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

 

Claude gets smarter, cheaper, and faster 
Anthropic have released Claude Sonnet 3.5, which is a lot stronger than their previous strongest model Claude 3 Opus and 20% of the price. Anthropic have 3 model sizes Opus (biggest), Sonnet, and Haiku, so this is their mid-size model getting a bump with a 3.5 Opus expected later this year. The model is pretty much even with gpt-4o, though slightly behind in the chatbot arena leaderboard, though it is better at coding and now my default coding assistant.

 

A new model is cool but it is their improvements in UI which are getting me to use Claude as my AI of choice (unless I need web search which they don't have). They released artifacts, which separates text output like code and documents into their own virtual files to make them much easier to review and manage. Today they also released projects which let you save documents and prompts for specific workflows and I like a ton more then OpenAI's gpts.

 

They are still behind in terms of breadth of tools (search, image generation, advanced data analytics) but their model strength is right up there and I feel like the gap is closing with OpenAI.

 

The wired article headlines "Perplexity is a bullshit machine" and "Perplexity plagiarized our story about how Perplexity is a bullshit machine" give you a sense of the vibe. A lot of people have been testing how and when perplexity scrapes websites and are finding that it is at odds with the companies stated policies.

 

But most importantly they also found that it would sometimes not visit the site at all and instead just hallucinate about the content. It feels like 'move fast and break things' is very much alive in the culture at Perplexity. Always remember to double check the sources if you're relying on output from Large Language Models.

 

I wrote an post on linked in with an overview if you want to learn more.

Music studios sue AI music generators

Suno and Udio are two of the most popular AI music generators and the big music studios have banded together to sue them into oblivion, or maybe just to the negotiating table. My bet is oblivion, to make any future negotiations more lucrative.

 

There is a strong narrative from a lot of creative professionals that "AI companies train models off our life's work that can create cheap substitutes and hurt our livelihoods". AI companies responses are basically "Our models don't copy, they learn".

 

I think both sides are right - one way to view large language models is as large compression algorithms. They compress massive amounts of data into (relatively) small files and you can use those files to get back the gist of things, and fragments of the actual form. It takes a lot of work from AI companies to stop their models reproducing training data verbatim (they call it regurgitation) and that smells a lot like copyright infringement.

 

On the other hand, copyright is pretty clear that you can't copyright ideas, only the final form of works. I've no idea where all the lawsuits will end up but I wouldn't be surprised if we find it to be different in various jurisdictions. It will be fascinating how the concept of copyright evolves in an AI world.

 

---

 

It isn't AI related but I hope Apple get absolutely hammered by the EU and the rest of the world follows suit and breaks down the walled garden known as the App store.