Andrej Karpathy
Andrej Karpathy used a weekend project, HackerTube, to rethink discovery on YouTube. He says Claude filtered roughly 43M Hacker News comments spanning twenty years into 6,454 repeatedly recommended videos, in about two hours. The site requires no login and keeps state in the local browser; thumbnails use representative frames from the middle of each video rather than optimized upload artwork. He also offers the comment-analysis data as a zip so readers can ask Claude to build their own version. He used the Mac desktop Claude Code app with Opus 5.5 on Medium effort, helping with a Google Cloud API key, starting a VPS and buying a domain. He then gave Claude the server IP and asked it to update, harden, deploy and optimize the code, reporting that it handled the work. His question is what previously lengthy projects can now be built quickly; the timing and results describe his own exercise.
AI practitioner and interviewer Peter Yang
Peter Yang chooses different tools by his own experience: ChatGPT plus Codex for browser and computer use, Claude for design, video editing and taste, and Grok Bot for always-on agents and scheduled jobs. He considers Gemini underrated for writing but explicitly says he needs more testing. Upcoming Behind the Craft guests include OpenAI VP of Product Ashley Alexander on caring for parents with cancer and where AI can help, and Anthropic designer Nate Parrott with a live example of avoiding generic AI design. He also previews Matt Pocock on engineering fundamentals and software factories, and Jacob Sansbury, whose team size and Claude spending are claims in the announcement. Yang emphasizes interviewing practitioners about topics he cares about. Another post quotes a Google Labs product director on whether a product needs a human at its interface: creation, learning, entertainment and consumption still offer reasons for direct interaction, while some products could primarily be operated by an agent. His tool ranking is personal preference and the interviews are previews, not released episodes or general benchmarks.
Vercel CEO Guillermo Rauch
Guillermo Rauch argues for ownership of intelligence tools, including the harness, data, connections and runtime. He wants local models to keep proliferating and users to retain options even when the largest models remain impractical locally. He links open source with owning one’s agents. In another update, he says Gemini is doing well on two high-complexity projects he is testing, while joking about its highly enthusiastic manner. The ownership argument is his position and the Gemini report is personal experience; the post supplies neither project details nor comparative results.
Y Combinator CEO Garry Tan
Garry Tan highlights Particle-1, claiming that it matches much larger models on agentic retrieval at 1/23 of the price. His post links to more material but does not specify the benchmark, comparator or billing conditions, so the statement remains an attributed claim.
Game builder Thariq
Thariq added a taekwondo master who fights with fire to his game and changed its art from 3D models to generated sprites. He also made a highlight intro video and hopes to have a vertical slice of the core experience the following week. These are development updates and a target, not an announcement of a finished game.
Replit CEO Amjad Masad
Amjad Masad warns users of subsidized AI subscriptions to prepare for the terms to change. He expresses this strongly but identifies no particular subscription, price-change date or contract rule. In another post he describes Replit’s feedback widget: enabling it on a site turns user feedback into prompts for agents to build from. He presents it as a one-toggle feedback entry point. The text does not say changes automatically reach production, so collecting feedback as prompts should not be read as removing review or directly deploying it.
Builder Zara Zhang
Zara Zhang proposes a visual twist on Linktree: make the link tree an actual tree, with a linked demonstration. The brief post establishes that idea without specifying interactions, release status or results.
Box CEO Aaron Levie
Aaron Levie argues that more roles will shift from agents helping people work incrementally faster toward people helping agents do their work. He describes a move from individual contributors with AI assistance to managers of teams of agents. He identifies information-heavy work including coding, cybersecurity, law, biotech and engineering. His reasoning is that parallel agent work can loosen the old relationship between hours worked and output. He expects people and organizations that reorganize around this sooner to see compounding gains; this is his forecast, not proof that every role can parallelize without limits.
Unsupervised Learning — Ep 95: Modal CEO on the Future of Inference, the Chaos of the GPU Market & Origins of the Sandbox Bet
Key takeaway: Durable differentiation in inference may come from a composable infrastructure platform, rather than merely serving cheaper tokens.
Modal founder and CEO Erik Bernhardsson positions the company as a compute and storage foundation for AI applications. After seven years at Spotify, he is now combining fragmented GPU supply, rapid startup and multiple workloads into tools developers can use directly. As models and agent architectures change, he focuses on building blocks people will keep needing.
GPUs are not standardized commodities. Modal works with 25 clouds, with procurement still involving WhatsApp groups, meetings in hotel lobbies and personal introductions. Location, disk, networking, memory and reliability all vary; when a supplier fails, the platform must remove problematic capacity and handle tickets. He says, “It’s very far from an efficient market.” Public prices may also differ from actual procurement prices.
Fast startup reflects investment underneath the product. The team built a file system, container runtime and scheduler, uses tools such as gVisor for isolation, and added CPU/GPU snapshots and fast transfers between nodes. Bernhardsson sees token-only serving as vulnerable to price competition, with more durable differentiation in combining inference, sandboxes, training and batch processing. The initial sandbox prototype took only a few days in 2023, but growing demand required repeated rewrites for scale, speed and reliability.
Agents also change product compatibility. They need fast feedback, clear errors and documentation just as developers do, but old code in training data keeps leading them to deprecated features. Removing documentation and waiting three months does not remove that problem. Modal has therefore adjusted deprecation windows and continuously has models write, build and run code to find where agents struggle.
Senior engineers still own supervision. Bernhardsson has observed especially strong adoption among senior colleagues who can supply extensive tests and rapidly express their ideas through agents; reliability and scaling tail risks still require people. On GPU commitments, he warns that an upfront payment of roughly 30% can leave future bills dependent on later fundraising, so current cash is not the full cost. These procurement, engineering and market observations are his own account, emphasizing constraints and continuing responsibility rather than supporting price or investment predictions.