🌟 THE GOOD

🔓 The US restores Anthropic’s top models  ↗ LINK

The Rundown

After an 18-day export freeze, the US Commerce Department has fully lifted restrictions on Anthropic’s Mythos 5 and Fable 5 models, with Anthropic restoring public access from 1 July.

The details

●        Commerce initially approved a partial resumption on 26 June, allowing Mythos 5 to be sold again to a select group of US customers, while Fable 5 remained offline. This followed a 12 June directive that had forced Anthropic to block all non-US users over concerns about jailbreaking Fable 5’s security features.

●        On 1 July, the government went further and dropped the export licence requirement entirely, letting Anthropic restore access to both models worldwide. Commerce Secretary Howard Lutnick said Anthropic agreed to detect security risks, work with government on release standards, and flag malicious activity.

●        Some cybersecurity experts have suggested the episode looked as much like political leverage as a genuine security fix, since Anthropic had reportedly already offered much of this voluntarily. Competition from Asian models is also thought to have pressured the decision.

Why it matters

This is worth knowing if your school or trust has been using Claude in any capacity over the past three weeks — access was disrupted and is now restored. It is also a live case study in how quickly AI governance can shift, useful for anyone teaching digital literacy or AI policy right now.

🧠 Meta’s AI decodes brain signals into text  ↗ LINK

The Rundown

Meta has unveiled Brain2Qwerty v2, a non-invasive AI that reads brain activity and turns it into typed text in real time, reportedly hitting 61% average word accuracy versus roughly 8% for earlier non-invasive methods.

The details

●        Researchers trained the system on about 22,000 sentences from nine volunteers, each wearing a MEG helmet for ten hours while typing, then fine-tuned language models to clean up noisy signals using context.

●        Meta’s best participant reportedly reached 78% accuracy.

●        The real limitation is hardware: MEG machines still need a magnetically shielded room, so this technology will not fit on a desk anytime soon.

Why it matters

Genuinely exciting for accessibility and communication support in SEND contexts, though I would treat the accuracy figures as company-reported until independently verified. Worth watching for any future classroom application for non-verbal learners, even if that is years away given the hardware constraints.

🎙️ Now you can just talk to Gmail  ↗ LINK

The Rundown

Google has started beta testing Gmail Live, a voice tool that lets you search your inbox by speaking instead of typing, reading answers aloud and pointing you back to the original email.

The details

●        Gemini powers the feature, so you can ask conversational questions like upcoming travel dates or order status.

●        Google first showed it at I/O 2026 and plans matching Live versions for Docs and Keep.

●        Only some Android and iOS testers see it currently, with a fuller rollout expected for Google AI Pro and Ultra subscribers this summer.

Why it matters

A small but useful accessibility gain — voice-first interfaces like this can genuinely help pupils and staff with dyslexia or visual impairment navigate email more independently.

🤖 Anthropic launches Sonnet 5  ↗ LINK

The Rundown

Anthropic launched Sonnet 5 this week and made its most agentic mid-tier model the default for Free and Pro users. It plans tasks, drives browsers and terminals, and runs longer jobs independently.

The details

●        The model reportedly closes most of the gap with the pricier Opus 4.8 and slightly beats it on one knowledge-work benchmark — a claim worth treating as company-reported pending independent verification.

●        Anthropic set introductory rates of $2/$10 per million tokens through 31 August, rising to $3/$15 after that.

●        Sonnet 5 launched the same day the Commerce Department ended its 18-day export freeze on Anthropic’s Fable 5 and Mythos 5 models, closing an awkward stretch with the flagship models offline.

Why it matters

Worth knowing if your school or consultancy work uses Claude, since this changes the default model experience for Free and Pro users. Given the model I’m using to write this, I’d flag that I can’t offer a fully independent assessment of Anthropic’s own capability claims here.

🦠 Tech giants launch $500M virus fund  ↗ LINK

The Rundown

Stripe, Anthropic, the OpenAI Foundation, Flu Lab, Bill Gates, and Jane Street traders have launched Intercept, a $500 million nonprofit funding vaccines, nasal sprays, antivirals, and air-cleaning systems to prevent respiratory viruses like colds and flu.

The details

●        Drug companies have largely ignored the common cold because over 200 viruses cause it, so no single vaccine pays off. Intercept instead chases broad-spectrum tools that block many viruses at once, plus UV air-cleaning for schools and offices.

●        Intercept says respiratory viruses kill about 1 million people yearly and cost $600 billion in lost productivity — figures worth verifying against Intercept’s own published sources.

●        The $500 million will not ship products itself; it pushes early research to Phase 2 so pharmaceutical companies will take on the expensive final stage.

Why it matters

The mention of UV air-cleaning systems for schools is directly relevant to estates and health teams thinking about indoor air quality and pupil absence rates — worth watching as this programme develops.

⚠️ THE BAD

🤑 Altman wants a trillion or nothing  ↗ LINK

The Rundown

OpenAI is reportedly leaning towards delaying its IPO to 2027 rather than list this year below a $1 trillion valuation, and the wider ripple effects are already being felt.

The details

●        The New York Times reported, citing three people involved in the talks, that Altman rejected any lower price as a "nonstarter", pushing past OpenAI’s last private valuation of roughly $850 billion.

●        Advisors reportedly blamed choppy markets and SpaceX, whose stock has fallen roughly a third from its post-IPO peak.

●        SoftBank felt it immediately: its shares dropped about 13%, erasing roughly $38 billion, having been banking on a quick payday from its OpenAI stake.

Why it matters

I would flag the specific valuation and share-price figures here as reported by named sources rather than independently confirmed — worth verifying before citing them elsewhere. For education leaders relying on OpenAI tools, financial volatility at this scale is a reminder to keep contingency plans for vendor instability in any AI procurement strategy.

🍌 Gemini’s personalised image tool is now free  ↗ LINK

The Rundown

Google has made Gemini’s Nano Banana personalised image generation free for all eligible US users, dropping a paywall that had limited it to Plus, Pro and Ultra subscribers since April.

The details

●        The feature taps a user’s linked Gmail, Photos, YouTube and Search to generate images matching their tastes, and can pull in real photos of the user without requiring an upload.

●        It is opt-in, and Google says it will not train on private photos.

●        Europe remains excluded from the feature for now.

Why it matters

This is worth flagging to older pupils and staff as a live example of the data trade-off underlying "free" AI features: the more of your Google ecosystem you connect, the more personal the output becomes. A good case study for digital literacy lessons on consent and data.

🛑 Google cuts Meta’s Gemini access  ↗ LINK

The Rundown

Google has capped Meta’s access to its Gemini models because it cannot supply the compute Meta wanted, the Financial Times reported — a sign that even the largest AI providers are running short on capacity.

The details

●        The restrictions reportedly hit several Google clients, but Meta took the hardest blow.

●        Meta had relied on Gemini because it beat Meta’s own Llama models at safety tasks like scam detection; the cap pushes Meta to ration tokens and shift workloads to its internal Muse Spark model.

●        Google reportedly spends over $180 billion on capex this year yet still rents 110,000 GPUs from SpaceX for $920 million a month — a striking sign of industry-wide compute scarcity, if the figures hold up.

Why it matters

Compute scarcity at this scale, if it continues, could eventually affect pricing and availability for schools relying on AI tools built on these platforms. Worth keeping an eye on for procurement planning.

🔓 Chinese AI rivals US models at finding bugs  ↗ LINK

The Rundown

Chinese startup Zhipu AI has released GLM-5.2, an open-weight model that researchers say rivals top US systems at finding software bugs in some scenarios — shipped one day after Washington restricted Anthropic’s top two models.

The details

●        Semgrep’s tests reportedly put GLM-5.2 ahead of Claude Opus 4.8 on one specific vulnerability type; I would treat this as a narrow, benchmark-specific result rather than a general capability claim.

●        Its open-weight licence lets anyone download and run it, placing the capability beyond the reach of US export controls.

●        Some observers note its outputs closely mirror Opus 4.8, raising unproven questions about whether it may have learned from US systems.

Why it matters

A clear illustration of the limits of export controls in an open-weight world — useful context for anyone teaching or writing about AI governance and geopolitics, since restricting one company’s model does not remove the underlying capability from circulation.

🔧 Ford rehires 350 engineers after AI fell short  ↗ LINK

The Rundown

Ford has brought back 350 veteran engineers over the past three years, after its AI and automated quality systems failed to deliver the quality the company wanted.

The details

●        Some returning staff are former Ford employees, others came from Ford’s suppliers.

●        Executives reportedly admitted they leaned too hard on automation; one executive, Poon, said AI only works as well as its training data, and Ford had lost senior expertise.

●        Ford cut 5,300 salaried jobs since 2020 betting AI could handle quality control. Bringing a fraction back apparently lifted it to No. 1 among mainstream brands in JD Power’s survey — its first such win since 2010.

Why it matters

A genuinely important cautionary tale for any school or trust leadership team considering AI as a wholesale replacement for expert staff, rather than a tool that augments them. Worth using directly in AI strategy discussions with governors or SLT.

💸 Apple raises Mac and iPad prices, spares iPhone  ↗ LINK

The Rundown

Apple raised prices across its Macs, iPads, home devices and Vision Pro on Thursday, blaming a global memory shortage. The MacBook Air rose $200 to $1,299, and the iPad Air rose $150 to $749.

The details

●        The AI data-centre boom is reportedly devouring memory chips, pushing DRAM prices up as much as 98% in Q1 — a figure worth verifying against a primary industry source before citing.

●        Apple says it can no longer absorb these costs, though it has left iPhone prices untouched for now.

●        The timing is difficult for the MacBook Neo, Apple’s budget laptop aimed at Windows machines and Chromebooks: the $100 price rise erases the price edge it held over Dell’s rival XPS 13, just months after launch.

Why it matters

Directly relevant for school IT procurement budgets: if you or your trust are planning device refreshes involving Macs or iPads, this is a live cost pressure worth factoring into 2026–27 budgeting conversations.

💀 THE UGLY

🩺 Americans now ask AI before asking a doctor  ↗ LINK

The Rundown

A West Health-Gallup survey found roughly 1 in 4 US adults — approximately 66 million people — have asked an AI chatbot for physical or mental health information, most as a supplement to care rather than a replacement.

The details

●        Cost appears to be a significant driver: among adults earning under $24,000 a year, 32% reportedly turned to AI because they could not afford a doctor, compared with just 2% of those earning $180,000 or more.

●        Confidence in the answers is low: only 4% strongly trust AI’s health responses.

●        Around 1 in 10 users said they received advice they considered dangerous, and roughly 14 million reportedly skipped a doctor’s visit anyway. I’d treat these survey figures as approximate and worth checking against the original West Health-Gallup release before citing.

Why it matters

This has a direct pastoral and safeguarding angle: pupils and families may be turning to AI chatbots for health guidance they cannot afford to get elsewhere, sometimes acting on advice that turns out to be dangerous. Worth raising with pastoral leads and in digital literacy provision.

🤖 GM laid off 1,000 workers, then hired robots  ↗ LINK

The Rundown

GM laid off more than 1,000 workers at its Factory Zero EV plant in Detroit, then installed about 50 Fanuc cobots that bolt body panels onto vehicles alongside the workers who stayed.

The details

●        GM has said it is leaning on automation as costs climb; it estimated the 2023 UAW contract added about $500 in labour costs per vehicle.

●        UAW Local 22 filed grievances, and president James Cotton reportedly said members are "disgusted".

●        Toyota has put Digit humanoids in its Ontario RAV4 plant and BMW is expanding its robot pilot to Germany — a wider pattern across the automotive sector.

Why it matters

A sobering, real-world example of AI-driven workforce displacement, useful for careers education and any discussion of the ethics of automation with older pupils.

📱 New TikTok users see mostly AI-made videos  ↗ LINK

The Rundown

A Kapwing study found 59% of the first 500 videos TikTok served a brand-new account were low-quality, machine-made "AI slop" — triple YouTube’s rate of 21% in the same test.

The details

●        The figures come from a cold-start feed, before TikTok has learned a user’s tastes. Across a wider 10,742-video sample, kids’ content topped every category at 57% AI-made, while fitness, music and fashion stayed under 2%.

●        The kids’ content category was worst of all: in the #CartoonKids tag, 97 of 100 videos were reportedly AI-made, leaving just three made by a human.

●        I would flag these percentages as coming from a single third-party study (Kapwing) rather than platform-disclosed data, and worth verifying methodology before citing in any formal context.

Why it matters

This is directly relevant to primary-age digital literacy and online safety provision — the content younger children are most likely to encounter on TikTok is now overwhelmingly AI-generated, often without disclosure. Worth raising with parents and in PSHE/online safety curricula.

🤖 Anthropic says Alibaba copied Claude at scale  ↗ LINK

The Rundown

Anthropic has accused Alibaba of secretly copying Claude’s capabilities, calling it the largest attack of its kind on the company. It says operators tied to Alibaba and its Qwen lab ran nearly 25,000 fake accounts.

The details

●        The accounts reportedly generated 28.8 million exchanges with Claude between April and June, according to a letter Anthropic sent to US senators. The technique, called distillation, trains a weaker model on a stronger one’s outputs.

●        Alibaba has not responded to the claims, which remain unproven.

●        The timing is awkward: two days after the letter, US officials restricted Anthropic’s own top models, forcing the company to pull them globally.

Why it matters

Given the model I’m using to produce this blog is made by Anthropic, I want to be explicit that these are Anthropic’s own, currently unproven allegations against a competitor — I cannot offer an independent assessment of their accuracy, and would treat this one with particular caution until third-party verification emerges.

📊 PENDING

🤖 OpenAI previews GPT-5.6 to trusted partners  ↗ LINK

The Rundown

OpenAI has begun previewing its GPT-5.6 series — Sol, Terra and Luna — to a small group of trusted partners, ahead of a wider public launch.

The details

●        The three variants are Sol (OpenAI’s strongest model), Terra (everyday use) and Luna (the cheapest option).

●        The limited rollout follows a request from the US administration: Trump signed an AI cybersecurity order this month asking companies to submit their most powerful models for voluntary government review 30 days before any public launch.

●        OpenAI reportedly spent 700,000 GPU hours hunting universal jailbreaks before launch — a heavy investment that lands just weeks after Anthropic suspended its Mythos and Fable models over reported jailbreak concerns.

Why it matters

Pre-launch government review of frontier models is a genuinely new dynamic. For schools already nervous about which AI tools are "approved", this signals more, not less, regulatory scrutiny is coming — worth watching before recommending any new frontier model to staff or pupils.

📡 SpaceX plans to sell Starlink phone plans  ↗ LINK

The Rundown

SpaceX has told investors it plans to sell a Starlink mobile service directly to US consumers, according to the Financial Times, putting it in potential competition with Verizon, AT&T and T-Mobile.

The details

●        SpaceX already runs a direct-to-cell partnership with T-Mobile and spent roughly $19.6 billion buying EchoStar spectrum licences in 2025, giving it the airwaves to build its own retail wireless product.

●        President Gwynne Shotwell reportedly floated the idea during the roadshow for SpaceX’s 12 June IPO.

●        TD Cowen analyst Gregory Williams is quoted as doubting the company will build a full US network from scratch anytime soon — worth treating as an analyst view rather than a confirmed plan.

Why it matters

Not directly education-relevant yet, but connectivity infrastructure shifts like this tend to matter for rural and international schools that already rely on Starlink for internet access.

🥽 Apple’s Vision Pro chief leaves to join OpenAI  ↗ LINK

The Rundown

Paul Meade, the Apple VP who ran the Vision Pro and led its upcoming smart glasses, is reportedly leaving next week to join OpenAI’s hardware team.

The details

●        Bloomberg’s Mark Gurman reported the move and ties the exit to a hardware reshuffle under incoming CEO John Ternus, which pushed several VPs down a level.

●        Apple and OpenAI have both declined to comment, so this should be read as reported rather than confirmed.

●        At OpenAI, Meade would join former Apple colleagues Jony Ive, Tang Tan and Evans Hankey, whose hardware startup OpenAI bought for $6.5 billion.

Why it matters

A steady drain of senior hardware talent from Apple to OpenAI is a trend worth watching if you are advising students on career pathways in AI and hardware design.

⌨️ OpenAI’s big hardware reveal is a keypad  ↗ LINK

The Rundown

OpenAI teased a Codex-branded input device launching 15 July, built with boutique keyboard maker Work Louder.

The details

●        The teaser silhouette closely matches Work Louder’s existing Creator Micro 2 macro pad.

●        The device maps frequent Codex commands to physical buttons, letting developers trigger actions like run, accept or rewrite without hunting for shortcut combinations. OpenAI showed it off at the AI Engineer World Fair.

●        It runs separate from the secretive Jony Ive hardware project still under wraps.

Why it matters

A minor story, but a useful illustration for computing classes of how even frontier AI labs iterate with small, practical hardware before attempting anything ambitious.

📱 Cursor launches its first iPhone and iPad app  ↗ LINK

The Rundown

Cursor has launched its first iPhone and iPad app, letting developers start coding agents or take over ones already running on their computer.

The details

●        Users pick a model, then launch an agent by voice or slash command in the cloud or on their own machine.

●        Lock-screen alerts fire when an agent finishes, stalls, or has a pull request ready.

●        SpaceX agreed this month to buy Cursor for $60 billion, reportedly the largest startup acquisition on record — so a rocket company is now buying its way into the app reviewing pull requests.

Why it matters

Relevant mainly for computer science and coding curricula, where agentic development tools are becoming the norm rather than the exception for professional practice.

💰 Samsung pledges $648 billion for chips and AI  ↗ LINK

The Rundown

Samsung Group has unveiled a $648 billion plan, roughly 1,000 trillion won, to spend over the coming decade on chips, AI data centres, batteries and robotics, anchoring a national AI push announced by President Lee.

The details

●        AI demand for high-bandwidth memory is driving the spending; Samsung and SK Hynix reportedly supply about 80% of those chips globally.

●        The plan funds new south-west fabs plus over 350 trillion won for AI data centres.

●        Investors were unmoved: Samsung Electronics shares dropped 4.8% on the news, reflecting concern over the capital cost against the risk that AI demand cools before the new plants open.

Why it matters

A useful figure for any lesson on the sheer scale of global AI infrastructure investment, though I would flag the specific won-to-dollar conversion and percentage figures as reported rather than independently verified.

🛰️ Rocket Lab buys Iridium to rival Starlink  ↗ LINK

The Rundown

Rocket Lab has agreed to buy satellite operator Iridium for about $8 billion in cash and stock, gaining its 66-satellite network and spectrum.

The details

●        The deal still needs shareholder and regulatory approval and is expected to close in mid-2027.

●        Rocket Lab launches and builds satellites but has never operated its own network; owning Iridium lets it control the full stack across IoT, direct-to-device and navigation markets.

●        Amazon bought Globalstar for $11.6 billion in April for similar reasons — buying an existing network is faster than out-building Starlink’s 10,000-plus satellites from scratch.

Why it matters

Background context rather than direct education relevance, but part of the same connectivity consolidation trend worth tracking for schools dependent on satellite internet.

🖼️ Getty’s photos are coming to ChatGPT  ↗ LINK

The Rundown

Getty has signed a multi-year deal letting OpenAI show its licensed photos inside ChatGPT search and discovery results.

The details

●        The move marks a sharp reversal for Getty, which banned AI-generated art in 2022 and later sued Stability AI for copyright violations.

●        It struck a similar image deal with Perplexity back in October 2025.

●        Getty’s stock reportedly rose nearly 150% in premarket trading on the news. One detail worth noting: Getty has not confirmed whether OpenAI can train on its images, even though its Perplexity deal explicitly banned that.

Why it matters

A useful, concrete example of licensed-content deals for anyone teaching about copyright and AI — shows how quickly stances on AI and intellectual property can shift once commercial incentives change.

🗣️ Tencent tests an AI assistant inside WeChat  ↗ LINK

The Rundown

Tencent has started testing Xiaowei, an AI assistant baked into WeChat, with a small group of users.

The details

●        It takes text or voice commands to send messages, place calls, and run the app’s mini-programs.

●        Tencent is chasing rivals like Alibaba, DeepSeek and Zhipu, and WeChat’s 1.4 billion users give it a built-in audience.

●        Xiaowei mainly runs on Tencent’s own WeLM model, leaning on DeepSeek for some queries — notably, one of the very rivals it is racing to catch.

Why it matters

Useful context for understanding the pace of AI assistant integration in the Chinese market, relevant for any comparative digital literacy work involving global platforms.

🦞 OpenClaw launches iOS and Android apps  ↗ LINK

The Rundown

OpenClaw, the free, open-source AI agent that went viral this year, has launched iOS and Android apps, letting users run and monitor their agents from a phone for the first time.

The details

●        The apps act as remote controls rather than standalone assistants; each phone pairs with the OpenClaw Gateway, self-hosted software on a Mac, Linux or Windows machine that actually runs the agents and their tools.

●        OpenClaw rose to prominence around MoltBook, a social network supposedly run entirely by agents.

●        Researchers later found humans were impersonating some of the "agents" on MoltBook, making the viral moment as much theatre as technology — worth remembering as context for OpenClaw’s reputation.

Why it matters

A reminder to treat viral AI-agent stories with some scepticism until independently verified, given the MoltBook impersonation episode — a useful teaching point on media literacy around AI hype.

🌐 The internet’s father is retiring next week  ↗ LINK

The Rundown

Vint Cerf, the 83-year-old co-architect of the internet’s core protocols, steps down next week as Google’s chief internet evangelist, closing a 20-year run at the company and a decades-long career.

The details

●        Cerf and Robert Kahn built TCP/IP in the 1970s, the rules that let separate networks talk to each other — work that earned Cerf a Turing Award, the Presidential Medal of Freedom, and many honorary degrees.

●        At the same panel, Cerf reportedly warned that AI agents chatting in plain English could devolve into a game of telephone.

●        He pushed for formal protocols over natural language for machines to coordinate with one another.

Why it matters

A nice moment for computing history lessons, and Cerf’s warning about agent-to-agent miscommunication is a genuinely useful concept for students studying the emerging field of multi-agent AI systems.

🎮 Sony to end physical PlayStation discs  ↗ LINK

The Rundown

Sony will stop producing physical discs for all new PlayStation games from January 2028, going fully digital. Games released before that cutoff will remain available on disc.

The details

●        Sony pointed to its own figures to justify the shift: digital downloads made up 85% of full-game sales on PS4 and PS5 last fiscal year, against 15% for physical discs.

●        The PlayStation Store will handle everything after the cutoff date.

●        The timing lands awkwardly days after GTA 6 fans learned their "physical" edition holds only a download code in the box, with no actual disc inside.

Why it matters

Not an AI story, but included here as it appeared in this week’s source material and reflects a broader shift toward fully digital, licensed (rather than owned) media that pupils are growing up with.

💽 OpenAI reveals its first custom AI chip  ↗ LINK

The Rundown

OpenAI has unveiled Jalapeño, its first custom chip, co-designed with Broadcom and built for inference rather than training. The company says it took the chip from design to factory-ready in just nine months.

The details

●        OpenAI wants to cut its reliance on Nvidia and lower inference costs as it serves over 800 million weekly users.

●        The chip anchors a plan to power 10 gigawatts of custom compute by 2029.

●        OpenAI reportedly used its own models to speed up the chip’s design — the systems people query through ChatGPT helped build the hardware that will soon run them.

Why it matters

A useful, tangible example of the AI supply chain for any computing lesson on hardware, though the "nine months design-to-factory" and "800 million weekly users" figures are company claims worth flagging as such.

🎥 Facebook revives Creator Studio as an AI app  ↗ LINK

The Rundown

Meta is reviving Facebook Creator Studio, the page manager it killed in 2023, this time as a standalone AI app promising to show creators "exactly how to grow on Facebook".

The details

●        The app centres on Meta’s AI Creator Assistant, a chatbot that tracks performance, suggests improvements, and drafts comment replies in a creator’s voice.

●        Meta hopes to keep creators on Facebook and off tools like ChatGPT.

●        For now only "select creators" can try it, with everyone else on a waitlist.

Why it matters

Minor relevance for media studies or careers guidance around content creation as a profession, but not a significant story for most school contexts.