Thought #9 - Never ending Updates and Unexpected Dolphins

Hi lovely humans,

A little over a week’s worth of releases and news, and it seems no one really slowed down (so apologies for the length of this newsletter). OpenAI have released 3 new models, Anthropic and Google also have new features. The news is everything from talking to dolphins to the first meeting of the UK AI Energy Council.

On the subject of energy, being polite to AI is costing millions (according to Sam Altman, CEO of OpenAI on Twitter):

What We’ve Been Up To

We’ve been pretty busy (and have some exciting resources, platform updates and news to share). But more importantly, we’ve had a new intern join!

Carys has joined as our Learning and Research Intern. She is a qualified teacher and studying towards her MA in Comparative Literatures and Cultures at the University of Bristol. We’re excited for her to apply her education background to our platform development.

AI New Releases

OpenAI

OpenAI have been busy as usual with 3 new models being rolled out:

o3 and o4-mini

The newest reasoning models (replacing o1 and o3-mini in ChatGPT for Plus, Pro and Team users). They are designed to think for longer, and they definitely do (which means they feel awkward if you use ChatGPT to actually chat).

Where they seem to excel is using images in their chain of thought reasoning (rather than just text like previous models).

We’re also seeing a lot of people complaining about the hallucinations in o3, and for coding both models have returned to adding “previous code the same” to the output (which is really annoying for vibe coding…).

We’re finding them both a little underwhelming, and will continue to use GPT4o for our usual ChatGPT usage, but we’ll keep you updated if that changes.

GPT4.1

Why is it 4.1 when we’ve already got 4.5? Our guess is 4.5 is still in preview to help them train the model, but 4.1 is ready to go. But who really knows?

GPT4.1 is only available using the APIs at the minute - we’ve swapped over in Dotly and it has been an easy change (and will be cheaper).

GPT4.1 is said to be much better at coding than 4o, and 4.1-mini performs much better than 4o-mini at all of the metrics.

We expect this to rolled out to ChatGPT users in the coming weeks.

Preparedness Framework

OpenAI have also released their updated framework for measuring new models and ensuring safety (big focus on biological weapons, cybersecurity and AI self-improvement).

One update to flag is the new ability for OpenAI to adjust its safety requirements if a competitor releases a high-risk system without comparable safeguards, provided that such adjustments do not meaningfully increase the overall risk of severe harm and are publicly acknowledged.​

So safety but not at too much of a risk to innovation.

Anthropic

In case you missed it, Anthropic, again similar to ChatGPT, have announced a new tier - Max at $100 per month for more usage. Unlike ChatGPT Pro, I am a bit tempted by this one as I do often hit the limits on Claude.

Research

For those on Max (not us), Anthropic have released Research (similar to Perplexity and OpenAI and Google’s DeepResearch). Focusing on research seems to be where all the big players are going.

One interesting feature of Anthropic’s version is the integration to Google Workspace.

Slick New Projects Interface

While this was rolled out a few weeks ago, we’ve only been really using it this week. But Anthropic has really upped the interface for their projects (where you can use the same files in several chats). You can now drag and drop files into the knowledge, and easily delete them.

New paper - Our Approach to Understanding and Addressing AI Harms

Similar to OpenAI’s announcement, but this is much more focused on AI harms (than big massive world ending safety). There was also a greater focus on the impacts on humans.

Google

Veo 2 Video Creation - rolled out in UK

Veo 2 can now be accessed via the Google AI Studio (but you must give Google Drive access to use). It is getting pretty good reviews, even compared to OpenAI’s Sora.

Google Workspace Flows

One of the big benefits of Gemini and Google AI in general is the seamless integration to Google Workspace - they have been working hard and rolled out some new agents which can really speed up using Gmail, Sheets, etc.

CanvaAI

It’s been all the rage on Reddit to click the envelopes on Canva - if you happened to miss this (and you’re signed in) you can still see what is going to be rolled out on their What’s New page.

The new features in the graphic design platform (used by a lot of marketing teams and startups instead of Adobe) involve AI. Everything from better prompting to bulk creating social images. We love to see a really hands on useful application of AI tools, and Canva seem to responding to user requests for useful features.

AI News

Lots and lots going on - mostly good news for a change, with some warnings to ensure responsible AI use:

• NHS in West London trialling AI app for skin cancer detection, aiming to save 10,000s of hours in waiting times - BBC.

• Google’s Gemini is being used to talk to Dolphins.

• UK Government is using a new AI tool to digitise housing records.

• Meta (Facebook) are coming for EU users data for their AI.

• First meeting of the AI Energy Council - the meeting was about how to generate more power for AI, rather than save energy with AI.

• Context.ai is joining OpenAI to work on model evaluations.

• Cursor (an AI coding agent) lost users because their own chatbot told users about a policy which didn’t exist.

• A Microsoft tool for screen recording called ‘a privacy nightmare’ has been rolled out outside the EU despite being paused last year - BBC.

• Lots going on in the world of AI chips (the hardware needed to run this tech) -NVIDIA stock has fallen 5% since an announcement that Huawei have a new chip [this is an impact of tariffs

Also worth a read

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

• Meta’s Llama3 was trained on (pirated) books rather than just the internet, including Gerry Adams’ book - the Atlantic.

• People are forming relationships with AI chatbots from therapists to wives - Guardian.

• Interesting to see a deepfaked debated for Keanu Reeves and Elon Musk being shared by Musk fans as true (on X formerly known as Twitter, unsurprisingly) - Indian Telegraph.

• (Or a watch) Sam Altman did a TED Talk at TED2025.

LinkedIn AI Poll

We were interested in whether people cared about regulation vs innovation, and people have come down very much on the pro-regulation (eg EU AI Act) side. Only 14% were pro-innovation (eg US).

Perhaps this has to do with the current state of the world, or our proximity to the EU over the US, but we weren’t expecting for such a strong result.

The comments were mostly around how regulation is needed but needs to implemented correctly.

We didn’t run a poll this week because of the Easter holidays in the UK - but we’ll continue with our usual from this week.

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 going to start sharing some of our new AI resources with you.

Laura
Always learning