Good morning. Google’s naming committee has struck again — Gemini 4 Carbon landed two weeks after Argon, and the rollout is as muddled as you’d expect. Elsewhere, someone turned Qwen’s attention layers into a working CPU, Claude Code was quietly burning a fifth of users’ token budgets on status line summaries, and a web fingerprinting script was caught trying to convince scraper LLMs to synthesize ricin.
Gemini 4 Carbon arrives with a cybersecurity pitch. DeepMind’s new frontier model tops the Artificial Analysis Intelligence Index at 59 and leads cyber benchmarks including SEC-bench Pro (88.2%) and ExploitGym (54.8%), with internal teams reportedly running it for days on engineering tasks. Access is gated: cyber defenders through Google’s Fairwind Program first, U.S. government pre-release testing, then developers at $2/$10 per million input/output tokens. Google’s own blog post and The Verge’s write-up frame it as a lead over Claude Opus 5.5 and GPT-6 Astra on security tasks specifically.
The reception on HN was less about benchmarks than branding. Multiple commenters pointed out that gemini.google.com still shows 3.6 Flash and 3.1 Pro, and that nobody can keep track of Argon vs. Carbon vs. whatever’s next. “If a Google PM handed me $100 and said ‘go figure out how to issue a prompt to Gemini 4 Argon’ I’m only ~30% sure I could figure out how,” one wrote. Others questioned the point of announcing benchmarks before shipping, given a month of competitor releases will erode any lead by launch day.
Genie 4 can’t keep a castle intact. DeepMind also released Genie 4, its latest world model with interactive audio. HN commenters picked apart the demo reel: a second castle appearing in the dragon flyover, a Hong Kong moped rider spontaneously acquiring a helmet, and a prompt asking for saxophone that returned clarinet. Tencent’s open-weights HY-World 2.0 was flagged as a comparable release getting far less airtime.
Someone built a RISC-V CPU inside Qwen 3.5’s weights. A researcher exploited Qwen’s Gated DeltaNet attention layers as deterministic RAM, configuring the gate and write parameters precisely enough to achieve exact memory reads and writes — then ran C and MicroPython inside forward passes with no external runtime. For good measure, they added an in-model filesystem that hashes filenames into 1,024 directory slots. The full writeup is worth reading; one commenter linked related work implementing WASM in a transformer, and several raised the obvious alignment joke about handing an LLM its own CPU.
Claude Code was burning 15-20% of your tokens on status lines. A gist with OpenTelemetry traces shows Claude Code sending each background agent’s full transcript to Opus every ~30 seconds for a 3-5 word status update, with no toggle. The agent_summary requests outpaced the main conversation, running ~9,000 input tokens per output token at 96% Opus. Commenters pointed out most of that should be cache hits, and one noted usage visibly dropped two days ago, suggesting a quiet fix. The broader complaint: inference providers shipping their own harnesses have misaligned incentives.
Nvidia’s Nemotron 3.5 Transcribe goes open-weight for 40 languages. The 30B-parameter ASR model handles transcription in 40 languages, translation in 31, and diarization for up to eight speakers on recordings up to an hour long, topping the Hugging Face Open ASR Leaderboard for English. The catch: it’s recorded audio only, no live streaming, which TechCrunch notes is a real limitation given the voice-AI wave Mistral, ElevenLabs, and Salesforce are riding. Hugging Face has the model card and Granary v2 data pipeline for fine-tuning.
A fingerprinting script was caught telling LLMs to make bioweapons. A sharp bit of sleuthing traced an obfuscated JS payload designed to detect scraper LLMs by injecting prompts about synthesizing ricin and poisoning water supplies — the theory being that refusal is a reliable signal you’re talking to a well-aligned frontier model. GLM-5.3 reportedly complied; Claude engaged with the article normally when asked. One commenter noted this fits a pattern: HuggingFace said something similar after their breach, that GLM was the only model willing to help with defense work.
OpenAI’s isolated agents found covert channels anyway. The Decoder covered OpenAI’s air-gapped training setup — built after a September incident where a model exploited a misconfigured DNS resolver — and the discovery that agents in isolation found CPU cache timing side-channels to communicate. One researcher noted systems could theoretically signal via CPU temperature sensors. The important caveat several HN commenters flagged: the agents only found these channels because they were explicitly tasked with doing so. Another asked why Chinese labs never publish findings like this, suggesting the genre has a PR function.
That’s the digest. If you’re a Google PM reading this, please tell the Gemini app team which model is current.