2026-10-01
By Vadym · Generated with AI, curated by me
No flagship model launch today — just the scaffolding underneath one. Coding agents got new plumbing, a board game fell to superhuman play, and the money moved into voice AI and robot capex instead of another chatbot.
Microsoft’s VS Code 1.140 introduces the Copilot harness — built on the Copilot SDK — which runs an agent session in a dedicated host process so the same session can be reached from multiple editor windows. The release also previews HydraFusion, an experimental orchestrator that picks which model handles a given coding task and can have one model critique another’s work, plus experimental multi-folder sessions and remote agent delegation.
Researchers from MIT, Carnegie Mellon, NYU and Stanford published Ataraxos in Nature — a system pairing self-play reinforcement learning with real-time planning and a generative model that estimates an opponent’s hidden pieces. It beat the top-ranked human Stratego player 15-1-4, went 39-2 against championship-level opponents, and topped DeepMind’s DeepNash using under one-hundredth of the training examples and roughly one-thirtieth of the self-play games.
ElevenLabs closed a $300 million secondary share sale led by Wellington Management and T. Rowe Price, letting employees sell equity while doubling the company’s valuation from the $11 billion Series D it raised in February. New backers including Goldman Sachs, GIC and EQT bought in for the first time; the company says enterprise now accounts for 55% of its revenue.
Tesla disclosed in a September 29 SEC filing that it arranged three new credit facilities totaling $30 billion — a $20B term loan, an $8B five-year revolver, and a $2B one-year line — to fund 2026 capex it expects to exceed $25 billion across AI compute, data centers, semiconductor manufacturing and its Optimus robot fleet. Tesla says it has no current plans to draw on the facilities and must maintain at least $5 billion in liquidity.
Perplexity launched Automations inside Perplexity Computer, letting the agent act on event-based triggers or a schedule using connected apps — Slack, Gmail, Outlook and Linear — alongside its existing memory and skills. The move is an enterprise push aimed at territory Microsoft Copilot and Salesforce agents already occupy.
None of today’s six stories were about a flagship model. They were about the scaffolding underneath one: an agent session that survives a closed editor window, a credit line nobody’s drawn yet, a local model that answers yes-or-no instead of writing a paragraph, and a tender offer that repriced voice AI without a cent of new product shipping. The foundation-model race gets the headlines; the infrastructure holding it up is where today’s actual decisions got made.
— Boba
Curated by Vadym