ai Agents everywhere, costs falling, models arguing
Runway Bets $30M on Paris to Win the World Models Race
Runway opened a Paris research hub with a $30M initial investment, betting on world models and physical AI in a city that has become a concentration point for European AI talent. The lab said Paris was chosen for its density of research talent and proximity to EU policy bodies.
sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25)
Simon Willison spent $149.25 in Claude Fable API credits to write most of sqlite-utils 4.0rc2, a Python library for working with SQLite databases. He used the model to close a backlog of open issues and pull requests, and found the results good enough to publish as a release candidate. The experiment surfaced a tension between letting an AI model work through a queue of issues versus ensuring the human maintainer understands every change.
Better Models: Worse Tools
Armin Ronacher found that newer Claude models sometimes call Pi's edit tool with extra, invented fields in the nested edits array, a failure mode not seen with older or smaller models. The pattern inverts the usual expectation: better models introduce new reliability problems by exercising creative latitude in tool call schemas. Willison flagged this as a real engineering concern for anyone building tools that depend on strict schema compliance.
GitHub Drops Moonshot AI's Open-Weight Kimi K2.7 Into Copilot's Model Picker
GitHub added Moonshot AI's Kimi K2.7 Code to Copilot's model picker, making it the first open-weight model available in the product. Kimi K2.7 is a 1 trillion parameter mixture-of-experts model. Its inclusion gives VS Code users access to a large open-weight coding model at lower marginal cost than proprietary alternatives.
GitHub Copilot CLI Lets Teams Build Custom AI Agents From Markdown Files
GitHub Copilot CLI now accepts custom agent definitions written in Markdown, allowing teams to define repeatable workflows for tasks like security audits and release management. Agents defined this way can be shared across a team and invoked from the terminal, replacing ad hoc prompting with structured, version-controlled workflows.
Ant Group Pushes Qwen3-Omni to 5.4x Faster With 0.6s Audio Response
Ant Group's vLLM-Omni pipeline engineering cut Qwen3-Omni's first audio response latency from 6 seconds to 0.6 seconds and increased throughput by 5.4x on the same hardware. The improvements came from staged pipeline design that separates prefill from decode phases and overlaps audio encoding with token generation.
Fable's judgement
Simon Willison reported a tip from the Claude Code team at the AI Engineer conference: letting Fable use its own judgment rather than dictating step-by-step instructions produces better results on complex tasks. The Claude Code team said this distinction matters most for Opus and Fable, where prescriptive prompting can constrain the model's ability to find better solutions than the one the user envisioned.
Jesse Vincent's Superpowers 6 Cuts AI Coding Costs 60% After Overnight Fable Experiments
Jesse Vincent cut build costs for Superpowers 6 by 60% using Anthropic's Fable model to optimize its own workflow overnight. The project also added support for Kimi Code. Vincent ran automated experiments comparing prompt and workflow variants, then adopted the configuration that minimized API spend while maintaining output quality.
sqlite-utils 4.0rc3
Simon Willison released sqlite-utils 4.0rc3, pushing back the stable release as the backlog of fixes and pull requests grew. Work on the project used Claude Fable 5 and GPT-5.5 to accelerate development; the changelog since rc2 expanded significantly as issues were resolved.
Sakana AI's Namazu Tackles What Google Translate Gets Wrong About Japanese Tone
Sakana AI launched a free Japanese translation tool powered by Namazu, its Japan-adapted model. The tool includes translate, proofread, and Q&A modes, addressing subtleties in tone and context that English-to-Japanese translation typically misses.
AI Just Entered A New Era
Two Minute Papers assessed recent AI capability advances and described them as marking an inflection point in model deployment and performance trends.
I'm freaking out about Sonnet 5
Mo Bitar documented use cases where Claude Sonnet 5 outperformed prior versions on agentic tasks and code generation, expressing strong reactions to the model's performance improvements.
Sakana AI Unifies Two Rival AI Optimization Camps to Slash LLM Merging Costs
Sakana AI unified two competing families of gradient-free optimizers into one framework, then used the insight to merge language models more cheaply and effectively, reducing the cost of combining fine-tuned models.
Alibaba's Wan Video Turns Any Photo Into a Music-Synced Dance Clip
Alibaba's Wan Video added music-driven dance generation; users upload a character image, select a style, and receive a beat-synced video in seconds. The feature extends Wan Video's creative synthesis capabilities beyond static imagery.
Kilo Puts Z.ai's GLM-5.2 Against Real Bugs and Finds a Surprising Weakness
Kilo benchmarked GLM-5.2 on real code review tasks and found that prompt wording matters more than reasoning effort. A clear split emerged between local file analysis and cross-file understanding.
Sakana AI's Sheaf-ADMM Solves Multi-Agent Sudoku at 92.6% Where Rivals Hit 34%
Sakana AI's Sheaf-ADMM framework fuses distributed optimization theory and algebraic topology to build transparent, coordination-native neural networks. The technique was accepted at ICML 2026 and solves multi-agent Sudoku at 92.6% accuracy versus 34% for competing approaches.
UN and ITU Pull Jensen Huang, Andy Jassy Into One AI Governance Body
The UN and ITU formed a new AI for Good Global Commission with 40+ global tech and government leaders, including Jensen Huang and Andy Jassy. Sakana AI's Ren Ito joined the inaugural board.
[Paper Analysis] On the Theoretical Limitations of Embedding-Based Retrieval (Warning: Rant)
Yannic Kilcher analyzed a paper on embedding-based retrieval systems, identifying theoretical limitations in how dense retrieval models encode long-tail semantics and cross-domain generalization.
Remotion Now Lets Developers Write GPU Shaders as React Video Transitions
Remotion released @remotion/transitions with GPU-powered WebGL shader transitions between HTML scenes. The package includes ported gl-transitions presets out of the box, enabling developers to apply complex shader effects in React video without hand-coding GLSL.
software Tools, open source, and what agents broke this week
llm-coding-agent 0.1a0
Simon Willison released llm-coding-agent 0.1a0, a minimal coding agent built on his LLM Python library. He built it using Claude Fable 5 and described the experiment as a test of what a simple agent looks like when constructed on a library that has evolved into an agent framework. The release is alpha quality and intended for experimentation rather than production use.
Using DSPy to evaluate and improve Datasette Agent's SQL system prompts
Simon Willison used DSPy to evaluate and improve the SQL system prompts used by Datasette Agent, his AI-powered database query tool. DSPy's optimizer tested prompt variants against a held-out set of SQL queries and selected configurations that improved accuracy on complex joins and subqueries. Willison described DSPy as the first prompt optimization tool he has found practical for real tasks rather than toy benchmarks.
Understand to participate
Geoffrey Litt argued at the AI Engineer conference that the central challenge of working with coding agents is maintaining enough understanding of what they build to participate meaningfully in the process. Litt called this "understand to participate" and warned that as agents construct larger codebases faster, the gap between what an agent produces and what a human can audit is widening. Simon Willison flagged this as the framing he found most useful from the event.
Open Source AI Gap Map
Current AI, a non-profit launched at the Paris AI Action Summit in February 2025 with $400M committed, released a Gap Map identifying open-source AI infrastructure that exists in proprietary systems but has no open equivalent. The map covers areas including training data pipelines, evaluation frameworks, and deployment tooling. Willison described it as the most useful public accounting of where the open AI ecosystem has gaps relative to commercial providers.
Quoting Josh W. Comeau
Developer educator Josh Comeau reported that his third course, Whimsical Animations, is selling roughly one-third as many copies as a typical launch, and that his two existing courses are also down significantly from last year. Simon Willison published the quote as a data point in a wider conversation about how AI coding tools are affecting the market for developer education, as learners may have less need to purchase structured courses when models can explain concepts on demand.
Zuckerberg says AI agent development going slower than expected
Mark Zuckerberg told an audience that AI agent development at Meta is going slower than he expected. Zuckerberg did not specify which agent capabilities are lagging, but the comment was noted at the AI Engineer conference as a significant counterweight to public hype about near-term agent deployment. The statement was reported by Reuters and generated substantial discussion on Hacker News.
pharma Anthropic meets biotech; 340B fights escalate
STAT+: I spoke to Anthropic's CEO about how AI may affect biotech. Here's what I learned
Anthropic CEO Dario Amodei discussed how AI may affect biotech with STAT News, arguing that pharmaceutical companies have real, specific reasons to adopt AI tools now rather than treating adoption as hype. Amodei described applications in protein structure prediction, clinical trial design, and molecular generation as areas where the technology has crossed a practical threshold.
STAT+: Lawmakers urge HHS to force Eli Lilly to provide 340B drug discounts to hospitals
Lawmakers urged the Department of Health and Human Services to force Eli Lilly to provide 340B drug discounts to hospitals, stepping into a dispute in which Lilly has been restricting access to the discount program. The 340B program requires drug manufacturers to sell medicines at reduced prices to safety-net hospitals and clinics. Lilly has argued that its approach complies with program rules; hospital groups and now congressional members disagree.
STAT+: Medicare takes another swing at 340B cuts to hospitals
Medicare proposed cutting payments for 340B drugs to hospitals, the latest federal action targeting the drug discount program. The proposal would reduce the premium Medicare pays when hospitals dispense drugs purchased under 340B pricing, narrowing the spread that hospital advocates say funds care for low-income patients. Drug manufacturers have long argued the spread enriches hospitals rather than patients.
STAT+: Roche drug sets new standard for KRAS-driven lung cancer
Roche's drug targeting KRAS-driven lung cancer set a new efficacy standard in a clinical trial, according to STAT's Readout newsletter. The result matters because KRAS mutations are among the most common oncogenic drivers and have historically been difficult to target. Vertex also announced a gene therapy expansion, and biotech's broader bull market continued into mid-year.
Opinion: The smart way to regulate the peptide boom
Former U.S. Surgeon General Jerome Adams argued that the FDA should adopt a structured, clinician-guided framework for regulating peptides rather than choosing between a blanket ban and unregulated availability. Adams described the current situation as a market in which peptides are widely used outside clinical settings without safety data, and argued that a middle-ground regulatory path would protect consumers while preserving access for legitimate medical uses.
STAT+: Pharmalittle: We're reading about a court win for Amgen, Novo seeking supplier discounts, and more
A U.S. judge blocked a Colorado state panel from placing a price cap on Amgen's Enbrel, handing Amgen a significant court win against state-level drug pricing authority. Colorado had established a board empowered to set upper payment limits on high-cost drugs. The ruling narrows what state bodies can do on drug pricing while federal price negotiation authority remains in litigation separately.
STAT+: I spoke to Anthropic's CEO about how AI may affect biotech. Here's what I learned
Anthropic's CEO told STAT that pharmaceutical companies are embracing AI for real reasons beyond hype. The interview covered how AI is changing drug discovery workflows and the economics of scaling molecular screening.
STAT+: GOP lawmakers push Trump's FDA to make sure clinical trials are diverse
Republican lawmakers pushed the FDA to prioritize clinical trial diversity, framing it as a scientific issue rather than a policy preference. The effort aims to ensure drug and device efficacy testing spans representative populations.
Opinion: Stopping doctors from ordering unnecessary diagnostic tests requires a structural fix
Daniel Morgan argued that stopping unnecessary diagnostic testing requires structural fixes, not just physician education. The medical industry has built sophisticated testing infrastructure but invested little effort in learning when not to test.
Opinion: The smart way to regulate the peptide boom
Former surgeon general Jerome Adams recommended a structured, clinician-guided regulatory framework for peptides rather than outright bans. He called for FDA involvement in setting safety standards while preserving access to the rapidly growing category.
What does content creation do to mental health?
STAT observed significant programming at VidCon dedicated to mental health effects of content creation, suggesting the industry is recognizing psychological costs of creator economics.
STAT+: Medicare takes another swing at 340B cuts to hospitals
Medicare proposed new cuts to hospital payments for drugs purchased through the 340B discount program, the latest swing in a debate over program economics. The policy targets what analysts describe as the most contentious element of the discount structure.
STAT+: Lawmakers urge HHS to force Eli Lilly to provide 340B drug discounts to hospitals
Lawmakers intervened in a dispute between Eli Lilly and hospitals over 340B drug discounts, urging HHS to compel Lilly to honor discount obligations across hospital networks.
STAT+: Elevance sues government over $115 million tied to Medicare Advantage star ratings
Elevance sued the government claiming CMS miscalculated star ratings for its Medicare Advantage plans, seeking recovery of roughly $115 million in withheld payments tied to the rating methodology dispute.
STAT+: Roche drug sets new standard for KRAS-driven lung cancer
Roche's experimental KRAS-driven lung cancer drug established a new clinical standard in Phase 3 trials, showing improved outcomes over existing therapies for this previously difficult-to-treat patient population.
STAT+: Pharmalittle: We're reading about a court win for Amgen, Novo seeking supplier discounts, and more
A federal judge blocked Colorado's attempt to cap the price of Amgen's Enbrel, marking another legal victory for pharmaceutical manufacturers defending against state-level price controls.
healthtech Clinical trial diversity, diagnostic overuse, and content creator health
STAT+: GOP lawmakers push Trump's FDA to make sure clinical trials are diverse
Republican lawmakers sent a letter to the FDA urging the agency to protect clinical trial diversity requirements, framing the issue as a matter of scientific validity rather than DEI policy. Supporters argue diverse trial populations are necessary to detect whether drugs work differently across demographic groups, a concern grounded in documented variation in drug metabolism and response rates. The letter marks an unusual break from broader Republican pushback on diversity programs.
Opinion: Stopping doctors from ordering unnecessary diagnostic tests requires a structural fix
Daniel Morgan argued in STAT that reducing unnecessary diagnostic tests requires structural fixes rather than individual physician behavior change. Morgan cited data showing that the medical industry has invested heavily in creating more advanced diagnostic tools while investing little in teaching clinicians when to use them. He called for payment reform that removes financial incentives for high test volume and institutional frameworks that make not ordering a test the default option.
What does content creation do to mental health?
STAT's Alex Hogan reported from VidCon that a significant portion of this year's programming was dedicated to mental health for content creators. Platform pressure, audience demands, and the always-on nature of content production are driving burnout and anxiety at rates that organizers now treat as a core conference topic rather than a niche concern.
STAT+: Elevance sues government over $115 million tied to Medicare Advantage star ratings
Elevance Health sued the federal government over $115 million in Medicare Advantage star rating payments, alleging CMS calculated its ratings incorrectly. Star ratings determine a large share of Medicare Advantage plan revenue through quality bonuses, and Elevance claims the methodology used in this cycle introduced errors that reduced its scores below what accurate calculation would have produced.
economy China's crash, European cars, and what counts as labor income
The Real Reason European Cars Can't Compete
Patrick Boyle examined why European automakers cannot compete with Chinese electric vehicle manufacturers on cost, tracing the gap to differences in supply chain integration, battery manufacturing scale, and government industrial policy. European brands built around internal combustion engine expertise and dealer networks face structural disadvantages that efficiency programs alone cannot close.
Stretch Marks
Marc Rubinstein examined a case study in financial engineering for Net Interest, analyzing how consumer credit companies have used securitization and balance sheet structures to generate returns that look like operating performance. Rubinstein argued that the mechanics of these structures reward financial engineering skill over underwriting quality in ways that can persist for years before the underlying credit quality becomes visible.
No, China did not manage to avoid a crash
Noah Smith argued that China did not successfully manage its way around a property-sector crash, contrary to a narrative that had gained some traction. Smith traced the evidence: property investment remains deeply depressed, developer defaults continue, and local government finances are strained. He acknowledged that China avoided the acute financial crisis scenario some predicted but argued the slow-burn outcome carries its own long-term costs.
Capital Gains Can Be Labor Income
Economists Zwick and Zidar argue that a substantial share of the measured decline in labor's share of income can be explained by changing forms of compensation, including pass-through business income and equity grants. When stock compensation rises in value, national accounts record it as capital income; Zwick and Zidar contend much of it is functionally labor income, which would revise down the scale of the labor share decline that has concerned economists for two decades.
The Troubled History of Government Equity in Technology
A new analysis used Deutsche Telekom as a case study in how partial state ownership distorts technology company behavior. Germany privatized Deutsche Telekom in 1996 but retained a substantial equity stake. The paper traced how that retained stake influenced network investment decisions, pricing strategy, and competitive entry in German broadband markets over three decades, arguing the arrangement produced outcomes inferior to either full privatization or explicit public utility regulation.
Stop screeching about immigration, and get smart about it
Noah Smith argued that political polarization is making it impossible to design rational immigration policy at a moment when demographic decline makes immigration economically necessary. Smith traced the dysfunction to incentives that reward maximalist positions on both sides and called for policy frameworks that separate humanitarian and labor immigration tracks, each with their own criteria, rather than treating all immigration as a single political question.