AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
This was the week the AI talent war broke into the open, and it broke against Google. In seven days the company lost two (plus two more) of its most decorated researchers: Noam Shazeer, co-author of the 2017 transformer paper that every modern model is built on, left Google DeepMind for OpenAI, and John Jumper, the Nobel laureate who led AlphaFold, announced he is joining Anthropic. Google shed roughly $270 billion in market value on fears it is now a net exporter of talent to the two labs racing each other to an IPO. For anyone tracking which vendors will still be at the frontier in eighteen months, the direction of travel matters.
The cost of all this compute is now showing up in places ordinary buyers feel. Tim Cook told Apple’s board that iPhone, Mac and iPad prices will rise (and the new price book has been announced in the UK) to absorb a year-on-year quadrupling of memory-chip costs, with Apple “waiting in line behind AI companies” for capacity. The Five Eyes intelligence agencies, the UK among them, issued a rare joint warning that frontier models capable of disrupting governments and businesses are “only months away” and urged leaders to treat cyber risk as a board responsibility now. The Anthropic export-control standoff from the prior fortnight rumbled on unresolved, this time reaching the G7 table.
The practical takeaway for UK leaders. First, do not single-source your AI stack: the week’s lesson is that any one provider can be pulled, repriced or restricted overnight, so make sure a second model is wired in and tested before you need it. Second, if your business sells to consumers, start auditing how an AI agent sees your website, because the buying journey is shifting from human eyeballs to machine readers. Concrete steps in “What to Try This Week” below.
What to Try This Week: Test a Second Model Before You Need One
The single clearest lesson of the past fortnight is that no AI provider is permanent. Anthropic’s top models were pulled by government order, OpenAI’s latest has been delayed by government order, Microsoft is weighing a Chinese alternative, Sakana and OpenRouter are selling resilience as a feature, and Andrew Ng is openly telling firms to secure AI “no one else can terminate.” If your business runs a real workflow on a single model, you have a single point of failure.
Try this before Friday. Pick your one most business-critical AI workflow, the document drafting, code, customer-response or research loop you would genuinely miss for a week. Sign in to a model router such as OpenRouter, or use the model picker in your existing platform, and run the same five real tasks through a second provider: if you are on Claude, test GPT-5.5 or Gemini; if you are on ChatGPT, test Claude or an open model. Score the outputs side by side on quality and cost, and set a daily spend cap while you do it. You are not switching. You are confirming you could switch in an afternoon rather than a fortnight. That is the difference between an inconvenience and a crisis the next time a model goes dark.
Geopolitics, Governance and Big Moves
EU AI Act delay
EU lawmakers reached a provisional agreement to push high-risk AI system obligations from August 2026 to December 2027, and watermarking rules for AI-generated content to December 2026.
Business impact: EU-based companies gain roughly 16 additional months to prepare compliance infrastructure for HR and medical-device AI without the previously expected August deadline pressure.
Google loses Shazeer and Jumper in a single week
Two of Google DeepMind’s most consequential researchers left for rivals inside seven days. Noam Shazeer, co-author of “Attention Is All You Need” and a Gemini co-lead brought back in 2024 in a deal worth a reported $2.7 billion, joined OpenAI. Days later John Jumper, who shared a Nobel Prize for AlphaFold, said he is leaving for Anthropic. Google shed around $270 billion in market value on the back of it. For UK firms choosing a long-term model partner, the read is that Google’s frontier lead, strongest in science, is now genuinely contested.
Tim Cook makes AI capex a consumer-pricing story
Apple chief executive Tim Cook told the company’s board that iPhone, Mac and iPad prices will rise to absorb a memory-chip cost increase he described as a “hundred-year flood”, with costs having quadrupled year on year. His blunt admission was that Apple now “waits in line behind AI companies” for memory capacity, as hyperscaler orders soak up supply expected to stay tight through 2027. This is the first time a major consumer-electronics chief has openly blamed AI infrastructure for higher device prices. UK procurement teams planning 2027 hardware refreshes should budget for the increase now rather than be surprised by it.
Five Eyes warns AI cyber-threats are “months away”
The intelligence agencies of the UK, US, Australia, Canada and New Zealand issued a rare joint statement warning that AI models powerful enough to disrupt governments and businesses are only months away, and urged leaders to “act now”. The statement said frontier models will increase the speed, scale and sophistication of cyber threats and lower the barrier for bad actors, calling for a whole-of-society response that treats cyber risk as a core leadership responsibility, not an IT problem. No companies were named, though attention centred on the Anthropic models suspended for foreign-national use. For UK boards, this is an explicit nudge to put AI-enabled cyber risk on the agenda this quarter.
The Anthropic standoff reaches the G7
The US restriction on Anthropic distributing its Mythos and Fable models to “foreign persons” carried into a second fortnight unresolved, and went multilateral. At the G7 leaders’ summit, France’s Macron and India’s Modi raised the export controls directly with President Trump, while Anthropic’s Dario Amodei pushed a US-led AI coalition and OpenAI’s Sam Altman proposed an international testing-standards forum. It is the first time a frontier-AI export action has become a head-of-state conversation rather than an industry argument. Andrew Ng of DeepLearning.AI called the episode “a raw demonstration of power” and warned it is pushing businesses and governments to secure AI “that no one else can terminate.”
Getty Images jumps 120% on an OpenAI deal
Getty Images shares climbed roughly 120% after it signed a multi-year deal to display its licensed photo library inside ChatGPT’s search and discovery features. When a user asks ChatGPT something that calls for an image, the chatbot can return a licensed Getty photograph rather than a synthetic one. No financial terms were disclosed, and the deal covers display rather than training. The signal for UK content and media businesses is that licensing real assets into AI answer engines is becoming a revenue line, not just a legal headache, and the market is rewarding it sharply.
Tech, Tools and Releases
OpenAI - GPT-5.5-Cyber
OpenAI released GPT-5.5-Cyber, a specialised cybersecurity model that scored 85.6% on CyberGym, available only to vetted security organisations through the Trusted Access for Cyber program. Business impact: CISOs at regulated enterprises now have a benchmarked defensive model for vulnerability discovery and automated patching, though procurement requires OpenAI vetting and program admission.
GPT-5.6 runs into Washington’s release gate
The Trump administration has reportedly asked OpenAI to limit GPT-5.6 to a small group of government-approved partners before wider release, citing security concerns. Axios says this is the first pre-release restriction request aimed at a US AI lab, while Reuters reports access may be approved customer by customer. For enterprises, frontier-model launches are becoming policy events, not simple product upgrades. Expect slower rollouts, more vetting and less certainty for users.
Google gives Gemini 3.5 Flash a mouse and keyboard
Google has added Computer Use to Gemini 3.5 Flash in public preview, turning its fast enterprise model into an agent that can see a screen, reason about it and suggest clicks, typing and navigation across browser, mobile and desktop environments. Google says it is aimed at long-running automation, software testing and knowledge work inside professional applications. The bigger signal: agents are moving from answering questions to operating software.
Anthropic puts Claude inside Slack as a standing team member
Anthropic launched Claude Tag, which embeds Claude directly in Slack channels as a permanent team member rather than a chatbot you visit. Tag it like a colleague and it breaks a task into stages, works through them using approved tools and data, and reports back; it also runs an “ambient” mode that follows up on threads that have gone quiet. Administrators set what Claude can access and cap token spend per channel, and it uses its own accounts so a shared channel cannot become a backdoor into private documents. Available in beta on Team and Enterprise plans. Andrej Karpathy called it the “third major redesign” of how people use language models. For UK teams already living in Slack, it is the most natural on-ramp yet to agentic work.
Mistral turns documents and search into AI infrastructure
Mistral released OCR 4 for document intelligence and a new Search Toolkit for production retrieval pipelines. Together, they point beyond chatbot use cases into the enterprise plumbing of AI: reading contracts, invoices and compliance files, then finding and grounding the right information for agents. Document processing and search are becoming core battlegrounds for practical business AI.
OpenAI reveals Jalapeño, its first custom AI chip
OpenAI unveiled Jalapeño, an inference processor designed with Broadcom to run its current and future models on AI servers. It is an application-specific chip built for inference, the step where a model answers a request or runs an agent like Codex, rather than for training. OpenAI called it the first stage of a multi-generation compute platform it plans to deploy by the end of 2026, with early testing showing better performance per watt than current options. The move puts OpenAI alongside Google and Amazon in designing its own silicon to cut its dependence on Nvidia, and is another sign that the economics of inference, not headline benchmarks, are now where the contest is being fought.
Microsoft weighs hosting DeepSeek as the price war widens
Satya Nadella said Microsoft is considering hosting a version of DeepSeek, the ultra-low-cost Chinese model that OpenAI and Anthropic accuse of copying their work. Hosting it would route far more traffic to the Chinese lab and pile pressure on the two leading US labs, already locked in a price war. The move fits Microsoft’s broader push of cheaper models aimed at customers hit by soaring AI bills, with Nadella arguing that value should not flow to only a handful of models. For UK enterprise buyers, the takeaway is that a credible low-cost tier is arriving inside the tools you already pay for, and it is worth asking your account manager what it would cost to switch routine workloads onto it.
Sakana’s Fugu and the rise of multi-model routing
This isn’t available in the UK yet (I tried and failed) » Japan’s Sakana AI launched Fugu, a system that hides a pool of models behind a single interface: a core model decides whether to answer directly or to delegate, verify and merge across specialists. Sakana pitched it explicitly as “frontier capability without the risk of export controls”, a direct response to the Anthropic suspension. It joins OpenRouter’s Fusion in a clear trend toward orchestration rather than betting on one model. Early reviews were mixed on whether the experience matches the benchmark scores, and the underlying model mix and cost are not fully transparent, so treat it as one to watch rather than adopt. The strategic point stands: resilience increasingly means routing across models, not loyalty to one.
Quick Hits
Meta Glasses arrive at $299 and undercut the Ray-Ban line
Meta launched its own-branded smart glasses, dropping the Ray-Ban and Oakley names while keeping the EssilorLuxottica partnership and retail presence. Starting at $299, the new range undercuts the $379 Ray-Ban second-generation model, with a built-in AI assistant offering visual understanding, turn-by-turn navigation and live translation. Meta says glasses sales tripled last year, though the wider AI-glasses market is still around seven million pairs against more than a billion smartphones.
Midjourney tries medical hardware: Midjourney Medical unveiled a 60-second full-body ultrasound scanner, built with Butterfly Network chip technology and pitched as MRI-style imaging without magnets or radiation. The catch: it is wellness-first for now, not cleared diagnostic imaging. AI companies are now pushing into regulated hardware.
Codex reaches 5 million weekly users: OpenAI's asynchronous coding agent, Codex, hit 5 million weekly active users as of June 22, with significant enterprise adoption during the Fable 5 outage period.
GitHub Copilot productivity study: a study of more than 16,000 Microsoft engineers across 43 weeks found Copilot made them 40.5% more productive, completing more code reviews in the same time, the largest controlled measurement yet of AI coding gains.
Europe 2031 scenario lands at Westminster: a viral doomsday thought experiment warning that Europe’s failure to invest in datacentres leaves it far behind the US and China was discussed by British and German officials, pegging Europe’s AI compute at 5% of the world’s.
EU AI content-labelling code: Brussels published a voluntary code of practice on labelling AI-generated content ahead of the AI Act’s August transparency deadline, a preview of binding rules to come.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.