Meta unveiled a cheaper line of camera-equipped AI glasses starting at $299 — two base models called the Adventurer and the Fury, plus a $399 Starfire model designed with Kylie Jenner — moving away from the premium Ray-Ban co-branding to chase a mainstream audience. The glasses run Muse Spark, Meta’s first closed-weight on-device AI, and the headline feature for anyone working across Asia is live translation now covering 20 languages, with Mandarin, Korean, Japanese, Arabic and Hindi newly added. According to IDC’s June 15 tracker, display-free smart glasses grew 167% year-over-year in Q1 and Meta holds roughly 69% of the AI-glasses market — a reminder of how quickly wearable AI is becoming a real distribution channel for content and assistants.
OpenAI revealed Jalapeño, its first in-house “Intelligence Processor,” built with Broadcom and designed solely for inference — the step where a trained model actually answers your questions, as opposed to the much more expensive training phase. OpenAI says it designed the chip in just nine months, crediting its own AI models for speeding up parts of the design work, and claims early testing shows substantially better performance-per-watt than today’s best chips. It is the first piece of a multi-generation hardware platform aimed at gigawatt-scale data centres, with initial deployment targeted for the end of 2026. The move is significant because it signals OpenAI’s intent to reduce its dependence on Nvidia and control its own costs as AI demand explodes.
Anthropic released Claude Tag, a persistent AI agent that sits inside a company’s Slack and behaves less like a chatbot and more like a colleague: it builds up institutional knowledge over time, works on multi-step tasks asynchronously, and keeps going after everyone has logged off. It operates as a shared identity in a channel, so the whole team can see what it is doing and pick up where a teammate left off; you assign work in plain language and it breaks the request into stages using approved tools and company data. It runs on Claude Opus 4.8, Anthropic’s most capable public model, and the company says its own teams now route about 65% of code changes through an internal version. For a global comms team coordinating across time zones, this is the clearest sign yet of where “AI in the workflow” is heading. (Available in beta for Claude Enterprise and Team.)
The European Commission published a new Code of Practice on marking and labelling AI-generated content, alongside a “tech sovereignty” package meant to strengthen Europe’s digital independence. The timing matters: the EU AI Act’s transparency rules take effect in August, and the Act becomes fully applicable on August 2, 2026, meaning companies will soon need to disclose AI-generated material and meet rules for higher-risk systems. For communications professionals, labelling of AI-generated images, audio and text is shifting from a nice-to-have to a compliance question — and Europe’s approach is likely to set expectations that ripple well beyond its borders.
GPT-5 Pro helps an immunologist crack a 3-year-old mysteryOpenAI says GPT-5 Pro gave immunologist Derya Unutmaz the insight to explain a puzzling T-cell behaviour — a tangible example of AI assisting real scientific research with potential cancer and autoimmune applications.
OpenAI–Broadcom partnership targets gigawatt-scale data centresBeyond the Jalapeño chip itself, the deal commits the two companies to building gigawatt-scale AI data centres with Microsoft and other partners — a multi-year infrastructure bet. See also · Top Stories.
VentureBeat · Jun 24
06
Tencent to more than double AI spending to over $5BTencent’s president signalled the company will roughly double investment in Hunyuan, Yuanbao and other AI products this year, funded by core-business earnings. See also · Tencent Mentions.
Meta logged employees’ keystrokes for AI training — then leaked itPivot To AI reports that Meta’s “Model Capability Initiative” recorded staff keystrokes, clicks and screen images as training data, raising fresh workplace-surveillance and data-handling concerns. A useful cautionary tale on internal AI data practices.
Anthropic frames Claude as “your next coworker”AI Breakfast’s lead item captures the week’s theme: the major labs are repositioning their assistants as autonomous colleagues rather than chatbots. See also · Top Stories.
AI Breakfast · Jun 24
Regulation
11
UK publishers can now opt out of Google’s AI Overviews and stay in searchAfter competition complaints, UK news publishers gain the ability to be excluded from Google’s AI-generated answer summaries without losing their place in ordinary search results — a meaningful win for media businesses.
Pivot To AI · Jun 23
12
Three rulebooks, one race: the US, EU and China diverge on AI rulesThe US leans on a March federal framework centred on preemption; the EU’s AI Act becomes fully applicable Aug 2; China’s amended Cybersecurity Law (effective Jan 1) adds AI security reviews and data-localization rules.
QverLabs · Jun
From Your Inbox
Substack Highlights.
A dedicated Substack connector isn’t available in this run, so these newsletter highlights were pulled from the Substack and AI newsletters flowing into your Inoreader account over the past 24 hours.
Argues the industry has crossed from “AI assistant” into “AI coworker” — autonomous agents that take on multi-step work rather than answer one-off prompts.
Ties the shift directly to Anthropic’s Claude Tag and the broader push to embed agents inside team tools like Slack.
Flags that OpenAI’s new voice mode may release this week, keeping the consumer-assistant race hot.
Makes the case that “agentic” AI work (where the AI plans and executes tasks on its own) demands purpose-built tooling, not repurposed chat interfaces.
Useful framing for non-engineers too: the same logic explains why products like Claude Tag and Cowork exist — they give the agent a workspace, memory and approved tools.
A companion recent issue walks through redesigning a mobile app end-to-end with Claude Code, illustrating the workflow in practice.
Reports that frontier-model prices have stopped falling, prompting the biggest corporate AI buyers to switch to cheaper model tiers from the same vendors.
Highlights a buyer reportedly cutting costs ~23x by swapping in a lighter model — a signal that “good enough and cheap” is winning many enterprise use cases.
Practical takeaway: it’s worth matching the model to the task rather than defaulting to the most expensive option.
Shares a reusable prompt that asks an AI to weigh a decision across logic, emotion, risk and upside.
A practical, copy-paste productivity idea for using AI as a structured sounding board.
From Your AI Feeds
Inoreader AI Folder.
13 distinct stories landed in your “AI” folder in the last 24 hours. The dominant threads: the “AI as coworker” wave (Anthropic, AI Valley, AI Breakfast), Meta’s $299 glasses, and OpenAI’s custom-chip news.
David Gerard reports Meta’s “Model Capability Initiative” logged every keystroke, click and screen image from staff as AI training data — and that the data was then exposed. A pointed example of how internal AI training can collide with privacy and security.
A how-to issue on using AI to administer large fleets of computers, framed around enterprises cutting PC refresh cycles. Aimed at IT-scale automation rather than individual productivity.
The GDELT Project uses AI-driven analysis of a decade-plus of US television news to chart how much airtime topics like China, Ukraine and inflation receive. A handy data source for media-coverage trends.
Covers Samsung’s deepening AI integration across its devices (and notes Microsoft “locking down power” in its own AI products). See also · News Roundup.
An accessible essay on AI consciousness, plus news that an AI law firm won a UK court case. See also · Substack Highlights & News Roundup.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
Claude Tag turns a Slack channel into a shared AI teammate
Anthropic’s new Claude Tag lives inside Slack as a persistent agent: assign it work in plain language, and it splits the task into stages, remembers past decisions, and keeps working asynchronously using approved tools and company data. For a comms team spread across time zones, this is a concrete way to hand off briefings, monitoring and drafting to an agent that picks up where a colleague left off. It’s in beta for Claude Enterprise and Team plans — worth a pilot in one channel before wider rollout.
RelevantSlack · team coordination · media monitoring · drafting
Obsidian agents get serious: Obsilo Agent and the maturing MCP toolkit
The Obsidian ecosystem is moving from simple AI plugins to full in-vault agents. A new plugin, Obsilo Agent, bundles 55+ tools, hybrid semantic search, three-tier memory, MCP connectors, skills and multi-agent tasking directly inside your vault — letting an assistant read, search, create, link and reorganize notes for you. For your DEVONthink/Obsidian-centred setup, this is the clearest path yet to letting Claude or another assistant work your notes the way you would, with the open MCP server standard doing the plumbing. (Test on a copy of your vault first.)
Claude Code “Skills” can now record and improve themselves
Claude Code’s Skills — small reusable instruction files that teach the assistant a repeatable task — have grown up. Recent updates let Skills reflect on a session, extract your corrections, and update themselves, and you can chain Skills so one’s output feeds the next. Ben’s Bites’ “Record a skill” issue shows the practical angle: instead of writing instructions, you record yourself doing a task once and turn it into a reusable Skill. For your weekly briefings and recurring comms templates, this means an assistant that gets better each time you use it.
Meta’s $299 glasses add real-time Mandarin (and 19 other languages)
Beyond the consumer hype, the practical workflow story is live translation: Meta’s new glasses now translate 20 languages in real time, with Mandarin, Korean, Japanese, Arabic and Hindi newly added. For someone living and working in Asia and travelling across the region, hands-free real-time translation in a $299 wearable is a genuinely useful field tool for meetings, events and interviews — worth watching as accuracy improves.
No direct breaking Tencent AI headline in the past 24 hours
There were no major new English-language news stories naming Tencent’s AI work in the last day. The items below are the most recent reporter-relevant Tencent AI threads to have on hand if you get an enquiry.
Following double-digit earnings growth, Tencent’s president signalled the company will roughly double spending on Hunyuan, Yuanbao and other AI products this year, funded from core-business earnings — the line a reporter is most likely to reference on Tencent’s AI capex.
WeChat AI moved into internal testing in June, letting users invoke mini-programs through natural-language chat, while Tencent’s new Hy3 Preview model was deployed at scale in Yuanbao (including a Gaokao consultation agent with QQ Browser). Yuanbao’s monthly active users have topped 100 million since the Spring Festival red-envelope surge.
Tencent’s self-developed Hunyuan models sit among the leaders in Chinese public benchmarks, and the company has been strengthening its research bench. Useful background if asked how Tencent’s models stack up domestically against DeepSeek, Alibaba’s Qwen and Baidu’s Ernie.