picozeta

Hi,

I am on the mid-plan (paid) and using the web interface and in the last weeks the quality of Claude's output quality decreased dramatically. The model used is Sonnet 4.5 and my workflow is copy/pasting code and inspecting the diffs.

It changes code it shouldn't change, even makes syntax errors (like a character different than the one provided in a source file).

I know these LLMs change all the time, and this is just an anecdote, but I am interested what others think.

Small anecdote:

When doing let's call it "mundane" coding, I am able to answer a big chunk of questions using ChatGPT only without having to use a search machine to "ask the internet". If you are experienced in the domain you can give useful hints to the LLM (e.g. you know which libraries to use, what's critical, etc.) and really get extremely useful answers. And I am not talking about common areas, but highly specialized ones.

I think that's pretty remarkable, as doing coding without Internet access is ... challenging to say the least.

... or fat client era?

At the moment Chat-GPT and other instruct-based models show what's possible with modern LLM models. Although most SOTA models need cloud-compute to predict, it's feasible to assume that they might work on beefy standard desktops in the next 5 to 10 years (IMHO).

Now, historically we had:

  - thin-client main frame architecture (1970s - 1980s)
  - fat-client "home computers" (1980s - 2010s)
  - thin-client SaaS software platforms (2010s - 2025s)
  - fat-client LLM inference engines (2025s - ?)

In particular I think there will be a lot of ethical questions and legal work for companies to sell LLMs as SaaS, and because of fear to "recommend stuff against the status-quo", they might be inferior to "open" (unconstrained) models and that might be just possible for private persons (at first).

Just my 2 cents, what do you think?

Currently we are living in a thin-client epoch where many important compute stuff is done via SaaS (Google, Amazon, Wikipedia, StackOverflow, Dall-E, ...), but due to hardware and AI advancements it might be possible to compress much of stuff that matters into reasonable size (stable diffsion is around 4G and mobile phones might have 2T in the next 10 years).

Though important stuff like

- getting images (via iterations of stable diffusion) - derving facts (via GPT iterations and wikipedia) - mapping (via just compressing openstreetmap in 100G) - math (via open source jupyter iterations)

What do you think?