Hi Everyone,

We always like you to get the most value out of the newsletter, in addition to the email version of the newsletter there is a supporting website where you can log on and get to all the previous versions. You can also, share, and comment and do all the usual things.

Thanks for reading.

Paul

Table of Contents

• New chat GPT ‘models’

  • Memory

  • Strawberry the ‘new model’

• A summer of visual tools

  • Midjourney

  • Leonardo.ai

  • Ideogram

  • Adobe

  • Runway ML

  • Why is any of this of interest?

• Some links to have a look at

• The Current State of AI in Organisations

  • Three-Point Action Plan for Getting Started

• LLama 3.2..

  • New Models

  • Capabilities and Performance

  • Technical Innovations

  • Availability and Partnerships

  • Focus on Responsible AI

New chat GPT ‘models’

OpenAI rolled out a few features in the closing part of the summer. We still have no information on chatGPT 5, which will probably be called chatGP o2…

Memory

Announced in February, the memory feature was released into general use. The basic idea is that the AI ‘learns’ and remembers the most important things to you, and uses this in its framework of responses.

In practice, for me at least, it seems to have made absolutely no difference.

Strawberry the ‘new model’

Openai released a new ‘very limited’ access model which was codenamed / rumoured under the name ‘strawberry’. Its official name is chatGPT o1.

1) limited to 50 queries per week

2) very limited user interface , no file upload, no images, no audio

3) better reasoning model

The model data set looks to be the same knowledge cutoff date (sept 2023) but either fine tuning, or an additional ‘expert’ has been added to the interface. This causes the AI to spend much more time ‘reasoning’ (inference time) and follow a much more comprehensive chain of thought. What it is actually doing is not shown

The results, quite mixed. For example, when given the prompt below to claude.ai

produce a plan to organise a conference on the future of agentic Ai in the west midlands, in November 2024, be specific what do I need to do, spend some time <thinking> and the produce a detailed and specific step by step plan for me

The well known inclusion of a <thinking> block to force forwards reasoning produced almost identical results, both produced a very good starting point for the plan. What is really apparent though is that if you compare this to a year ago, they are both much further in capability than chatgpt 3.5.

Conversely, my favourite training course prompt (meal planning for the week), it clearly did a better job. Avoiding all the pitfalls of mixing measurements, not adding things up. But the ‘thinking messages’, read like the loading screen prompts from an old version of ‘The Sims’, including the following

In conclusion, I have been able to get almost exactly the same results just using chain of thought prompting. Is this some smoke and mirrors, or a fundamental improvement. Somewhere in between.

Just a word of caution. Remember the ‘jagged edge’, I would suggest this thinking model is more prone to being tried on complex tests by yourself, on a couple of my tests it has been 100% wrong. Use it to supercharge what you can do, not do what you can’t do.