Good morning. Anthropic is everywhere this morning — filing for an IPO, publishing a recursive self-improvement essay, and shipping an open-source security harness — which together read like a coordinated pre-IPO narrative. Elsewhere, a Berkeley failure-rate spike is reopening the debate over what LLMs are doing to students, and a quietly important arXiv paper asks whether transformers have been carrying unnecessary baggage for years.
Anthropic’s IPO pitch comes into focus. Daniela Amodei told TechCrunch that public markets are the natural home for the capital frontier labs need, citing $47 billion in annualized revenue in May (up from roughly $9 billion at the end of 2025) and an oversubscribed $65 billion raise at a $965 billion valuation. She brushed off enterprise ROI concerns as a learning-curve issue and defended the choice to rent compute rather than build, including a $1.25 billion/month deal with xAI that’s striking given the rivalry. As we noted Monday, the S-1 is already filed — now the roadshow narrative is taking shape.
Anthropic argues AI is already accelerating AI. A new Anthropic essay on recursive self-improvement claims engineers there now ship 8x more code per day and that AI task-completion horizons are doubling every four months. HN commenters weren’t buying it: lines-of-code is a notoriously bad productivity metric, the timing reads as IPO promotion, and several found the “if we don’t, someone else will” safety framing morally hollow. One commenter put it bluntly: regardless of whether Anthropic can build self-improving AI, the question is whether they should be allowed to without strict supervision.
And Anthropic open-sources a security harness. The company released a reference implementation for autonomous vulnerability discovery and patching with Claude, focused on C/C++ memory bugs via Docker and ASAN. The framing on HN is that it’s a “shop jig” — a starter kit pointing users toward the hosted Claude Security product, made obvious by the repo’s “not maintained, not accepting contributions” notice. One commenter noted the cost-of-building-your-own equation has shifted enough in two years that the harness itself is less interesting than the token economics behind it.
A QKV ablation paper questions transformer orthodoxy. An ICML 2026 paper, Do transformers need three projections?, finds that sharing keys and values (Q-K=V) cuts the KV cache by 50% at a 3.1% perplexity cost, and stacks with grouped query attention for up to 96.9% cache reduction. The catch, raised repeatedly on HN: the 1.2B model was trained on only 10B tokens, well below chinchilla-optimal, so whether the result holds at modern scales is unknown. One commenter suspects the exact attention mechanism matters less than how efficiently it runs on GPUs — the same story as the ReLU/GELU debate.
Berkeley’s CS failure rates jump. The Daily Cal reports that CS 10 and CS 61A saw spring 2026 failure rates of 35.3% and 10.6%, far above the department’s 7% guideline, with professor Dan Garcia blaming LLM-enabled cheating and homework crutch use leaving students unprepared for in-person exams. Nearly 30 CS 10 students were caught cheating. HN commenters flagged a buried lede: 1,300+ UC faculty are also petitioning to restore SAT/ACT requirements, citing math preparedness collapse since the tests were dropped. One CS professor described catching ChatGPT’d project submissions just by asking students to explain their architecture choices.
Quick hits. TSMC’s C.C. Wei told The Verge the company “can only support so much” AI chip demand, with US capacity expansion described as taking a “very long time” despite the $165 billion Arizona buildout. Sam Altman, Dario Amodei, Mustafa Suleyman, and Demis Hassabis signed an open letter — a follow-up to yesterday’s biosecurity push — asking Congress to mandate that synthetic DNA/RNA manufacturers screen orders for dangerous pathogen sequences. And MIT Tech Review reports that AI content in federal filings jumped from 1% to 18% over the past few years, with hallucinated citations now a routine problem and pro se litigation rising from 11% to 16.8% of cases — though using AI doesn’t appear to improve plaintiffs’ win rates.
That’s the morning. Watch whether Anthropic’s IPO filing dictates the tempo of these self-promotional safety essays — the pattern is getting hard to miss.