Good morning. The open-weight arms race got another entry overnight: Alibaba announced Qwen 3.8 at 2.4T parameters, clearly timed to blunt Moonshot’s Kimi K3 moment. Meanwhile Kimi is now so overloaded Moonshot has paused new signups, Claude reportedly cracked a 1939 math conjecture (maybe), and a solo developer shipped a whisper.cpp replacement that puts most funded projects to shame.
Qwen 3.8 arrives to answer Kimi. Alibaba announced Qwen 3.8, a 2.4T-parameter open-weight model, with a preview live on Alibaba Cloud ahead of the weights release. The HN thread reads the timing as a direct response to Kimi K3’s 2.8T announcement last week. Community reception is mixed: fans of the smaller Qwen 3 MoE and dense variants are hoping for local-deployable sizes, while one daily user of Qwen 3.7 Pro called it unusable for coding compared to DeepSeek V4 Pro, citing stuck loops and poor debugging.
Kimi K3 sells out its own GPUs. Moonshot has temporarily paused new subscriptions after 48 hours of demand exceeded capacity, and plans to split membership into general-use and coding tiers. The HN reaction was unusually warm for a service outage, mostly because Moonshot said the quiet part out loud rather than silently throttling limits the way Google has been accused of doing. Users are reporting strong code review performance despite the current slowness — one benchmark placed K3 19th on one-shot coding but 3rd on agentic coding, matching the pattern where Chinese models compensate for weaker one-shot reasoning with better tool iteration.
Claude Fable produces a Jacobian counterexample — allegedly. Mathematician Levent Alpoge posted on Twitter that Claude Fable generated a counterexample to the Jacobian Conjecture, a problem open since 1939 and famous for its graveyard of flawed proofs. The HN thread is understandably cautious — Wikipedia still lists it as open, and announcing via Twitter rather than arXiv is an odd choice for a result of this magnitude. One commenter noted that the conjecture’s long history of subtly-wrong published proofs may have given the LLM enough scaffolding to synthesize a valid counterexample. Independent verification is the whole ball game here.
OpenAI trims Codex context to 272k. A merged hotfix drops the Codex context window from 372k tokens to 272k. The HN reaction split predictably: some users are fine because compaction handles it, others argue compaction is lossy in ways that break complex multi-step work. A recurring point across the thread is that model quality degrades well before the technical limit — several commenters said they manually clear around 30-40% of Opus’s 1M window because performance falls off a cliff past that.
Two impressive speech stacks, both from small teams. transcribe.cpp is a ggml-based whisper.cpp replacement supporting 60+ ASR models across 16 families, with Vulkan/Metal/CUDA acceleration, bindings for four languages, and WER-validated model conversions. The HN thread kept returning to disbelief that it’s one developer’s work — the same person who built the Handy app. Separately, Moonshine AI released Moonshine Micro, a full VAD + STT + TTS pipeline that runs in 470KB of RAM on an $0.80 RP2350 microcontroller, MIT-licensed. Local voice is quietly becoming a solved problem at both ends of the hardware spectrum.
Study: AI advice made people less accurate but more confident. French and Italian researchers found that giving subjects access to AI advice dropped accuracy from 27% to 9% while pushing confidence from 30% to 76%, with willingness to admit uncertainty collapsing from 44% to 3% — a pattern Wharton researchers have called “cognitive surrender.” TechNextWeb has the writeup. The HN thread is skeptical of the methodology: researchers deliberately used a model known to fail on the test questions, incentives were $0.10 per correct answer, and the article never links the paper directly. As one commenter put it, the same effect probably shows up with any authoritative wrong-answer source.
Nvidia’s Japan tour. TechCrunch reports that Jensen Huang’s two-day Tokyo trip produced commitments from Fanuc, Yaskawa, and Kawasaki Heavy to build on Nvidia’s Cosmos models, plus a “Vera Rubin” data center slated for 2028 and partnerships around Noetra, Japan’s ~$6.2B sovereign AI initiative involving SoftBank, Sony, and Honda. Also worth a look: Alibaba’s T-Head Semiconductor quietly open-sourced a Triton fork for its PPU accelerators, with FP8/FP16/BF16 Tensor Core support and a new tl.aiu_load API — another sign that Chinese hardware stacks are building serious compiler tooling around domestic silicon.
That’s the morning. Whether Qwen 3.8 lands as a genuine Kimi competitor or a rushed announcement will come down to the weights release — expect the pelican benchmarks by the weekend.