GPT-5.6 Bridges a 30-Year Gap in Convex Optimization

🔬 RESEARCH

GPT-5.6 Bridges a 30-Year Gap in Convex Optimization

OpenAI's GPT-5.6 reportedly used a single prompt to close a long-standing open problem in convex optimization — not reproducing a known proof, but generating a novel construction that human mathematicians could verify but hadn't discovered. The Reddit math community is picking it apart and the consensus is holding: this isn't a hallucination dressed up as insight, it's a genuine contribution to the field. If this replicates across other open problems, we're watching the researcher-model relationship fundamentally shift from "AI checks my work" to "AI proposes things I missed." (481 likes | 311 RTs) Read more →

Fable 5 vs. GPT-5.6 Sol: Does /goal Actually Help on NP-Hard Problems?

Charles Azam runs a rigorous head-to-head between Fable 5 and GPT-5.6 Sol on an NP-hard problem, specifically testing whether Anthropic's /goal feature gives a meaningful reasoning edge. The results cut through marketing noise — independent benchmarks like this are worth ten vendor blog posts. If you're choosing between frontier models for hard optimization tasks, this is the comparison to read. (205 likes | 102 RTs) Read more →

💡 INSIGHT

Stack Overflow's Traffic Collapse, in One Graph

The data is brutal. A Stack Overflow query visualization shows the platform's activity falling off a cliff post-LLM — the clearest quantitative evidence that AI coding assistants have fundamentally rewired how developers seek help. This isn't a gradual decline; it's a step function. The implications for knowledge preservation are real: if nobody's asking or answering on public forums, where does the next generation of training data come from? (352 likes | 408 RTs) Read more →

Anthropic Makes Fable 5 a Permanent Model: Simon Willison covers the news that Fable 5 is now a permanent offering rather than a limited preview. The signal: Anthropic is confident enough in its differentiation to give developers a stable target. Update your model configs accordingly. Read more →

Kaiser Nurses Say AI Surveillance Is Making Patient Care Worse: Front-line nurses at Kaiser report that AI monitoring tools are degrading care quality — charting takes longer, clinical judgment gets second-guessed by algorithms, and patients notice. A critical counterpoint for anyone building enterprise AI: your deployment metrics might look great while the humans in the loop are drowning. (152 likes | 115 RTs) Read more →

First City Bans Undisclosed AI Images in Real Estate Ads: Mayor Mamdani's office just made it illegal for landlords to use AI-generated images in property listings without disclosure. It's the first municipal-level regulation of its kind — and likely a bellwether for disclosure requirements across every industry where AI-generated visuals touch purchasing decisions. (90 likes | 35 RTs) Read more →

🔧 TOOL

SQLite Query Explainer: An AI-powered tool that translates SQLite EXPLAIN output into plain English. If you're building local-first apps or debugging embedded database performance, paste your slow query and get a human-readable breakdown of what the query planner is actually doing. Immediately useful. Read more →

Vercel AI SDK 6.0.230 Fixes Non-Streaming Tool Call Ordering: The Vercel AI SDK patches input event ordering for non-streaming tool calls — subtle but critical if you're building human-in-the-loop agent workflows where event timing determines whether the UX feels broken or smooth. Update if you're using tool calls with onToolCall. Read more →

llama.cpp Gets OpenCL Transpose Optimization for Q4_K Models: llama.cpp b10064 adds coalesced-read OpenCL transpose ops for Q4_K quantized models — continued performance work for anyone running local inference on non-NVIDIA hardware. The AMD and Intel GPU inference story keeps getting better. Read more →

🏗️ BUILD

Full Speech Recognition + TTS in Under 500KB

Moonshine Micro packs complete speech recognition and text-to-speech into less than 500 kilobytes. That's small enough to run on microcontrollers, IoT sensors, and anywhere network connectivity is a luxury. On-device voice AI just became possible on hardware that costs less than a coffee — the deployment surface for voice interfaces just expanded by orders of magnitude. (200 likes | 25 RTs) Read more →

Spot Birds Not Golf: Vision Models for Hobbyist Projects: Simon Willison demonstrates using vision models to identify birds instead of golf balls — a playful but practical showcase of how accessible computer vision has become for personal projects. If you've been waiting for permission to point a model at something weird, this is it. Read more →

bbarit-agent: MIT-Licensed Terminal Coding Agent in a Single Rust Binary: A self-hostable terminal coding agent that connects to 15+ LLM providers and 1,000+ models — all in one Rust binary. For teams that need the Claude Code / Codex CLI experience but can't use cloud-hosted agents for compliance reasons, this is the MIT-licensed alternative. (34 likes | 10 RTs) Read more →

📝 TECHNIQUE

Turn a Spare Mac into a Claude Code-Controlled Workstation: A step-by-step guide to dedicating a Mac as an autonomous Claude Code machine — SSH access, permissions, persistent sessions, the works. This is the setup many developers want but haven't bothered configuring. Instructions are concrete enough to have it running this weekend. (166 likes | 125 RTs) Read more →

🎓 MODEL LITERACY

Proof Discovery vs. Proof Verification: GPT-5.6's convex optimization result crosses a critical line — from AI-as-calculator (verifying known proofs, checking steps humans already wrote) to AI-as-collaborator (proposing novel mathematical objects humans hadn't found). Verification means confirming that a proof is correct; discovery means generating the proof itself. Most "AI does math" claims are verification dressed up as discovery — the model reproduces known results or finds proofs for problems humans already solved differently. When a model proposes a construction that closes a 30-year open gap, that's genuine discovery. Understanding this distinction helps you evaluate which AI math claims represent real capability advances and which are parlor tricks.

⚡ QUICK LINKS

Next.js Canary 89: Packages Claude Code skills as a plugin, fixes Turbopack SSR for soft navigations. Link
Quixote: New AI tooling project flagged by Willison — worth watching. Link
Free Claude Code Skills Collection: Community-curated skill library gaining traction. (36 likes) Link
Anthropic TypeScript SDK v0.112.3: Documentation improvements, keeps the SDK current. Link
Even Latent Space Says "Not Much Happened Today": The meta-signal that confirms a quieter news cycle. Link

🎯 PICK OF THE DAY

Closing a 30-year-old gap in convex optimization isn't impressive because a model did math. It's impressive because it suggests frontier LLMs are now generating novel mathematical objects that human researchers can verify but hadn't discovered — fundamentally changing the researcher-tool relationship. Previous "AI math" milestones were mostly about speed (proving things faster) or scale (checking more cases). This is categorically different: GPT-5.6 proposed a construction that experts in the field hadn't found in three decades of trying. The verification came from humans; the discovery came from the model. If this pattern holds across other open problems, we're entering an era where the bottleneck in mathematical research shifts from "can we find the proof?" to "can we ask the right question?" The researchers who learn to prompt effectively become force multipliers. The ones who dismiss this as "just autocomplete" will watch their colleagues publish first. Read more →

Until next time ✌️