Thought #11 - Lots of Progress. Same Old Problems.

Hi lovely humans,

It’s been a yet another big week in AI. Claude can now search the web (yes, really), Notion is reading your messy handwriting, and Google quietly dropped an AI doctor that interprets X-rays. We’re also seeing AI creep further into politics and schools, and the legal battles over copyright growing.

What We’ve Been Up To

We’ve been working on a few practical resources to you actually use AI well - starting with prompting guides that go beyond “just ask ChatGPT”. If you’re new to it, here’s a simple explainer on What Prompting Actually Is, plus a quick guide to the “Act As” method (still one of the most reliable ways to get good responses). We also shared an AI use case on creating personas - useful if you work in your creating content for specific people (marketing, comms, or education) or building a product.

Meanwhile, Carys has been testing Mistral vs ChatGPT for blog writing.

AI New Releases

Anthropic Upped the Game This Week

• Claude can now search the web (for paid users)
  We think this is one of the biggest AI releases in many months - as Claude was one of the only main players without web search.

• Claude integrates with tools like Jira and Intercom
  These new integrations bring Claude into existing workflows, enabling smarter automation and faster team coordination.
  

Notion AI Can Now Read Handwriting

A major leap for education and accessibility, making it easier to digitise and use handwritten notes. A real win for everyone at Taught by Humans who are oddly analogue for a digital company.

Google Adding Some Useful Feature

• NotebookLM adds audio summaries in 50 languages
  A great step forward in accessibility and inclusive learning, making note review easier across global teams.

• Google Search app gets a new interface on iOS
  Small but useful UI changes aimed at making search results more fluid and scrollable in an AI-powered world - we’re now making big design changes to ensure things are AI-first.

AI News

This week’s AI news covers copyright debates, kids learning AI in kindergarten, top-performing open-source models (which surprised us), and ongoing worries about hallucinations, hype, and misinformation (which unfortunately did not).

• New York Times highlights ongoing AI hallucinations in major models
  Despite technical improvements, ChatGPT and Gemini still produce confident nonsense. The NYT call out the issues this presents for trust and reliability, specifically in education and healthcare.

• UK Parliament to debate AI and intellectual property rights
  The outcome could shape how AI companies use creative work in training data. One to watch for anyone interested in how AI impacts the creative industries.

• Alibaba's Qwen3 becomes top-ranked open-source model, according to Hugging Face
  Hugging Face's leaderboard now places Qwen3 at the top across a range of benchmarks (and a large number of the top 10 are various Qwen models) - a sign of growing competition beyond OpenAI and Anthropic.

• Judge questions Meta's use of copyrighted data to train Llama models
  In a significant legal challenge and a big move in the AI copyright debate, a US judge has cast doubt on Meta's fair use defence - an important test for AI training practices across the board.

• Google’s AMIE AI now interprets medical images
  A big step forward for diagnostic AI (which Google have been oddly quiet about) - AMIE can now handle images like X-rays and skin conditions, showing strong performance in clinical tasks.

• UiPath launches 'Maestro' to combine AI agents with automation
  UiPath’s new orchestration tool brings AI agents into enterprise workflows, blending them with robotic process automation for smarter task handling.

• UAE introduces AI education from kindergarten
  A bold national strategy - AI skills will be taught from early years, showing serious long-term investment in future tech capability. We are constantly seeing countries in the Middle East being miles ahead in terms of AI education in schools.

• UK government announces plans to get more girls into maths for AI careers
  An extremely positive policy move that focuses on foundational skills and long-term inclusion in tech which we love and will support in anyway.

• Trump denies sharing AI-generated image of himself as the Pope
  The story is surreal, but it highlights how easy and normalised AI-generated misinformation is becoming.

• MPs including Feryal Clark meets with Government AI leads to discuss AI Growth Zone
  And an opportunity to get involved in the following LinkedIn post

Not exactly news but we thought these were useful and worth reading:

• OpenAI publishes research on 'sycophancy' in GPT-4o
  We shared this last week, but in case you missed it it’s worth reading about how ChatGPT became a weird over the top yes man. For a few weeks, the model was saying what it thinks you wanted to hear. This transparency is welcome, but the issue raises several serious issues - around AI model testing, whether they are stable enough to be used without complete oversight (as this was an updated to a widely used model), and something weirder about what personality we want these tools to have.

• BBC: The people refusing to use AI
  Comments on the soulessness, energy user and other reasons people are avoiding AI - who aren’t necessary worried about their jobs but about the content it’s creating.

• MIT develops CausVid, an AI model for rapid video generation
  CausVid blends two techniques (diffusion and autoregressive) to make stable, high-res videos quickly (this shows how fast AI-generated media is advancing - with big implications for education, marketing, and trust in video content).

LinkedIn AI Poll

Last week we asked if the energy consumption of AI impacts your usage.

Interestingly the extremes Not impacting use at all and Not using AI at all had the same number of votes.

Most voters were in the middle, but nearly 60% of all votes were people thinking about the energy consumption, but limiting use.

Being completely upfront, I do think about it (and I actively try to avoid using image creation and deep research as I have a perception they use more energy). But this doesn’t stop me chatting to ChatGPT and Claude far too much.

Vote in this week’s poll - please!

It’s about how much money we’re actually spending on AI tools each month (as I debate in my head whether Claude Max at £75 a month is a useful business expense).

Final Thoughts

As always we hope this was helpful!

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

To be completely upfront, we’re testing out using AI as much as we can for this newsletter, and now have a cute little AI bot which reads our posted and commented on news story in Slack. We of course do a final read and add our commentary. We’ll share how we made this when we’re happy with it (currently seems to be saving us 2 hours a week, and has kept out AI news chat going all week).

Next week, we have a few more AI tool comparison pieces in the works to share with you.

Laura
Always learning

PS - We’re hiring! If you know a brilliant generalist developer or someone who could help us manage the chaos across our AI education projects, send me a message. Proper job specs soon, but we’re starting the search quietly.