Simon Willison Ships condense-json 1.0 to Squeeze Your Context Windows

🔧 TOOL

Simon Willison Ships condense-json 1.0 to Squeeze Your Context Windows

condense-json hits 1.0 — a utility that compresses JSON payloads specifically for LLM consumption, stripping whitespace, shortening keys, and restructuring nested objects to minimize token count without losing semantic content. If you're piping API responses, database dumps, or config files into prompts, this is the kind of unsexy infrastructure that saves you real money at scale. Structured data is one of the biggest hidden token sinks in agentic workflows — now there's a dedicated tool for it. Read more →

🔬 RESEARCH

Laguna S2.1, Inkling, and Kimi K3 Redraw the Open-Weight Map

Interconnects' Nathan Lambert maps how three open-weight models — Laguna S2.1, Inkling, and Kimi K3 — are reshaping the Pareto frontier for open models, with each occupying a different cost-performance sweet spot that closed-model providers used to own exclusively. The real story isn't any single model — it's that the capacity to train frontier-competitive models is proliferating well beyond the Big 3. For teams evaluating model choices, the frontier chart in this piece is worth ten benchmark tables. Read more →

💡 INSIGHT

Federal Judge Lets Minnesota's AI Nudification Ban Stand Against xAI

A federal judge denied xAI's request to block Minnesota's ban on AI-generated non-consensual intimate imagery, rejecting the company's First Amendment challenge. This is the first time a state-level AI content regulation has survived serious legal scrutiny from a frontier lab — and at least a dozen other states have similar bills in committee. The precedent matters: courts are signaling that narrowly targeted AI regulation can pass constitutional muster, even when well-funded challengers show up. Read more →

📝 TECHNIQUE

The Habsburg Frog: A Personal Benchmark That Exposes Real Model Differences. "Generate an SVG of a frog with a Habsburg jaw" — it sounds absurd, but this personal benchmark is a surprisingly sharp test of spatial reasoning, instruction-following, and creative interpretation across frontier models. The results page shows how wildly outputs diverge when models have to combine precise anatomical knowledge with vector graphics generation — the kind of compound task that flat benchmark scores completely miss. Worth bookmarking as a quick sanity check when evaluating new models. (83 likes | 42 RTs) Read more →

🎓 MODEL LITERACY

Pareto Frontier in Model Evaluation: Today's open-model roundup ranks Laguna S2.1, Inkling, and Kimi K3 on the Pareto frontier — the curve where no model dominates another on every metric at once. A model is "on the frontier" if improving one dimension (say, coding ability) would require sacrificing another (say, cost or latency). Models below the frontier are strictly worse than at least one frontier model on all measured axes. Understanding this concept explains why "best model" is always the wrong question — the real question is which tradeoffs you're willing to accept for your use case.

⚡ QUICK LINKS

• The Open Letter Tracker: Simon Willison catalogues which AI petitions actually moved the needle — and which were pure performance art. Link
• Hank Green Admits His AI Usage Has Become 'Not Healthy': The most high-profile acknowledgment yet that LLM interaction creates genuine dependency patterns, not just productivity habits. Link
• AI Poster Wins Ohio State Fair Contest — Again: Four years after the Colorado incident, competition rules still haven't caught up. (65 likes | 38 RTs) Link

🎯 PICK OF THE DAY

Minnesota's surviving ban reveals that state-level AI regulation is quietly assembling the patchwork legal framework federal lawmakers can't agree on. xAI brought a serious First Amendment challenge to Minnesota's nudification ban — and lost. That matters far beyond one state's content law. While Congress spins its wheels on comprehensive AI legislation, individual states are passing narrowly targeted bills that are surviving judicial review. The irony is that xAI's challenge may have actually hardened the precedent it tried to kill: a federal judge has now explicitly ruled that this category of AI regulation doesn't violate free speech protections. For the dozen-plus states with similar bills in the pipeline, that's a green light. The emerging picture isn't the top-down federal framework that industry lobbyists could negotiate with — it's a patchwork of state laws, each tested in court, each setting precedent. Compliance teams at frontier labs should be mapping state legislative calendars, not waiting for Congress. Read more →

Until next time ✌️