Thought #4 - Too Many New Updates - Can AI Slow Down for a Second?

Hello lovely humans,

We had every intention of sending this last week, but AI developments need to pause so we can catch up on testing them out and writing about them...

We also had a wonderfully insightful newsletter on the skills needed for AI, but so much has gone on, we’re going to focus on the new developments and save that for next time.

Keeping up with all the goings on in AI can be a full-time job some weeks, don’t worry even we struggle and it is our job. Hopefully we can help get the important information across without overwhelming you.

(This is much longer than our usual newsletter, but reading it means you’ll be up to date with AI in 2025)

The DL on DeepSeek

DeepSeek, a Chinese company, made headlines in January with a new AI model. It was claimed to outperform competitors like OpenAI while being significantly cheaper, sparking global discussion.

If you struggled to keep up with all the noise around DeepSeek, then this educational blog by our Founder’s Associate Tom is for you. It’s long but covers everything (including what it actually is, why the US stock market fell, whether we trust it).

Tom is (in his words) not an AI expert, so he has been sure to explain all terms and background needed to follow what happened. I might be biased but I think it’s an excellent blog, and everyone should read it.

All Things OpenAI

It feels like OpenAI has been releasing something new every week this year. It has been a lot to keep up with so we thought we’d summarise.

Sora - AI Created Videos

Last year OpenAI teased everyone with Sora - a beyond impressive AI video generation tool. They then rolled it out in the US, leaving everyone else a bit put out. But finally, we can access it in the UK.

I had a quick go trying to create a video.

To summarise - I was very impressed with the technology, not necessarily with my first attempt. It feels like there is a bit of a learning curve (and some real creative skills required). We’re going to keep playing, and maybe read some of the learning guides.

GPT4.5 - a new model

For a bit of context - ChatGPT is a conversational chatbot interface around a bunch of different AI models OpenAI has created. These models can be used in ChatGPT by prompting (using spoken or written human language) or via the APIs using code.

Last week a new model was announced - GPT-4.5. Many new AI models (like ChatGPT o3 and DeepSeek R1) are designed for reasoning - solving complex tasks step by step. GPT-4.5 is different - it focuses on more natural responses, improved emotional intelligence, and reducing 'hallucinations' (where the AI confidently makes up false information).

Initially GPT-4.5 was not available for ChatGPT users (but developers could access it using the API).

We haven’t had a chance to play around with the API yet, but one of our trusted expert Tia Collard from fiftyminds has been sharing her views on LinkedIn

But since yesterday (Thu 6th March 2025), paid users of ChatGPT can select and use GPT-4.5. We played around quickly, and we like it better than the o3 models for general chatting (but it seems a lot more wordy than GPT4o). We’ll keep you updated with any tips.

Deep Research

Last week Deep Research was also rolled out in the UK - this is specifically for researching a topic and writing a report. I’ve tried it out and shared my thoughts in this blog.

One thing this release makes us think is how much OpenAI is moving towards what Perplexity is already doing (more on Perplexity’s updates at the bottom).

Incase you missed some updates in January

• o3-mini - a new model with “advanced reasoning”. We mentioned it briefly in the last newsletter - it doesn’t use the memories about you saved in ChatGPT so we initially struggled. But as we get used to it, and for specific tasks, we’re starting to see the value.

• Tasks - by selecting GPT-4o with scheduled tasks, you can ask ChatGPT to follow up later. We might be doing it wrong but we haven’t got this to be particularly useful (or reliable) yet.

• Updated user custom instructions - we quite like this one, the ability to tell ChatGPT more about yourself and how to talk to you (we also do this with Dotly!)

Anthropic is Also Ramping Up

Anthropic is not as well known as OpenAI, but a very important contender in the space. They were created by former OpenAI employees, who left because of safety concerns. Whether this is true, or a great marketing ploy (or both), their models and chatbot Claude, focus on safety and are often preferred by those who care about data privacy.

Now, they are making headlines again by announcing a Series E funding round with a $61.5 billion valuation.

Claude Sonnet 3.7

It has been awhile since a new model was released by Anthropic (we’ve had several updates on 3.5). Claude Sonnet 3.7 is their new reasoning model, and it comes with writing style choices (from some presets or creating your own by providing writing samples).

We find this new Claude model works well with our usual prompting style (unlike with ChatGPT’s new models where we usually have to change how we use it), and we are finding the conversation style more human. Like the other reasoning models, it will output it’s thinking for the user as it goes.

Other Anthropic Releases

• Claude Code is an AI agent for coding. It needs to be downloaded and runs in the terminal / command line (which adds to the coolness of this in my opinion).

• Integration to Google Drive - they have also rolled out a handy integration to Google Drive which makes using their Projects feature much more effective.

Overall we’re big fans of Claude, and find it much better than ChatGPT, especially for anything technical.

Make It Fair

On 25th February, newspaper across the UK shared the same front page with a call Make It Fair.

The call is for UK creatives to be protected, over the Big Tech companies which they claim the UK Government are siding with. It comes from AI companies training their models on information such as articles, blogs, images, other creative content available online without having to pay or even credit the creator.

How these discussions impact copyright rules will could have a big knock on effect on the current AI models we use. Similarly, there are important implications for the future of the creative industries, and jobs in general.

If you want to find out more News Media UK have the details.

Gemini 2.0

Gemini is the model (and chatbot) for Google’s Generative AI (created by DeepMind, and previously called Bard).

Google have seemed like the slow horse in this AI race with Bard being almost unusable, and Gemini 1.0 seemingly okay. But with Gemini 2.0 (both the API and the Google Workspace integration) it seems they finally have a contender.

Tom has been testing out Gemini 2.0 and comparing it to the free version of ChatGPT (for fairness). He seems impressed so far (although he’s excited for Gemini to help with his Google Sheets formulas).

If you are using the API for basic tasks which don’t need advanced reasoning, then it easily comes out as the cheapest.

Perplexity

Perplexity is a conversational chatbot which focuses specifically on web searching and research (kind of what OpenAI is rolling out with new ChatGPT features). You can try for free without logging in. We still find them quicker and easier to use for research than ChatGPT’s deep research, but the overall report is probably more logical from deep research.

They also have a waitlist for their AI Agent Comet.

Perplexity Business Fellows

I am now one! Complete honesty, I’m sure they let everyone who applied in. But I am really keen to learn more about Perplexity (and the lectures are from industry leaders like Jenson Huang CEO of NVIDIA and Ali Ghodsi CEO of Databricks). I think this is a great opportunity to learn new skills and use cases to share with our customers.

Perplexity Investment Fund

Perplexity themselves are valued at around $9 billion, having closed a round in December which included Softbank (pyments.com). Very impressive for a 2 year old company.

They have also announced a $50 million investment fund for early stage AI companies in the US. It will be interesting to see the type of AI companies they choose to invest in.

Phew - we’re sure we’ve missed something, we’re only human after all. And we haven’t even touched on regulation or other important topics.

But we hope this helps you feel confident in the AI world. And maybe inspires you to start experimenting with these tools.

If you know anyone else who’d find this useful, please share!

Laura - always learning