# Brain-to-Text Breakthrough, LongCat-2.0, and FDA-Cleared LLM Care *By AI High Signal Digest • June 30, 2026* Meta advanced non-invasive brain-to-text decoding, Meituan launched a 1.6T coding model tied to domestic Chinese compute, and UpDoc brought patient-facing LLMs into an FDA-cleared care workflow. The brief also covers new agent products, faster RL infrastructure, and corporate moves around evaluation and AI commerce. ## Top Stories *Why it matters: today’s biggest updates showed AI advancing at once in frontier research, domestic model building, and regulated care.* - **Meta pushed non-invasive brain-to-text forward.** Brain2Qwerty v2 is described by Meta as its highest-performing end-to-end non-invasive brain-to-text decoder, moving beyond character-level output to word and semantic decoding [^1]. Meta says it trained the system on ~22,000 sentences from 9 MEG-recorded volunteers; it reached 61% average word accuracy and 78% for the best participant, and Meta released training code for v1/v2 [^2][^3]. The stated goal is restoring communication for people with brain lesions or disorders [^1]. - **Meituan unveiled LongCat-2.0 as a very large coding model.** The MoE model has 1.6T parameters with ~48B active and 1M context, and Meituan says it was built for agentic coding [^4]. Its published benchmark claims include 59.5 on SWE-bench Pro versus GPT-5.5 at 58.6 [^4]. Third-party posts said the model was trained on 35T tokens and 50K Chinese chips, and explicitly framed the run as evidence of progress on domestic hardware [^5][^6]. - **UpDoc brought LLMs into an FDA-cleared workflow.** One report described UpDoc as receiving the first FDA clearance for a medical device built on patient-facing LLMs, allowing AI management of Type 2 diabetes insulin between doctor visits [^7]. Within physician-set parameters, the system can contact patients, adjust insulin doses, order follow-up tests, and log decisions; the report said this followed a Stanford Medicine clinical trial [^7]. ## Research & Innovation *Why it matters: the strongest technical work focused less on bigger models and more on making search, training, and evaluation scale better.* - **SPIRAL trains search plus aggregation end-to-end.** The paper reports up to 11x better scaling efficiency and up to 15% gains over GRPO on math reasoning when search and synthesis are scaled together [^8]. - **Snowflake open-sourced Arctic RL infrastructure.** ZoRRo is reported to deliver up to 6x actor-update acceleration and 3.5x end-to-end training speedup, while Arctic-Text2SQL-R2 scored 48.7 on Snowflake’s enterprise SQL benchmark, ahead of Gemini 3.1 Pro and Claude 4.7 under the tested conditions [^9][^10]. - **Qwen’s new coding-agent paper focuses on verification horizons.** The work argues reward signals such as tests, LLM judges, and execution traces each stop tracking real correctness beyond a certain horizon, so verification has to co-evolve with the agent [^11]. ## Products & Launches *Why it matters: new launches increasingly center on cheaper multi-model agents and better control over where they run.* - **Devin Fusion** introduced a hybrid harness where a smaller 'sidekick' agent runs in parallel with the frontier agent; Cognition says it cuts the cost of Fable-level intelligence by 35% and can reroute work mid-session as tasks become harder [^12][^13][^14]. - **Cursor for iOS** lets users launch always-on cloud agents or remotely control agents running on their computers, with Live Activities notifications and phone-based review of demos and diffs before merging PRs [^15][^16]. - **Claude in Microsoft Foundry** is now generally available on Azure, giving customers access to Claude Opus 4.8 and Haiku 4.5 with Azure identity, networking, billing, governance, prompt caching, and thinking support [^17][^18][^19]. ## Industry Moves *Why it matters: companies are starting to reorganize around agents as infrastructure, not just features.* - **Meituan is repositioning itself for agent commerce.** CEO Wang Xing said serving 'AI Agents' is becoming as important as consumers and merchants, and Xiaomei—powered by LongCat—will connect to Tencent’s Yuanbao for food delivery and local services inside chat [^20]. - **Arena hit a $100M annualized revenue run rate** eight months after launching its evaluation product, while expanding from preference voting into objective agent metrics such as task completion and hallucination rates in long, tool-using sessions [^21][^22]. - **Spotify’s engineering org is already heavily agent-assisted.** The company says 73% of PRs are AI-assisted, and adding a judge model raised PR success from roughly 25% to 80% after it rebuilt test automation around verification [^23][^24][^25]. ## Policy & Regulation *Why it matters: AI is moving closer to state power as well as regulated industries.* - **Marc Andreessen will join the Defense Policy Board at Pete Hegseth’s request;** commentary around the appointment framed it as another sign that AI, chips, software, and industrial capacity are being treated as strategic assets in U.S. defense thinking [^26]. ## Quick Takes *Why it matters: these smaller items still show where tooling and deployment are heading.* - DeepSeek says the formal V4 release is coming in mid-July with peak-hour pricing, while baseline pricing stays the same [^27][^28]. - OpenAI says Codex shortcuts are getting an upgrade on July 15 [^29]. - Claude Code’s next version will run subagents in the background by default [^30]. - Waymo says it has driven 220M+ fully autonomous miles with consistent safety performance as it expanded to more complex environments [^31]. --- ### Sources [^1]: [𝕏 post by @AIatMeta](https://x.com/AIatMeta/status/2071566924803395741) [^2]: [𝕏 post by @AIatMeta](https://x.com/AIatMeta/status/2071566927420588162) [^3]: [𝕏 post by @AIatMeta](https://x.com/AIatMeta/status/2071566934571954326) [^4]: [𝕏 post by @Meituan_LongCat](https://x.com/Meituan_LongCat/status/2071783587205308721) [^5]: [𝕏 post by @iScienceLuvr](https://x.com/iScienceLuvr/status/2071814463037005865) [^6]: [𝕏 post by @Yuchenj_UW](https://x.com/Yuchenj_UW/status/2071791961250656715) [^7]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2071676839152452019) [^8]: [𝕏 post by @askalphaxiv](https://x.com/askalphaxiv/status/2071283713401409665) [^9]: [𝕏 post by @StasBekman](https://x.com/StasBekman/status/2071628398234087642) [^10]: [𝕏 post by @StasBekman](https://x.com/StasBekman/status/2071628964171543039) [^11]: [𝕏 post by @omarsar0](https://x.com/omarsar0/status/2071763405049123258) [^12]: [𝕏 post by @cognition](https://x.com/cognition/status/2071624568465490170) [^13]: [𝕏 post by @cognition](https://x.com/cognition/status/2071624574270157074) [^14]: [𝕏 post by @cognition](https://x.com/cognition/status/2071624577420009771) [^15]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2071641103191998810) [^16]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2071641104869691671) [^17]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2071653958905467027) [^18]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2071697437136486585) [^19]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2071697439485215150) [^20]: [𝕏 post by @TechBuzzChina](https://x.com/TechBuzzChina/status/2061690347978117624) [^21]: [𝕏 post by @arena](https://x.com/arena/status/2071630464583151727) [^22]: [𝕏 post by @ml_angelopoulos](https://x.com/ml_angelopoulos/status/2071629882057228680) [^23]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2071671418245492926) [^24]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2071671421630320713) [^25]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2071671423480025201) [^26]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2071674050657886378) [^27]: [𝕏 post by @teortaxesTex](https://x.com/teortaxesTex/status/2071537369631625608) [^28]: [𝕏 post by @synthwavedd](https://x.com/synthwavedd/status/2071562051697180681) [^29]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2071639953927438440) [^30]: [𝕏 post by @bcherny](https://x.com/bcherny/status/2071647677591466098) [^31]: [𝕏 post by @Waymo](https://x.com/Waymo/status/2069812796267643194)