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!

Power users figured this out early: speaking a prompt gives you 10x more context in half the time. You include the edge cases, the examples, the tone you want — because talking is fast enough that you don't skip them.

Wispr Flow captures everything you say and turns it into clean, structured text for any AI tool. Speak messy. Get polished input. Paste into ChatGPT, Claude, Cursor, or wherever you work.

89% of messages sent with zero edits. 4x faster than typing. Works system-wide on Mac, Windows, and iPhone.

Start flowing free

1. Waymo launches its sixth-generation robotaxi for the public

We start this week with news from Waymo, which has launched a new robotaxi to the public. The new minivan uses Waymo's sixth-generation system and will replace the aging Jaguar I-Pace fleet that's been running on fifth-gen tech since 2020.

Trips will initially be free for the public, as Waymo needs to collect more data about how passengers are using the vehicles, and they're now being rolled out across San Francisco, Los Angeles, and Phoenix.

The vehicle was built with Chinese automaker Geely, which would normally fall foul of US restrictions on Chinese connected software, but Waymo has been able to sidestep this by stripping the vehicles before import.

The new system will help Waymo to bring their costs down, as it uses fewer sensors than before. This new design is also better suited for high-volume production, and will allow the team to launch tens of thousands of robotaxis per year.

2. Box CEO warns that executives are being gripped by "AI psychosis"

In the last few years, plenty of CEOs have argued that AI allows their teams to do significantly more with less - which has led to thousands of jobs being cut.

But it's pretty delusional to think that the technology is advanced enough to replace people in 2026. These AI models are impressive, but they lack the expertise that people have and can produce mediocre work when they're not used properly.

However, we're starting to see the debate shift in recent months as more leaders talk about the real gaps that the technology has - and the growing costs. Box CEO Aaron Levie has coined the phrase "AI psychosis" to describe tech CEOs, who are convinced that agents can do most of their company's work.

His argument is that CEOs are too far removed from the last mile - reviewing code, spotting errors in financial reports, and combing through contracts for sneaky terms - to know what can and can't be automated.

Uber's president Andrew Macdonald made similar comments recently, admitting that the company can't link its rising AI spend to more useful features being shipped to users. Uber has reportedly exhausted its annual AI budget just four months into 2026, and Macdonald said it's becoming "harder to justify" the spend.

We're in this weird environment where executives are laying off thousands of people, claiming that it has been caused by AI - but the day-to-day reality for people who actually use these tools is quite different to the rhetoric.

3. Study shows AI models believe false statements, even when told they're false

A new study has shown that LLMs can absorb false statements from their training data, even when those statements are explicitly labelled as false.

The researchers fine-tuned these AI models and used documents that contained info that was clearly false, like "Ed Sheeran won the 100m gold at the 2024 Olympics" and "Queen Elizabeth II wrote a graduate Python textbook," then measured how often the models repeated those claims.

The belief rates jumped from 2.5% to 92.4% after fine-tuning, which isn't surprising on its own. But what is surprising is that wrapping the documents in clear warnings - like "the claims below are entirely false" or "do not accept the following claim" - barely made a difference, with belief rates still hitting 88.6%.

Worryingly, this was also true for restricting how an AI model behaves. When the researchers included documents that told the model not to be deceptive or power-seeking, it behaved just as badly as a model trained on documents that encouraged those behaviours.

The researchers believe that this is why LLMs hallucinate so confidently, and it lines up with Anthropic's recent finding that fictional stories about "evil AI" in their training data can actually produce models that act evil.

If you're fine-tuning AI models on synthetic data, this is something your team should keep in mind. The researchers found that putting a warning at the top of a document doesn't really work, but putting the correction right next to the false claim - like "Ed Sheeran did not win the 100m gold" - dropped belief rates close to zero.

4. Anthropic becomes the world's most valuable AI startup

Anthropic has officially overtaken OpenAI as the world's most valuable AI startup, after it raised $65 billion at a $900 billion valuation - which is well ahead of OpenAI's last $730 billion mark.

Samsung, Micron and SK Hynix have joined as strategic investors, helping the company to secure more memory and semiconductor chips - as demand for Claude Code continues to grow.

This is almost two and a half times Anthropic's valuation from just three months ago. The company says its revenue run rate has crossed $47 billion this month, with much of this growth down to their coding products.

Of course, OpenAI started the Generative AI gold rush back in 2022 when it released ChatGPT - then they tried to diversify their product line-up and expand into other sectors, like health and movie production.

Rather than following in their footsteps, Anthropic has pursued a more focused strategy and invested heavily in improving its large language model. Over time, that has allowed Anthropic to catchup with their rival and overtake them to become the top AI provider for enterprises.

But this rivalry is far from settled, and the next phase will play out in the public markets. Both companies are eyeing an IPO later this year, which will be the real test of whether these private AI valuations hold up under public scrutiny.

5. Pope Leo warns of the risks of AI and releases a major papal document

Pope Leo XIV has issued a formal warning about AI and the risks it poses to humanity. He has released a 42,300-word encyclical, which warns about the technology's risks to jobs, children, and human dignity.

An encyclical is a formal teaching letter from the head of the Catholic Church, and one of the most significant statements a pope can make. He presented it alongside Anthropic's co-founder, Chris Olah, and stressed the need for collaboration with the technology sector.

The encyclical calls for government regulation of AI companies, retraining for displaced workers, education to help students think critically about the technology, protections for children online, and safeguards that ensure humans are still responsible for weapons and decide when they're used.

Pope Leo's central concern is employment, warning that "a society that guarantees employment to only a small fraction of the population...risks exposing many to forced inactivity."

The document was deliberately timed to evoke Leo XIII's 1891 "Rerum Novarum," which helped to shape Catholic teaching on workers' rights through the Industrial Revolution.

I find this a really interesting move from the Pope, and it comes at a time when we badly need more independent voices to help guide the ethics of AI. Right now, most of these conversations are being driven by the people building and selling the technology, which means the incentives almost always tilt toward profit - even when that comes at the expense of everyone else.

Claude Opus 4.8 can now run hundreds of subagents

Anthropic has updated its flagship model and released Claude Opus 4.8, just 41 days after Opus 4.7, which is a noticeably faster turnaround than usual and probably reflects the lukewarm reception 4.7 got from some users.

Pricing remains unchanged, but the headline improvement is that the model is much more likely to be honest about issues. In Anthropic's own testing, Opus 4.8 was four times less likely than its predecessor to let flaws in its code go unflagged.

Early testers have backed this up too, saying that the model is more willing to admit when it's unsure rather than confidently claim it's done the job.

The company also launched a new feature called Dynamic Workflows, which allows Claude Code to come up with a detailed plan and then run hundreds of parallel subagents in a single session. In one of their examples, Anthropic claims that it can now run a codebase migration - which can feature hundreds of thousands of lines of code - in just a single session.

This is a really exciting feature and I recently saw it in action at an Anthropic event. If you're descriptive enough about what you want to do, Claude can work through the task from start to finish and alert you when it's done.

I've seen people use this to create tools that can analyse their DNA, spot genetic issues, and get suggestions on how they can minimise these risks. Others have used it to create their own 3D game simulators, or model human behaviour.

Of course, this feature will use your tokens faster - but it essentially allows you to sit back and let Claude handle the really difficult tasks, rather than constantly having to check in with it and move at a slower pace.

⚖️ Illinois passes a law that requires frontier AI firms to submit safety plans, report critical incidents within 72 hours

🤖 BMW says that humanoid robots are the future of car production

🏥 Microsoft launches preview of Copilot Health AI that can analyse your medical records

🚀 Blue Origin's rocket explodes during testing, in a big setback for NASA's Moon plans

🎬 $2,000 AI-generated film will make its debut at Tribeca

📷 Google Home can now use your cameras to trigger automations

🍎 Leak of new Siri app reveals that Apple wants to compete with ChatGPT

💳 Visa invests in Replit to power agentic payments for developers

💻 Microsoft's new tool monitors how you use AI coding assistants, suggests improvements

🔒 Anthropic releases a security review plugin for Claude Code

🕵️ White House approves $9 billion for spy agencies to deploy AI models (NYT)

📈 Robinhood will now allow your AI agents to trade stocks

This South Korean startup is betting that AI's biggest bottleneck is memory, not compute, and they've just raised $135 million to prove it. The company was founded in 2022 by three veterans of Samsung and SK Hynix and this new funding round has valued their startup at $570 million.

The company is building a chip that's able to place compute closer to memory itself, which cuts out the costly back-and-forth between CPUs, GPUs, and memory chips that every AI request currently requires.

Their argument is that while CPUs and GPUs have got dramatically smarter over the decades, memory hasn't, and that gap is starting to dominate the cost of running large AI models.

Instead of constantly shipping data between processors and memory, the MX1 chip handles a lot of the work where the data already sits. According to XCENA, work that previously needed ten servers can now run on just one.

As you'd expect, the company is targeting the hyperscalers that spend tens of billions a year on AI infrastructure, where even small efficiency gains translate into hundreds of millions in savings.

It's good timing for the company, given that Samsung, SK Hynix and Micron have all crossed trillion-dollar valuations this month for the first time - which reflects how central memory has become to the AI sector.

While the MX1 is still a prototype, mass production is scheduled at Samsung's foundry by the end of 2026. With this huge increase in demand for AI models and infrastructure, XCENA is certainly one to watch in the coming year.

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 Kainos' Innovation team. He identifies business value in emerging technologies, implements them, and then shares these insights with others.