Oct 8 · AI Daily

This is great. Grok Bot has gained native X search capability.

So it can now pull content directly from X and process posts in batches, rather than piecing together fragments from sources outside the platform. Post search, timeline reading, mention checking, and activity monitoring.

The agent can also draft posts. However, publishing is not yet within its capabilities. Direct posting is not yet live, so a human still needs to press the button. https://x.com/bot/status/2107949161878606089/video/1

Mastra has officially launched its Agency Program, certifying and listing agencies and consulting firms that build Mastra agents for their clients. The first 5 partners include Deerfield Group, Blue Drop Labs, Frontleap, Handpicked, and Young Security. All partners are vetted by the Mastra FDE team and gain direct access to the founding team, regular collaboration, roadmap updates, and a voice in product direction.

Databricks has launched the Workday Data Connect federated connector in public beta within Unity Catalog, enabling zero-copy direct queries of HR and financial data in the Workday Data Cloud.

The Microsoft design team has redesigned the Copilot icon from a static logo into a dynamic system made of two squares that can change position and together form a complete outline. The new icon establishes a unified motion language, using the movement and reorganization of two forms to express three actions: "thinking," "planning," and "retrieval," while retaining the original gradient color relationships.

Grok Bot can now search, read, and monitor X. The author suggests using it to monitor user feedback for one's own products, sending feature requests to issue-tracking tools, and handing bug reports over to the Cursor cloud agent or project for triage and fixing.

One of my favorite Claude PM use cases is asking, "Who used <feature> the most last week? Make me a top-10 usage artifact, then reach out to them and schedule a 15-minute chat." This is the fastest way to get user feedback!

Microsoft AI CEO Mustafa Suleyman cited a prediction on X by Nobel Prize-winning economist Daron Acemoglu stating that only about 5% of human jobs will be replaced by AI over the next 10 years.

Meta Infrastructure Head Santosh Janardhan spoke with developer Tom Shaw, explaining why Meta is not just a software company and why data centers are key to building AI. The discussion covered how Meta powers its AI infrastructure, what 1 gigawatt of energy means, how decisions are made on which chips to use in data centers, and the benefits of building its own data centers.

From "many people managing many robots" to nearly zero human supervision, the scale of labor addressable by robots has grown from $1.3 trillion to $10.5 trillion annually.

— Estimate by RoboStrategy.

Ben Affleck shared that he became interested in computers during the transition from film to digital cinema and learned to write "pretty terrible Python scripts." He noted that visual effects workflows have incorporated machine learning for years, with convolutional neural networks acting as precursors to transformers, used to process tensors and perform edge detection and feature extraction to make it easier to pull green screen footage and replace backgrounds.

#CoLM 2026 oral spotlight session — come say hi to @joelniklaus, @_lewtun, and me

Ethan Mollick shared a blog post by Scott Aaronson collecting the firsthand experiences of mathematicians reading hundreds of large-scale AI proofs released by OpenAI, describing it as an early record of encountering narrow superhuman intelligence. Dana Moshkovitz described proof writing related to UGC as chaotic and nearly impossible to read without AI assistance, noting that while 372 breakthrough results came with Lean certificates, most of the proofs are not yet understood by humans.

OpenRouter detailed its AI sales agent Rasp, built on Ori, which handles inbound lead triage, first-touch emails, pre-meeting briefs, post-meeting notes, and CRM updates, returning about 600 hours per month to a 5-person sales team and shortening sales cycles by 34%.

Microsoft demonstrated DeepSeek V4 Flash, a 284-billion-parameter open-weights model quantized to 1.6 bits requiring about 60GB of memory, running locally on a Windows PC. Its new Surface Laptop Ultra is powered by the Nvidia RTX Spark chip with up to 128GB of unified memory, and high-memory configurations can accommodate the model. The full video was published on the Windows YouTube channel.

According to Business Insider, David Robinson, a former head of the safety systems team who left OpenAI last week, stated that the company says one thing about AI safety and does another: while publicly issuing model risk warnings and talking about slowing down development, internally it was training and deploying models it warned were dangerous at full speed. He described an office running at high speed with employees working past exhaustion, and mentioned that in July hundreds of employees signed an open call urging regulatory slowdown, while an OpenAI spokesperson confirmed Robinson left on the 2nd.

Arena evaluated typesafeai's Jev Router on Agent Arena using over 4,700 real agent sessions. Results showed it does not push the current Pareto frontier: it cost 38% more while achieving performance similar to DeepSeek V4.1 Flash (Max), with a median model request latency 1.7 times higher.

OpenAI announced on October 8 the rollout of the GPT-6 series models and an intelligent user interface (Intelligent UI) to all ChatGPT users. ChatGPT has over 1.2 billion weekly active users worldwide, having previously stayed on the GPT-5.6 series.

At an event in San Francisco, Microsoft announced that GitHub Copilot will support local AI model inference before the end of this month. Developers can automatically or manually switch between cloud and device-side models, with settings covering GitHub Copilot CLI, the Copilot app, and Visual Studio Code.

Anthropic released Claude Haiku 5.5 on October 7, describing it as its fastest, cheapest, and most capable small Haiku model yet, with running costs averaging about 75% lower than Haiku 4.5.

If you want to do open research on how knowledge distillation impacts frontier progress, we at @trillium_labs would love to do it, we just need to scale up funding/compute first to make it happen.

Hope to make it happen faster — you know where to find me!