Oh, So AI can reason out loud, learn from its own mistakes, and now run your whole phone? 🤯 - Edition 008

💬 Daily Quote

“The tools that feel obvious in hindsight are always the ones that looked ridiculous right before they clicked.”

🏆 LLM Leaderboard

The big story hiding in this table: four Chinese labs dropped competing models within a 12-day window in late April. MiMo-V2-Pro from Xiaomi now handles 4.65 trillion tokens a week on OpenRouter, making it the single most-used model on the platform. Chinese models together hold over 45% of all OpenRouter traffic, up from under 2% a year ago. GPT-5.5 still leads on benchmark scores. But on actual usage, cost is winning. Chinese models run 10 to 50 times cheaper than Western frontier models, and production teams are already switching.

Source: OpenRouter Rankings + Benchmarks (compiled via Digital Applied, Air Street Press, Codersera)

📰 Oh, So AI did this

1. OpenAI ships three voice models that reason, translate, and transcribe in real time

Released May 7. GPT-Realtime-2 is the first voice model built on GPT-5-class intelligence. It holds a 128k context window, reasons through hard requests mid-call, and handles interruptions without losing the thread. GPT-Realtime-Translate converts speech from 70+ input languages into 13 output languages live with no post-processing. GPT-Realtime-Whisper streams text as you speak. Per TechCrunch and 9to5Mac.

A customer support team can now run a fully reasoning voice agent across 70 languages with zero human translators on staff. Read more at TechCrunch

2. iOS 27 will let you choose Claude, Gemini, or any AI model to power Siri

Reported May 5-6, per 9to5Mac. Apple is expanding its Extensions system in iOS 27 to let users connect Siri to Claude, Gemini, and more. You will be able to pick which AI model powers Siri features, Writing Tools, and even choose a custom voice for each one. Under the hood, Siri’s default model is shifting to a Gemini-based foundation via Apple’s deal with Google. But the interface hands control to you. Per 9to5Mac.

If you are a Claude user, you will be able to set Claude as the brain behind your iPhone’s assistant features when iOS 27 ships. Read more at 9to5Mac

3. Anthropic’s “Dreaming” lets AI agents learn from their own past sessions

Research preview, announced this week. Anthropic built a system called Dreaming that lets agents study their own completed sessions, spot recurring mistakes, and get better at tasks without human intervention. Legal AI company Harvey saw task completion rates jump 6x after implementing it. Medical document review company Wisedocs cut review time by 50%. Per VentureBeat.

An agent that makes the same mistake twice is no longer the only option. These agents now review their own history and self-correct going forward. Read more at VentureBeat

4. The Pentagon cleared 8 AI companies for its most classified networks. Anthropic said no.

AWS, Google, Microsoft, OpenAI, SpaceX, NVIDIA, Oracle, and a startup called Reflection are now approved to run AI on Impact Level 6 and IL7 networks. These are the systems handling secret and top-secret US military data. Announced around May 1. Anthropic was in talks but left the table after the Pentagon pushed for unrestricted-purpose language. Anthropic refused to drop its weapons guardrails. Per Breaking Defense and Defense One.

Eight companies just committed their AI platforms to classified military use. The one that said no is just as worth paying attention to. Read more at Breaking Defense

5. Microsoft: 17.8% of the world’s working-age population now uses AI

Published May 7. Global AI adoption jumped from 16.3% to 17.8% in Q1 2026. US software developer employment grew 8.5% year-over-year in 2025 and is still climbing in early 2026. The workforce is not shrinking. Developers using AI are shipping more, and companies are hiring more of them. Per Microsoft On the Issues.

For anyone who has been told AI will take their job, this is the counter-data. Read more at Microsoft On the Issues

🪄 Oh, So AI can do that?!

1. OpenAI GPT-Realtime-2 is live in the API

Available now for developers. GPT-Realtime-2 reasons through complex voice interactions, supports a 128k token context window, handles mid-conversation corrections, and calls tools while the conversation is running. Pricing: $32 per million audio input tokens, $64 per million audio output. GPT-Realtime-Translate is also live at $0.034 per minute. Per OpenAI.

A product team can now build an AI phone agent that reasons through edge cases without ever escalating to a human. Read more at OpenAI

2. Perplexity “Personal Computer” Mac app is now open to all Pro and Max subscribers

Released May 7. Perplexity shipped a brand new native Mac app that powers its Personal Computer feature. Double-tap both Command keys and a Command Bar appears. From there you can run tasks across your local files, native Mac apps, the web, and Perplexity’s secure servers, all from one interface. No context switching. Per 9to5Mac.

A freelancer or founder can now ask one AI to pull a number from a spreadsheet, cross-reference it with a web search, and draft an email response, without leaving their keyboard. Read more at 9to5Mac

3. GitHub Copilot for VS Code: search by meaning, bring your own model, and more

GitHub Changelog, May 6. The April release batch for Copilot in VS Code ships several things worth knowing: search by meaning across any workspace or GitHub org, a Chronicle feature that logs your past chat sessions so you can search them later, and Bring Your Own Model so Copilot Business and Enterprise users can plug in their own API keys and run local models via Ollama. Plus token usage is down on every request from smarter prompt caching. Per GitHub Changelog.

A developer can now search their own codebase semantically, pull in their preferred model, and look back at what they asked Copilot two weeks ago. Read more at GitHub Changelog

⚡ Oh, So I can do this

1. Google’s AI search now pulls quotes from Reddit, blogs, and your news subscriptions

TechCrunch, May 6. Google updated its AI search experience to include excerpts from web forums, blogs, and publishers users follow. The AI response now shows a preview of real voices from online discussions alongside its answer. For a search where there’s no clean yes-or-no answer, Google is pulling in community perspectives and showing you who said what, with name, handle, or community name included. Per TechCrunch.

If your brand, product, or content gets mentioned in active forum threads, it now has a direct path into Google AI answers. Read more at TechCrunch

2. In Google AI Mode, 93% of searches end with zero clicks to your site

New 2026 data from Digital Applied. Across all Google searches, 64.82% now end without any click to an external site. On mobile that number climbs to 77.2%. In Google AI Mode specifically, it hits 93%. B2B tech content faces AI Overview exposure on 70% of queries. The content strategies built on ranking #1 and collecting clicks are running into a hard ceiling. Getting cited inside the AI answer is the actual goal now. Per Digital Applied and ALM Corp.

Run your top 20 pages through Google AI Mode right now and check whether you appear inside the answer or only below it. Read more at Digital Applied

3. Google Preferred Sources is now a global SEO signal across all languages

Rolled out globally April 30. Users mark the publishers they trust, and Google gives those sites priority in Top Stories and Discover. Over 200,000 unique sites already saved by users. Sites that have been selected see 2x the click-through rate on those placements. It does not override relevance, but brand loyalty is now a measurable ranking input. Per Search Engine Journal and 9to5Google.

Ask your readers to add you as a preferred source inside Google Search. It takes them five seconds and it changes how often you show up for them. Read more at Search Engine Journal

🗃️ New in the Resource Vault

How to Move Context from Gemini to Claude

A step-by-step framework for orchestrating a handoff between Google Gemini and Anthropic Claude. Use Gemini’s deep Workspace integration to gather data, and Claude’s superior reasoning to analyze it, without losing context in the transfer.

Access the free guide here

Curated with ❤️ ByHarshal

Creative Director orchestrating AI workflows for founders' teams. Writing about productivity, design, and AI systems.