Key Takeaways
- Builders report AI coding agents are converging on Claude, while one developer now runs an agent nearly 24/7 and expects human review to shift toward product behavior and UI.
- A solo developer shipped 15 Mole Windows releases in under a month, with test code growing to 6.7万 lines and functional code only 30% larger.
- A bootstrapped product went from $19 MRR a year ago to a $1M valuation, and Revid added frame-level object editing for AI video.
- Infrastructure costs and platform friction remain live concerns, from cheap VPS alternatives to iCloud's large-video export experience.
1. AI Coding Agents and Developer Workflow
- Jon Yongfook said that by mid-2026 people were still using multiple AI coding agents, but usage has now largely converged on Claude; he asked whether he is in a "Claude bubble" and noted he has only used Claude and never tried other tools 1. Tony Dinh said that a year ago he could not keep an agent running reliably for five minutes or trust it to decide next steps, but his agent now works almost 24/7; he argues for trusting the agent while building frameworks, guardrails, and quality control around it, similar to collaborating with a person 2. He also expects that in about 12 months he may no longer do code review, with human review reserved for product behavior and UI, meaning taste 3.
- Tony Dinh predicted that SWE jobs will not be replaced by AI by 2030, but for a different reason: AI writes code and builds systems far better than 99% of humans, tokens will become more expensive rather than cheaper, and working with AI will require significant spending, so saving costs may mean hiring human teams; he said he hopes this does not come true 1. Tw93 argued that with code barriers lower, builders should do more non-scalable things to create product differentiation, and said large models raise the default product level to 60 points, making the old passing grade equivalent to today's zero 2.
2. Product Shipping and Bootstrapped Growth
- Tw93 said the Windows version of Mole launched nearly a month ago and has sold to dozens of countries; during that period it shipped 15 official releases from 0.10.0 to 0.21.2, created 20 tags, made 1,084 commits, added 62,000 lines of C# and XAML, deleted 12,000 lines, and now has 96,000 lines of functional code 1. Test classes rose from 2,032 to 3,302, an increase of 1,270, and test code grew from 41,000 to 67,000 lines, leaving functional code only 30% larger than test code; the author called this normal in the AI era, said 482 Windows applications have been tested, and plans to continue 2. New Mole features mostly came from user suggestions: multi-select batch deletion on the analysis page, parallel updates for multiple apps, uninstalling while keeping the app and clearing only data, uninstall scanning about 5x faster, and reduced UAC pop-up interruptions 3. The status page now matches the Mac version, showing battery levels for Bluetooth headphones and keyboards, and the menu bar mascots added four characters; the author said selling Windows software is far harder than Mac software and that dozens of similar products were tried to fill out cleanup and uninstall capabilities 4.
- A repost on Tibo's account said Bazzly grew from $19 MRR a year ago to a company now valued at $1 million, thanking co-founders Tibo and Tom and crediting years of persistence 1. The same account highlighted a new Revid feature: users can pause a video, drag a box around an object, and enter a modification request, or stack multiple notes and send them together, with the AI editor receiving the frame with box-selection marks 2. Quoting that feature, the account called it "this is dope" and said that compared with other AI video tools, Revid's major advantage is editing capability and that it is very strong as an editor; the sample video was generated by Revid and managed by Squad 3.
3. Infrastructure, Platforms, and Costs
- @levelsio argued that companies should stay as lean as possible and said moving from an expensive PaaS to self-hosted VPS is typically 100 to 1,000 times cheaper, listing Vultr at $2.50/month, AWS Lightsail at $3.50/month, DigitalOcean from $4/month, Linode from $5/month, and Hetzner from $7/month; he recommended pairing these with free Tailscale and Cloudflare Tunnel and closing all firewall ports 1. Separately, he filed a feature request with Termius for two-way syncing of open connections and tabs between MacOS and iOS 2.
- Jon Yongfook criticized the experience of extracting large video files from iCloud Photos: dragging to the desktop or exporting shows no loading indicator, forcing users to guess, which he attributed to a San Francisco-style assumption that everyone has a 10 Gbps connection 1. Jack Ellis said Ember looks polished and unlike rough AI output because the developer is experienced and invested more than $30,000 worth of time before releasing it for free 2.
4. Platform Signals and Community Notes
- Tony Dinh said X is under attack from bot accounts posting low-quality AI-generated content, all from verified accounts with more than 30,000 followers 1. Jon Yongfook raised an open question: who will build the sandbox for these sandboxes 2. Ben Tossell said Ramp advertises heavily across podcasts, Broadway, and other channels, but he did not understand its actual business, so he built an interactive visualization site to map Ramp's marketing moves 3.
- Tw93 said he has developed the native Markdown note app MiaoYan for Mac, iPad, and iPhone since 2020, focused on speed and lightness; notes are plain Markdown files that agents can edit, and the miao CLI lets agents search, read, and add notes 1. Derek Sivers recommended the book 我在北京送快递, saying its first half presents the situation of unskilled labor in China and the ending turns when the author quits and becomes an insightful writer 2.
