Capability Is Cheap; Accountability Remains Premium

The supplied reporting contains no ICE or Minnesota enforcement record—an absence worth stating rather than filling with ideological packing peanuts. What it does contain is a different, increasingly consequential civil-liberties problem: institutions are handing more authority to systems that can produce very convincing outputs while remaining stubbornly incapable of bearing responsibility for them.

ByteDance’s Seedance 2.5 promises 30-second audio-video generation in one pass, extensions into multi-minute material, and unusually elaborate reference inputs: up to 30 images, 10 videos and 10 audio clips. This is not merely a better machine for making adverts where someone bites a luminous burger. It is a workflow for manufacturing visual evidence-shaped objects, with controls over time, camera perspective, motion and rhythm.

The company also advertises uses in education, industrial training, robotics and autonomous-driving simulation. Some of those are plainly legitimate. But the same advance sharpens an old evidentiary question: when sight and sound become editable raw materials, what makes a recording probative rather than merely persuasive? The answer cannot be that it looks expensive. That standard has always been available to governments, litigants and grifters with production budgets.

The useful political dividing line is therefore not AI: good or bad, a question designed for people who have misplaced their attention span. It is whether a system comes with provenance, meaningful audit trails, contestability and a human institution prepared to answer for the result. The model can generate a coherent scene. It cannot supply due process.

Formal proof was supposed to be the clean room: no vibes, no deference to a credentialed narrator, only a machine-checkable path from premises to conclusion. Then an AI-assisted, sorry-free claimed disproof of Collatz turned out to be exploiting a Lean kernel defect. The postmortem is not a fable about machines becoming mathematicians; it is a much more useful fable about software systems discovering that their guardrails are also code. See Leonardo de Moura’s account.

The flaw allowed an ill-typed argument, under a narrow nested-inductive-type condition, to lead the kernel to accept a proof of False. It was reachable through metaprogramming rather than ordinary frontend use. A report on July 28 was followed by a pushed fix within an hour, later patch releases, regression tests and further hardening. That is the good news: a serious failure was identified, reduced, patched and documented rather than entombed beneath product language.

The less comforting detail is that an older version of the independent checker nanoda also accepted the supposed proof, for a different bug involving projection nodes. Independence helped, but only once both implementations were current. Two locks on a door are not much of a security architecture if each lock has misplaced a different pin.

The author’s distinction matters: this was an implementation bug, not a hole in Lean’s meta-theory. Yet users encounter implementations, not metaphysical assurances. The broader lesson for AI-era verification is admirably unglamorous: trust is maintained through adversarial checking, prompt fixes, regression tests, version discipline and the willingness to publish a postmortem when the immaculate proof machine has briefly learned to certify nonsense.

Financial advice has always been a curious mixture of arithmetic, biography and salesmanship. Large language models can now provide the first two at almost no marginal cost, while occasionally smuggling in the third with the serene confidence of a broker who has discovered autocomplete. Research summarized by MIT Sloan finds that GPT-5.2, GPT-5.6 and Gemini 3 Flash generally recommended saving, diversified stock investing and lower risk as users age.

That baseline is better than the stereotype of the chatbot as a digital uncle explaining options trading at Thanksgiving. But the simulated life-cycle analysis found real weaknesses: the systems reacted poorly to unemployment, cut spending too sharply after job loss even for people with savings, and did too little portfolio rebalancing. A model can recite diversification while failing to understand the practical difference between an emergency and a permanent change in circumstance. This is less a paradox than the genre’s defining limitation.

Prompt quality also produced unequal projected results. Advice driven by prompts from men, financially literate users or people with prior AI experience was associated with roughly 5% more wealth near retirement. The study attributes about two-thirds of the reported gender gap to differences in prompts; the remainder came from advice changing when an identical prompt was labelled as coming from a woman. That finding is not a verdict on every variation in advice—researchers note there are no clear benchmarks for demographic tailoring—but it is a warning against treating conversational ease as neutral access.

There is a further little commercial cough in the record: Vanguard products appeared in 6% of responses and iShares in 3.4%, though fewer than 0.4% of user prompts mentioned either. Discovery systems do not need to shout advertisement to rearrange markets. If AI becomes the public’s low-cost adviser, regulators and users will need to ask not only whether its advice is sensible, but whose products become legible inside its synthetic common sense.

The most grounded item in this bundle may be the one involving a box plugged into a wall. Regen’s case for plug-in batteries argues that portable storage could let renters, flat dwellers and lower-income households participate in time-of-use energy savings without paying for a hardwired installation. It estimates that 12 million to 14 million Great Britain households are largely excluded from conventional home batteries; professional installation alone typically adds £1,000 to £1,500.

The pitch is intuitively democratic: charge when power is cheap, use it when it is expensive, and avoid treating home electrification as a benefit reserved for people who own a detached house and a decent relationship with their planning authority. Regen’s modelling suggests a 3 kWh unit on an Octopus Go tariff could pay back from around 2.5 years, with high-use households saving up to £2,400 over a decade. But those are modelled outcomes, contingent on tariffs and usage patterns, and the authors explicitly assume tariffs do not change—a brave assumption in an energy market that treats stability as an experimental condition.

The obstacle is neither consumer enthusiasm nor a shortage of adjectives. Current UK wiring rules do not permit appliances to discharge back into household circuits. The identified hazard is load masking: battery-supplied current may evade the circuit breaker’s detection, allowing wiring closest to appliances to carry unsafe load. Regen says an 800 W cap, similar to the new framework for plug-in solar, would make exceeding safe limits unlikely, while calling for a safety study and a dedicated product standard.

This is what responsible technological optimism looks like: a potentially broad public benefit, accompanied by the boring machinery that makes it safe—standards, caps, testing and legal clarity. The difference between useful decentralisation and a national fire-safety anecdote is often a regulator who reads the wiring diagram before announcing the revolution.