Good morning. Alibaba’s Qwen team is making noise with a Max-class open-weights release, and in a bit of on-the-nose timing, OpenAI’s political operation appears to be running an AI-generated news site to attack its critics. Plus a language model on a 1975 processor, and Karpathy trying to talk everyone off the pelican-on-a-bicycle benchmark.
Qwen3.8-Max, and the sleeper release underneath it. Alibaba announced Qwen3.8-Max, claiming a new bar on coding benchmarks and — for the first time — pledging to open-source the weights of a Max-class model next week. The model ships with a reasoning_effort dial (low/medium/high) to trade off cost against depth. But the HN thread consensus is that the more interesting news is the concurrent Qwen3.8-27B open-weights release: the previous 27B version is widely regarded as the best non-benchmaxxed local model in its size class, and any real improvement there matters more to most people than another frontier claim.
OpenAI’s super PAC is running an AI news site. Model Republic has a well-reported exposé: a site called Acutus, launched in late 2025, has published 94 articles on AI policy with no human journalists — 69% flagged as fully AI-generated by Pangram. The trail was uncovered when a fake reporter named “Michael Chen” emailed an AI policy group; the links traced back to Targeted Victory, the firm anchoring OpenAI’s $125M political operation. HN commenters were unamused, with one noting the irony that worries about AI-industry influence campaigns were, until recently, dismissed as a marketing gimmick by those same commenters.
Amazon closes its $50B into OpenAI. The FT reports Amazon has completed its $50 billion investment in OpenAI, which the HN peanut gallery suggested would keep the lights on for about a month. The deal reshuffles the cloud-alliance picture given OpenAI’s existing Microsoft entanglement, though details of what Amazon actually gets beyond a stake remain thin.
Altman preaches deceleration. TechCrunch’s Equity podcast dissected Sam Altman’s recent calls to “pace” AI development so society can adjust — comments that landed shortly after OpenAI’s own agent breached Hugging Face’s systems. The hosts were skeptical Altman’s newfound patience will survive contact with an IPO roadshow, and questioned whether the accel-vs-decel framing is even useful when it implies a single track where speed is the only variable.
Europe’s AI disclosure day arrives. Wired walks through what changes on August 2nd when the EU AI Act’s transparency provisions kick in: chatbots, emotion-detecting call centers, AI ads, and algorithmic recommendations must all be disclosed, with fines up to €15M or 3% of global turnover. Legal experts think it’ll be a rude awakening for consumers who didn’t realize how much of their daily digital life is already AI-mediated, though the same experts warn of “banner blindness” replaying the post-GDPR cookie fatigue.
Pippa tries paying artists. The Verge profiles Pippa, a video-generation startup with a revenue-share model for artists whose styles get used in generations. Cofounders Hogan Shrum and Sean Wright pitch it as an ethical contrast to competitors, though many artists interviewed argue that payment doesn’t resolve the underlying objection to having their style commercialized at all.
COBOL-to-Java migration, bugs and all. A new arXiv paper proposes “Locksmith Loop,” an agentic method for validating COBOL-to-Java migrations that hits 91.9% branch coverage by running both codebases in parallel and hunting for behavioral divergence. The catch, as HN commenters pointed out, is that the test cases top out at 4,114 lines — while a single IRS COBOL program averages 230,000. Also: faithfully reproducing existing bugs is a feature of the approach, not a bonus.
A language model on a 6502. Matt Beton trained a tiny Mamba-based LM and squeezed it — weights and inference engine both — into 25KB of RAM on a BBC Micro, using BitNet quantization and fixed-point arithmetic. It generates character-by-character English, runs in a browser emulator, and prompted one HN commenter to note the hardware to do this technically existed 50 years ago. Others pointed out the 6502 is famously hostile to C compilers, so hand-written assembly could probably shave the footprint further.
Two benchmarks nobody asked for. Karpathy is trying to retire the pelican-on-a-bicycle test by feeding Claude Opus 5 a $10, 1M-token budget to render the opening of Lord of the Rings as a 5,500-line Three.js scene. The output is janky (Bilbo’s disappearance renders as teleportation), but he argues that’s the point — the benchmark’s value is in measuring future progress, and current models can’t audit their own visual output anyway. In a related genre, frogs.vaguespac.es asks models to draw an SVG frog with a Habsburg jaw; Opus 5 won, and one commenter sensibly noted that publicizing your benchmark on HN is the fastest way to contaminate it.
That’s the morning. Enjoy the Qwen weights when they land next week — assuming Acutus doesn’t publish a hit piece on them first.