Skip to content

Latest commit

 

History

History
323 lines (213 loc) · 21 KB

File metadata and controls

323 lines (213 loc) · 21 KB

Humanity's Last Exam, CAIS, and the moving goalpost

Why this page exists: DiffusionGemma's standout HLE score (~12%, beating its AR twin on no-tools) is the main reason diffusion-llm-mcp exists as a separate fleet repo. You cannot understand that routing decision without understanding what HLE is, who built it, and what incentives sit behind it.

Primary sources: lastexam.ai · arXiv:2501.14249 (Nature, 2026) · CAIS · DiffusionGemma model card HLE scores


1. What HLE is

Humanity's Last Exam (HLE) is a multi-modal benchmark of ~2,500 expert-level academic questions across 100+ subjects — mathematics, physics, chemistry, biology, medicine, computer science, humanities, social sciences.

Property Detail
Authors Center for AI Safety (CAIS) + Scale AI + global contributor consortium
Curation ~1,000 subject-matter experts, 500+ institutions, 50 countries
Filter Questions had to defeat frontier LLMs at submission time to advance
Answer format ~80% exact-match, ~20% multi-choice (5+ options), ~10% multimodal
Anti-lookup Designed to resist quick web retrieval; unambiguous verifiable answers
Branding "The final closed-ended academic benchmark of its kind"
Prize $500K pool for contributors

It exists because MMLU saturated. Models hit 90%+ on standard evals; the field could no longer measure frontier capability on the leaderboards everyone watched. HLE was built to be the new wall.


2. Who CAIS is

The Center for AI Safety (CAIS), led by Dan Hendrycks, is an AI safety research organization — not an "anti-AI" lobby in the sense of opposing all machine learning, but an institution whose reputation and funding are tied to AI risk, capability surprise, and the gap between current systems and safe superintelligence.

Hendrycks also created MATH (2021), MMLU, and other benchmarks that became the field's scoreboard — then watched models saturate them.

His HLE launch quote is the mission statement in plain English:

"When I released the MATH benchmark in 2021, the best model scored less than 10%; few predicted that scores higher than 90% would be achieved just three years later. Right now, Humanity's Last Exam shows there are still expert questions models cannot answer. We will see how long that lasts."

Read that last sentence carefully. It is not neutral science. It is a timed bet against closure.


3. "Anti-AI" — what that actually means here

Nobody at CAIS is burning GPUs. "Anti-AI" in this context means anti-satisfaction — a structural skepticism toward the claim that current LLM trajectories constitute "real" reasoning or near-AGI.

Surface position Underlying claim
"HLE measures expert frontier" Saturated benchmarks lied; this one won't (yet)
"Models still fail expert questions" Therefore current AI is not there
"We will see how long that lasts" The gap is temporary; our job is to measure its retreat
Safety framing Capability gains without alignment = existential risk

This is the benchmark priesthood: credentialed institutions that build the walls models must fail, publish the failure as news, and rebuild the wall when models climb it.

CAIS is the Hendrycks branch. François Chollet (ARC-AGI) is the parallel branch in the Python-prof league. Emily Bender (stochastic parrots) is the linguistic branch. Ed Zitron is the newsletter branch. Different vocabularies; same geometry.


4. Human essentialism — the philosophy hiding in the rubric

Human essentialism (in this doc's sense): the implicit belief that some cognitive capacity is uniquely, essentially human, and that AI systems demonstrating surface competence on benchmarks have not captured it.

HLE encodes this assumption in its design:

  1. Expert humans wrote the questions — authority flows from credentialed humans, not from task structure alone.
  2. LLM failure at curation time was a selection criterion — the benchmark is defined by what machines could not do when it was built.
  3. "Humanity's Last Exam" — the name itself claims a finality, a category ("humanity") that stands apart from machine performance.
  4. Low model scores are interpreted as "significant gap" — not as "early technology on a hard test," but as evidence about the nature of machine cognition.

The God-of-the-Gaps structure, adapted to AI:

Theology AI benchmarking
God lives in unexplained phenomena "Real intelligence" lives in unexplained benchmark failures
Science explains X Models score 40% on X
"God is in Y now" "That doesn't count — saturated / gamed / not real reasoning"
Repeat Ship benchmark Y

Human essentialism does not require hatred of machines. It requires that the remaining gap always be interpreted as proof of a categorical difference — not as a temporary engineering deficit on a moving frontier.

The stochastic parrot thesis is the folk version: machines cannot genuinely recombine; they only retrieve. Shakespeare also retrieved — every play before him — and recombined. The essentialist move is to call human recombination "understanding" and machine recombination "pattern matching," regardless of output equivalence on hard tests.


5. Goalpost moving — documented pattern

HLE is not the first wall. It is the current wall.

Benchmark Era Design intent What happened
MATH 2021 Hard math for models >90% in ~3 years
MMLU 2021+ Broad knowledge Saturated; models cluster 85–90%+
ARC-AGI-1 2019 Fluid intelligence; novel puzzles o3 → 75–85%
ARC-AGI-2 2025 Adversarially harder; frontier models → single digits Active
ARC-AGI-3 2026 Interactive agentic; AI <1%, humans 100% Active
HLE 2025–26 Expert academic; stump frontier models Climbing — GPT-5 class 40–60%, open models low double digits
HLE-2 (projected) When HLE saturates Inevitable under current incentives

Goalpost moving is not a moral failure. It is a measurement strategy when the field outruns evals. The problem is epistemic, not ethical: treating any particular gap as permanent evidence about paradigm limits, or as proof that "real intelligence" remains human-only.

Chollet designed ARC-AGI-1 as a binary fluid-intelligence test: near-zero or human-level, nothing between. o3 landed in between at 75–85%. Response: ARC-AGI-2. The textbook author who insists LLMs cannot reason keeps issuing harder exams to the thing he says cannot pass.

Hendrycks built MATH, watched it die, built HLE, and bet publicly on how long HLE will live. Time for another goalpost shift is not cynicism — it is the stated product roadmap.


6. HLE scores that matter for this repo

From DiffusionGemma model card (Entropy-Bounded sampler):

Benchmark DiffusionGemma 26B A4B Gemma 4 26B A4B
MMLU Pro 77.6% 82.6%
HLE no tools 11.0% 8.7%
HLE with search 11.9% 17.2%
GPQA Diamond 73.2% 82.3%

Reading this correctly:

  • On saturated evals, the speed-optimized diffusion model loses to its AR sibling — expected.
  • On HLE no-tools — the exam CAIS built to resist AI — diffusion wins by 2.3 points.
  • On HLE with search, AR wins — tool loops and agentic retrieval favor autoregressive serving stacks.

So the anti-saturation benchmark is exactly where the "worse" paradigm over-indexes. The wall CAIS built is already cracking from the direction they were not optimizing for — a v1 diffusion model optimized for throughput, not for leaderboard farming.


7. Why the fleet routes on HLE, not MMLU

If you optimize for… You build… You miss…
MMLU / saturated leaderboards Bigger AR, better fine-tunes Paradigm hops that ace anti-saturation evals
HLE / frontier reasoning Models that revise under constraint Streaming chat UX
Catch-them-all Separate repos per paradigm Single-point-of-failure hype pivots

diffusion-llm-mcp exists because routing on MMLU alone would have written off DiffusionGemma — and HLE says that was wrong for batch/frontier-shaped work.

This does not mean dLLM is AGI or that CAIS is wrong about safety. It means:

  1. The gap's address is negotiable — which benchmark you watch determines which paradigm looks "ahead."
  2. Benchmark priests have incentives for the gap to survive — not necessarily dishonest, but structurally loaded.
  3. Capability surprises arrive from unanticipated architectures — slightly Yudkowsky-shaped; see NARRATIVE.md.

8. Fair accounting — what CAIS gets right

Do not strawman this. CAIS and HLE contributors did real work:

  • Saturation is real. MMLU at 90%+ was measuring less than it used to.
  • Expert curation matters. HLE questions are hard, verified, and cross-disciplinary.
  • Safety research needs hard evals. Capability measurement and risk assessment are linked.
  • The gap is still enormous. 11% vs near-100% human expert performance is not AGI.
  • HLE-Verified and critique exist. The benchmark is contested; good — science should be.

The fleet critique is narrower: do not let the priesthood's choice of gap dictate your inference architecture. Watch which gaps close, from which direction.


9. The stadium perimeter — how far can goalposts move?

Think of the whole eval landscape as a stadium. The field is everything we can meaningfully test: closed-ended expert questions, puzzles, agentic environments, economic tasks, physical embodiment. Goalposts are the painted lines the priesthood moves whenever models start scoring.

You can move them a long time. You cannot move them forever.

Walls you eventually hit

Perimeter wall What stops the move
Human calibration ceiling ARC-AGI-3 already pins this: humans must stay at ~100% or the benchmark is useless. Harder walls that humans cannot clear destroy the "Humanity" brand.
Verifiability HLE requires unambiguous, gradable answers. Push into open-ended judgment, aesthetics, politics — automation breaks; the priesthood loses its instrument.
Expert supply Each wall needs credentialed question-writers. Finite pool, finite time, $500K prize pools do not scale to quarterly goalpost releases.
Build latency o3 cracked ARC-AGI-1 faster than ARC-AGI-2 shipped. If capability outruns wall construction, the game looks rigged — not sacred.
"That doesn't count" Brute force, memorization, scaling, diffusion, test contamination — infinite regress of disqualifiers. The epistemic perimeter: when the only winning move is redefining victory, you left science for theology.
Deployment reality The outer stadium is not HLE. It is what people pay for. Code shipped, drugs designed, taxes filed, robots that do not fall over. A priesthood that only tends benchmark gaps while the economy reorganizes around models has moved the posts off the field entirely.
Degenerate privacy Held-out private sets, secret leaderboards, "trust our number." Moves the post into fog — still essentialist, no longer falsifiable in public.

The sequence so far (same stadium, new paint)

MMLU saturated     →  paint HLE on the field
HLE climbing       →  paint HLE-2 (projected)
ARC-AGI-1 fell     →  paint ARC-AGI-2, then 3 (interactive)
Still losing?      →  agentic, embodied, adversarial, private

Each move is legitimate measurement innovation and goalpost relocation at once. Both can be true.

When you hit the perimeter

You do not hit one brick wall. You hit a squeeze from four sides:

  1. Top: human calibration — cannot make tests harder than humans without abandoning the comparison.
  2. Sides: verifiability and expert supply — cannot grade or staff arbitrary hardness.
  3. Bottom: deployment — the real world grades you whether or not CAIS publishes.
  4. Back: disqualification regress — if every closure is "not real intelligence," the claim becomes unfalsifiable.

At that point the priesthood does not admit defeat. It changes stadiums — from capability measurement to alignment impossibility, from public benchmarks to policy, from "models fail HLE" to "even if they pass, we get one shot." Different game, same essentialism.


10. God of the Gaps redux — yes, with upgrades

Yes. Structurally it is the same move theology made for centuries:

Classic God of the Gaps AI essentialism (HLE edition)
God explains what science cannot "Real reasoning" explains what benchmarks cannot
Science closes a gap Models score 40% on HLE
God retreats to a new mystery "That doesn't count — saturated, gamed, not fluid intelligence"
Mystery = evidence of divinity Low score = evidence of human cognitive uniqueness
Permanent soul Permanent humanity

The redux part: we have seen this movie. The gaps closed before. MATH was a gap. It is a worksheet now.

The upgrade part: modern benchmark priests are more self-aware than old God-of-the-Gaps apologists. Hendrycks literally said we will see how long that lasts. He knows the gap is temporary. The essentialist interpretation of the gap — that what remains is proof of categorical human superiority, not a moving engineering frontier — still ships in the branding, the safety narrative, and the discourse around every new wall.

So: God of the Gaps redux, but with a countdown timer. The theology admits the miracles might stop holding. The essentialism often does not.

What is not God of the Gaps

Measuring real capability gaps is not theology. 11% vs human expert performance is a real cliff. The mistake is inferring permanent category difference from a temporary scoreboard — or letting the priesthood's choice of scoreboard dictate which architectures you build.

DiffusionGemma on HLE is the counter-sermon: the gap closed from the wrong aisle (speed-optimized diffusion v1), on the right exam (anti-saturation). The stadium is bigger than the painted lines.


11. No True Scotsman — the soft weapon before the stick

Before you hit the stadium perimeter, before the Butlerian trajectory, the priesthood has a softer fallacy: No True Scotsman.

"No true Scotsman would put sugar on his porridge."
"Angus puts sugar on his porridge."
"Well, no true Scotsman would."

In AI essentialism:

Claim Counterexample Scotsman move
"LLMs cannot reason" o3 on ARC-AGI, HLE climbing "That's not real reasoning — brute force / scaling"
"Diffusion models are worse" dLLM beats AR on HLE no-tools "That's not the right benchmark / saturated soon"
"AI only retrieves" Novel synthesis at scale "Not true recombination — stochastic parrot"
"Fluid intelligence requires X" Model passes v1, fails v2 "No true fluid intelligence without Y"

Every counterexample triggers a criteria rewrite that excludes the positive result without updating the theory. The category ("real intelligence," "true reasoning," "genuine understanding") shrinks as models expand until it becomes unfalsifiable — which is the point, if your identity is tied to the gap.

Connection to God of the Gaps: No True Scotsman is how the gap relocates verbally when it closes empirically. God of the Gaps is where you put the mystery. No True Scotsman is how you deny the mystery was ever touched.

Our docs already catalog the disqualifiers: brute-forced, gamed, memorized, scaled, contaminated, not fluid, not agentic enough. Each is a kilt adjustment.


12. The Butlerian trajectory — argumentum ad baculum at the perimeter

When goalposts hit the stadium walls and No True Scotsman runs out of credible kilt, the discourse escalates. Not always fallacy — sometimes policy. Often both.

Butlerian (from Herbert's Dune): the Butlerian Jihad — humanity's violent rejection of "thinking machines," a cultural taboo: Thou shalt not make a machine in the likeness of a human mind. In 2026 discourse: Pope Leo XIV's AI encyclical, Butlerian Jihad manifestos, calls to ban classes of systems, criminalize training runs, delist capabilities, treaty the GPUs.

Argumentum ad baculum — argument to the stick. Accept the conclusion or face consequences: regulation, liability, compute caps, licensing, export controls, "pause" enforced by states, existential doom if you don't comply.

The escalation ladder

1. Saturated benchmark        ("MMLU solved — not impressive")
2. Anti-saturation wall       (HLE, ARC-AGI-2)
3. No True Scotsman           ("not REAL reasoning")
4. Stadium perimeter          (human ceiling, verifiability, deployment reality)
5. Butlerian trajectory     (outlaw the likeness; constrain by force)

Steps 1–4 are epistemic. Step 5 is political. The jump happens when epistemic moves stop persuading the people who build things — but essentialism still needs to win.

Soft (still on the field) Hard (Butlerian / ad baculum)
"Models fail HLE" "Certain models may not be trained"
"Not true intelligence" "Systems in likeness of human mind prohibited"
"We will see how long that lasts" "Moratorium until alignment solved"
Private held-out tests Export controls on H100s
Benchmark priesthood Licensing priesthood

Not every ad baculum is a fallacy. If the threat is the actual instrument — EU AI Act, liability law, treaty — and the argument is "comply or face legal force," that's policy, not logic. The fallacy version: accept that current AI is not real intelligence, or accept catastrophic risk — using doom as a substitute for engaging with HLE scores from diffusion models you dismissed on MMLU.

Why the fleet cares

diffusion-llm-mcp is a builder's response to steps 1–4: route on the benchmark the priests chose (HLE), not the one they said mattered until it didn't (MMLU). Catch the paradigm before the Butlerian layer decides which paradigms are legally thinkable.

If step 5 arrives in force, the question is no longer "does dLLM beat AR on HLE?" but "are we allowed to run it?" Open weights, local inference, fleet repos — Goliath in the garage — are structurally on the other side of that trajectory from closed API essentialism.

The old chestnuts stack: God of the Gaps (where the mystery lives) → No True Scotsman (deny it was touched) → goalpost paint (move the line) → stadium perimeter (run out of field) → Butlerian ad baculum (change the game to force).

We are between 3 and 4 today. Step 5 is not science fiction in regulatory discourse — it is the default ending of essentialist arguments that cannot close empirically.


13. Glossary

Term Meaning here
CAIS Center for AI Safety — Hendrycks; HLE co-author
HLE Humanity's Last Exam — anti-saturation expert benchmark
AI of the Gaps Capability attributed to whatever benchmarks still show low scores
Human essentialism Treating remaining AI failures as categorical, not temporal
Goalpost moving New benchmark when old one saturates — rational, but often presented as permanent verdict
Benchmark priesthood Institutions whose influence scales with the gap remaining
Stadium perimeter Limits on how far harder benchmarks can move before the game breaks
God of the Gaps redux Locating essential humanity in whatever AI still fails — with a self-aware countdown
No True Scotsman Redefining "real intelligence" to exclude each counterexample
Butlerian trajectory Dune-style rejection of thinking machines → bans, taboo, regulatory force
Argumentum ad baculum Appeal to stick — comply or face law, doom, or deprivation
Anti-saturation eval Test designed because prior leaderboards stopped discriminating

14. Related reading in this repo

Deep archive (MCD): mcp-central-docs/projects/diffusiongemma/ASSESSMENT.md — full Chollet essay, stochastic parrot debunk, Yudkowsky footnote.


God of the Gaps → No True Scotsman → moving posts → stadium walls → Butlerian stick. We are between the walls and the stick.