Software, AI and the tools I use every day. First-hand analysis of what's happening — what changes in practice, and when it isn't worth it.
24 notes
Three TechCrunch Disrupt 2026 announcements are ticket promos, but the agenda they reveal works as a leading indicator of what you'll be asked to build next year. How to read it without buying the hype.
Opus 5.5 landed on AI Gateway and Copilot, but it brings two changes that turn valid requests into HTTP 400: forced tool use and fixed thinking budgets are gone. What to check before you switch models.
OpenAI added Sol and Luna to the GPT-6 family, with Copilot and AI Gateway already serving them. What changed isn't model quality: it's that the cheap tier is now enough for real agentic work.
The probing that preceded the Hugging Face hack ran for two months and nobody saw it. The practical takeaway: your telemetry is built for human attackers, not for agents that explore slowly.
Meta bundled the paid features of its three apps and put a generation quota on Instagram, Facebook and WhatsApp. For anyone building software, the news isn't the price: it's that the meter is already installed.
Nvidia puts one of its own on stage to explain why robotics still hasn't had its mass-adoption leap. The useful takeaway isn't about robots: it's about how you design software that acts in the real world instead of just answering.
The repetitive work around an event is a pipeline with no technical owner. How to turn a runbook into code, what a small team gains from it, and when you're better off doing it by hand.
The TechCrunch announcement gives you a deadline and zero numbers. How to evaluate a booth as an acquisition channel, what to ask before September 18, and when to skip the edition.
An email app buying a notetaker isn't a product move: it's a race to capture the raw context agents need to be useful. Here's what that means if you build software.
Three agenda posts from TechCrunch Disrupt 2026 and one analysis from Latent Space point the same way: if you build on foundation models, your edge is moving from the product to the customer's context.
Jacob Coxon's resignation and the security incident report change the risk calculus for anyone putting AI agents into production. Not because of the apocalypse: because of what already fails today.
Shopify is migrating its mobile apps from React Native to Swift and Kotlin. Cross-platform didn't die — what changed is who pays the cost of the abstraction. How to read it if you're picking a mobile stack.
Vercel's Agents API integration takes the agent loop and session state off your plate, but ties you to two vendors for a single feature. What your architecture gains, and what it loses.
Mann isn't remaking Heat: he freezes it as canon and extends it backward and forward. It's the same architectural decision you make when an old system still brings in revenue and you can't touch it.
OpenAI stopped selling generation and started selling directed editing. For anyone producing assets every day, that changes the workflow more than any visual quality improvement.
Meta is trialing robots to swap cables and power servers on and off in Iowa and Virginia. The useful takeaway isn't fear of replacement: it's that physical maintenance of the cloud becomes programmable, and your app has to survive a software-driven blackout.
OpenAI published an AI-generated solution to the Navier–Stokes problem with a formal proof in Lean. What matters isn't the math result: it's that mechanical verification and massive compute change which tasks are worth delegating.
Cymphony raised $30 million to secure agents inside the enterprise. The signal isn't the amount: it's that the identity and permission model you use today was never designed for something acting at machine speed.
Atoms might be getting into the robotaxi business, and there are still no dates, no fleet, no architecture. What we do have are two nearby signals — Roland inside the DAW and Anthropic's Labs team — that say plenty about where AI belongs in a product.
OpenAI published internal data on how coding agents are changing its research. The interesting part isn't that they speed things up: it's that the metric shifted from lines written to experiments per week.
The cost of producing text collapsed; the cost of reading it didn't. That changes who foots the bill on a PR, an ADR or an issue — and why your team started skimming everything you send.
Ollie promises not to train on your data and GitHub simply stops exposing identities. One is policy, the other is schema. For anyone designing APIs, the difference decides what happens the day something goes wrong.
The Thinking Machines round and Vercel's new Basic tier tell the same story from both ends: capital is overflowing at the top, every build minute is being squeezed at the bottom. What to do about it if you work on this infrastructure every day.
TechCrunch closes its Side Events call for Disrupt 2026 today. The signal isn't the deadline: it's that a conference's value has moved to the parallel events the community itself organizes.