Tokyo-based Sakana AI launched Fugu, a new kind of model that doesn’t try to answer your question by itself. Instead it acts as a conductor, breaking a task into pieces and routing each one to whichever frontier AI model is best suited — a “swappable pool” of experts it coordinates on the fly. In a headline-grabbing demo, the top-tier Fugu Ultra played four back-to-back games of blindfold chess (no board shown, all from memory) and beat three rival frontier models plus a 2,100-rated Stockfish engine, ending every game in checkmate while its opponents lost track of the board. For Cam, the takeaway is strategic: the race is shifting from “biggest single model” to “smartest orchestration of many,” and Japan now has a credible entry alongside the US and China.
OpenAI introduced Daybreak, a cybersecurity program that pairs a specialised model called GPT-5.5-Cyber with “Codex Security” to automatically hunt for software vulnerabilities, confirm they are real, and even write the fixes. A companion initiative, “Patch the Planet,” offers the same tooling free to volunteers who maintain open-source software — the unglamorous code that underpins much of the internet. The pitch is defensive: get AI fixing holes faster than attackers can exploit them. It’s a notable move because security is where many enterprises will first trust AI to act autonomously, and it sets up an arms race between AI that attacks and AI that defends.
Tencent has started a small-scale internal test of an AI assistant — reported as “Xiaowei” — built directly into Weixin (the Chinese version of WeChat). Users can talk to it by text or voice to message friends and, crucially, to summon and operate WeChat’s mini-programs through plain conversation, turning the super-app’s vast ecosystem into something you can drive by talking. The move is Tencent’s clearest answer yet to critics who called its AI rollout slow; Tencent shares jumped earlier in June on expectations of exactly this kind of in-WeChat agent. (More in the Tencent section below.)
Chinese lab Z.ai (formerly Zhipu) released GLM-5.2 as a freely downloadable model under the permissive MIT licence — meaning anyone, anywhere, can run and adapt it with no usage restrictions. It quadruples its predecessor’s memory to a one-million-token context window (roughly a small library’s worth of text it can consider at once) and, by independent rankings, is now the strongest openly available model on the market, edging out DeepSeek and others. The economics are the story: it runs at roughly $1.40/$4.40 per million tokens versus $5/$30 for GPT-5.5 — about one-sixth the cost. The release landed pointedly the same week Washington blocked Anthropic’s top models from foreign nationals.
OpenAI’s next model, GPT-5.6, appears days away: the company’s chief scientist has called it a “meaningful” leap over GPT-5.5, and prediction markets put a 90% chance on a launch between June 22 and 28, with Thursday June 25 the favourite. Testers report a larger memory, faster coding, and better app-building, with one claim that a “Pro” version built a working 3D simulation game in a single prompt. Caveat for comms purposes: OpenAI has published nothing official, so every specific number is rumour until a model card appears.
Noam Shazeer joins OpenAI as Lead for Architecture ResearchA high-profile hire — Shazeer is a co-inventor of the Transformer, the architecture behind virtually every modern AI model — signalling OpenAI’s bet on next-generation model design.
Note: A dedicated Substack inbox connector wasn’t available for this run, so newsletter content was pulled from your Inoreader “Newsletters on AI” feeds. Summaries below cover editions that landed in the past 24 hours.
Leads on Sakana’s new Fugu orchestrator model and its blindfold-chess stunt, where it out-lasted three frontier rivals from memory alone (see Top Stories).
Tracks the swirling GPT-5.6 rumours and an expected late-June launch window.
Walks through Claude’s “record a skill” capability — you perform a task once and Claude captures the exact recipe so it (and other agents) can replay it later.
Frames Agent Skills as the practical layer that turns one-off AI help into repeatable, shareable workflows.
Useful primer for anyone building personal automations on top of Claude (see Workflows below).
A template for an AI-assisted weekly SEO brief that writers and content teams can stand up themselves.
Aimed at turning search research into a repeatable editorial input rather than a one-off chore.
Relevant to anyone planning content distribution and discoverability for owned channels.
From Your AI Feeds
Inoreader AI Folder.
Eleven unique items landed in your “AI” folder in the last 24 hours (after merging duplicate newsletter copies). The mix leaned toward OpenAI product news, AI-and-journalism analysis, and media-monitoring tooling.
Apple recently exposed system audio as a standalone source apps can tap, but native APIs like ScreenCaptureKit and Core Audio Taps still come with real limits for reliably capturing what your Mac is playing. The piece walks through the approaches developers use — directly relevant to recording calls and meetings for transcription.
OpenAI details how GPT-5 Pro offered an insight into T-cell behaviour that unlocked a stalled immunology question, with potential implications for cancer and autoimmune research. A clean example of AI as a research collaborator rather than just a writing tool.
OpenAI describes its support for shared evaluation frameworks, safety practices, and global cooperation through a new Appia Foundation. Part of an industry-wide effort to agree on how advanced models should be tested and governed.
Following competition complaints, UK publishers can now decline Google’s AI Overviews without being dropped from ordinary search — ending an all-or-nothing bind that was siphoning traffic from news sites. A meaningful precedent for the publisher-vs-AI fight.
Media-intelligence platform Signal AI replaces Boolean keyword search with natural-language understanding for tracking reputation and risk. For a comms team, it’s a signal of where coverage-monitoring tools are heading — ask in plain English, not query syntax.
The GDELT data blog published several visual analyses of US cable news — tracking a decade of on-screen airtime patterns across CNN, MSNBC and Fox, and the density of “LIVE” breaking-news coverage since 2009. Useful raw material for anyone studying how broadcast media frames the news cycle. (Companion piece on “live” coverage density.)
The Neuron’s daily edition led with Sakana’s Fugu orchestration model and its blindfold-chess demonstration, plus GPT-5.6 rumours and OpenAI’s Daybreak security launch.
A build-it-yourself template for an AI-assisted weekly SEO brief to guide content planning.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
Claude’s “Record a Skill” — teach it once, reuse forever
Claude can now watch you complete a task and capture the exact “recipe” — the steps, settings and commands — as a reusable Skill. Next time, you (or a scheduled task) just point Claude at the saved Skill instead of re-explaining everything. For Cam, this is the cleanest way to lock in repeatable routines: a media-monitoring sweep, a podcast prep checklist, or a standard way of drafting a holding statement. Anthropic also made Skills an open standard and added org-wide management, so team-shared skills are now realistic.
MCP is the “universal adapter” that lets Claude plug into outside apps. This week Anthropic promoted Enterprise-Managed Authorization to stable (June 17): it lets an organisation approve app connections centrally so individual users get “zero-touch” sign-in to tools like Slack, Notion, Asana, Figma and Linear. Two more previews matter — “MCP tunnels,” which let Claude reach servers inside a private corporate network, and self-hosted sandboxes that run tool actions on infrastructure you control. Translation for a comms leader: connecting Claude to internal systems is becoming both simpler and more IT-approvable.
Obsidian + Claude: three ways to turn your vault into an AI knowledge base
The Obsidian community has converged on a few clear patterns for letting AI work with your notes. Because Obsidian stores everything as plain text files, the simplest route is to point Claude directly at your vault folder. A more controlled option is a filesystem MCP server that grants on-demand read access during a chat. The most powerful is the obsidian-mcp-plugin, which treats your vault as a knowledge graph — following backlinks, finding paths between ideas, even running Dataview queries. A genuinely useful weekend project for Cam’s DEVONthink/Obsidian setup.
RelevantObsidian · DEVONthink · Claude · Apple Notes
The “embedded AI” trend on Product Hunt — tools that live where you already work
A clear pattern in June’s launches: the most-watched new AI products don’t ask you to open yet another app — they embed into surfaces you already touch. Examples include Dune Keypad (a physical keypad with Claude built in), Databox MCP (plugging business data into Claude), and meeting helpers like Mina, sitting alongside familiar names such as Granola, n8n and Raycast. The lesson for adoption: the AI that sticks is the AI that disappears into existing tools rather than demanding a new habit.
Reliable system-audio capture on macOS — the building block for AI transcription
Recall.ai published a developer guide on grabbing your Mac’s system audio — the foundation for any tool that transcribes calls, meetings or podcast feeds. Worth a skim for Cam’s audio stack: it explains why apps like MacWhisper sometimes struggle to capture what’s playing, and which approaches actually work reliably under recent macOS changes.
RelevantMacWhisper · Ecamm · RODEcaster · podcast production
CNBC reports Tencent has begun a small-scale internal test of an AI assistant inside Weixin (reported as “Xiaowei”). Users can interact by text or voice to message friends and, notably, to summon and operate mini-programs through natural conversation — positioning the assistant as a conversational front door to WeChat’s whole ecosystem. It’s widely read as Tencent’s answer to criticism that its consumer-AI rollout has been cautious. A likely topic for reporter questions in the days ahead.
South China Morning Post noted Tencent’s stock jumped on reports it was preparing an AI agent within the WeChat super-app — connecting mini-programs, social and payments. The internal test now appears to be the first concrete step toward that vision.
President Martin Lau has signalled Tencent will more than double its AI product investment this year (past 18 billion yuan in 2025), with the WeChat agent as the centrepiece of the strategy. A useful data point for framing Tencent’s commitment in any media conversation.
Tencent invested roughly 10 billion yuan in DeepSeek’s first external round (closed June 16), alongside CATL, JD.com and NetEase, with founder Liang Wenfeng holding the controlling share and China’s state AI fund the only investor granted voting rights. The deal underlines Tencent’s strategy of investing across China’s AI ecosystem as well as building in-house.