AI Builders
Digest

The Takeaway: Real creativity starts by emptying yourself out and building a personal compass first, then letting AI act as a virtual writers' room instead of thinking for you.

Four-time Emmy-winning filmmaker and technologist Fred Graver has moved from National Lampoon and David Letterman through Norman Lear, Disney Imagineers, VH1/MTV, early Twitter and Microsoft conversation design to rebuilding production from the ground up with AI. His core insight: show business has always succeeded by knowing what the audience wants before they do. AI accelerates the first two stages (ideation and production) but cannot supply the unique thing inside you.

Graver breaks the process into three looping phases: ideation and architecture (producing the best possible script), production (everything from storyboards to editing, including hybrid AI animation), and reaching the audience (marketing and distribution). The non-negotiable first step is spending a day or two dumping everything you know about who you are and what you want to explore onto a page by hand or voice before you ever open an AI. Then you feed those documents into Claude or ChatGPT and treat it as a sharp, questioning team. “You have to empty out. You have to do this yourself as a human being. Then the nice part is… now there’s AI.”

He is building Compound-style systems that mix industry expertise (from Selznick to Tarantino) with personal culture documents so AI can mediate disagreements between writers and marketers using historical solutions and reduce the personal drama that usually derails projects. For the next generation of creators his advice is blunt: if you have something to say and the addictive drive to keep saying it, AI can help you say it better; without a compass you simply follow the model toward the mean.

“Success is knowing what the audience wants before the audience knows it.” AI can map the audience and platforms; the improvisational quirks that make characters endearing still come from you.

Google VP Josh Woodward announced Gemini 4 has moved from the Argon version two weeks ago to Carbon. Internal vibes are strong; it is rolling out first to cyber defenders before wider availability. Swyx noted Gemini sitting at #1 on AAv5, prompting him to reopen Antigravity for an early review and calling the simultaneous resurgence of Meta, Grok and GDM a bizarre counter-example to the bitter lesson and a win for model neurodiversity.

Woodward also shipped a deep Gemini–Strava integration that turns years of raw workout logs into a personalized coaching plan.

Former Claude Code member Thariq observed a strange split: technical people feel their skills are becoming useless while non-technical people desperately wish they were more technical so they could direct agents better. The same dynamic is playing out in creative fields—artists worry about AI, yet he wishes he could simply sketch animation ideas and character concepts for an agent to flesh out instead of hunting for references. His sharper point: “courage in shorter supply than genius.”

Vercel CEO Guillermo Rauch announced a partnership with Suhail to serve Particle’s models and inference through the Vercel AI Gateway. Particle’s mission is to make intelligence as accessible and abundant as electricity. Box CEO Aaron Levie noted that agents will consume 1,000× more tokens than humans do today, requiring continued progress on closed frontier models alongside more capable, secure and efficient open-weight models; he is excited about what Particle is building. Y Combinator President Garry Tan added that Particle aims to be the frontier for the smartest model at the lowest dollar per unit of intelligence.

Anthropic’s Boris Cherny shared results from Utah’s pilot giving Claude Code to 34 developers across 10 state agencies: 80% had a working environment within 30 minutes and 71% reported saving 20+ hours by the end of the four-week trial. Concrete evidence of AI coding tools delivering value in real government workflows.

Vercel CEO Guillermo Rauch outlined how Vercel CI attacks the AI coding bottleneck: ① fast, stable infrastructure across 20 worldwide clusters; ② Turborepo’s intelligent skip-and-cache to cut redundant work; ③ Agentic CI that lets agents inspect, abort, trigger and replay runs. He urged teams to have agents migrate to Turborepo and teased more intelligent hooks that will give CI the nuance of an engineer’s judgment.

Rauch also described AI Gateway model aliases as “DNS for models,” enabling traffic steering, freeing teams from stuck old versions, cutting costs and removing decision paralysis.

The official Claude account launched a gallery of ready-made artifacts inside Claude for Legal covering contract review, regulatory horizon scans, cite checks and more. Users start from built-in examples, add their own documents and connectors, then share across the business.

Peter Steinberger of OpenClaw announced the web UI now works offline with smart WebSocket compression and reconnect logic. He got substantial work done on a flight to Vienna despite spotty airplane Wi-Fi, proving the tool is genuinely usable in real conditions.

Peter Yang argues world models (Genie 4 and peers) are still in the GPT-3.5 / “Will Smith eating spaghetti” era—impressive yet not ready. Once they reach GPT-5/6 quality they will transform movies (a director generates an entire dragon-battle VFX sequence from one frame) and games (anyone can build visually and sonically stunning personal titles). He linked a full Genie 4 tutorial showing his own generations.

Former Cursor, Notion and Stripe designer Ryo Lu published a long reflection returning to Engelbart’s 1962 vision of augmenting human intellect and Licklider’s human-computer symbiosis. He criticizes the current story that over-emphasizes automation and agent output while neglecting the relationship between person and tool. The real goal should be systems that help you discover what you actually mean, not systems that take a request and make you wait outside. He wants AI to expand what people can understand, express and change without asking permission.

Replit CEO Amjad Masad noted that GPT 6.1 Sol has taken the lead on the newly released ViBench 2.0 prompt-to-app benchmark.

OpenAI’s Thibault Sottiaux highlighted “Text your dot,” letting users message the AI like any contact, leave voice notes or even issue requests via Siri. He also clarified the team is carefully working on bringing the best of Chat and Work capabilities together later this year, proceeding deliberately rather than rushing a full rollout.

SPC General Partner Aditya Agarwal shared an LP letter co-written with collaborators that distills a decade of frontier investing lessons on spotting talent. They accept the cliché that early-stage investing is 90% about the founder, yet find the statement lacking falsifiability and repeatability, so they mapped concrete dimensions of excellence and how those dimensions translate into companies.