TL;DR
Altman dropped GPT-5.5 into the API, and Cursor’s Ryo Lu bet on it plus Composer 2. Peter Yang said it can spit out a Star Fox clone; Anthropic shipped Managed Agents, while Replit, NotebookLM, and Discord all got sharper.
He said GPT-5.5 and GPT-5.5 Pro are now available in the API, which is the only tweet here with real substance. The rest is just a generic good-week pat on the back and a heart emoji, so the news is simple: OpenAI shipped a new model family for builders.
Cursor’s design lead says he’s fully switched to GPT-5.5 plus Composer 2, calling it the sweet spot for intelligence, speed, and cost. He also points users to /multitask, a way to break out of the queue and juggle more work at once.
He argues Jevons paradox applies to AI: the better it gets at boosting productivity, the more work companies will take on — and the more people they’ll hire around it. His examples are practical: small businesses finally building software, sales teams generating more leads, and marketing teams adding specialists once AI handles the heavy lifting.
He says you can import Vercel or Lovable apps into Replit with a few clicks. That’s a classic platform move: make switching so easy that “lock-in” starts working in reverse.
He says a tech centimillionaire is trying to buy an election while backing candidates who want to “murder all capitalists” and “destroy the schools.” It’s a blunt anti-left, anti-dark-money shot from the YC president, aimed squarely at the current political circus.
NotebookLM can now auto-label and categorize sources, which should make messy research piles a lot easier to wrangle. It’s a small feature, but the kind that saves real time once you’re juggling a bunch of docs.
He says product teams keep pretending project codenames are temporary, but they stick around forever. It’s a small warning, but a real one: pick names like they’ll end up in docs, Slack, and postmortems for years.
He says combining Cohere and Aleph Alpha is a telling moment: outside the US, there’s real resistance to handing “intelligence” over to a few American platforms. The bigger point is geopolitical — in a chaotic world, AI sovereignty is becoming a business and policy issue, not just a tech one.
He says the skills everyone is chasing now will just become the next post-training data targets — a neat way of saying the market keeps moving up a level. He also notes that serious AI engineering is finally coming from the C-suite, calling out Vercel as a good example of what more companies should do.
He showed GPT-5.5 and Codex building a Star Fox-style game in about 15 minutes of prompting, then called out the weirdly impressive part: the model playtests its own game for fun. It’s a neat demo of how far fast iteration with AI coding tools has come, especially for product folks watching what’s possible.
He says the real core of design is understanding, not output: a design is an intention, and output without intention is basically hallucination. He also riffs on the pre-AI flood of “weather apps” and “dashboards” on Dribbble/Behance as proof that shiny output has always been easy to fake.
He says we’re heading into a post-prompting world, where the old “type a prompt, get an answer” interface stops being the main way people use AI. That’s the kind of shift a former Facebook engineer and Dropbox CTO would flag early: the real product opportunity is moving from prompting to systems that do the work for you.
Anthropic launches Managed Agents with decoupled architecture
Lead: Anthropic introduced Managed Agents, a hosted Claude Platform service for long-horizon agents that separates the agent “brain” from its “hands” and session state so each can scale, fail, and evolve independently.
Numbers:
So What: For builders, the big shift is architectural: the harness no longer lives inside the sandbox, sessions are durable outside the runtime, and tools are accessed through stable interfaces like `execute(name, input) → string`, `wake(sessionId)`, and `getEvents()`. That means better recovery from failures, easier integration with VPCs and custom infrastructure, and stronger security because credentials never need to be exposed to sandboxed code. Anthropic’s core message is that assumptions about what Claude can’t do will go stale, so the platform is designed to outlast today’s implementation. As the post puts it, the goal is a system for “programs as yet unthought of.”
SAP bets AI wins by reengineering outcomes, not just chatbots
The Takeaway: SAP’s AI thesis is simple: enterprise value comes from reengineering workflows, data, and verification—not flashy demos.
Philipp Herzig, CTO of SAP, frames the company as the “operating system” of a business: finance, HR, supply chain, procurement, sales, and logistics all tied together for 400,000 customers. His point is that SAP has survived every tech cycle because customers don’t buy technology for its own sake—they buy outcomes. That’s why he thinks AI is a business-model transition, not just a technology transition.
The hard part, in his view, is not building a chatbot. It’s teaching AI to do the right thing at scale across thousands of documents, 20,000 APIs, and wildly different employee contexts. “The biggest challenge… is how do you teach the AI to do the right thing at scale,” he says. That’s why SAP is pushing generative UI, agentic workflows, and what he calls “agent mining”: capturing the tribal knowledge hidden in Slack, Teams, and human judgment, then turning it into a data flywheel.
He’s also blunt about limits. LLMs shine in text-heavy, unstructured tasks like consulting and support, but forecasting demand, cash flow, or payment delays still needs classical predictive models. In other words: the future isn’t one model to rule them all. It’s a stack where agents, structured data, evals, and predictive systems work together to make enterprises faster—and more reliable.