Thought #17 - Thinking Machines, Teaching People

Hi lovely humans,

The news this week is very education -ocused - UK Dept of Education releasing a big doc on AI for Education, still lots of updates on the big announcements from the PM last week, an AI Skills Hub launched by InnovateUK and PWC, and sadly a report on university students using AI to cheat.

A nice reminder of how important education on AI in general, but also safely using AI in education and work, is going to be in the coming years.

Elsewhere Mistral has released some new model, we’re very interested in Netflix new data architecture and Apple and Anthropic have released some papers worth a read.

Our Week in AI

Interestingly - I have a very human based week (new hires, interviewing, lots of events) so have been using AI less than usual.

I have been using Perplexity for some research for a grant project - in our polls this has been the least used AI tool so far. It is a specific AI tool for research and web searching. They do excellent, well referenced research. And have been rolling out dataset links (eg to the SEC and academic papers).

Worth trying out, especially as the other major AI chatbots have all been releasing Deep Research mode (which is basically a Perplexity competitor).

AI New Releases

OpenAI’s new o3 model is cheaper and faster
The updated o3 API is now significantly cheaper, with some useful tweaks to the Responses API as well. Infrastructure plays like this are OpenAI’s real power moves - especially as they eye more government work.

Mistral launches Magistral - a new reasoning model
Positioned as Europe’s answer to high-performance LLMs, Magistral is smaller, open-weight, and focused on reasoning tasks. A promising step, but early days.

Mistral also announces Mistral Compute
Their own compute platform - again, a clear move into infrastructure. Everyone’s building their own stack now.

Revolut is launching an AI financial assistant
A real-world consumer tool - coming soon to your app, apparently. It’s a move towards AI as the default interface for financial services. One to watch.

Netflix shares their data model for ML pipelines
Standardising how teams build ML systems. More useful for technical teams, but a reminder that under-the-surface tools shape everything else.

AI News

OpenAI for Government
A $200 million Department of Defense contract and a new push to support governments using OpenAI’s models. No real surprise - but worth noting how fast this is scaling.

PwC and BridgeAI launch AI Skills Hub
A new UK initiative to support AI skills in business. Still very growth-sector focused, but good to see practical steps on skills.

UK government guidance on AI in education
A solid overview of risks and opportunities, mainly aimed at schools (a good single source of existing Ofsted and DfE resources).

Apple: The illusion of reasoning
If you’re into how LLMs “think” (or don’t), this paper breaks it down. The paper discusses how reasoning models just use pattern recognition - I’m a bit shocked this has been as big as it was because I personally can’t understand what else anyone thought they were doing.

Thousands of students caught using AI to cheat
A student survey suggests cheating is widespread. Not exactly surprising, but a reminder that we still haven’t figured out how AI fits into assessment - and a call for to start educating students on how to use AI responsibly

UK workers told to ‘embrace AI or be left behind’
A strong message from government, but without much support behind it. Still too focused on future talent - what about today’s workforce?

Amazon CEO: AI means fewer people
Blunt but honest. Amazon’s using thousands of AI agents internally and expects to need fewer staff. One of the clearest signals yet about real workforce change.

Google moves away from ScaleAI after Meta buys a stake
A big shake-up in AI data labelling. Google pulling back, Meta buying in. It’s not just about tools - it’s about who controls the training data.

Reddit launches AI ad tools
Advertisers can now use AI to understand Reddit conversations. Useful, yes - but feels like a privacy headache waiting to happen.

AMD releases new AI chips to rival Nvidia
Hardware news usually flies under the radar, but this matters. If AMD can offer real performance at lower cost, it’ll shift how teams train and deploy models.

AI is improving protein design (again)
One for the science crowd - continued advances in generative protein folding. Not flashy, but this is the kind of work that quietly changes the world.

Not Quite News, But Worth a Read (or Listen or Watch)

• Claude Code: How Anthropic teams actually use it
  Not a product launch, but a useful insight into how one company uses Claude internally for development. Less hype, more practical application.

• Alan Turing Institute x LEGO: GenAI and children
  Fun but thoughtful. Looking at how children interact with GenAI tools - a nice angle on AI literacy from an early age.

• AI meal planner takes off in the Netherlands
  Built on Microsoft tools, used by the biggest supermarket chain, and solving a real everyday problem. Feels like the kind of AI we want more of.

LinkedIn AI Poll

Last week we asked about how often people are using Microsoft Copilot - a large portion have never used it. This is unsurprising as it is the most expensive paid version (on top of a Microsoft 365 licence), and the only real edge is the ability to work with Microsoft products.

Vote in this week’s poll - please!

This week we want to know how much you are using Claude (my favourite) - already looking different from previous week’s polls.

Final Thoughts

As always we hope this was helpful!

Feel free to share this with anyone who might find it useful.

Next week, I am on holidays so fingers crossed the AI is quiet.

Laura
Always learning