MAI-Thinking-1 is now available in Microsoft Foundry
Today's top 20 insights for PM Builders, ranked by relevance from X, Blogs, LinkedIn, and YouTube.
MAI-Thinking-1 is now available in Microsoft Foundry
#1 đť•Ź
Mustafa Suleyman says MAI-Thinking-1 is his group’s first reasoning model, built from scratch and now available in Microsoft Foundry.
#2 đť•Ź
Sebastian Raschka shared links to Meta AI’s introduction of Muse Glimmer and the `meta-models` Hugging Face page for Muse-Glimmer-30B, emphasizing that it is real—not an April 1st joke.
Also covered by: @Fireship
#3 đť•Ź
Google DeepMind announced SL2T, its sign language-to-text model powering new Android features for Deaf and hard of hearing users. Starting with American Sign Language-to-English on Pixel 11, users can sign directly into Gboard and Live Transcribe instead of typing.
Also covered by: @Sundar Pichai
#4 đť•Ź
Sundar Pichai announced that the Pixel 11 lineup is here, designed for Gemini Intelligence. New features include Rambler for natural voice input, Magic Capture for effortless photos, and HiLight, which subtly glows for important calls when the phone is face down.
#5 đť•Ź
Philipp Schmid says a small Gemini API update now lets builders combine Google Maps and Google Search tools with Gemini to build location apps.
#6 đť•Ź
LlamaIndex 🦙 said it released ExtractBench the previous day, benchmarking 14 systems across 370 enterprise documents. It reported that frontier VLMs scored 8.9–35.8% F1 on the longest documents as recall collapsed, while its iterative Agentic Plus tier achieved 96.1% F1 on long-list tasks and was the only system to hold performance flat as documents grew.
#7 đť•Ź
Claude announced that the Claude for Chrome side panel now runs the same Claude Cowork session as the desktop, web, and mobile apps, letting users start in a tab and continue elsewhere through account-linked sessions.
#8 đź“ť OpenAI News
How enterprises put AI to work - OpenAI published two reports showing enterprise AI is shifting from assistance to execution: as of June Codex produced 64% of combined Codex and ChatGPT output tokens, and frontier firms (top 10% by usage) now generate 8.3× as many output tokens per active user as typical firms (up from 2.6× in January). Frontier adopters use advanced capabilities far more—21% of weekly active users at frontier firms use Plugins (vs. 9% at typical firms) and 19% use skills (vs. 3%)—Codex weekly users grew 108× in legal, 41× in sales, 41× in recruiting, and 26× in marketing (vs. 5× in engineering), and OpenAI recommends connecting agents to company context/tools, strong permissions and governance, and shared workflows to scale adoption.
#10 in
Guillermo Rauch recommended trying `npx sandbox@latest sh`, saying Sandbox now includes a customizable default set of pre-installed tools and feels faster than a local machine.
Also covered by: @Guillermo Rauch
#11 đť•Ź
Cognition says Grok 4.6 shows particular strength within Devin in thorough code exploration and root cause analysis before modifying code.
Also covered by: @Cognition
#12 đť•Ź
Harrison Chase demonstrated a Managed Deep Agent that scans Hacker News and optionally X, drafts three posts, saves them to durable memory, and sends them to Slack. The linked video covers Slack channel integration, custom tools, and memory.
#13 đť•Ź
Santiago shared three highlights on Deepgram’s Flux TTS, focusing on context retention, streaming, and handling interruptions in AI-powered phone agents. He recommends that voice-agent builders check it out.
#14 ▶️
AgentMail: The Email Inbox Built for AI Agents
SyntaxGTM
AgentMail is used to create an API-first email inbox for a Syntax Cinema request-form agent, with messages, threads, labels, and attachments handled as API resources and replies generated through a locally running Codex instance.
- AgentMail is positioned separately from transactional email services such as Postmark and Resend: Postmark primarily sends email, while Resend recently added an inbound receiving feature; AgentMail provides inbox-oriented API resources for sending, receiving, threads, labels, and attachments.
- In the Syntax Cinema flow, a Netlify form submission sends an email to a programmatically created AgentMail inbox; the local Codex-connected agent receives the request and sends an email response back to the requester.
- The speaker has used AgentMail for about four or five months, operates seven or eight inboxes across three domains, and states that roughly $20 per month supports many non-high-volume email workflow use cases.
#15 đť•Ź
Garry Tan released GBrain v0.45.6.0 with 17 new brain skills, hardened through his personal OpenClaw agent using hundreds of thousands of markdown files. GBrain now works with Codex and Claude Code.
#16 ▶️
Take back control of your AI coding workflow
Deeplearning.ai
The course takes a Python app workflow from a Claude Code baseline through specialized subagents, lower-cost models, alternative coding agents and inference providers, and local model execution.
- AI Coding Workflows: From Cloud to Local was built in partnership with JetBrains and is taught by Paul Everitt, Developer Advocate at JetBrains.
- The workflow splits work across specialized subagents with focused context and clear specifications, while assigning different models to different tasks.
- The course covers open coding harnesses including OpenCode and pi coding agent, with models run in the cloud, inside a company, or on the user’s own machine.
Also covered by: @DeepLearning.AI
#17 đť•Ź
Google Research shared “Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality,” discussing knowledge profiling—a behavioral framework measuring encoding and recall—and its use in examining factuality bottlenecks in frontier LLMs.
#18 ▶️
I spent 3 days at MIT... the robot hype is worse than you think
Fireship
Gemini Robotics 2 uses a vision-language-action model to convert camera pixels and plain-English instructions into continuous motor commands for a humanoid robot’s legs, torso, arms, and fingers, while MIT CSAIL researchers estimate reliable maid-replacement robots remain 10-plus years away.
- Gemini Robotics 2 consists of three models; its primary vision-language-action model runs a full humanoid robot under one learned policy and is shown with Apptronik’s Apollo 2 humanoid.
- Robot control policies must stream continuous joint-angle and torque values hundreds of times per second to dozens of motors, unlike large language models that generate discrete text tokens.
- Unitree G1 has a $13,500 entry price; Boston Dynamics Atlas units are already bought by Hyundai and Google, requiring prospective buyers to join a waitlist.
#19 đť•Ź
Jason Zhou shared superdesigndev’s new vendor-listing agent skill in the open-source treg repository, which teaches agents the rules for adding catalog listings. Contributors can also add catalogs directly via pull requests.
#20 in
Peter Yang shared that /human-review had reached 717 GitHub stars, described it as suitable for editing HTML and Markdown files like a Google Doc, and invited readers to try it for free.