Key Takeaways
- Anthropic's global affairs lead told the IMF that Claude is expected to take on 80-90% of frontier model training next year, up from 3% in April.
- Arvid Kahl argues AI makes public codebases easy to audit, so developers should scan and fix vulnerabilities before others find them.
- @levelsio is investing in what he calls remaining AI-era moats, including paid city-data APIs, community growth, X revenue, and a large subscriber base.
- Jon Yongfook questions proprietary-data moats when data is exposed through APIs, and describes a common indie SaaS playbook.
- Jack Ellis disputes a claim that Codex leaked an env file, saying the user published it to Cloudflare Pages.
- Ben Tossell detailed a full day building an agent setup with email updates, a shared task list, and a work log.
1. AI Models and Infrastructure
- Anthropic global affairs head Michael Sellitto told the IMF that the company expects Claude to handle 80-90% of frontier model training next year, compared with 3% in April, according to a message cited by Jack Ellis. The figure is a forward-looking company projection, not an independently verified measurement. — via 1
- Jack Ellis disputed a claim that Codex leaked an env file, saying the user had published the env file to Cloudflare Pages themselves. He warned against blindly trusting LLMs and said developers are responsible for the code they ship. — via 1 2
- Ben Tossell said he spent a full day building and debugging his agent configuration: a @bot now opens threads in @get_bb_app for formal coding work, sends bot updates by email, and updates a shared task list and work log. He described the setup as a Linear-style “vibe-slopped superhuman + todos,” said the bot can be tagged in comments, and said the email layer is based on @RhysSullivan's Executor, with domain, expiry, host, and linked local repository added. — via 1 2
2. Developer Security and AI-Era Moats
- Arvid Kahl argued that AI makes it easy for anyone to audit public codebases for vulnerabilities and security issues, so developers must continuously run such scans and fix problems early or they will eventually be found and exploited. He said closed-source code carries less risk, but front-end code and its API calls can still serve as an entry point for investigation. — via 1
- @levelsio said he is chasing the remaining “moats” of the AI era, listing paid city-data APIs at $2,490 to $9,990 per year, a $1 community membership to expand network effects, about $15,000 per month in X revenue, and 187,279 combined blog email and Telegram subscribers. He also listed eight offline meetups in the next 30 days, hotel-booking product improvements including a “longevity hotel” filter, an annual vibe-coding game event with about 1,000 participants and more than 1 million players this year, a 1990s retro PC mini replica, and about $1,000 per month in merchandise revenue. On capital, he said he holds $SPCX stock after Cursor was acquired and plans to keep it, bought IB01 ETF bonds to hedge AI-bubble risk, and is considering more land and coastal property; he also invests in startups but questioned which startups are still worth backing when large AI and tech companies absorb everything, calling robotics interesting but easy to enter. He said he is not good at or willing to pursue regulatory-license barriers and would sell parts of projects to increase investable capital. — via 1 2
- Jon Yongfook questioned the “proprietary data moat”: if data is provided through an API, terms of service alone are hard-pressed to stop scraping, especially when the scraper is in a different jurisdiction. He also summarized a 2026 indie-developer playbook: complain on X that a mature SaaS product is too expensive, get agreement and use that to “validate demand,” then build a competitor 83% cheaper; he called the play quite viable. — via 1 2
3. Product Launches and Workflow Experiments
- Ben Tossell relayed that WebGPU support is coming soon to uzu, with an early preview available to try. He also relayed that Delegations is now generally available, allowing teams to hand work from existing tools to Droid, and shared an open-source Claude Code-related project that fixes copy-paste, links, and other interactions so they do not affect other UX. — via 1 2 3
- @levelsio said a tweet helped push a product's filter response time down further to 87 milliseconds, and added a “Don't want to join? Tell me why” entry on the checkout page that opens a feedback window. He also relayed that letting Claude run freely on a Hetzner server with an empty Stripe account and a 30-minute improvement loop produced a small app. — via 1 2 3
- Jack Ellis said he completed new infrastructure for analytics data ingestion, new infrastructure and code for customer-site installation and debugging, a data-model upgrade after extensive benchmarking, and the launch of new marketing-tool infrastructure in one day, work he said would previously have taken weeks. — via 1
