How 1.5 engineers got Stripe Kai in two weeks
Today's top 10 insights for PM Builders, ranked by relevance from LinkedIn, X, and YouTube.
How 1.5 engineers got Stripe Kai in two weeks
#1 in
Claire Vo recapped a How I AI episode in which Sharadh Krishnamurthy demonstrated Kai, Stripe’s “company brain,” which the post says 1.5 engineers and 2 weeks helped deliver. The episode covered projects as governance, skill routing and telemetry, and a skills platform described as working for 10k teammates.
Also covered by: @claire vo đź–¤
#2 in
Udi Menkes recapped Amplitude’s Wave, a production product that identifies issues through analytics, opens PRs, deploys approved changes, and measures results—in one example, opening a PR within 15 minutes. Personalization increased suggestion-card engagement by 160% in an A/B test, while other changes raised successful searches from 92% to 98%, onboarding completion from 6% to 9%, and blog sign-ups by about 45 per week.
#3 đť•Ź
There's An AI For That recapped VS Code Agents, which can plan coding tasks, edit files, run commands, and check results inside the editor while supporting separate sessions for parallel work.
#4 ▶️
5 open source tools that replaced my $320/mo AI stack...
Fireship
A self-hosted developer AI stack combines Ollama for local models, Nine Router for provider routing and fallbacks, Headroom for reversible context compression, Diffy for visual LLM workflows, and Open Hands for autonomous coding agents.
- Ollama provides a command-line interface and API to download and run open-weight large language models locally, keeping prompts private and making inference cost zero; frontier-sized models require substantially more hardware than small models.
- Nine Router exposes an OpenAI-compatible local proxy endpoint, routes requests across model-provider fallback tiers, and can move from a Claude Max subscription to a cheaper OpenAI model and then free providers such as Chinese models or Vertex trial credits; it also tracks usage and compresses tool output.
- Headroom compresses tool outputs, log files, and other context before billable input tokens reach a model provider while caching the original content locally for retrieval; Diffy exposes drag-and-drop visual workflows as APIs, and Open Hands can run self-hosted autonomous coding agents with OpenAI, Anthropic, or Ollama-hosted local LLMs.
#5 đť•Ź
Teresa Torres said an unhandled third-party data issue that causes an agent to quote a $0 price should be caught through production-trace error analysis—but the real fix is defensive code for handling third-party data, not treating it as an agent error.
#6 đť•Ź
Peter Yang shared a 24-minute tutorial on building four games using GPT-6 Astra, Blender, and Godot.
#7 đť•Ź
plori, a plori.ai product, gives an AI agent its own cloud computer, keeps files between sessions, and runs tasks on schedules or when webhooks arrive.
#8 đť•Ź
Guillermo Rauch argued that Linux could prevail across desktop, mobile, server, IoT, robotics, and embedded computing as programmable agents drive personalized experiences, use open-source training data, and orchestrate work in Linux cloud sandboxes. He also cited democratized design and collective global security investment as advantages.
#9 đť•Ź
Peter Yang shared “4 product principles” attributed to @suekhim, including “Never tell the learner the answer,” and said they’re also useful as parenting advice. The post directs readers to watch the full episode.
#10 đť•Ź
Julien Chaumond recapped supportive reactions to an intent to join forces, saying 99% were positive and highlighting praise for NVIDIA’s support of Hugging Face and the open-source AI ecosystem. He said future actions would show the work will continue and called it a chance to make open-source AI bigger.