When Capability Outruns Review

For nine years, the transformer has been the agreeable landlord of modern AI: expensive, occasionally forgetful, but impossible to evict because every major tenant has built around it. MIT Technology Review reports growing pressure on dense attention, whose cost rises with the amount of text and whose ability to keep track of lots of information is not exactly its party trick. Four proposed alternatives are now being pitched as routes to faster, more efficient, perhaps smarter systems. The underlying roundup is here.

At the same moment, OpenAI says its unreleased Astra model produced solutions to ten long-standing mathematics problems, some reportedly unresolved for decades. That could mean machines are becoming serious collaborators in discovery; it could also mean the public has been asked to applaud before seeing the proofs, the model, or independent validation. The supplied reporting does not identify the problems or establish peer review, which is not a minor footnote when the product being demonstrated is mathematical correctness. The Verge’s account is appropriately interested and appropriately incomplete.

The common tension is less mystical than the press release implies. Models are getting better at producing candidate answers, while the cost of knowing whether an answer is good remains stubbornly human, domain-specific, and slow. Architecture may change; verification still has to clock in on Monday.

The industry’s preferred answer to every AI constraint remains a larger building full of increasingly apologetic electricity. Nature notes that global data centres used about 485 terawatt-hours last year, roughly Germany’s electricity use, with the IEA expecting that figure to double by 2030. Amazon, Alphabet, Microsoft, Meta and Oracle are expected to spend more than US$600 billion collectively on AI infrastructure this year. Capital expenditure has become the closest thing the sector has to a theology. Read the argument for a different research posture.

The authors’ practical point is more interesting than a generic plea to consume less compute. Research labs do not necessarily need the same infrastructure as consumer services serving millions. Open-weight models on institutional servers can offer more control, and the article says a mini-fridge-sized rack can support roughly 50 simultaneous prompt users. That will not replace frontier training or every hosted service. It does, however, challenge the convenient assumption that every university workflow requires renting a sliver of someone else’s hyperscale empire.

Local deployment is not inherently virtuous; it transfers operational responsibility along with control. But it makes the trade-off visible: hardware procurement, patching, access policy, model selection, power, and accountability all return to the institution. The cloud did not eliminate these chores. It merely bundled them into a monthly invoice and called the arrangement innovation.

Local models also have a less pastoral constituency. Genians, a South Korean security firm, reported that the North Korea-linked Kimsuky actor operated local LLM environments using tools including Ollama, GPT4All and Msty, alongside phishing and Git-based command-and-control infrastructure. The reported activity included RAG-style search over documents, AI-assisted decoy material, speech-to-text tooling and experimentation with code-editing tools. Genians found no evidence that the group was training its own models. The Register has the technical reporting and its attribution caveats.

That distinction matters. The threat is not necessarily a state actor building a secret frontier lab under a mountain, which at least has the decency to fit a familiar movie plot. It is the much more ordinary prospect of capable existing components being assembled into operational workflows: better lures, faster triage of stolen material, and less data sent to external AI services. Commodity capability lowers the cost of competence for both legitimate teams and hostile ones.

For defenders, this weakens one of the lazier anti-phishing instincts: judging danger by awkward prose or ugly formatting. If the reported campaigns used polished, AI-produced decoys in some cases, then linguistic vibes are no longer a control. Detection needs to follow behavior after execution—LNK files launching PowerShell, persistence attempts, and external communications. In security, as in infrastructure, locality is a design choice with consequences, not a moral alignment.

Meanwhile, software organizations are discovering that generating a change and absorbing a change are entirely different throughput problems. A widely shared opinion essay imagines an AI-assisted team opening a 25,000-line pull request after a weekend, then argues that implementation is now cheap while architectural judgment is not. It is a hypothetical, not labor-market research, but any senior engineer who has met an inexplicably enthusiastic abstraction will recognize the species. The original argument is here.

The useful diagnosis is not that code volume is bad. Plenty of excellent changes are large; plenty of catastrophic ones fit in a three-line configuration diff. The problem is reversibility and comprehension. An LLM can add schema in minutes, but once production data lives there, removal means migration plans, failure modes, foreign keys, users, and the familiar discovery that a supposedly temporary field has become constitutional law.

This is why the AI economy may reward people who can frame a problem, constrain a system, inspect output, and refuse a bad premise before it enters main. It may also make institutions that have neglected review culture fail faster, because an agent can now manufacture complexity at a rate previously reserved for mergers. The frontier model, the local rack, the mathematical conjecture, and the phishing kit all point to the same unsentimental conclusion: capability is proliferating faster than judgment. The scarce resource is not tokens. It is responsible refusal.