Right. Too often people conflate (a) risk-loving entrepreneurs and their marketing claims with the (b) realities of usage patterns and associated value-add.

As an example, look at how cars are advertised. If you only paid attention to marketing, you would think everyone is zipping around winding mountain roads in their SUVs, loaded up with backcountry gear. This is not accurate, but nonetheless SUVs are dominant.

Just this week a library got deprecated.

Open source of course.

So what's my response to that deprecating? Maintaining it myself? Nope finding another library.

You always depend on something...

> Maintaining it myself?

You say that like it's an absurd idea, but in fact this is what most companies would do.

I can maintain basic code no issue but not if it becomes complex or security relevant.

And I have worked in plenty of companies I'm the open source guy in these companies and me or my teams never had the capacity to do so

> I think 384gb of ram is surprisingly reasonable tbh.

> 200-300$/month are already 7k in 3 years.

Except at current crazy rates of improvement, cloud based models will in reality likely be ~50x better, and you'll still have the same system.

I've had the same system (M2 64GB MacBook Pro) for three years.

2.5 years ago it could just about run LLaMA 1, and that model sucked.

Today it can run Mistral Small 3.1, Gemma 3 27B, Llama 3.3 70B - same exact hardware, but those models are competitive with the best available cloud-hosted model from two years ago (GPT-4).

The best hosted models (o3, Claude 4, Gemini 2.5 etc) are still way better than the best models I can run on my 3-year-old laptop, but the rate of improvements for those local models (on the same system) has been truly incredible.

I'm surprised that it's even possible running big models locally.

I agree we will see how this plays out but I hope models might start to become more efficient and it might not matter that much for certain things to run some parts locally.

I could imagine a LLM model with a lot less languages and optimized for one programming language to happen. Like 'generaten your model'

Yes LLMs are a funny workload. They require high amounts if processing but are very bursty.

Therefore using your own bare metal is a low of expensive redundancy.

For the cloud provider they can utilise the GPU to make it pay. They can also subsidise it with VC money :)

We now have wall printers based on UV paint.

3D models can be generated quite well already. Good enough for a sculpture.

Did you try Claude to just extract it or to write a python script to do so?

I type fast.

I don't think I could Mimik a LLM. But I'm prompting fast so that helps

You can actually have dementia without the symptoms if your interconnect is strong.

Like knowing multiple languages/ multiple words for one thing.

The way the input doesn't match the output should imply that it's not just statistics.

As soon as compression happens, optimization happens which can lead to rules/learning of principles which got feed by statistics.

That's "just" more statistics though.

Are you good in math definitions or is this an opinion?

For me a compressed model learning rules through statistics is not statistics anymore.

Physic rules are not statistics.

Of course they are. Force has a strong correlation with mass times acceleration. Objects at rest have a high chance of being observed to remain at rest. And so on.

Statistic is not the same as constant equations.

The equations were discovered by experiment and curve-fitting.

We do the same thing.

We have long term memory and short term.

Context is short term.

The still long and expensive training phase embeds the long term memory.

Hmm, interesting.

One of the problems programmers have is loading a problem into working memory. It can take an hour. An interruption, a phone call, or a meeting can mean that you have to start over (or, if not completely over, you still have to redo part of it). This is a standard programmer complaint about interruptions.

It's interesting that LLMs may have a similar issue.

We do not do the same thing.

Right now inference doesn't cascade into training.

In biology, inference and training are not so decoupled.

What we do may be analogous but absolutely not the same

Yes if course after all ML and LLM are not made out of brain.

But do I really have to say colloquial?