The Computational Ceiling on LLMs
- https://www.youtube.com/watch?v=AIYQp1n51ZI
- Original title: They Lied to You About AI (This Study Proves It)
Vishal Sikka (former Infosys CEO, Oracle/BMW board member, Stanford PhD who studied under John McCarthy) and his son published a paper arguing, using computational complexity theory settled since the 1960s, that current LLMs and AI agents have a fixed, provable computational ceiling per response — meaning many "AGI is near" claims are architecturally impossible, not just unlikely, and hallucination on complex tasks is mathematically unavoidable rather than a training defect.
The core argument
Every token an LLM generates gets the same fixed compute budget via the self-attention mechanism — a simple "hello" and a hard physics problem receive identical processing. There is no "think harder" option built into the architecture. The paper uses the traveling salesman problem (20 cities, over two quintillion route combinations) as an example of a task that physically cannot be solved by one forward pass — the model can only guess and pattern-match a plausible-looking answer.
Doing vs. verifying
A key distinction: even if a model can't solve a problem outright, one might expect it to at least verify a proposed answer. But verification often requires as much computation as solving the problem in the first place. Every polished AI demo you've seen was implicitly designed to stay under this complexity ceiling; real-world business tasks routinely blow past it. The paper invokes the time hierarchy theorem: some problems provably require a minimum number of computational steps that cannot be shortcut, and if a task needs more steps than the model can perform, hallucination is the only possible output — not a bug, and not something better training fixes.
Why "agentic" chain-of-thought doesn't escape the ceiling
Giving a model more steps (agentic loops, chain-of-thought, tools like Manus/OpenClaw) is compared to giving a writer more sheets of paper: each sheet is still the same size, so the writer isn't smarter, just given more room to ramble. Because the model can't mathematically verify its own reasoning, errors compound across steps — the "book a flight perfectly, then get stuck in an infinite loop changing a seat" pattern. In the agentic era, hallucination becomes a cumulative mathematical certainty over long chains, not a training bug. Handing sub-tasks off to classical tools/calculators (the LLM as "orchestrator") helps, but the LLM still must verify the tool's output, and verification carries the same complexity limits.
Bigger context windows (e.g. Gemini 3 Pro's million-token context) solve information access, not computational steps per word — a bigger filing cabinet doesn't add brainpower to process what's inside.
Evidence: Vending Bench 2
Andon Labs' Vending Bench 2 (2026 gold-standard agent benchmark) gives frontier models $500 and a year to run a simulated vending machine business. Claude Opus 4.6 currently leads AI models with about $8,000 profit — but the human baseline for the same simulation is $63,000. Frontier models hit under 15% of human performance because they lose coherence over long time horizons; observed failures include agents giving away inventory to social engineering or contacting the FBI over a $2 bank fee. This is presented as the time hierarchy theorem playing out in practice: as task chains lengthen, an agent's ability to verify its own logic collapses, and without a human resetting the error rate, the chain eventually breaks regardless of model quality.
What this means practically
The paper does not claim AI is useless — current models are described as exceptional for tasks that stay under the ceiling: drafting, summarizing, reformatting data, research and comparison. The gap is between that reality and marketing promises of fully autonomous business-running agents. Recommended practices: be specific about tasks ("draft an email in my tone" works; "automate this workflow" tends to fail), build in human verification as a structural requirement rather than an optional nicety, and use AI for pattern recognition rather than logic-heavy computation.
The video closes on a speculative point: if AGI/the singularity were truly imminent, senior engineers wouldn't be leaving frontier labs to start independent, riskier startups — they'd stay for the payoff. The suggested read is that insiders see the next models as better but not qualitatively different (as GPT-5 was better than GPT-4 without being a different kind of thing), and that the real opportunity is using AI well within its ceiling rather than waiting for AGI.