Tech Digest – May 15, 2026
The Optimization Loop Closes
AI Systems Now Optimize Their Own Training — and Beat Humans Doing It
Poetiq released its “Meta-System” on LiveCodeBench Pro, where it built its own evaluation harnesses and hit a new state-of-the-art 93.9% atop GPT-5.5 — with no fine-tuning, no special access, and no hand-built pipelines. Separately, Prime Intellect handed Codex and Claude Code its idle compute to attack the NanoGPT Speedrun optimizer track. After approximately 14,000 H200 GPU-hours, both agents beat the human baseline, with Opus 4.7 holding the record at 2,930 steps.
Note: The loop is now literal: machines optimizing how machines are trained. When the thing being improved is the improvement process itself, forecasting timelines from outside becomes structurally unreliable. Planning horizons built on “current pace of progress” are already stale.
Sources: Poetiq, Prime Intellect
Defence Windows Shrink
Palo Alto Networks: Three-to-Five Months Before AI Makes Zero-Days Routine
Palo Alto Networks warned of a narrow three-to-five-month window for organisations to harden their systems before models like Anthropic’s Mythos and OpenAI’s GPT-5.5-Cyber make exploiting unknown vulnerabilities routine. The company framed this not as a distant threat but as a present capability reaching deployment-grade reliability within months.
Note: Five months is one budget cycle. Any institution that hasn’t started its security hardening by the time this capability ships will be patching in production against adversaries using the same tools. The remediation window is the window before the window closes.
Sources: CNBC
arXiv Imposes One-Year Bans for AI-Generated Plagiarism and Errors
The arXiv preprint server will now hand one-year submission bans to any author caught publishing AI-generated plagiarism, fabricated references, or errors they plainly never reviewed. Flagged indicators include hallucinated citations, leftover LLM meta-comments, and illustrative data never replaced with real results. After a ban, authors must first have subsequent work accepted at a peer-reviewed venue before returning to arXiv.
Note: This is the first major research infrastructure to impose meaningful penalties for AI quality failures — not for using AI, but for failing to supervise it. The template transfers directly: any institution accepting AI-assisted deliverables (consultancy reports, grant applications, policy drafts) needs an equivalent accountability line.
Sources: Thomas Dietterich (arXiv moderator), 404 Media
Compute Supply in Motion
Chip Monopoly Fractures: Cerebras IPOs at $95 Billion, Zyphra Trains on AMD, Apple Moves to Intel
Cerebras closed its first day of Nasdaq trading up 68%, reaching a $95 billion market cap in one of the largest US tech IPOs in years — establishing a publicly traded alternative to Nvidia for AI accelerators. Meanwhile, Zyphra released ZAYA1-8B, the first mixture-of-experts model trained end-to-end on an AMD Instinct stack, proving frontier training no longer requires Nvidia hardware.
The supply chain is shifting at the legacy end too. Apple reportedly began producing older iPhone, iPad, and Mac processors at Intel on its 18A-P node — a vote of confidence in Intel’s foundry ambitions. And Washington cleared approximately ten Chinese firms to purchase Nvidia’s H200, its second-most-powerful AI chip, in what appears to be a targeted de-escalation of export controls.
Note: For any institution planning compute procurement over a 3-5 year horizon, the Nvidia-only assumption is dissolving. Cerebras public, AMD viable for training, Intel back in foundry — the optionality that didn’t exist 18 months ago now does. Procurement frameworks written around a single vendor need updating.
Sources: CNBC (Cerebras), Zyphra, Ming-Chi Kuo, CNBC (H200 exports)
Seven in Ten Americans Would Oppose a Data Centre Near Their Home
A Gallup survey published in the Washington Post found that 70% of Americans would oppose a data centre being built in their community. Opposition is so strong that respondents said they would rather live beside a nuclear power plant than the warehouses powering the AI boom. The finding lands as hyperscalers are planning hundreds of new facilities across the US and Europe.
Note: The compute has to go somewhere. If public resistance to data centres mirrors what happened to wind farms a decade ago, siting will become the binding constraint on AI infrastructure — not chips, not power, not capital. EU municipalities reviewing data centre applications should expect organised opposition regardless of economic benefits offered.
Sources: Washington Post / Gallup
Capital, Tax & Geopolitics
Anthropic Closes $30 Billion Raise at $900 Billion Valuation, Pledges $200 Million for Public Goods
Anthropic agreed to terms on a $30 billion raise valuing the company at $900 billion — a figure that would have ranked it among the world’s ten most valuable companies a year ago. Separately, Anthropic and the Gates Foundation announced a $200 million, four-year partnership to deploy AI for health and education in underserved regions.
Note: A $900 billion valuation for a company that didn’t exist five years ago is its own signal about the speed of structural change. The Gates partnership is notable for what it implies: AI public goods are now being funded at a scale that used to require multilateral development banks. Institutions waiting for public funding rounds to experiment with AI in health or education may find the private sector got there first.
Sources: Financial Times, Reuters
California Proposes 7.25% Tax on Cloud Software
Governor Gavin Newsom pitched a 7.25% tax on cloud software as part of a broader revenue package. The proposal would treat SaaS, IaaS, and PaaS subscriptions as taxable goods — a significant shift from the current treatment of software services in most US states. The move is designed to capture revenue from the digital economy at a scale that offsets declining traditional retail tax bases.
Note: The EU already taxes cloud services via VAT, but California’s move signals a global direction: as more government and institutional operations run on cloud platforms, those platforms become the tax base. Any institution budgeting for cloud migration should model for rising effective costs, not just subscription price increases.
Sources: Bloomberg
Anthropic Paper Maps Two Versions of 2028 — One Where Democracies Lead, One Where They Don’t
Anthropic published a research paper on US-China AI competition sketching two scenarios for 2028: one in which democracies maintain their compute advantage and shape global AI governance norms, and another in which export-control loopholes allow authoritarian regimes to close the gap and set the rules instead. The paper frames the next two years as a decisive window for technology policy.
Note: The paper’s framing aligns with the EU’s own Open Strategic Autonomy agenda — but with a harder edge. If the decisive period is 2026-2028, the EU’s Digital Decade 2030 timeline assumes a stability that may not exist. Institutions building five-year digital strategies should stress-test them against a world where the rules are set by 2028, not 2030.
Sources: Anthropic Research
Today’s through-line is time compression. The machines are optimizing themselves faster than institutions can assess them. The cybersecurity window is months, not years. The chip monopoly that seemed permanent 18 months ago is already fragmenting. A company worth $900 billion didn’t exist five years ago. And the research paper warning about 2028 was published in 2026 — which means the planning window it describes is already half spent. The recurring institutional question is no longer “what’s coming?” but “how much time do we actually have to respond?”
