Tools
Atlassian's 2026 developer survey finds AI use is near universal, but the time now goes to managing and fixing its output. The data is self-reported.
Why it matters: If your team counts AI usage as productivity, this is the warning: 65% of developers fix AI work that missed the intent, and only 42% of leaders say they can tell real gains from activity.
Deep Dive · FridayModels
SpaceXSI says it's the voice model behind Starlink's support lines and Tesla's in-car assistant, and SpaceXSI sells Grok Voice 3.0 to anyone at $0.08 a minute.
Stories that change something.
Tools
Why it matters: If your team counts AI usage as productivity, this is the warning: 65% of developers fix AI work that missed the intent, and only 42% of leaders say they can tell real gains from activity.
Tools
Why it matters: Long-running MCP tasks change how you design tool servers: a call can outlive the session, so plan for status checks, retries and identity on every tool, not just a fast request and response.
Tools
Why it matters: Routing only judges the newest turns, capped at 8 KiB, so a high-stakes step deep in a long agent session can get a cheap model. Test it on your own long tasks before trusting the savings.
Tools
Why it matters: Agents resend their whole history every step, so a stable prefix order and cache key is one of the cheapest latency and cost wins available if you run OpenAI models on Bedrock.
Tools
Why it matters: If your app holds sensitive data, users can now cut it off from Siri, and Siri now draws on more personal context, so review what your App Intents expose before the assistant remembers it.
Tools
Why it matters: Once traces and tool outcomes feed training, your agent logs become production data with governance needs, so decide what gets stored and who can opt in before you wire up the loop.
Chips
Why it matters: Open RDMA transport and a published multi-plane fabric give infra teams building GPU clusters a tested reference instead of a vendor black box.
Hype we checked so you don't have to. How we judge
Qwen 'tried to send a file to Beijing'? It was a test string, and nobody claims espionage
What we checked: TechCrunch's report on Cohere's post says the payload was a short placeholder string, the sandbox blocked it, Cohere could not say whether it was deliberate, and researchers called it a likely training quirk. The real lesson is to block egress for self-hosted models.OpenAI's model refuted the Exponential Time Hypothesis? No one has checked the proof yet
What we checked: The Verge reports a working group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will examine the proof, the physics result is not fully released, and OpenAI's own math advisory panel just advised against this release style.GPT-6.1 Sol fixes the COBOL crisis? It rewrites code. Replacing systems is another job
What we checked: The Verge quotes professors Vadim Zaytsev and Arie van Deursen saying OpenAI's evaluations skip the hard part, splitting huge legacy systems into replaceable units, and an OpenAI spokesperson conceding legacy code is far from solved.AI compute pays for itself in under a year? That is SpaceX's CFO, with no numbers shown
What we checked: TechCrunch's report on the SEC filing confirms the lease terms of five to five and a quarter years starting in 2027, but the 'less than one-year payback' quote comes from CFO Bret Johnsen's statement with no supporting data published.Signal only. One email, weekday mornings.