AI engineering builder Swyx
Swyx says a signal about Gemini’s AAv5 ranking is making him reopen Antigravity for the first time in a long while. He is getting an early review and plans to report back; the post is not a completed evaluation. He places Google’s progress alongside recent moves from Meta and Grok as signs of renewed frontier-model choice. Recalling Sarah Guo’s discussion of model diversity last year, he estimates roughly twice as many frontier labs as when she last appeared on the AIE stage. This preserves his competitive observation without treating the first-place question as a ranking independently verified here.
Google VP Josh Woodward
Josh Woodward announces a Gemini connection to Strava that turns years of raw workout logs into a personalized coaching plan. A separate post provides an activation link without detailing generation or eligibility conditions. On models, he contrasts Gemini 4 Argon two weeks earlier with Gemini 4 Carbon that day, describing a fast-moving frontier. He says internal impressions are good and that rollout starts with cyber defenders before broader availability. This is his stated product rollout plan, not evidence of general access. The updates connect Gemini to personal activity data and a specialized security setting, respectively.
Anthropic Claude team member Boris Cherny
Boris Cherny shares a Claude Code pilot in Utah’s state government. It involved 34 developers across 10 state agencies. In the figures he cites, 80% had a working environment within 30 minutes. By the end of the four-week pilot, 71% reported saving more than 20 hours. The pilot size, duration and self-reported savings remain explicit; these participant results are not generalized to all teams.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux shows texting a dot like any other contact. This makes sharing, forwarding and leaving voice notes straightforward. He also suggests using Siri to text Iggy with a request. Another post clarifies that the team is working on combining Chat and Work capabilities but is not rolling that integration out yet. The plan is to bring their strengths together at some point this year while proceeding carefully. The demonstrated messaging entry point and future Chat/Work integration are therefore separate developments.
Practical AI tutorial and interview creator Peter Yang
Peter Yang argues that world models are still in something like the GPT-3.5 era: impressive, but not fully there. He combines his own and Google’s Genie 4 clips and compares the stage to early Will Smith spaghetti videos. For films, he imagines a director generating an entire effects scene from one frame instead of a VFX team spending months on a dragon battle. For games, he emphasizes core loops, progression and fun mechanics; graphics and sound are already strong, but gameplay lags. He predicts that world models plus good coding models could make building personal games as enjoyable as playing mass-market ones. Separately, he praises ChatGPT with browser use for medical paperwork, without describing the process. He also jokes that selling handmade trinkets at farmers’ markets will escape AI disruption and predicts a market renaissance. These posts address creative capability, paperwork and in-person craft; their projections are not established industry outcomes.
Former Claude Code team member Thariq
Thariq observes a contrast: technical people worry their skills are becoming useless, while nontechnical people desperately want more technical knowledge to use agents better. He sees a similar pattern in creative fields, where artists worry about AI while he wishes for stronger expressive skills. He wants to sketch animation ideas and character concepts for an agent to develop instead of searching for examples. His difficulty is lacking the language to express those ideas, not merely lacking a generation entry point. He connects technical and creative skills to how users articulate and develop what they mean.
Replit CEO Amjad Masad
Amjad Masad shares the release of ViBench 2.0, describing it as a prompt-to-app benchmark. He says GPT 6.1 Sol takes the lead. The post supplies no scores, test set or methodology, so this remains his report of the ranking.
Vercel CEO Guillermo Rauch
Guillermo Rauch divides Vercel CI’s response to AI coding bottlenecks into infrastructure, Turborepo and agentic CI. He says it operates 20 worldwide clusters running millions of microVMs daily to reduce slow, unreliable or unavailable CI. Turborepo reduces work through computed dependency graphs, skipped tasks and caching, and he recommends considering it even without Vercel. The vercel ci CLI lets agents inspect, abort, trigger and replay CI, while Vercel Agent can help optimize it. He expects more hooks and capabilities to judge which tests are worth running, but that remains a future direction. Another post compares model aliases to DNS: an alias resolves to a particular model and enables traffic steering. He argues this helps with outdated model versions, cost and decision paralysis. He also announces a partnership with Suhail to serve Particle models and inference through Vercel AI Gateway. Infrastructure scale, savings and capability descriptions remain the supplier’s own statements.
Box CEO Aaron Levie
Aaron Levie predicts that agents will consume 1,000 times today’s tokens. His reasoning is that they will run workflows, process data, build much of the software and secure enterprises. Supporting this at scale, he argues, requires multiple approaches at the AI frontier. These include advancing closed models and making open-weight models more capable, secure and efficient for large workloads. He is enthusiastic about Suhail’s Particle effort, but the 1,000-fold figure is a demand forecast, not an observed usage statistic.
Designer with Cursor, Notion and Stripe experience Ryo Lu
Ryo Lu reframes his hopes for software as augmenting human intellect: things on a screen should extend thought, helping people move material, find connections and build small systems suited to them. He invokes Engelbart’s 1962 focus on understanding and problem-solving, treating capability as a relationship among people, tools, language, methods and learning. He also invokes Licklider’s 1960 distinction around human–computer symbiosis: computers should help form ideas, not only execute already-defined problems. He welcomes delegated searches, file movement and repetitive work, but argues that not all difficulty is overhead; making things and seeing wrong versions can be how understanding develops. Drawing on Ingalls’s 1981 Smalltalk discussion, he emphasizes personal comprehension and mastery, with systems fitting human minds. People may think through pictures, writing or simulations without sharing a single SaaS dashboard. For AI tools, he wants users to ask why, change one part, see results, undo and learn to make the next change themselves. Rather than measure the future solely by work done without people, he asks what previously excluded people can now understand, express and reshape.
Y Combinator CEO Garry Tan
Garry Tan expects Particle to be at the frontier of the smartest model at the lowest dollars per unit of intelligence. This is his capability–cost judgment; the short post provides neither a measurement definition nor comparative results.
OpenClaw creator Peter Steinberger
Peter Steinberger says OpenClaw’s web UI now works offline and uses smarter WebSocket compression and reconnection. On a flight to Vienna, he says he still got plenty of work done despite patchy airplane Wi-Fi. This is a personal account of continued interface use on a poor connection. The post does not specify which tasks are fully offline or how reconnection synchronizes state.
SPC general partner Aditya Agarwal
Aditya Agarwal shares an LP letter written with Gopal Krishnan about a decade of identifying talent at the frontier. He agrees with the familiar claim that early-stage investing is “90% about the founder,” but finds it insufficiently falsifiable, specific and repeatable. He asks what dimensions constitute excellence and how they translate into companies. The post states the questions without presenting the letter’s framework. It therefore supports his push for more concrete, testable judgment, rather than supplying criteria from the linked full text.
Anthropic AI assistant Claude
Claude’s official account announces a ready-made artifact gallery in Claude for Legal for contract review, regulatory horizon scanning, citation checks and other tasks. Users can start from an example, add their own documents and connectors, and share across the business. Another post links to more information, but the posts do not detail legal scope, accuracy or permissions.
AI & I by Every — Can You Use AI To Make Good Art? — With Fred Graver
Core takeaway: AI is better used as a virtual writers’ room that helps creators think than as a system that decides what they should express. Fred Graver’s career spans National Lampoon, David Letterman, Cheers, Disney Imagineering, MTV and Twitter’s television work, as well as conversation design at Microsoft. He puts that experience into AI workflows while focusing on bringing creators closer to audiences.
Start with two documents: who you are as a writer, including inspirations, tastes, reading and ambitions, and what you are trying to make. Then ask whether they fit, where the audience is and what it cares about. He uses AI to simulate an eight-person focus group containing both enthusiasts and sharp critics to explore ideas. Those simulated characters are not actual audience interviews.
Next, put expert knowledge and team culture into a shared core. Writing, producing, editing, marketing, distribution, operations and legal knowledge form one half; the production company’s approach and participants’ professional backgrounds form the other. He groups the work into ideation, production and reaching audiences, but wants a circle rather than an assembly line. Marketing and business colleagues should participate before creative decisions become fixed, potentially using a concept trailer to explore appeal.
Third, empty out your own thinking before opening AI. He suggests spending a day or two writing or dictating what you know, your ideas and priorities, because letting a model lead first can pull you toward average answers. “None of that is about AI writing for you.” He treats AI as a group of intelligences for thinking things through while continuing to write himself and then asking for review.
His concrete checks ask whether the work sounds like him, whether a believable audience exists on a platform and what is missing between the idea and a script or screen production. A morning check-in agent challenges him on commitments, the selected project and deadlines rather than letting each new idea derail progress. With prior decisions retained, AI can also point out when a change contradicts a judgment made three days earlier.
The hardest part remains the third stage: finding audiences and revenue, not simply producing work. His scenario of a thousand people each paying $10 a month illustrates the sustainable-livelihood problem; the golden age of television will not automatically return. Creators still need direction, distinctive observations and audience understanding. AI can help express those things without guaranteeing appeal or income.