning states and tool-building are interesting engineering, but if the generality lives in the scaffolding and the prompting rather than the system, it is the developers being general, not the model.

Watch the body and the world. Human-like generality grew up inside sensorimotor experience and causal interaction. World-model architectures that learn from sensory input, simulate consequences and plan remain early, with no agreed path from pure language prediction to persistent understanding. A decade of robotics, as one sceptic frames it, would tell us more than another doubling of benchmark scores.

Watch independent replication on your own work. Contamination, harness effects and budget choices move headline numbers by tens of points. The questions that matter for a UK business or charity are local and concrete: does it handle your caseload, your spreadsheets, your beneficiaries, your edge cases, repeatedly, with an audit trail? Nobody's era announcement answers that. Your own evaluation does.

The balanced verdict is the one the evidence actually supports. Something remarkable is compounding: agents complete longer tasks, mathematics and coding curves bend hard, efficiency per action reaches human parity and beyond. That is worth taking seriously, budgeting for, and governing now. But human-like generality, the fluid, embodied, common-sense kind, faces measured barriers that more of the same may not cross. AGI-then-ASI assumes the ladder continues. The alternative is that we are climbing fast toward a ceiling that needs new science to break through. Both futures reward the same posture: pilot powerful agents where the trail is clean, govern them where stakes are high, and judge generality by learning speed in the unknown, not scores on the known.