Good morning. Today’s thread is about perception gaps — between what AI insiders think the technology can do and what everyone else sees, between what agents produce and what actually works, and between what companies say and what they’re doing in forests in Finland. Karpathy kicked off a debate about whether the top 0.2% of users even agree with each other, DHH showed what happens when you don’t read the code, and a Berkeley study claims 10 minutes with ChatGPT is enough to dent your willingness to think hard.
Karpathy on the widening capability gap. Andrej Karpathy argued on X that LLM understanding is stratified into tiers — roughly 6 billion people with no exposure, 20 million professionals seeing agents complete weeks of work, and 5,000 insiders watching swarms do frontier math and cyberattacks. HN was skeptical of the middle tier’s supposed consensus, with one commenter noting there’s wide disagreement even among heavy users about whether agents can really handle large codebases without creating maintenance debt. Another pushed back that if the gap were truly monumental, Karpathy could just point to the monument instead of waving at “vertical professionals.”
DHH: you still have to think. DHH’s post on cross-language rewrites of Campfire Once into Rust, Elixir, and Go found that agents quietly dropped Redis, broke backwards compatibility, and made inconsistent design choices across versions — rendering any performance comparison between languages meaningless. The irony noted by commenters: “you have to think” is now a mildly controversial position. One put it well — workslop is cheap to produce and expensive to debunk, which is why it keeps winning.
Ten minutes of AI, measurably less grit. A UC Berkeley study with 1,222 participants found that people who used ChatGPT on math problems gave up faster once the tool was taken away, compared to controls who never had it. HN was skeptical of the methodology — several commenters wanted a non-AI tool control group (calculators, textbooks, human tutors) to see whether this is really about AI or just about any assistance. One analogy that stuck: give Group A a hammer and a banana, give Group B just a banana, then take the hammers away halfway through and measure frustration.
Terence Tao on Math 2.0. Tao’s Caltech slides frame open problems as “lighthouses not destinations” and argue AI’s real contribution to math should be new principles, not just more proofs. His cancer-medicine analogy — that you’d want at least one human to understand a drug’s mechanism before taking it — drew sharp pushback from a working drug-discovery researcher who pointed out that medicine runs almost entirely on half-understood empirical methods, including most of general anesthesia. Another commenter read the whole talk as a grieving process: juniors used to earn Fields Medals by solving proofs and get invited to collaborate, and that path is narrowing.
500 billion tokens to decompile Modern Warfare 2. A developer spent three months and ~500B tokens orchestrating Claude and Codex agents via Discord and GitHub to decompile what’s widely assumed to be Call of Duty: MW2, hitting ~80% before a corporate takedown ended the project. The HN consensus was that targeting byte-matching assembly output (rather than functional equivalence) is what blew the budget by orders of magnitude — one commenter estimated the same result could’ve cost 1/100th as much. The more interesting question raised in comments: if AI makes decompilation trivial, does all proprietary software eventually move server-side?
Sakana’s recursive self-improvement claim. Sakana AI and UC Berkeley introduced MASS, a system that optimizes multi-agent workflows and distills their outputs into self-supervision signals, pitched as a step toward recursive self-improvement for frontier problems where humans can’t evaluate. HN wasn’t buying the framing — top commenters alleged the released code doesn’t match the paper, and that what’s actually being optimized is multi-agent communication, a narrow dimension dressed up in RSI language. One caveman-voice comment summed up the skepticism: “Caveman make ring of rocks. Now rock think self correct. But mammoth not cooked somehow?”
Google clear-cut Finnish forest without permits. Finland’s licensing authority has ordered Google to halt construction on two data centers after its subsidiary cleared ~530 hectares in Muhos and Kajaani without finishing environmental impact assessments. Google’s response — that it had “fallen short of its own high standards” — drew the obvious rejoinder on HN: it wasn’t a standards failure, it was a legal violation. Related: Amazon and Microsoft are dropping NDAs in municipal data-center negotiations, which TechCrunch frames as recognition that the industry’s communication problem may be bigger than its project problem.
Silicon Valley adopts “super intelligence.” The Verge reports that Musk, Benioff, Huang, and now Nadella have all picked up Trump’s preferred “super intelligence” terminology, following an executive order mandating it across federal agencies and Trump’s declaration that users of the term “AI” are “THE ENEMY.” We noticed Nadella’s usage yesterday in his emergency-brake post; the pattern is now broad enough to be a pattern.
Voice AI isn’t there yet. Execs at PolyAI and Otter told TechCrunch that even with full-duplex models, voice AI still trips on reasoning latency, keyword capture, speaker ID, and emotional range — all the things that keep callers reaching for the human-escalation button. The implicit bet: the ChatGPT moment arrives when users stop asking for a human.
That’s the roundup. If today’s theme has a through-line, it’s that the gap between “the demo works” and “the thing works” is where all the actual arguments are happening right now.