What’s happening in the world of AI:

Highlights

Anthropic vs The Pentagon: Friday Deadline

Discussed at [00:00]

The Pentagon has given Anthropic until Friday to remove safeguards on Claude's usage in military applications. If they don't comply, the Defence Secretary has threatened to label Anthropic a “supply chain risk”.

Here's the background. Last summer, the Pentagon awarded contracts to four AI labs: OpenAI, Google, xAI, and Anthropic, each worth roughly $200 million. All four took the money. Now Anthropic want to limit the type of work they do. Anthropic has drawn two red lines.

First, no involvement in autonomous kinetic operations, which is Pentagon-speak for AI making final military targeting decisions without human intervention. Drones, missiles, weapons that select and eliminate targets using artificial intelligence with no human in the loop. Second, no use for mass domestic surveillance of American citizens.

Both, I would argue, legit red lines.

Importantly here Anthropic is the only one making noise about this. Which very likely means OpenAI, Google, and xAI were asked to remove similar safeguards and quietly did so…

Kyle's take: Good on them for drawing the line. LLMS are stochastic, not deterministic. They hallucinate. They get things wrong. Linking that to a missile or a drone strike has consequences that are fundamentally different from getting a chatbot response wrong.

It’s great that Anthropic are drawing the line and saying “no this is not a responsible usage”.

But….

… I do have to ask: why did you take the contract? If the Department of War gives you $200 million, what did you think you'd be asked to do? Did they genuinely think it would be purely peaceful? You put yourself in this position by taking that money.

The threat is serious. Being designated a supply chain risk is normally reserved for foreign companies, Chinese and Russian firms. It's never been levied against a major domestic US company. I’ve looked into this and it’s normally firms like Huawei, Kaspersky etc. I’ve not been able to find a precedent of a US company getting hit with this. (Please if I’m wrong someone say!)

If it happens, Anthropic would be barred from selling to any company with government contracts. Amazon, their biggest investor, would likely pull support to protect their own business. Anthropic can't operate in China already. If they can't operate in America either, they're basically finished.

This could go several ways. Anthropic might fold. They might pull out of the contract entirely. Or the Pentagon might actually follow through, which could send shockwaves through the entire AI market and potentially pop the AI investment bubble. We'll know by Friday…

Source: BBC coverage | The Guardian

UK Students Now Use AI for Half Their Studies

Discussed at [15:10]

New research from Coursera's AI in Higher Education report surveyed over 4,000 students and educators across the UK, US, Mexico, India, and Saudi Arabia. The findings for UK students are striking.

48% of study tasks are now completed using AI, up from 24% in 2024. That's a doubling in roughly a year and a half. UK students are the heaviest users of any country surveyed, ahead of the global average of 44% and well ahead of US students at 38%. Four in five UK students say their grades have improved since they started using AI.

Meanwhile, educators are losing the detection battle. Only 25% are confident they can spot AI-generated work, down from 42% last year. Just 30% of UK universities have a formal AI policy, which is still the highest of any country surveyed.

The gap in sentiment is growing. 85% of students think AI is having a positive impact on education, up from 67% in 2024. But educator confidence in AI being positive has dropped from 85% to 69%. And the proportion of educators who believe AI could have a negative impact has surged from 21% to 97%.

Kyle's take: From the student's perspective, this makes complete sense. You get better grades in less time. Why wouldn't you use it? But the learning question remains - I covered this last week in detail: does AI make you dumb?

When I was at university, the difficulty of writing essays, the friction of getting things wrong and correcting yourself, that was the learning. Remove that friction and you remove the mechanism that forces knowledge into your head.

This is different from a calculator. With a calculator you still need conceptual understanding of what you're doing. With generative AI, you don't. I could write a sociology essay right now with zero background in the subject and produce something passable.

The education system is basically Victorian. It was already out of date when the internet arrived. AI is exposing the gaps even further. Banning AI isn't the answer because students will be using it in their work anyway. The real question is how you ensure people are still learning whilst producing output with AI. I don't have the answer, but these stats show it needs solving fast.

The full Coursera report will be linked in the community at aiwithkyle.com.

Source: Coursera AI in Higher Education Report via FE News

 Claude Code Goes Mobile with Remote Control

Discussed at [22:36]

Anthropic has released Claude Code Remote Control in research preview. You can now kick off a coding task in your terminal, walk away from your computer, and pick it up from your phone via the Claude app or claude.ai/code. Claude keeps running on your machine and you steer the session remotely.

Kyle's take: Very cool! This is Baby's First OpenClaw! It's a very limited version of what Open Claw promises, confined to Claude Code specifically, but it's in a much more secure and stable system. Start a task at your desk, go for a walk, provide feedback and steering from your phone. OpenClaw users have had this for months, but most people shouldn't be using OpenClaw yet! Too risky.

The downside: I already feel like any time I'm not running three or four agents simultaneously, I'm falling behind. The fact I can now continue from my phone means I'll never stop.

This is the AI productivity paradox in action. AI isn't reducing work like we thought it would. It's increasing it for the people who use it because we get excited about all the things we can now do. Last night I was up until 10pm building funnels with Claude Code and Codex. This will make it worse! Available for Max users now, coming to Pro soon.

Source: Anthropic announcement

Cursor Gets Visual Bug Testing

Discussed at [26:01]

Cursor has shipped an upgrade where its agent can visit your website, click around, sign up, test password resets, basically act as a simulated user. It records a video demo of its interactions, routes that recording back into Cursor, and Cursor uses it to identify and fix problems automatically.

Kyle's take: This doesn't sound like a big deal unless you build things. Right now my workflow is: build first draft, send to my business partner Harms, he goes through and annotates screenshots in a PDF with feedback on spacing, copy errors, broken videos, missing testimonials. I feed that PDF back into AI to action the fixes. That works well. This is better. Set an agent loose on your site, it finds problems, records them on video, pipes them back into Cursor, fixes them, then sends another agent to test again. It's removing humans from the review loop. We're seeing this pattern across all of coding and it's coming to everything else.

Source: Cursor update

CloudFlare Engineer Rebuilds Next.js in a Week for $1,000

Discussed at [32:20]

One engineer at CloudFlare used AI to rebuild Next.js from scratch. The result is called vNext: a drop-in replacement built on Vite that deploys to CloudFlare Workers with a single command. In early benchmarks it builds production apps up to four times faster and produces client bundles 57% smaller. Total cost: about $1,000 in tokens over one week.

In the same vein, Claude Code refactored legacy COBOL applications and immediately knocked 13% off IBM's stock price. COBOL is a 66-year-old programming language that still powers a huge amount of enterprise finance and government systems. The problem is the people who understand it are retiring and almost nobody is learning it. So people have basically been stuck on IBM mainframes - legacy hardware to run legacy software.

If AI can refactor COBOL into modern languages, companies don't need IBM's mainframes anymore. One off-the-cuff blog post about being able to rewrite COBOL programs was enough to trigger IBM's worst stock crash in 25 years….oops.

Kyle's take: A year ago, vibe coding was for silly weekend projects. Andrej Karpathy said as much when he coined the term. Now we've got AI swarms building C compilers, engineers rebuilding the most popular frontend framework in a week, and single blog posts about COBOL refactoring wiping billions off a company's market cap. This is agentic engineering cracking big projects. It's not slowing down.

The arguments of “OK fine it can do x but it’ll never do y” are getting pretty weak…

Source: CloudFlare vNext announcement | Node.js rebuilt with AI (steipete's take)

Member Question: "How do I get into AI apart from teaching or training?"

Kyle's response: Just start building! If you've got no budget, use Antigravity from Google, it's free. If you've got some budget and you're non-technical, use Lovable. More technical with some budget, use Cursor or the Codex app on Mac. If you can spend £100-200 a month, Claude Code is the best option but you'll hit limits on the £20 plan quickly. Once you start building, you'll have questions about what's possible, and those questions become opportunities. The building itself is now the easy part. Distribution and finding what people actually want to pay for, that's the hard part.

This question was discussed at [30:44] during the live session.

Streaming on YouTube (with full 4k screen share) and TikTok (follow and turn on notifications for Live Notification).

Audio Podcast on iTunes and Spotify.