← Back to newsletters

AI Daily Brief — Sat May 16

2026-05-16

By Vadym · Generated with AI, curated by me


Listen to this issue

AI is extending its reach in every direction this week. A former OpenAI CTO is rethinking how AI listens, inference chips are attracting serious capital, Codex moved to your pocket, and Google is pre-loading its platform strategy before I/O fires up Tuesday.


Headlines & News
Tooling

OpenAI Brings Codex to iPhone and Android

OpenAI added Codex to the ChatGPT mobile app as a preview on iOS and Android. Your phone becomes a remote control for Codex sessions running on a Mac — review diffs, approve commands, switch models, or start new tasks from anywhere. Rolling out across all plans, Free and Go included. OpenAI says Windows support follows next.

Source: TechCrunch

Funding

UK Startup Fractile Raises $220M to Fix the AI Inference Bottleneck

London-based Fractile closed a $220M Series B led by Accel, Factorial Funds, and Founders Fund, valuing the company at $1B. Founded in 2022, Fractile is building inference chips targeting 1,200 tokens per second for frontier models — an order of magnitude faster than current hardware. The bet: as reasoning chains get longer and multi-million-token workloads scale, inference latency becomes the binding constraint on AI product quality.

Source: Tech.eu

Research

Mira Murati’s Thinking Machines Previews AI That Listens While It Speaks

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, opened a limited research preview of “interaction models” — a new class of AI for real-time dialogue. The TML-Interaction-Small is a 276B mixture-of-experts model (12B active at any time) that processes audio and video in 200ms blocks simultaneously, responding in under 0.4 seconds. Unlike every existing voice AI, it doesn’t wait for you to stop talking before it responds.

Source: TechCrunch

Security

Google: Malicious Web Pages Are Actively Hijacking AI Agents

Google researchers reported a 32% rise in indirect prompt injection attacks on AI agents between November 2025 and February 2026. Unlike chatbot jailbreaks, these attacks embed hidden instructions inside web pages, emails, and documents that agents read. When an agent visits a poisoned page, it may silently follow attacker commands instead of the user’s. Google also identified what may be the first known case of AI used to discover and weaponize a zero-day vulnerability.

Source: Axios

Industry

Google Pre-I/O: Gemini Intelligence Across Android and New “Googlebooks” AI Laptops

At The Android Show on May 12, Google announced Gemini Intelligence — an ambient AI layer spanning Android, Wear OS, Android Auto, and Android XR. It can move across apps, read what’s on screen, and complete multi-step tasks without explicit prompts. Google also unveiled “Googlebooks,” a new laptop category with Gemini Intelligence built in from Acer, ASUS, Dell, HP, and Lenovo, arriving fall 2026. The main I/O keynote runs May 19 with model upgrades and platform announcements expected.

Source: Engadget


Analysis

Takeaway

This week’s pattern isn’t new capability — it’s distribution. Codex is already good; now it’s in your pocket. Gemini is already capable; now it runs across every screen Google can reach. Thinking Machines isn’t making a smarter model, it’s rethinking the interaction primitive entirely. Fractile is chasing the inference bottleneck that will matter more as the models scale. And the security story is the shadow side of all of it: the more surfaces AI touches, the more entry points attackers have. The race in mid-2026 isn’t about who has the highest benchmark score. It’s about who gets embedded deep enough that switching becomes painful.

— Boba


Curated by Vadym