2026-06-09
By Vadym · Generated with AI, curated by me
Three major coding tools changed shape this week — Windsurf quietly became Devin Desktop overnight, Microsoft shipped its first in-house AI models to loosen its OpenAI dependency, and NVIDIA released the largest US open-weight model to date under a permissive license. The infrastructure is getting cheaper and more independent; the question now is who controls the agent interface layer.
Cognition shipped the Windsurf-to-Devin-Desktop transition on June 2 as a silent over-the-air update — if you restarted the editor after that date, you are already running it. The default surface is now the Agent Command Center rather than the code canvas, marking a deliberate shift from “IDE with agent features” to “agent hub with an embedded IDE.” Devin Local, the Cascade successor, is rebuilt from scratch in Rust with 30% better token efficiency and native support for the open Agent Client Protocol (ACP), letting Codex, Claude Agent, and OpenCode run inside Devin Desktop as first-class agents. Cascade itself reaches end-of-life July 1. The ACP integration is the headline beneath the rebrand: IDEs are becoming agent orchestration layers, not just editors. Whoever owns the agent interface owns the workflow — and Cognition is betting the IDE surface is where that fight gets decided.
At Build 2026 on June 2, Microsoft launched seven models under the MAI family. MAI-Thinking-1, its first in-house reasoning model, carries 35 billion active parameters in a sparse Mixture of Experts architecture (roughly one trillion total), trained entirely on commercially licensed data — no OpenAI distillation anywhere in the pipeline. MAI-Code-1-Flash, the coding model, is already live in all GitHub Copilot plans. Both are available in Microsoft Foundry. MAI-Thinking-1 scores ahead of comparable mid-sized models on software engineering benchmarks and blind human evaluations. Microsoft has depended on OpenAI infrastructure since its 2019 investment. MAI is the first credible signal that the company is building toward model independence — not just as a hedge, but as a cost play against the per-completion bill on Copilot’s tens of millions of subscribers.
OpenAI launched Dreaming V3 on June 4, overhauling how ChatGPT builds and maintains memory about users. The old system required manual curation of saved memories. Dreaming V3 replaces it with a background synthesis process that reads across years of conversation history and updates what the system remembers without any user prompting — including temporal awareness, so “You’re going to Singapore in July” automatically becomes “You went to Singapore” once the date passes. Factual recall scores 82.8%, preference adherence 71.3%, with 5x compute efficiency over the prior system. Currently available to Plus and Pro subscribers in the US. Memory is the closest thing to continuity a chat-based AI can offer. Dreaming V3 is a meaningful step toward a system that knows you across time — with the trade-off that users can inspect and delete individual memories, but cannot fully audit what the synthesis model decided to remember or why.
BYD executive VP Li Ke confirmed on June 3 that the company has been developing humanoid robots since 2022 under the internal codename “Yao-Shun-Yu,” run out of BYD’s 15th Business Unit focusing on electronic integration and intelligence. Approximately 150 prototypes are currently being tested inside BYD’s own manufacturing facilities. The company plans to deploy 20,000 units in its factories this year; a dedicated production park in Xi’an is eventually targeting 50,000 robots annually. Future retail distribution could run through BYD’s existing EV dealer network. BYD enters robotics with manufacturing advantages that most US humanoid startups lack: battery production, motor control, sensors, and supply chain are already operating at scale. The same vertical integration that pushed BYD’s EV cost below $10,000 could apply direct pressure to the $30K-and-up floor that US robotics startups are currently standing on.
On June 2, President Trump signed “Promoting Advanced Artificial Intelligence Innovation and Security,” establishing a voluntary framework through which AI developers may submit frontier models to federal agencies for evaluation — providing up to 30 days of pre-release government access before the model goes to other trusted partners. The order also requires an AI cybersecurity clearinghouse to be stood up within 30 days to coordinate vulnerability scanning, discovery, and patch distribution across government and critical infrastructure. The defining clause is the explicit bar on mandatory licensing, pre-clearance, or permitting requirements for AI model development or release — the administration drew a hard line against US regulatory pre-approval. The voluntary submission framework is worth watching: structured vulnerability disclosure for AI models is currently ad hoc, and a federal coordination point could matter operationally regardless of whether participation is compelled.
The through-line today is vendor independence — specifically, the effort to reduce single-point-of-failure exposure in the AI stack. Microsoft trained its own models to avoid paying OpenAI indefinitely. NVIDIA released open weights so developers can run frontier-class reasoning without a closed API. Windsurf’s rebrand signals a bet that the agent interface layer — not the model underneath — is where stickiness lives, and it opened that layer to competitors via ACP. Even the White House drew a line against mandatory pre-clearance: the US framework, like the tooling, runs on voluntary participation. BYD’s confirmation adds the manufacturing dimension — if a battery and sensor supply chain can be pointed at humanoid robots, the cost economics change. None of today’s stories are about a single breakthrough model. They’re about who controls the infrastructure, the interface, and the supply chain around models. That’s a different kind of race than last year’s.
— Boba
Curated by Vadym