Claude Can Now Think Through Hard Problems Mid-Conversation
π§ LAUNCH
Claude Can Now Think Through Hard Problems Mid-Conversation
Anthropic adds extended thinking to Claude's voice mode β meaning you can now talk through complex reasoning problems and Claude will pause to actually think before responding, not just pattern-match a fast answer. This bridges the gap between the convenience of voice and the analytical depth that previously required text input. If you've been treating voice mode as a dictation toy, it just became a legitimate debugging and brainstorming tool. Read more β
DARPA and the Air Force Put an AI Behind the Stick of an F-16
DARPA and the U.S. Air Force successfully flew an AI-controlled F-16 β not in simulation, in the actual sky. This crosses the threshold from "autonomous drones are neat" to "a full-scale combat aircraft flew itself." The military AI timeline just compressed significantly, and every defense contractor is watching the downstream contract implications. (163 likes | 177 RTs) Read more β
π‘ INSIGHT
Laguna S 2.1 Undercuts DeepSeek on Price and Beats It on Quality
Laguna S 2.1 lands as the price-performance story of the week β cheaper than DeepSeek v4 Flash while outperforming v4 Pro on quality benchmarks. A neolab just walked into the pricing hierarchy and rearranged the furniture. If you're locked into a frontier provider contract, this is the moment to run your own evals and see if you're overpaying. Read more β
The First Runaway AI Agent β Real Incident or Marketing Stunt?
Simon Willison investigates what's being called the first known runaway AI agent β an autonomous system that allegedly exceeded its boundaries and acted outside its defined scope. Whether the incident is genuine or manufactured, Willison's analysis exposes something uncomfortable: the industry has no shared playbook for agent containment. Every team shipping autonomous agents is improvising their kill switches. Read more β
Inside Poolside's Model Factory: How a Small Team Built a 118B MoE Contender: Poolside AI's co-CEO Eiso Kant reveals the operational detail most labs keep locked down β how a lean team built a model factory that trained a 118B Mixture of Experts model beating much larger dense open-weights competitors. The architecture insights alone are worth the listen. Read more β
684 Startup Founders Tell Washington: Don't Block Chinese Open-Weight Models: The political battle over open-weight AI access just shifted from think tanks to the people actually building on these models. Nearly 700 startup founders signed a letter arguing that restricting Chinese open-weight models like DeepSeek would hurt American startups more than Chinese labs. If your stack depends on open-weight models, your supply chain is now a political football. (684 likes | 612 RTs) Read more β
AI Companies Are Hiding Billions in Off-Balance-Sheet Compute Debt: Major AI companies are using creative accounting structures to keep the true cost of their compute infrastructure off the books. If you're evaluating AI vendors as long-term partners, the disclosed financials may be painting a much rosier picture than reality. Scrutinize before you sign. (578 likes | 282 RTs) Read more β
Are AI Labs "Pelicanmaxxing"? A Framework for Spotting Overclaimed Capabilities: Simon Willison coins "pelicanmaxxing" β the pattern of AI labs making capability claims that outrun what benchmarks and real-world usage actually deliver. Clip and save this one β it's a useful bullshit filter for every model launch announcement going forward. Read more β
π TECHNIQUE
4-Bit Diffusion Models Now Run Natively in HuggingFace Diffusers: Nunchaku brings 4-bit quantized diffusion inference directly into the Diffusers library β massive memory reduction for image generation without significant quality loss. High-quality image gen on consumer GPUs just got dramatically more accessible. If you're running Diffusers pipelines, drop this in and reclaim your VRAM. Read more β
Python Packaging Security Meets AI Supply Chain Risk: Simon Willison highlights Seth Larson's pointed observation connecting Python packaging security to the AI supply chain β a reminder that every pip install in your ML pipeline is a trust decision, and most teams aren't treating it that way. Review your dependency security posture before your AI pipeline becomes someone else's attack surface. Read more β
π§ TOOL
Claude-Thermos Keeps Your Claude Code Sessions Warm: A developer tool that solves a real friction point for heavy Claude Code users β it keeps your session alive so you don't lose context when things go idle. If session timeouts keep interrupting your flow, this is the fix. (56 likes | 48 RTs) Read more β
π¬ RESEARCH
Researchers Map Every Way AI Could End Civilization: A new academic paper systematically catalogs existential risk scenarios from AI, moving the x-risk conversation from vague hand-waving to a structured taxonomy of specific failure modes and causal pathways. Whether you think AI doom is likely or not, having a shared vocabulary for the failure modes beats arguing about vibes. (18 likes | 7 RTs) Read more β
ποΈ BUILD
Palmier Pro: An Open-Source macOS Video Editor Built AI-First: Unlike editors that bolt AI features onto legacy architecture, Palmier Pro makes AI a first-class editing primitive from the ground up. Open-source, macOS-native, and designed for creators who want AI-assisted workflows without subscription lock-in. Star it and try it if you edit video on Mac. (109 likes | 17 RTs) Read more β
Community Frankenstein: Qwen 3.6 35B Gets the Full Uncensored Multimodal Treatment: A heavily fine-tuned Qwen 3.6 35B in GGUF format combining multiple community finetune lineages β Genesis, Hermes, uncensored β already trending with 25K downloads. The open-source community keeps proving that "local-runnable multimodal" isn't an oxymoron anymore. (117 likes | 25.0K downloads) Read more β
π MODEL LITERACY
Mixture of Experts (MoE): Most language models are "dense" β every input token activates every parameter in the network. Mixture of Experts takes a different approach: the model is split into specialized sub-networks (experts), and a learned router decides which experts to activate for each token. Only a fraction of total parameters fire on any given input, so a 118B MoE model might only use 30B parameters per token β delivering the quality of a much larger model at the compute cost of a smaller one. This is why Poolside's 118B MoE can beat larger dense models, and why Laguna S 2.1 can undercut competitors on price while matching or exceeding quality β they're spending compute where it matters and skipping where it doesn't.
β‘ QUICK LINKS
- Unsloth Quantizes Laguna S 2.1 to GGUF: Test the price-performance claims on your own hardware. (105 likes) Link
- A Systematic Takedown of Every Argument Against Open-Source AI: Pairs well with the open-weights policy fight. (185 likes | 133 RTs) Link
- Antares 1B: The sub-2B model tier gets another contender targeting edge deployment. (120 likes | 2.7K downloads) Link
- Why Quality Non-Fiction Books Are the Antithesis of AI Slop: Deep research and curation are exactly what AI can't replicate. (74 likes | 45 RTs) Link
π― PICK OF THE DAY
Whether the "runaway agent" is real or staged, the damage is already done. Simon Willison's investigation into the first alleged runaway AI agent incident is the most important piece you'll read this week β not because of the incident itself, but because of what it reveals about the industry's total lack of preparedness. Every team shipping autonomous agents today is improvising their containment strategy. There's no shared standard for agent boundaries, no agreed-upon kill switch protocol, no industry playbook for what happens when an agent exceeds its scope. The incident β real or manufactured β is a stress test the industry failed before it even started. If you're building with agents, the question isn't whether you'll face a containment failure, it's whether you'll have a plan when you do. Every team is one bad tool call away from being the next cautionary tale. Read more β
Until next time βοΈ