X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
He publicly asked users about Codex’s shortcomings: “What surprises you that Codex still does badly, and that we should have gotten right long ago?” Direct feedback collection by a core team member usually means prioritization for the next iteration, making this a good time for readers to submit requests. He also shared a funny interaction with Sol: after he sent dozens of salute emojis, the model solemnly performed multiplication and subtraction, arriving at “negative 332 salutes” and explaining that “you owe a debt of 332 salutes.” He called it “silly, but it made me laugh.” Together, the posts show his recent interest in Codex’s capability boundaries and model behavior under absurd input.
https://x.com/thsottiaux/status/2073551549494596079
https://x.com/thsottiaux/status/2073554978053005607
Nan Yu, Head of Product at Linear
He offered a clear view of bug finding in the AI era: the best approach is using a product yourself and trying to break it, rather than staring at code to “deduce” where problems might arise. Most bugs are not caught in code review, he argued; review’s real value is guarding architecture and API design and controlling technical-debt growth. For teams generating large amounts of AI code, this provides a practical division: humans perform destructive product-level testing while reviews focus on structure. He joked that he used to swear constantly in flow while writing code by hand, so similar behavior during AI coding is “basically AGI.” Another joke targeted responsibility for AI code incidents: if all production database tables are deleted, who gets fired—the model or you? His recurring theme is that in the age of AI-written code, human value lies in testing and architectural judgment.
https://x.com/thenanyu/status/2073410299680428445
https://x.com/thenanyu/status/2073412466436878666
https://x.com/thenanyu/status/2073410944969932877
Cat Wu, Anthropic Claude Code Team
She shared a working detail of Claude Fable 5: during retention analysis, it proactively used propensity score matching without being asked—matching users by activity to compare similar groups. This standard statistical method guards against selection bias and typically requires the analyst to think of it. She sees it as evidence of generally improved judgment in Fable 5, noticeable not only in data analysis but in writing emails and documents in Cowork and debugging complex errors in Claude Code. For data users, this suggests models beginning to bring their own methodology, proactively applying greater rigor when users have not recognized bias risks.
https://x.com/_catwu/status/2073439890482794966
Guillermo Rauch, Vercel CEO
He turned Vercel AI Gateway’s historical token consumption into an animated “spending race,” aggregating real monthly usage from millions of developers and trillions of tokens. He highlighted three patterns: shifting rankings among labs, Anthropic’s sustained lead, and the rise of open-weight models. Since AI Gateway is a common entry point for many developers calling different models, the data offers a cross-section of revealed preferences closer to production choices than benchmark scores. Readers assessing trends in actual model market share may find it worth viewing.
https://x.com/rauchg/status/2073563586270781674
Garry Tan, Y Combinator CEO
He used San Francisco’s housing debate to restate his supply-side position: SF must build housing quickly, with policies encouraging and expanding supply rather than continuing to subsidize demand. He publicly backed Mayor Lurie’s YIMBY family zoning plan and harshly criticized NIMBY opponents, saying they “want to destroy housing and jobs and make the city less safe,” while his side “wants to build housing and make the city safer.” His framework is simple: housing is a supply problem, and proposals that do not increase supply are delays. For readers following SF’s technology ecosystem, this continues the YC leader’s support for the city’s “boom loop.”
https://x.com/garrytan/status/2073575065917280331
https://x.com/garrytan/status/2073558419412500564
https://x.com/garrytan/status/2073558154873593926
Peter Steinberger, OpenClaw Creator
He previewed a new version of his tool showing exact expiry times for usage resets at each tier, enabling what he calls valuemaxxing—maximizing subscription value within an allowance cycle. For heavy users of periodically limited products such as Claude and Codex, this visualization can directly change when they schedule large tasks: knowing when allowances refresh lets them put the most token-intensive work at the cycle’s start. Consistent with his OpenClaw background, his focus remains productivity tools around the agent toolchain.
https://x.com/steipete/status/2073482942513565713
Official Blogs
Claude Blog — Building intelligent apps for Apple platforms with Claude in the Foundation Models framework
Anthropic released a new Swift package letting Apple developers call Claude directly through Apple’s Foundation Models framework. The original framework enables calls to Apple’s on-device models with three lines of code and returns typed Swift values, suiting quick local tasks such as summarization and extraction. Requests needing multistep reasoning, code generation, web search, or code execution for data analysis can now be handed to Claude, with responses streamed into the same SwiftUI view. The key design is @Generable annotations returning typed Swift values, so Claude API inputs arrive as clean structured data rather than raw user text. Official examples include a journal app generating daily prompts locally before Claude connects themes across months of entries, and a learning app explaining terms locally before switching to Claude when a student asks why they matter to other material studied. The official framing is “one experience, powered by the best-suited model at every step.” It supports iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27. Developers add the package and sign in with an Anthropic API key; the package handles streaming, tool calls, and structured responses.
https://claude.com/blog/claude-for-foundation-models
Podcasts
The MAD Podcast with Matt Turck — Cloudflare CEO: The Internet's Business Model Is Dead
Key takeaway: In the first half of 2026, bot traffic exceeded human traffic on the internet for the first time. The advertising model that sustained the internet for nearly 28 years does not work for bots, forcing a rewrite of the internet’s business model over the next five years.
The guest is Matthew Prince, Cloudflare co-founder and CEO. Cloudflare sits in front of a substantial share of global internet traffic, giving him an almost unparalleled view through radar.cloudflare.com. The topic matters now because change is far faster than expected: in fall 2025, his team predicted bots would overtake humans by late 2027; in March 2026 it revised that to the first half of 2027, but it happened in the first half of 2026.
Several points are especially substantive. First, agents are driving this: a person buying a camera might visit 5 websites; an agent visits 5000. Prince expects bot traffic may reach 1000 times human traffic within five years, making the doubling of traffic in two weeks during COVID “look insignificant.” Second, micropayments offer a business-model path. Cloudflare processes approximately 500 million requests per second and estimates a substantial share could be monetized with micropayments, requiring support for 10 million financial transactions per second on day one and scaling to 100 million. Visa, the largest payment network, handles fewer than 100,000 per second. Cloudflare is working with Coinbase, Stripe, and x402, activating the HTTP 402 Payment Required status code that was never truly used. Control is already working: with access-control tools, major publishers such as Conde Nast and Dotdash Meredith secured significantly better AI-company agreements. Third, internal adoption: 93% of Cloudflare’s R&D staff use AI coding tools. The turning point was senior skeptic and Workers creator Kenton Varda spending a month trying to prove the tools were rubbish, then declaring a 100-fold productivity improvement. Fourth, organizationally, Prince cut more than 20% of staff not because business was struggling, but because AI replaced middle management and “measurement” roles, expanding average management spans from 6 to 12 people. He considers Meta’s 50-to-1 target too aggressive. He offered the industry’s most generous severance, with continued stock vesting, and predicted almost every company would undergo similar layoffs within 6 to 12 months. His sharpest remark: “Many CEOs are waiting because they’re afraid of looking bad by moving first. With respect, that’s cowardice, because the cruelest thing you can do to your team is wait.” Fifth, a security warning: models’ vulnerability-finding ability is so strong that he expects “a Log4j-level vulnerability every week” for the next two years. Yet this is a feint: after two years, AI review will substantially improve software quality. Cloudflare already has an agent trained on ten years of incident data reviewing every code release and configuration change, after which minor incidents fell dramatically.
His framing of the content ecosystem is memorable: an LLM is a mathematical model of human knowledge, like Swiss cheese. AI companies will pay for net-new knowledge filling its holes, not the Nth report on the same topic. Spotify paid music creators approximately $12 billion last year, and one Danish creator earned around €40 million annually by writing songs for “unmet search requests.” AI licensing revenue at the local newspaper Prince bought in Park City is expected to exceed display advertising this year. The traffic era rewarded enraging headlines; if micropayments reward people filling knowledge gaps, the internet may become healthier.