Thought #37 - AI agents, safety guards, and space-bound servers

Hi lovely humans,

It’s been another week where AI felt a bit like scaffolding around the digital world - propping up productivity, poking into policy, and creeping into culture. OpenAI and AWS are deepening their ties, Google’s dreaming of space-based data centres, and Character.ai is finally setting age restrictions after tragic events. Closer to home, there's new UK government support for AI skills and a major expansion in NHS diagnostics. Meanwhile, signs of friction are showing: from job automation disputes to content moderation mishaps and another round of “AI is distorting the news” concerns.

And a study from IBM says 66% of businesses are seeing productivity gains from AI (which is a big contrast to what everyone else is reporting).

Our Week in AI

Remembering that all the tools have their own merit:

Cyber Essentials - ChatGPT

We are finally on the road to backing up our safe-by-design claims, and the first step is the self-certify Cyber Essentials.

I’m a big over-thinker and over-explainer, so this type of thing is my nightmare.

I did get a lovely human to take a look (massive thanks to Neal Riley!). But I also got ChatGPT to go over with me what it was looking for, and how to word my responses (a difficult thing for a chatty dyslexic person).

One tip: I had to use the prompt:

be critical, don’t just agree with me. Give things an importance score - needed, nice to have, something to think about for future.

Otherwise it seemed to simultaneously agree with everything, and have too many unimportant suggestions.

Coding Help - Claude

I ran out of my Claude Pro messages (for only an hour), but being impatient as ever I turned to ChatGPT.

I was trying to get Google Login running in our new local developer set up.

ChatGPT got me pretty far. Was able to debug I had copied the wrong URL to Google console. But it could not work out why the direct was going to the wrong place.

Once Claude was available again, it understood what was wrong straight away - just from a sentence or two.

Note: we have a really standard set up of a common tech stack, which is what Claude has been optimised for. My general advice, if it’s not sensitive, is to share you code with AI agents to get help.

AI New Releases

GitHub Copilot comes to Slack
You can now use GitHub Copilot directly in Slack, giving dev teams a way to generate, discuss and review code without toggling tools. It’s in beta, and you’ll need to join a waitlist.

GPT-OSS-Safeguard: OpenAI’s safety reasoning models
OpenAI has released two new models - GPT-OSS-20B and GPT-OSS-120B - as part of a research preview. They're designed to make safety more transparent and adjustable, letting users set their own policy rules and see how the models interpret them.

Perplexity launches Patents Agent
A new specialised agent can now help users search and understand patent information. A handy tool for researchers or anyone navigating IP-heavy sectors.

Grammarly rebrands as “Superhuman”
Grammarly is now positioning itself as a broader AI productivity suite. The name’s a bold one - but the tools include writing, meeting, and comms support.

Google’s October AI update
Google’s latest updates include Gemini for enterprise customers, cancer research collaborations, and an upgraded version of Flow, their filmmaking assistant. A mix of very practical and very ambitious.

Zoom partners with Scholhouse for AI tutoring
The new partnership will focus on AI-supported tutoring for schools and colleges - though exact tools and timelines are still emerging.

AI News

AWS and OpenAI announce multi-year strategic partnership
The deal includes deeper infrastructure collaboration, and suggests OpenAI is spreading its hosting across providers - possibly a play for resilience or bargaining power.

IBM: 66% of large EMEA businesses see productivity gains from AI
The findings differ notably from other studies. Larger organisations report higher gains than SMEs, highlighting the scale and support needed to implement AI well.

Character.ai restricts under-18s
After a tragic case involving a young user, the platform is now blocking under-18s from access. A stark reminder that AI companions aren’t neutral, and risks can be real.

UK launches new support to fill the £400bn AI skills gap
Support will include vouchers, bootcamps, and partnerships with training providers. Focus is on SMEs and helping leaders understand how to use (and not just deploy) AI.

NHS expands use of AI to spot liver disease
The AI tool is already being used in GP practices in Staffordshire and plans to roll out further. It flags people at risk before symptoms appear, speeding up treatment access.

Anthropic and Iceland launch national AI education pilot
Teachers across Iceland will gain access to Claude as part of a pilot - making it one of the first national-scale education AI rollouts.

Worldpay and OpenAI launch “Agentic Commerce Protocol”
The idea is to streamline online payments initiated by AI agents. Interesting for retail, but we’ll see how much autonomy merchants are comfortable handing over.

JP Morgan faces backlash for AI-powered performance review tool
Staff are reportedly frustrated with a tool that helps managers write reviews. It’s not the first time workplace AI has raised eyebrows around fairness and oversight.

UK government spending over £3.35bn on AI contracts
Sky News investigation shows a sharp rise in public sector AI contracts. The Met Office’s contract with Microsoft tops the list - though Taught by Humans is in there too (just a tiny bit).

Google plans AI data centres… in space
Yes, really. Still early days, but highlights the power and cooling demands of modern models. Also sounds like a sci-fi book pitch.

Perplexity served cease-and-desist by Amazon
Amazon is not happy about Perplexity’s agents “shopping for you” on its platform. Could kick off broader legal battles around agentic browsing.

Stability AI wins court case against Getty Images
A UK court ruled that Stability AI’s use of Getty images didn’t breach copyright. A controversial verdict, with critics saying it puts creators at a disadvantage.

BBC: AI, mortgages, and social media to join English school curriculum
These updates reflect a growing sense that schools must tackle AI literacy and financial awareness. We’ll keep an eye on how it’s implemented.

AI-generated music charts on Billboard
Xania Monet has become the first AI artist to hit the US charts. It’s not the tech novelty - it’s the promotion, packaging, and human marketing machine behind it.

YouTuber wrongly banned by AI moderators
Popular creator Enderman was removed by YouTube’s automated system and only reinstated after a public outcry. Another case for keeping humans in the loop.

Tech stocks wobble as fears of an AI bubble rise
Nvidia hit a $5 trillion valuation - but others warn of inflated expectations and a possible correction. Interpretations vary, but the volatility is worth watching.

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

• A former OpenAI safety lead warns: “Don’t trust its claims about erotica”
  A sobering op-ed from a former OpenAI safety lead, raising concerns about how product decisions are made - and how safety claims are being handled. Serious, and worth reading.
  (And for those without a subscription, a kind Redditor has shared the full text here.)

• AI at work: less hype, more output
  A thoughtful piece from Perplexity on where AI actually helps, especially in research and admin-heavy workflows. Useful for anyone rolling it out across teams.

LinkedIn AI Poll

Last week we asked if people were happy with OpenAI having their company knowledge, and the vast majority weren’t:

Vote in this week’s poll - please!

The Times is reporting 56% of people use AI for financial advice - so we want to know if you are?

Blogs and Resources

We’ve added a few new bits to our resources page - all short, simple, and designed to be actually useful. Recent additions include:

• Minimising hallucinations (and spotting them)

• Simple prompting methods that don’t need fancy frameworks

• How we’re using all the major AI tools right now

• Specific tips for getting more out of GitHub Copilot

• What to consider around data privacy when using AI at work

More coming soon - and if there’s something you’d like us to cover, just shout.

Final Thoughts

As always we hope this was helpful!

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

Next week, we’re praying for a slightly quieter week.

Laura
Always learning