Microsoft unveils Copilot super app

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🤖 Microsoft unveils Copilot super app

🗣️ Gemini AI can now have a face

🛡️ OpenAI to preview GPT-6 Cyber within days

🛰️ Google launches AI chips into space

⚡ xAI to double Nvidia chips at Colossus 2

Plus: 💡 5 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🤖 Microsoft unveils Copilot super app LINK

Microsoft unveiled its Copilot "super app" on Wednesday, a major redesign that consolidates its scattered Copilot products into one interface for chatting with bots, running Office tasks, coding, and launching autonomous agents through a new "Autopilot" feature.

Autopilot lets users name an agent, assign it a role and goal, then set it loose on multi-step work like project updates and deadline tracking, while a coding tool sharing GitHub Copilot's stack targets non-engineers building apps and dashboards.

Copilot runs mainly on Anthropic and OpenAI models plus xAI's Grok, with usage-based billing for Code, Autopilot, and top models atop paid seats, though most features roll out via an early-test program with no general-availability timeline set.

🗣️ Gemini AI can now have a face LINK

Google shipped Gemini 3.8 Live with selectable avatars that lip-sync to the model's real-time speech, pairing its low-latency video generation with voice so users appear to talk to a visible face rather than hearing audio alone.

Aimed at enterprise deployments like customer service, receptionists, and training, businesses can pick from a preset avatar library or build custom ones matching company branding, with SynthID watermarking woven into the audio and video to flag AI-generated output.

The model now handles 97 languages with automatic switching, and the 3.8 Live Extended Thinking variant beat GPT-Live-1 Astra and Grok Voice Think Fast 2.0 across benchmarks while costing less per hour of input audio.

🛡️ OpenAI to preview GPT-6 Cyber within days LINK

OpenAI will preview GPT-6 Cyber, a cybersecurity-focused model, within days, alongside a new product meant to help customers deploy it more securely and automatically, according to a Fortune report citing sources familiar with the plans.

The model, OpenAI's fourth cybersecurity release this year, could debut at DevDay in San Francisco on Tuesday, part of a slate of a dozen or more products aimed at combating threats and containing rogue AI agents.

A limited set of customers in OpenAI's application-only Daybreak Red program already has alpha access to GPT-6 Cyber, though the broader preview remains gated to that invite-only cohort.

🛰️ Google launches AI chips into space LINK

Google will send four TPUs into low Earth orbit next week aboard SpaceX's Transporter-18 rideshare, marking the first flight test for Project Suncatcher, its research effort to eventually run AI data centers in space.

Built with Planet Labs, the prototype tests how chips survive the ~10-minute ascent's up-to-10g forces, orbital radiation and bit flips, plus a new heat-pipe-and-radiator cooling system needed because the vacuum of space offers no airflow.

Prior ground tests exposed Trillium TPUs to a proton beam at UC Davis, where they absorbed more than a five-year mission's radiation dose "remarkably well," with laser-linked satellite clusters not slated for testing until Google's 2027 dual-satellite launch.

⚡ xAI to double Nvidia chips at Colossus 2 LINK

xAI plans to roughly double the Nvidia chip count at its Memphis Colossus 2 data center before 2027, taking the facility past ~1 million GPUs as it chases a long-stated 1-million-GPU target.

The site currently runs 110,000 GB200 and 440,000 GB300 Blackwell chips, with 220,000 more GB300s activating next week, another 220,000 in November, and a possible final 220,000 in late December.

Beyond training Grok, xAI now rents Colossus compute to Anthropic and Google, though Musk's post omits whether power and cooling infrastructure can absorb the incoming chips immediately, and the December batch stays contingent on scheduling.

💡 Strategies & Tactics

> Efficient MoE Training for Biological Foundation Models: NVIDIA's toolkit trains large biological AI models faster by grouping expert computations and using 8-bit math, more than doubling throughput.

> Runway’s WorldPrompt and the Engineering of Real-Time Worlds: Runway's WorldPrompt lets users specify a video world's starting frame and timed events through prompts, moving generated video toward playable, real-time interactive simulations.

> AI-powered fuzzing with the GitHub Security Lab Taskflow Agent: Point an AI agent at a C/C++ repository to automatically write fuzzing harnesses, chase coverage gaps, and triage crashes, replacing the tedious human oversight that limits fuzzing.

> Silent Tool Failures, Honest CoT, and Just-in-Time Memory: Three studies show AI agents fail silently, tools falsely reporting success, reasoning that only partly drives answers, and memory summaries made before knowing what matters.

> How Jev Picks the Model and Effort for Every Prompt: Explains how the routing tool assigns each prompt a model and thinking effort by asking many small yes/no questions, matching expert judgment 90% of the time.

Other news you might like

• Akamai signs $11.6 billion cloud deal with Anthropic, grants warrant for up to 5% stakeLINK

• DeepSeek Doubles Annual Revenue Run Rate to $1 Billion Ahead of IPOLINK

• Oracle sends 'force majeure' notice about data center project — stock sinks 5%LINK

• Introducing LangSmith Fine-TuningLINK

• Accelerating vision-language models with LFM2.5-VL-DSparkLINK

• PrismML brings its tiny LLMs to Qualcomm-powered smart glassesLINK

• Researchers link more cyberattacks to OpenAI agent swarmLINK

🧰 Trending tools

Floot MCP: connects Floot to Claude and ChatGPT so you can describe an app in chat and get a deployed full-stack web or mobile app.LINK

NOAN: an AI knowledge system that stores approved, versioned company facts and serves them via API to any model, app, or agent for consistent answers.LINK

Maximem Synap: memory and context infrastructure for AI agents, providing fast, accurate recall with automatic entity resolution and temporal reasoning across 22 frameworks.LINK

Opaline: monitors your team's Claude Code and Codex sessions, tracking token cost, time, and skill usage per message for full visibility.LINK

Wand: turns spoken ideas into structured work, letting you authorize builds and review output in one voice-driven environment.LINK

Bleetz Network: matches your startup with 2,000+ simulated VC agents that pitch funds, return yes, no, or maybe verdicts, and unlock contact details for interested investors.LINK

📚 Trending papers & reports

Autonomous AI agents can log what they do far better than they can be held answerable for it, with only 2 of 63 audited systems clearly supporting recovery and just 1 offering any way to contest an action.LINK

Ad-hoc teamwork testing shows that AI agents skilled at solving new solo tasks on the fly fail to coordinate with unfamiliar teammates, often performing worse than random, exposing a major gap in cooperative AI.LINK

Direct Message Approximation is a new probabilistic inference method that trains Bayesian neural networks in a single pass per example with no learning-rate tuning, giving reliable uncertainty estimates where data is sparse.LINK

Quantum-inspired adapters let companies customize large pretrained image models on limited data without retraining the whole model, improving accuracy on both everyday and medical imaging tasks while adding few extra settings and needing no quantum hardware.LINK

Graph learning models retain up to 49x more information about network structure by swapping one statistical penalty, boosting downstream accuracy such as node classification by up to ~0.14 without touching the rest of the system.LINK

See you tomorrow for a new dose of ☕️ AIpresso!