Summary
The slot's real signal is the open-weight model wars heating up: Kilo Code's thread on StepFun's Step 3.7 Flash and MiniMax M3, both shipped with open weights, is the standout. M3 is the interesting one for efficiency readers, claiming a 1M-token context built on MiniMax Sparse Attention (MSA) that cuts per-token compute at long context to roughly 1/20th of the prior generation, with 9x faster prefill and 15x faster decode. The second cluster is Anthropic's Logan Graham (5 tweets) reflecting on the Project Glasswing expansion to ~150 more orgs and the Claude Mythos cyberdefense preview, with a candid line that "patching is the bottleneck, not finding vulns." Computex supplies the hardware thread: Intel touting 18A high-volume ramp plus new rackscale AI infra, and NVIDIA pushing real-time AI-for-media (a 92%-accuracy synthetic video detector at 22ms). Everything else is noise: a long run of French techno-optimist political rants from @brivael, fraud-task-force PR, and a Kombai 2.0 "AI design engineer" plug. A coding-model bug-hunt benchmark (Opus 4.8 caught 10/15, Grok Build 0.1 9/15, Gemini 3.1 Pro just 2/15) is a fun datapoint worth one click.
Posts
- StepFun Step 3.7 Flash + MiniMax M3 open-weight drops (@kilocode · blog) (cluster of 4). Two Chinese labs shipped open weights the same week: Step 3.7 Flash (196B params, ~11B active per token, 256K window, Apache 2.0) and MiniMax M3 (1M context, native multimodal). M3's MiniMax Sparse Attention (MSA) is the headline efficiency claim, ~1/20th per-token compute at 1M context with 9x prefill and 15x decode speedups. Worth a summary page if the M3 report holds up. See inference-efficiency/MISA for the adjacent sparse-attention line.
- Project Glasswing expansion + Mythos reflections (@logangraham · Anthropic post) (cluster of 5). Anthropic extended Claude Mythos Preview to ~150 more orgs across 15+ countries. Graham's frank takeaways: Mythos finds plenty of vulns but patching is the real bottleneck, orgs can stand up model-powered cyberdefense in days, and "Mythos will look dumb in 6-12 months." Extends responsible-ai/Project Glasswing.
- Coding-model bug-hunt benchmark (@kilocode · results). Five frontier models, 15 planted bugs: Opus 4.8 caught 10/15, Grok Build 0.1 and Sonnet 4.6 9/15, GPT-5.5 8/15, Gemini 3.1 Pro 2/15. Grok found the hardest three-file mutation bug at ~1/3 the cost-per-catch. Small N, but a clean cost-vs-catch datapoint.
- Intel Computex AI announcements (@magicsilicon · Intel newsroom). 18A ramped to high volume, 14A on track, plus new rackscale AI infra for inference and agentic workloads on Xeon. Foundry-recovery narrative continues; relevant if you track non-NVIDIA inference silicon.
- NVIDIA AI for Media at Computex (@nvidia · NVIDIA blog). Real-time AI for live video: a synthetic-video detector at up to 92% accuracy and 22ms latency, plus RTX super-resolution upscaling 480p to 8K. Provenance/authenticity tooling moving into production media pipelines.
- Kombai 2.0 "AI design engineer" (@Scobleizer). Microsoft Build demo plug for a tool pitched as design-aware coding. Promo. Skip.
- Westmag American robot actuators (@Scobleizer). $11M a16z-led raise to build US actuators and drone motors, framed against reliance on Chinese parts. Hardware-supply-chain angle, not AI-core. Skip.
- French techno-optimist political thread (@brivael) (cluster of ~13). A long run of abundance/anti-collectivism rants citing the a16z Techno-Optimist Manifesto, plus Elon and political content. No AI substance. Skip.
- Political / fraud-task-force PR (@WHFraudTF, @SeanParnellASW, @spencerpratt) (cluster of 4). Medicaid-fraud enforcement and US political messaging that tripped the keyword filter. Skip.