AI Builders
Digest

Anthropic released Haiku 5.5, now available across AWS, Google Cloud, and Microsoft Azure. Compared with Haiku 4.5 (launched October 2025), the new model shows major gains on nearly all alignment evaluations and far fewer misaligned behaviors. Cache reads on Sonnet 5.5 are also halved to $0.10 per million tokens, making long-running work roughly 20% cheaper. Alex Albert noted the two columns are less than a year apart, yet Haiku 5.5 is both much faster and 75% cheaper.

Sam Altman announced that ChatGPT can now generate a custom UI for the user. He said he has been waiting for this feature for a long time and would hate to go back to the old version of Chat. The capability turns conversations into interactive interfaces, further blurring the line between chat and applications.

OpenAI’s Thibault Sottiaux reported that the team silently re-shipped Codex Cloud on day 3 and it is “pretty good now.” The update has landed across all accounts and he asked the community how it is performing so far. The earlier commitment to ship by EOD PST was met.

Google Labs introduced Playground, an experimental gaming platform that lets anyone create and play their own games with zero coding experience. The pitch: if you can think it, you can play it. It is currently available to users 18+ in the US and quickly drew more than 23k likes.

The Takeaway: A shared company agent beats everyone maintaining their own personal agent because collective context, continuous improvement, and zero individual maintenance overhead win over short-term personalization.

Every CEO Dan Shipper and Head of Platform Willie Williams walked through the full arc of the last eight months. The team first got fully OpenClaw-pilled (earlier versions called Claudius and Moltbot). Everyone ran their own bots; Slack briefly filled with dozens of agents talking to each other. Personalization was fun and educational, but security, reliability, permissions, and ongoing maintenance quickly became a tax. Non-technical teammates never got traction, and even builders abandoned their bots once better tools like Codex appeared.

The decisive insight: instead of every person investing daily effort into their own agent, put that energy into one company-wide agent that everyone benefits from. That agent learns the real workflows, improves with every use, and has become infrastructure the company would struggle without. They have now released it to subscribers.

Willie argued that in the workplace a single org-level agent is structural: it already holds a year of context that any newly introduced personal agent would take months to match. On the personal side the forces reverse—personalization is extremely powerful—and the industry is heading toward a clean split between company agents at work and personal agents at home.

On engineering management, Willie noted the job feels surprisingly familiar. The manager’s field of view is radically expanded by tools like Tend custom feeds that surface decisions and signals across Slack, Notion, and meetings. Yet the core remains a people job: removing bottlenecks and setting direction. Token usage actually drops because the highest-leverage moves are no longer writing code yourself.

Dan’s memorable line on character loyalty: “Mine is named Boo... but Boo will probably be dead. RIP Boo.”

Anthropic’s Cat Wu shared one of her favorite PM use cases for Claude: ask “who used <feature> the most last week? make me an artifact of the top 10 by usage, then reach out and schedule 15 min to chat.” It is the fastest way she has found to get real user feedback.

Claude Code’s Thariq observed that the most common failure case is people working outside their domain of expertise and not knowing how to be precise with prompts and plans, forcing many imprecise iteration turns. The upside: you can simply ask the model to teach you what you don’t know. Because agents make it easier to operate outside one’s expertise, almost everyone hits this situation at some point.

Vercel CEO Guillermo Rauch reminded the engineering community that every program can be ① hardened (edge cases, tightened inputs, error handling) and ② optimized (profiled, benchmarked, rewritten) essentially ad infinitum. Real costs exist in every direction: time, attention, opportunity. Even in an infinite-token world you still need to know when to stop and which trade-offs to accept. Agents will happily keep drilling no matter what.

Replit CEO Amjad Masad noted that desktop AI apps are powerful yet expose users to supply-chain attacks and catastrophic agent mistakes. Replit is building a desktop experience focused on security and reliability, partnering with Microsoft and becoming an early adopter of NVIDIA’s OpenShell.

SPC General Partner Aditya Agarwal argued that “I can build software” is becoming a weaker differentiator. He outlined three paths: 1) work on the thing that is eating software (AI research, infrastructure, tools); 2) team up with a world-class domain expert who already knows where the expensive mistakes hide; 3) shepherd agents to write vanilla code—the least compelling option in his view. The real leverage comes from pairing engineering skill with a technical frontier or a problem worth solving.

Amjad Masad clarified that Replit is not a consumer app; most creators are building actual businesses or working on one. He also noted that math, being the purest field, was bound to fall to AI first.

Box CEO Aaron Levie argued that the compute required for the next stage of AI will be insane: personal agents, agent swarms defending enterprises, agents reviewing all code for security, agents processing nearly all enterprise data in workflows, and background agents running 24/7. This demands both orders-of-magnitude growth in inference volume and the surrounding infrastructure—computers, networking, file systems. We are still in the early stages of that buildout.

Peter Yang noted that anything you ship can now be decompiled and rebuilt by AI. Making money from software will soon be extremely hard unless you possess proprietary data, distribution, or another durable edge. Traditional SaaS is especially vulnerable—expensive and full of human-centric features that agents simply do not need. He also predicted that AI has already solved images, music, and video, and gaming is next; Suno is already good enough that it may eventually overtake Spotify.

Peter Steinberger (OpenClaw) shared his perspective from OpenAI DevDay 2026: while everyone talks about agents, he has been exploring how teams can use them to collaborate more effectively.

Zara Zhang reminded builders: remember that coding is a means to an end. The end matters more than the means.