Tech Digest – August 2, 2026
After the Breakthrough
Mathematicians Concede as Prediction Markets Price a Millennium Prize Solution by 2028
The fallout from OpenAI Astra’s ten solved open problems — covered in yesterday’s digest — is moving from shock to quantification. Number theorist Daniel Litt formally conceded, four years early, his bet that AI couldn’t produce Annals-quality number theory under $100,000 per paper, calling it “a big deal.” Claude Fable, asked to grade the haul, estimated that any single result “would plausibly anchor a medal case” on the Fields scale. Prediction markets have moved accordingly: Manifold now prices an AI-solved Millennium Prize Problem at 31% by 2027 and 52% by 2028.
The proof style itself is notable. Columbia’s Henry Yuen observed that the writeups bury their technical crux under boilerplate, introduced “as if this were the obvious thing to do.” But the style is deliberate — frontier models excel at cross-field translation into verifiable constructions, meaning, as one mathematician put it, “brace for the upcoming deluge.” New upper bounds on sphere packing density, pushed to the Cohn–Elkies threshold connected to Viazovska’s Fields Medal work, were described as near “science fiction” by Stanford number theorist Jared Duker Lichtman. And these are, another observer reminded, “sub 10T models.” The 100-trillion-parameter successors and 1,000x training compute are due by 2030.
Note: Yesterday, ten solved problems. Today, coin-flip odds on a Millennium Prize within two years. For any institution that relies on specialised human expertise as a competitive advantage, the timeline for “AI can’t do what we do” just shortened in public — with a price tag attached.
Sources: Daniel Litt (X), Nabeel Qureshi (X), Manifold Markets, Jared Duker Lichtman (X), Henry Yuen (X)
“The Old Gods Are Being Slaughtered” — Mathematics Confronts Its Workforce Crisis
The professional fallout is crystallising. Mathematician Przemek Chojecki called it “the last straw” for academic mathematics, where specialists spend months per conjecture in silos only for “an amateur” to one-shot their life’s work. Cosmologist Will Kinney framed it as spiritual crisis: “The old gods are being slaughtered by the new machine god, and it must be like watching heaven being plundered.” Fernando Borretti catalogued the standard coping arguments in “Mathematics Without Mathematicians,” refuting each in turn — math is the dynamo of science, not a chess game, ending in “marvelous devices no human understands.”
A DeepMind researcher noted the practical absurdity: a Fields-worthy human result could turn AI-trivial before the medal is even awarded. Kirwin Hampshire described a “dark night of mathematics,” arguing that discovery was how humans touched the ineffable, and questioning whether foreclosing that for future mathematicians is itself a kind of harm.
Note: The Fields Medal takes two to four years from result to award. If AI can trivialise a breakthrough in that window, the incentive structure of academic mathematics — publish, prove, get tenure — dissolves. This isn’t a productivity story. It’s a question about what human specialisation is for.
Sources: Przemek Chojecki (X), Will Kinney (X), Fernando Borretti, Timothy Nguyen / DeepMind (X), Kirwin Hampshire (Substack)
Cybersecurity Under Pressure
Apple Caps Bug Reports After AI Slop Buries a $200,000 Exploit
Apple has introduced a submission cap and a 30-day cool-off period on its vulnerability reporting portal after AI-generated reports — mixing genuine security flaws with fabricated ones — overwhelmed its human review team. The casualties are real: Italian cybersecurity startup Bynario used GPT-5.5 to discover a macOS Screen Sharing vulnerability (CVE-2026-43760) that could give attackers root-level access, but couldn’t report it because Apple had already blocked further submissions. CEO Alfredo Pesoli estimated the flaw’s black-market value at $100,000–$200,000.
The paradox cuts both ways. In the same period, AI-assisted security updates carried five times the usual volume of genuine fixes. The problem isn’t that AI finds too few real vulnerabilities — it’s that it generates so many plausible-looking false ones that the human review pipeline collapses.
Note: The bug bounty model assumed human-speed reporting. That assumption just broke. Apple’s answer was a 30-day cool-off and a submission cap. Anyone running a vulnerability disclosure programme should be asking what their version of that cap looks like — before the queue makes the decision for them.
Sources: Financial Times, The Decoder
Platform Strategy & Vendor Risk
Anthropic Signals That Model Labs Will Compete With Their Customers
Anthropic product lead Jess Yan argued publicly that achieving maximum AI performance is “impossible” without tying the application harness and the model together. Venture capitalist Chetan Puttagunta decoded the implication: Anthropic “will compete with their customers.” The message, he noted, has been “broadcast both publicly and privately to CEOs, VCs, startups.” Anthropic’s Claude Managed Agents — now in public beta — puts the company in direct competition with integration layers like AWS Bedrock and LangChain, with switching costs rising sharply once an agent runs on Anthropic’s infrastructure.
Note: For anyone evaluating AI vendors, the question is no longer just “which model?” It’s whether the model provider will eventually replace the integrator you hired to deploy it.
Sources: Chetan Puttagunta (X), Behind the Craft / Jess Yan interview
Infrastructure & Energy
20 Million AI Chips and Counting — Data Centre Power to Quadruple by 2030
The physical infrastructure behind AI’s capabilities is scaling at a pace that matches the intelligence curve. Epoch AI estimates 20.2 million AI chips (H100-equivalents) have been sold through 2025, with total computing power growing at 3.4x per year — doubling roughly every seven months. The trajectory points to 200 million H100-equivalents by 2028. Data centre power consumption is on track to quadruple by 2030, with total industry investment reaching $1 trillion by 2029.
The energy squeeze is already repricing adjacent markets. Used electric vehicles in the US have appreciated 7% this year as war-driven supply disruptions push petrol to $4.10 per gallon, making EVs more attractive even as data centres compete for the same grid capacity.
Note: A trillion dollars in infrastructure commitments and quadrupling power demand by decade’s end aren’t forecasts — they’re construction schedules. Every grid capacity negotiation and data centre siting decision is already operating in this market.
Sources: New York Times, Epoch AI, CNBC
AI in Practice
MIT Finds AI Financial Advice “Surprisingly Good” — With a Catch
Lifetime financial simulations by MIT Sloan researchers found that following LLM advice would move most people closer to optimal financial behaviour — broader equity participation, age-adjusted portfolios, and larger savings buffers. The quality depends heavily on how the question is asked: replacing casual prompts with structured ones significantly improves the output. The study, which won the 2026 Swiss Finance Institute Outstanding Paper Award, also found a troubling disparity: advice generated from prompts written by women or by less financially literate users led to measurably less wealth accumulation over a lifetime.
Note: Half of Americans already ask AI for financial advice. The quality gap between good and bad prompts maps almost exactly onto existing financial literacy inequality. Public advisory services may find their most useful role isn’t competing with AI but teaching citizens to query it properly.
Sources: MIT Sloan, SSRN Working Paper
ByteDance Ships One-Pass Audio-Video Generation With Timestamp Editing
ByteDance released Seedance 2.5, which generates up to 30 seconds of synchronised audio-video in a single pass — dialogue, sound effects, and music co-processed in one latent space rather than stitched together after the fact. The model accepts up to 50 multimodal references across images, video, and audio, supports 10+ languages, and offers timestamp-level editing for precise control. Enterprise availability is set for August 7.
Note: Institutional content production — training videos, public service announcements, multilingual communications — still costs thousands per finished minute. One-pass generation with editing precision makes sub-hundred-euro production runs a near-term reality.
Sources: ByteDance Seed, The Decoder
Yesterday this digest reported ten solved mathematical problems. Today, the mathematicians conceded, the prediction markets moved, and the grief set in. Meanwhile, Apple can’t sort real security flaws from AI-generated hallucinations, model labs are moving to compete with their own customers, and the chip infrastructure powering all of it doubles every seven months. The common thread: the speed of AI capability is now outrunning the processes — academic, commercial, institutional — that were built to evaluate it.
