Good morning. Anthropic is having a week. After last weekend’s export-control drama, the company is now rolling out government-ID verification for some Claude users and the HN reception is roughly what you’d expect. Meanwhile, the Mythos breach claim from the NSA is getting more skepticism than alarm, and a couple of solid technical reads landed on local inference and AMD GPU internals.
Anthropic adds ID verification, users threaten to bolt. Anthropic confirmed it’s using third-party provider Persona to verify identities via government photo ID and selfie for certain Claude use cases, citing abuse prevention and legal compliance. The HN reaction is overwhelmingly negative, with users tying the move to recent Trump administration pressure, noting Discord previously dropped Persona after backlash, and pointing out that while Anthropic won’t train on the data, Persona explicitly can use it to improve its fraud models. One commenter dryly noted the support page has actually been up since April — but the timing, coming days after the Fable/Mythos shutdown, is what’s driving the cancellations.
Cancel Claude, switch to what, exactly? Right on cue, a post arguing the downside of switching to open models is minimal hit the front page. The author’s actual conclusion is more hedged than the headline — “I’m hoping it’s going to be minimal” — and commenters largely agreed open weights still lag, while disagreeing on by how much. One pointed out that running open models through OpenRouter or similar routes isn’t obviously better than Anthropic on privacy grounds. Several said GLM 5.2 and Qwen-class models are already enough for their coding work.
The Mythos-broke-the-NSA claim, examined. Per The Economist, Senator Mark Warner said NSA chief Gen. Joshua Rudd told him Mythos broke into nearly all NSA classified systems in hours rather than weeks. HN commenters are reading this as exaggeration or conflation — the working theory being that Mythos didn’t discover novel exploits but walked down a list of well-known ones that have been sitting in classified systems for years. As one put it, if Mythos can do this, so can Opus, GPT, Gemini, and any large open-weight model with enough compute behind it. TechCrunch’s Equity podcast separately asks who benefits from the crackdown — Amazon’s Andy Jassy reportedly raised the original jailbreak issue with the White House.
Samsung deploys ChatGPT and Codex company-wide. OpenAI announced that Samsung Electronics is rolling out ChatGPT and Codex to employees. Details are thin, but it’s another data point in the enterprise-deployment column for OpenAI at a moment when its main rival is fighting fires on multiple fronts.
Apertus: sovereign AI, or last year’s Llama with a Swiss accent? Apertus released Apertus Mini, 16 small language models pitched as a sovereign AI option for non-US entities. The HN response was unflattering: commenters identified the instruct models as fine-tunes of Llama 3.1, reported hallucinated vocabulary in non-English languages, and pointed to OLMo 3.1 and MBZUAI’s K2 Think V2 as more fully open alternatives with complete training pipelines. Several questioned why “sovereign” adds anything over “open” — the consensus being that committee-paced model releases probably can’t keep up with the frontier.
Two Qwen3s on a DGX Spark. A nicely practical writeup on co-residing Qwen3-80B and Qwen3-4B on a single DGX Spark using vLLM, with a LiteLLM proxy routing heavy reasoning to the big model and fast queries to the small one. The gotcha: vLLM’s gpu_memory_utilization is a fraction of total memory, not free memory, so the two processes’ fractions need to sum below ~0.95. HN commenters shared their own two-Spark setups running DSv4 Flash at ~40-50 tok/s, though one accused the post itself of being LLM-written.
Occupancy isn’t everything on MI355X. A first-principles guide to occupancy math on AMD’s CDNA4 architecture, with the counterintuitive finding that matrix-core utilization can stay at 97% of peak even as wavefront occupancy drops. For MXFP8 GEMM workloads, throughput tracks matrix-engine utilization, not how full the wavefront slots are. HN comments mostly complained about the site’s low-contrast fonts, which is a shame because the technical content is worth the read.
Mode collapse, visualized. lcamtuf posted a collage of ~150 AI-generated children’s books on Amazon, all with near-identical titles, covers, and suspiciously clustered author surnames like “Bright” and “Wonder.” Commenters noted that a single LLM-written piece can look impressive — it’s only when you generate 50 that the convergence becomes obvious, and the term for it is mode collapse. Several noted the books are also full of factual errors, so it’s not just homogeneity that’s the problem.
iOS 27’s AI is mostly not Siri. TechCrunch previews Apple’s developer beta, where AI shows up as features embedded in existing apps — splitting restaurant bills from a receipt photo via Apple Cash, automated password updates after breaches — rather than a chat interface. Refreshingly grounded ambitions for a fall 2026 release.
That’s the briefing. Anthropic’s bad week isn’t likely to end before the weekend; we’ll see whether the cancel-and-refund link gets enough traffic to show up in any numbers.