Here are three things I found interesting in the world of AI over the last week. 

 

Of all the tech giants Apple have been the quietest about AI, until this week when they announced Apple Intelligence. There's a lot of things I like about this strategy - on device, privacy first, AI when and where people want it. But it is all designed to reinforce barriers to competition and I think is a step towards "AI as a monopoly" instead of "AI as a commodity".

One of the big applications of AI is as a natural language interface to existing software, and I think Apple gets this better than most. The models they are using are pretty underpowered compared to the leading models and they have instead focused on building out a semantic index of everything Apple knows about you and using that to augment queries to the AI. It will use on device compute where it can and Private Cloud Compute when it can't. I think it's a great strategy, for Apple and anyone happy to be locked in to their ecosystem.

 

I doubt anyone but Apple apps will ever get access to that semantic index which means trying to ship AI products in the apple system is going to be a tough gig.

Predibase released a study on fine tuning effectiveness - fine tuning index

Fine tuning is when you take an existing LLM (llama 3, mistral, zephyr) and do a small amount of additional training on your own data. Predibase specialise in providing cloud infrastructure for people to do the training and also run the fine tuned models and they have just published an evaluation of over 700 fine tuning experiments across 32 specific tasks.

 

For 85% of tasks the fine tuned models were stronger than gpt-4 by 25-50%. They were also faster and cheaper on all tasks. There is a bunch of interesting tech (checkout lorax) which they use to optimise serving lots of fine tunings which I am keen for an reason to play with.

 

HuggingChat have just released a set of tools to augment their free chatbot (image generation, image editing, web search, url fetch, document parser, calculator). They are available on the Cohere c4ai-command-r-plus model and are a great example of what open models can do.

 

I've found that the gap between HuggingChat  and closed providers has been closing slowly. You can still get stronger AI through OpenaI, Claude, Anthropic but for a lot of tasks you don't need the strongest model available, and for commercially sensitive or strong privacy requirements there is nothing better than open source tech.