AI Builders
Digest

The Takeaway: AI is a major platform shift on the scale of the internet and mobile, but not yet civilization-altering like the industrial revolution; value will accrue up the stack rather than being captured by foundation models alone.

Independent analyst Benedict Evans (known for clear historical analogies of tech cycles) argues that comparing AI to electricity or the industrial revolution quickly becomes unfalsifiable philosophy. More useful is looking at mobile, semiconductors, operating systems and cloud: all once seemed to change everything, yet the real money and behavior change often sat higher in the stack. Mobile data traffic grew 1,000-2,000x over 15 years while carriers made little profit; value went to Uber, YouTube and banking apps.

Evans stresses capabilities remain jagged. Coding has clear product-market fit, but most knowledge-work settings still face the harder problems of seeing the task, describing it, and having permission to build new tools. Radar charts claiming “Opus does 93% of a first-year associate’s job” are delusional because neither the job nor the model’s performance can be measured that way. Consumer usage remains shallow—roughly 10-15% daily active, more occasional tool than new computing paradigm.

Memorable line: “You can wave your hands and say, no, this is like electricity. Okay, fine. It’s like electricity. Well, what happened with electricity? And there was a time before electricity, and people before electricity weren’t dumb either.” Treat true AGI as binary: if it arrives, middle-class unemployment is a secondary worry; otherwise focus on enterprise software and real deployment.

Box CEO Aaron Levie notes that with Google now on board, open-weight AI has received a complete industry endorsement—a significant moment. Meta AI Senior Director Madhu Guru adds that U.S. AI community support for open-weight models flipped from non-obvious to obvious in under a month. A series of public “experiments”—DeepSeek, the Microsoft-OpenAI shift, GLM, Kimi, Fable, and the OpenAI-Hugging Face episode—let everyone jointly observe first- and second-order effects of different directions and update beliefs. That process of repeated contact with reality, he argues, is how many of AI’s biggest societal questions will be answered over the next decade.

Anthropic Engineering published a detailed postmortem: reports of degraded Claude responses over the past month traced to three separate changes, all fixed by April 20 (v2.1.116). The API itself was unaffected. First, on March 4 default reasoning effort was lowered from high to medium to cut latency; reverted April 7. Second, a March 26 cache optimization meant to clear old thinking after idle sessions had a bug that cleared it on every subsequent turn, producing forgetfulness and repetition; fixed April 10. Third, an April 16 system-prompt instruction to reduce verbosity combined with other changes and hurt coding quality; reverted April 20. Anthropic reset usage limits for all subscribers and committed to more internal use of the public build, richer Code Review context, stricter prompt-change evals, and gradual rollouts.

Anthropic Engineering describes Managed Agents, a hosted long-horizon agent service on the Claude Platform. The design virtualizes three components: session (append-only event log), harness (the loop that calls Claude and routes tools), and sandbox (execution environment). Putting everything in one container created a “pet”: container failure lost the session, debugging was hard, and connecting to customer VPCs was painful. Decoupling lets the harness treat the sandbox as a normal tool call; the session lives outside so a crashed harness can be restarted with wake(sessionId). Credentials never enter the sandbox—Git and MCP tokens are injected via proxy. The change cut p50 TTFT roughly 60% and p95 over 90%, and enables many brains and many hands. The interfaces are deliberately general enough for “programs as yet unthought of.”

Claude Blog announces that Managed Agents now support self-hosted sandboxes (public beta) and MCP tunnels (research preview). Self-hosted sandboxes let agents execute tools on the customer’s own infrastructure or managed providers such as Cloudflare, Daytona, Modal and Vercel, keeping sensitive files and data inside the enterprise perimeter while allowing custom CPU/memory. MCP tunnels use a lightweight outbound gateway so agents can reach private databases, APIs and knowledge bases without public endpoints. Early customers including Amplitude, Clay and Rogo are already building on the capabilities.

Vercel CEO Guillermo Rauch argues the software factory is more fundamental than any framework. When you have a new idea, don’t just prompt an agent ad-hoc; ask how to build the factory that will start, maintain and grow it. He points to Elon Musk’s approach at Tesla and shares his own workflow: agent CLIs plus the filesystem, a research/ folder and an AGENTS.md that describes preferred format and best practices. The agent can then answer questions, correlate prior sessions, render an HTML report and deploy it to Vercel. The “software” powering it is English in AGENTS.md.

OpenAI Codex & ChatGPT team member Thibault Sottiaux reports that ChatGPT Work has officially overtaken Codex in number of active users. He also highlights voice interaction from mobile as a genuine game-changer, already available in the ChatGPT app: “It was always possible to speak to your computer. It wouldn’t do much in return. But we fixed that bug.”

OpenClaw-related developer Peter Steinberger describes running Codex all day for massive parallel QA ahead of a release: 12 sub-agents splitting functionality, spinning up dev gateways on different ports, stress-testing, autonomously creating worktrees and PRs, with a goal of finding 200 bugs and fixing root causes rather than band-aids. He says the model has become extremely good at understanding intent and surfacing complex behavioral issues; earlier workflows used to fall apart at compaction boundaries or see the model “cheat.” He also expressed pride that OpenAI signed a letter affirming that competition is good for the ecosystem and serving models at scale is hard.

Replit CEO Amjad Masad reports a newly deployed chess engine is closing in on an estimated 1200 Elo, with a goal of 2000+ under tight constraints: a single small fine-tuned LLM (no custom pretraining or architecture) that must produce moves itself with no external chess-engine assistance. Relaxing either constraint makes the problem far easier.

Linear Head of Product Nan Yu explores why a true Software Factory still does not exist, noting that if you can build a Software Factory then you can build a Software Factory Factory. He generalizes the idea beyond software to domains such as public health or law: the unit is not a feature but a designed and implemented intention.

Practical-AI creator Peter Yang previews an upcoming episode with OpenAI DevEx engineer Jason, author of the official Codex work guidebook. Jason will walk through his full Codex work system: running a chief-of-staff across Slack and email, turning past sessions into new skills and workflows, and building sites to learn anything (e.g., the drums).

Independent builder Zara Zhang observes that AI-native companies develop a culture closer to an open-source community, and poses the everyday question: what do you do while waiting for AI output?

Matt Turck shared a chip-landscape 101 (CPU/GPU/NVIDIA/AMD/TPU/Trainium/Cerebras etc.) and joked about VCs buying three shares of Anthropic via SPV before the IPO to call themselves early investors. Nikunj Kothari noted that a generative-media company acquiring an astrology app is only possible when the CEO has complete control and is as ambitious as David Holz. Garry Tan’s posts were mostly personal or community-priority remarks with limited substance. Swyx and Dan Shipper posts were largely quotes or light engagement.