Key Takeaways
- Executor v2 is out, described as a major rebuild aimed at covering all previously requested use cases, with Zapier/Retool-style capabilities driven entirely by code and exposed to agents.
- DeepFinance launches as a long-horizon benchmark for high-end finance work: 21 tasks built on real trading data by experts from Evercore, JPM, UBS, HBS and Wharton, each taking a human professional roughly 20–40 hours.
- levelsio shut down "Tommy," a 13-year-old $300/month Linode VPS hosting 100+ sites, citing Akamai's maintenance-mode stewardship and AI-era security concerns; he moved to Hetzner and split sites into isolation.
- levelsio also plans to publish a list of 10,000+ blocked AI spam reply accounts for others to bulk-block, and reports the reply sections quieted after his warning.
- Tw93 lays out a product engineer's AI design playbook: solve the problem > be easy to use > look good, with concrete rules on copy, numbers, color, type and consistency.
- Jon Yongfook argues narrow niches are the moat against broad AI coverage, and that third-party AI agent ordering is impractical at DoorDash's scale: hundreds of agent orders against ~330M monthly orders, about 0.0001%.
- A builder's Notion-to-bookshelf project maps page counts to spine widths, pulls covers from Open Library without an API key, and self-hosts images for CSP compliance.
1. AI Products and Benchmarks
- Executor v2 has been released and is described as a major product rebuild intended to cover every use case users had previously asked for. Its positioning is to make Zapier- and Retool-class capabilities fully code-driven and to expose all of that capability to agents. 1
- DeepFinance is presented as the first ultra-long-horizon benchmark for high-end finance work, with 21 questions built on real trading data by experts with Evercore, JPM, UBS, HBS and Wharton backgrounds; each question takes a human professional roughly 20–40 hours. The publisher argues existing finance benchmarks only test document Q&A or simple numerical correctness and cannot cover end-to-end workflows full of assumptions and conflicting evidence. 1
- levelsio says Anthropic launched a personal-finance test feature for US users that connects accounts with read-only access and cannot move or invest funds, and reads it as AI labs entering nearly every industry one by one; he cites his own Claude Code-built expense tracker as evidence. He also amplified a guess that AI labs may be systematically reverse-engineering software, apps and games on the internet with unguarded unreleased models, using reconstructed source for training while building guardrails for themselves. 1 2 3
2. Infrastructure and Platform Operations
- levelsio shut down "Tommy," a Linode VPS he ran for 13 years at $300/month hosting more than 100 sites; he says it served 300 million visits a year and 115,000 paying customers. He attributes the move to Linode entering maintenance mode after the Akamai acquisition and migrated to Hetzner, splitting and isolating each site, which he frames as mainly an AI-era cybersecurity decision rather than a performance one; a collaborator who worked with Tommy for nearly 12 years said it taught him Linux, sysadmin and performance optimization. 1 2
- levelsio says he will publish a list of more than 10,000 blocked AI spam reply accounts and submit it to X and third-party lists so others can bulk-block; he argues AI replies crowd out people who engage seriously and that the growth tactic does not work. He reports reply sections became noticeably quieter after his warning and showed a self-built tool for blocking X reply bots. 1 2 3 4 5
3. Product and Design Practice
- Tw93 argues from a product engineer's perspective that design is a finishing touch, with priority order: solves the problem > easy to use > looks good. On usability he advises walking the main flow as a non-technical novice, fixing friction and gaps so no manual is needed, and writing copy in local language conventions, referencing Apple, Windows and well-known sites, and letting the product speak where possible. 1
- On data and interaction, Tw93 says numbers must be understandable and readable, units should follow best practices, accuracy matters and validation cannot be skipped; confirmed actions should not trigger repeated dialogs, only irreversible effects warrant a second confirmation, and dangerous actions should avoid large red areas. Visually he recommends one accent color per page, avoiding orange, terracotta and old enterprise bright blue, using warm grays, at most four font sizes per page, and consistent line, letter and paragraph spacing with attention to contrast and accessibility. 1
- Tw93 also advises sparing use of italics, monospace, all caps and middle-dot joins; one verb per action across the product and few marketing urgency words; consistent, restrained buttons, spacing, radii, dividers, modals, animation and shadows, with flat pages by default and shadows only for genuinely floating modals, and concentric nested radii. For localization he says lay out for the longest copy first, prioritize meaning over resizing or truncation, test across device sizes, and let an agent produce three options to pick from and generate before/after images in a PDF for comparison. 1
- Jon Yongfook argues broad categories should shift to niches: Claude may cover law, design and finance broadly but is unlikely to reach something as narrow as "an AI for picking the right vegan fishing gear in Hudson County, New Jersey," so niche positioning is a moat. He questions third-party AI agent ordering as impractical, since chatting back and forth with an AI is more effort than a few taps in an app, with scheduled repeat orders the only plausible case and apps likely already having scheduling; DoorDash CEO Tony Xu cited hundreds of cumulative agent orders, which against roughly 330 million monthly DoorDash orders is about 0.0001%. 1 2
- Jon Yongfook frames software development as factory and artisan: the factory scales repeatable output through process, the artisan applies human judgment and experience to finer products, and good software needs both. Artisans own core features and their behavior and experience; the factory handles automatic bug fixes, auto-merges, auto-written docs, automatic API/MCP updates, rolling plugin generation for a plugin library, and the琐碎 functions artisans should not worry about. The division assumes software is ultimately used by people; fully agentic software would be "game over," but that is still far off. 1
4. Builder Experiments and Community Notes
- A user built a personal reading shelf modeled on Ben's bookshelf idea: ten years of reading records in Notion, pulled via API, with covers and page counts from Open Library for free and without an API key. Spine width maps to page count (90 pages at 26px, 700 pages at 72px), color comes from the cover, height is jittered by a slug hash, and covers are downloaded and self-hosted to satisfy CSP and avoid Open Library slowdowns; the builder once spent about $5 generating realistic spines for 140 books at roughly $0.03 each, but the image proportions were wrong and needed aspect-ratio and prompt fixes. 1
- Tibo argues that as programming costs approach zero, apps are expanding features and trying to cover everything. He says being in a founder-heavy community surfaced good new products, and he subscribed to @pkp_io's @vidkraken_com (declaring no conflict of interest) with a good experience so far; he describes a reciprocal loop where the other person subscribes to his paid community while he pays for their product. 1 2
- Nathan Barry acknowledges his newsletter still has room to improve and says @chenellco broke down five levers behind large newsletter growth on a show and applied the method to his newsletter. He also argues you cannot claim to be good at AI unless you can prove it with a revenue curve. 1 2
- Arvid Kahl amplified a satirical post listing claims that SaaS is dead, everyone will build their own, and AI will somehow kill SaaS, without specific arguments. 1
