Aug 16 · AI Daily

OpenAI Codex & ChatGPT team member Thibault Sottiaux announced that ChatGPT now supports quick restaurant reservations, along with a batch of other feature updates released at the same time.

Vercel CEO Guillermo Rauch announced that Vercel is the world's fastest AI Gateway infrastructure. This means latency for calling AI models through the Vercel platform will drop further, bringing direct value to developers building AI apps.

Google Labs experimental product Pomelli, a product photography tool for small businesses, now supports turning AI-generated product images into short videos or GIFs, helping merchants create marketing assets at very low cost.

Replit CEO Amjad Masad stated that even without plans to publish to the App Store, building apps for personal use through TestFlight is a great experience. This hints that Replit's mobile AI coding capabilities are maturing.

OpenClaw co-founder Peter Steinberger (former PSPDFKit founder, now an OpenAI partner project) shared two agent workflow improvements: first, adding instructions in a shared AGENTS.md file telling the AI to upload a video in every PR that changes UI state; second, the whole team migrated to building OpenClaw using OpenClaw itself, enabling agent sessions to be shared as URLs.

Peter Yang, an AI tutorial creator, deeply analyzed TweetSpamBot behavior models in X's open-source algorithm. The model analyzes behavior patterns of an account's recent 512 actions (posting bursts, quote-repost frequency, time intervals, etc.) but does not read the actual content of posts. This means savvy AI content factories can still mass-produce content via templated quote-reposts without being detected.

Linear product lead Nan Yu shared a perspective: we shouldn't use "jagged" to describe the capability distribution of AI. That would be like saying a dog's capabilities are "jagged" because they're better than humans at some tasks and worse at others—in reality, it's just "AI-shaped," with its own capability contour.

Meta AI Senior Director Madhu Guru (former Google Gemini/Veo/Nano lead) pointed out: when anyone can build products using AI, true differentiation comes from product feel, domain knowledge, distribution channels, and execution. He also cited the Jevons paradox as an analogy: steam engine efficiency actually increased coal demand—lowering software development costs via AI will lead to more software being built, including problems that were previously economically unviable.

FirstMark Capital VC and MAD Podcast host Matt Turck humorously described the contrast in workdays before and after AI: previously it was "decision-execution-execution-decision-execution..." until 10 PM; now it's "decision-decision-decision-decision-decision" until mental exhaustion sets in by 3 PM.

OpenAI Codex team member Thibault Sottiaux asked the community: what difficult problem did Codex solve for you this week? Where do you look to learn how to push the boundaries of AI coding? This hints that OpenAI is actively collecting real-world usage feedback on Codex.

Anthropic released a new Swift package enabling Apple developers to call Claude through Apple's Foundation Models framework to handle more complex workflows. The Foundation Models framework natively supports calling models from Swift, returning typed Swift values with guided generation in as few as three lines of code. Developers can use Apple's on-device models for quick local tasks while handing off complex reasoning to Claude, achieving an edge-cloud collaborative intelligent app architecture.

Show: The MAD Podcast with Matt Turck (2026-08-06)

Basis co-founder Mitch Troyanovsky shared deep insights on the methodology of building AI agents that run autonomously for hours or even days. Basis's agents can complete complex work like tax filing end-to-end. Key takeaways:

• Agent definition: AI reasoning with autonomous decision-making capabilities. Long-horizon means the agent needs to maintain coherence across time scales exceeding the context window.
• Memento analogy: LLMs have large working memory but no short-term/long-term memory, needing to write notes to their future selves like the protagonist in the movie Memento.
• Three milestone models: Opus 3 (first model to truly understand long context), o1 (reasoning breakthrough), o3 (proving reasoning training scales to improve quality).
• Why coding agents lead: Not just verifiable rewards, but also runtime-verifiable, pure-text environments that are easy to execute, plus massive amounts of high-quality training data.
• Behavior specs: Using markdown to define how an agent should behave, not as direct prompts given to the agent, but as standards to evaluate whether the agent followed the correct process—similar to process guidelines in human organizations.
• Process vs. outcome: You cannot just look at whether the final result is correct; you must verify the process—just like an accountant citing IRS primary sources instead of Wikipedia, even if the answer is the same.
• Context engineering as runtime training: Context should be treated as training data, except training happens at inference time, making the total volume much smaller than pre-training, where every token is more precious.

Cursor has been officially acquired by SpaceX, completing the acquisition process initiated in April of this year. Leveraging SpaceX's massive global GPU cluster, Cursor will build stronger and lower-cost models, providing more powerful models to customers at lower prices. Grok 4.6, released this Wednesday, is an early embodiment of the results of this collaboration.

Box CEO Aaron Levie commented that this is an "amazing outcome"—Cursor executed its applied AI strategy perfectly, breaking several mental models: the agentic coding market is much larger than expected, there remains massive space for innovation between users and underlying models, and Cursor's neutral model layer, post-training optimization, and GTM strategy form a complete applied AI paradigm.

Swyx (Smol AI / Latent Space Podcast) commented on Databricks completing its Series M financing at a $188B valuation (in "IPO is lava" mode), joking that the M stands for "we are going to kill so many meetings." This marks the continuing trend of climbing valuations for AI infrastructure companies.

Google VP Josh Woodward (VP, Google Labs / Gemini App / Google AI Studio) announced that Gemini 3.7 Flash is now live in the Gemini app. This is Google's new generation of fast reasoning models, open to consumer users.

Y Combinator CEO Garry Tan stated that using GStack (based on Claude Code) has vastly improved after Fable 5—"one-way-door" issues that previously required repeated confirmations in Claude Code can now be directly accepted in full with satisfaction. This indicates that Anthropic's Fable 5 model has made substantive progress in code generation quality and decision-making capabilities.

An analysis of 14,419 self-published e-books shows that AI-generated books are crowding out the human author market through volume rather than quality. From Q1 2023 to Q1 2026, total book volume grew 38.3x, while quarterly revenue grew only 8.9x. Across eight genres, per-book revenue for non-AI text books dropped in seven of them—even human works with no detected AI text were impacted by market dilution. This study turns previous speculation into quantifiable evidence, which may influence fair use judgments in copyright lawsuits.