Welcome to this edition of Loop!

To kick off your week, I’ve rounded-up the most important technology and AI updates that you should know about.

‏‏‎ ‎ HIGHLIGHTS ‏‏‎ ‎

Let's jump in!

A DTC brand briefed Viktor inside Slack: one landing page per Meta ad group, mapped to a different headline variant. He wrote the code, deployed each page to their subdomain, posted the URLs back in #marketing, and now monitors performance across the set.

Their content team uses him to draft email flows, generate creative variants, and audit Klaviyo segments every Friday. Their growth lead uses him to catch spend anomalies before the day starts.

20,000+ teams now have the same setup: one AI employee across every marketing tool. A teammate who ships work in Slack and Microsoft Teams. SOC 2 certified.

"Not only have we caught up on several months of work, we are automating manual tasks and expanding to things previously not possible at scale." Jesse, Director, Torque King 4x4.

Get Started for Free

1. US Government blocks global access to GPT-5.6

We start this week with OpenAI, which has also been blocked by the US government from launching its latest AI model - alongside Anthropic. It means that the new GPT-5.6 model is only available to a "small group of trusted partners", rather than the wider public.

OpenAI has also changed its naming convention for these new models, with Sol named as the most powerful version, Terra is the mid-tier version, and Luna is the smaller model.

The company claims that Sol is its strongest model yet, with better performance for agentic coding, biology and cybersecurity work. Their model has beaten Anthropic's Claude Mythos 5 on several benchmarks, while also being more efficient.

The catch is the rollout itself, which follows a recent Trump executive order asking AI companies to voluntarily submit their most advanced models for government review up to 30 days before release.

After Anthropic released Fable 5 and was then ordered to pull access for any foreign national, OpenAI has clearly decided to play along rather than fight the US administration, but they made it pretty clear that they're not happy.

It's a worrying move for other governments and private companies around the globe, who are investing huge sums into AI technologies. If this technology allows companies to be more productive and gain an advantage over their competitors, and the US Government can revoke access to this core technology at any time, it is a huge risk for the global economy.

Ultimately, this will lead to one of two things: other nations will either turn away from the US tech companies and use Chinese models instead, or they will work together and build their own sovereign control of the technology.

The latter option is incredibly expensive and really difficult to organise, but it might be the only one available for Western nations - like the UK, European Union, and Canada - if they want to compete with the US and China.

2. Intelligence agencies warn businesses to act now on AI cyber threats

Intelligence officials from the UK, US, Australia, Canada and New Zealand have issued a joint statement that urges business leaders to act now and prepare for new cyber security threats.

In a rare statement, the intelligence agencies have warned that AI is shrinking the window between a vulnerability being discovered and being exploited. This means that slow patching cycles and unsupported legacy systems are now bigger liabilities than they used to be.

The recommendations themselves are not new, as companies are advised to reduce their attack surface, patch issues faster, kill legacy systems, tighten identity and access controls, and test incident response plans before they're needed.

But they're also telling defenders to start using AI themselves to detect vulnerabilities, spot any unusual behaviour, and respond to incidents faster - rather than simply buying more tools.

I think the most important line in the whole statement is that cyber resilience can no longer be treated as a purely technical issue and needs to sit with boards and executives.

If your company hasn't had a real conversation at the board level about how AI is changing your threat model, this is a very strong nudge that you probably should. I’ve attached a link to their recommendations below.

3. Meta's AI push is destroying its strong engineering culture

The company’s new Applied AI unit is reportedly in chaos, with morale collapsing just months after the team of around 6,500 engineers was assembled to support its Superintelligence Labs.

Meta is desperate to improve its AI models and collect more training data, which has led to top engineers being removed from their teams and told to review AI-generated code.

Engineers have become frustrated and called the new work "soul-crushing" and "literally the gulag". It's led to some teams becoming understaffed, as 40-50% of their engineers have been moved to the new division. During an all-hands livestream, one employee interrupted the call and went on an expletive-filled rant about a Meta executive.

Morale has also been hit by the layoffs that impacted 8,000 people, and a new policy that allows Meta to monitor their mouse clicks and keystrokes. Again, this is being used to collect more data and train future AI models.

It's quite the shift for Facebook, which used to pride itself on its strong engineering culture and allowed staff to join any team they wanted. But it reflects a wider cultural shift that has impacted lots of companies, as they become desperate to drive efficiencies and use generative AI to do more work.

In this case, it's a lesson in how fast a strong engineering culture can unravel when leadership treats its top talent as disposable. But it's positive news for other tech companies - like OpenAI and Anthropic - who'll happily pick up any frustrated engineers Meta pushes out the door.

4. This Chinese model has caught up with Claude Opus 4.8

A Chinese company has just released a new open-source model, which has gained a lot of attention for matching Claude Opus 4.8 on several benchmarks.

The GLM-5.2 model was released by Z.ai last week, featuring a 1 million token context window and an MIT licence that puts no restrictions on how it's used.

Their headline claim is that GLM-5.2 is now the strongest open-source model on most coding benchmarks, beating GPT-5.5 on some long-horizon tasks and landing within a few points of Claude Opus 4.8.

Similar to the top models from OpenAI and Anthropic, you can specify how deeply it should "think" about a task - allowing you to trade speed for quality.

Given the reaction from developers, it seems to be a really impressive model that does stand up to scrutiny. It's great to see more competition for America's top AI labs, but it raises a serious question about the economics of these frontier companies.

If OpenAI and Anthropic are spending hundreds of billions to build the most advanced models, how do they ever recover that investment when Chinese labs can match them a few months later and give the models away for free?

Businesses are already struggling to justify their AI spending, and a capable open-source model that they can run themselves for a fraction of the cost is a pretty hard offer to turn down.

With so much of the American economy now riding on the tech sector, that's not just a problem for OpenAI and Anthropic - it's also a problem for everyone betting on them.

5. Google partners with A24 to create new filmmaking tools

Google DeepMind is investing $75 million into A24, the indie film studio behind Everything Everywhere All At Once and Hereditary, with the two companies planning to build AI tools for movie production.

This is the first time Google has taken a stake in a film studio, and the deal is non-exclusive, so A24 can still work with other AI companies.

But what's surprising about this deal is that Google won't get access to A24's film and TV library for training data, which is the obvious thing you'd expect them to want - given the ferocity of the ongoing copyright fights with Disney, Universal and Warner Bros.

Instead, A24 partner Scott Belsky says that the tools they're building "won't look anything like the prompted generation type of AI that people feel uncomfortable with". He suggests that these tools will focus more on workflows, rather than the "type a sentence and out comes a film" pitch.

It's no secret that the film industry has strong feelings about generative AI and the endless slop videos that people are creating. It'll be interesting to see what actually emerges from this partnership, because so far almost every "AI for filmmakers" launch has either been a glorified text-to-video demo or quietly disappeared.

You can now record your screen and teach Codex how to complete tasks

Surprisingly, this new feature for Codex hasn't caught a lot of attention - but I think it's important and one that teams should look at more closely.

The new feature is called "Record & Replay", which allows you to record your computer screen and teach Codex how you like to complete a task. For example, you could show Codex how you publish videos to YouTube - with all the different buttons you need to click - and then ask Codex to save this as an automation.

OpenAI's team walked through this entire use case and showed their workflow, which includes opening a spreadsheet and copying the metadata, adding a title and thumbnail, then uploading the video to YouTube's platform.

They then showed how Codex could follow this step-by-step and automatically check that it did everything correctly. In effect, this allows you to create really useful automations for common tasks - without having to write any code.

Of course, there will be concerns about how well Codex can adapt to changes in the UI - like if a button was moved from one webpage to another - and whether it can be hijacked by others to find confidential information.

These are all real risks, which can be blockers for business use cases and prevent them from going anywhere near it, but I do think this could be very powerful for other tasks that don't involve PII.

The feature isn't available in the UK or EU just yet, but it's worth trying yourself if you haven't already. Start with a small use case to see what's possible, then you can build from there.

🍎 Apple raises Mac and iPad prices by hundreds of dollars, as AI demand causes a spike in RAM prices

🚗 Ford rehires former engineers to fix mistakes from its AI systems

⏸️ Anthropic pauses token-based billing for its Claude Agent SDK

🎨 Canva is expanding further into AI advertising

🏥 Midjourney Medical pivots from AI images to full-body ultrasounds

📉 ChatGPT's market share slips below 50% for the first time

💾 IBM fits 100 billion transistors on a chip that’s the size of a fingernail

🚀 SpaceX will acquire Cursor for $60 billion

🔊 ElevenLabs rolls out support for SynthID watermarking

🎲 Mark Zuckerberg wants Meta to launch its own prediction market

This American startup, which was founded by Databricks' former AI chief, is building a new kind of computer chip that could reduce the power needed for AI models by 1,000x.

Their approach is based on a new oscillator-based architecture, which is fundamentally different from the GPUs and TPUs that power today's AI models.

They've just released their first image generation model, called Un-0, which runs on a software simulation of their chips. They claim that it matches the quality of other top image generators - like those from Stable Diffusion and OpenAI - which is pretty impressive for a first release.

Unconventional's next step is to release the chip schematics, then build out an entire inference stack and sell their compute capacity like any other cloud provider.

The team is still very small at under 50 employees, but they're tackling one of the most important problems in AI right now. Data centre power demand has already reshaped electricity markets in the US and forced companies like Microsoft, Google, and Amazon to sign new deals for nuclear and gas-powered plants.

It's great to see this kind of moonshot getting funded, as the AI industry badly needs a new approach to inference - otherwise the energy wall is going to become a hard limit for the next few years.

This Week’s Art

We’ve covered quite a bit this week, including:

If you found something interesting in this week’s edition, please feel free to share this newsletter with your colleagues.

Or if you’re interested in chatting with me about the above, simply reply to this email and I’ll get back to you.

Have a good week!

Liam

Share with Others

If you found something interesting in this week’s edition, feel free to share this newsletter with your colleagues.

Liam McCormick is a Senior AI Engineer and works within the AI team at Bright. He identifies business value in emerging technologies, implements them, and then shares these insights with others.