Thought 14 - The Week We Hit Pause (Sort Of, Maybe, Not Really)

Hi lovely humans,

It’s been a bit of a strange one this week. The biggest news is the unicorn builder.ai going through insolvency and the revelations about them impacting trust in AI products. On the policy side - the EU Commission might be pushing back the EU AI Act.

In terms of releases - the big news is Anthropic releasing Claude Sonnet 4. Meanwhile, AI voices are irritating train passengers, and even The Economist has declared we might be in the trough of disillusionment.

What We’ve Been Up To

Writing blogs - we’ve been hit by a bit of a writing bug at Taught by Humans’ this week (especially Carys).

She wondered if ChatGPT is a good teacher:

Contemplated what AI looks like in the next 5 years, based on recent reports:

While I have been thinking about whether our dependency on tech is worrying (especially with the builder.ai news):

AI New Releases

Anthropic

• Claude 4 quietly lands - No flashy livestream, just a strong update. Claude 4 is now Anthropic’s go-to model across tools, with notable improvements in reasoning and coding. We’re definitely seeing an improvement in coding ability, but we’re not convinced on its writing and chatting abilities - a proper write up and test coming soon.

• Claude Sonnet 4 also live on Perplexity - Perplexity were quick off the mark with this one.

OpenAI

• OpenAI Responses API gets an upgrade - Now includes options to structure and moderate outputs more easily, in line with their Model Customisation Platform (MCP). Small changes, but ones that might quietly improve safety and usability.

• WhatsApp meets ChatGPT - Not a tool, just a light-hearted post, but shows how personal and travel assistant-style prompts are becoming more visible in mainstream chat use.

• Codex is now on iOS - Now embedded in the ChatGPT iOS app, this could improve code-related chats, although early tests still show mixed results compared to Claude.

Google

• Google I/O 2025 - Gemini 2.5 Pro was announced, with major performance boosts, especially in education and logic tasks. Google Search now includes an ‘AI Mode’ for multimodal results, and AI Overviews are now reaching 1.5 billion users monthly.

Other New Releases

• Mistral adds OCR and PDF capabilities - A step forward for enterprise use. Document understanding is still a weak spot for many tools, so this could unlock use cases in sectors like law, finance and government. Mistral are really playing for the enterprise market.

• Notion expands its AI integrations - With 13 tools now supported, this marks a shift from a walled garden to a more connective ecosystem.

• Miro AI goes from demo to release - Clickable wireframes and live AI prototyping are now officially out of beta. While still basic, the combination of diagrams and workflows in one place could be genuinely useful for product and learning teams.

• Spring AI 1.0 released - A framework aimed at developers building production-ready AI apps. It supports multiple LLMs and integrates easily with vector databases like pgvector and Redis. Useful if you're building internal AI tools without wanting to go full-on ML engineer.

AI News

• Builder.ai’s financial troubles emerge - Once valued at $1.5bn, Builder.ai has reportedly hit funding issues, cancelled projects, and laid off staff. A reminder that glossy marketing and celebrity backing (Will.i.am was involved) don’t guarantee sustainable tech.

• EU delays AI Act enforcement - The EU has quietly pushed back parts of its flagship AI Act. Key rules for general-purpose AI models like GPT-4, originally expected to apply soon after summer, will now come into force next year. This gives companies more breathing room - but also extends the current period of uncertainty, especially as national regulators are still being set up.

• ChatGPT model “refuses” to shut down - A dramatic-sounding headline from The Independent, but the real story is about how newer models handle shutdown signals - useful for safety testing, not Skynet.

• Jony Ive joins OpenAI to ‘reimagine’ the computer - Possibly more design PR than product at this stage, but the idea is a new kind of AI-first hardware device. Given the personalities involved (Ive, Altman, and SoftBank), expect hype. Whether it’ll deliver something actually useful is another matter.

• UK-EU AI collaboration reboot - A small but potentially useful policy step, inviting British institutions to become gateways to EU compute resources.

• First public sector GenAI tool launches in Derby Council - An interesting milestone. We’ll be watching closely to see if it's genuinely useful or mostly a pilot press release.

• ScotRail's AI station announcements criticised by the public - Passengers prefer real voices. Raises questions about accessibility, clarity, and the role of AI in public-facing roles.

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

• We're in the AI Trough of Disillusionment - The Economist reports that corporate excitement has dimmed as costs rise, use cases under-deliver, and hallucinations persist. It’s not a crash - more a collective “hmm.” Worth reading for context, not just headlines.

• Can AIs understand emotions better than humans? - A joint study from Bern and Geneva put six LLMs through emotional intelligence tests. The models not only outperformed human responses, they also generated their own EI assessments. Fascinating - and raises big questions for education, therapy, and HR.

• Two paths for AI - A thoughtful piece from The New Yorker on the divide between those worried about superintelligence and existential risk, and those focused on immediate regulation and real-world harms. Useful context for anyone navigating the policy or ethics side of AI.

LinkedIn AI Poll

Last week we asked if people use different models within ChatGPT - and overwhelmingly people just start chatting to the default model (currently GPT-4o).

20% of people didn’t even know this was a thing.

Currently we use 4o for chatting and general use, prefer o3 for coding or really specific tasks (although I do switch to Claude for these). And we really like 4.1 for data analysis - especially how it presents the findings (but we do switch back to 4o to get a report if needed).

Watch this space for our guide on which models and tools to use!

Vote in this week’s poll - please!

We’re kicking off a series of polls to understand AI usage - starting with whether people are using Google’s Gemini models and tools.

Final Thoughts

As always we hope this was helpful!

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

Next week, hopefully less news and more resources from us!

Laura
Always learning