myzz's latest activity

The nice thing about learning to use LLMs is you can just ask them. They know a lot about their own capabilities and limitations. I’m a few weeks into my journey of properly learning to use local models, having previously only dabbled, and I’ve...

https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned TLDR or the takeaway from this is that LLMs do not function the way normal code does. Its much more opaque, non-deterministic, and at this point will "run" anything...

Credit Default Swaps where have I heard that before...

Not only the prompting but being able to tell if the output is complete garbage or not.

Good analogy actually - because when you do math with a calculator you have known inputs and expected outputs and it is deterministic. AI coding is none of those things. Not yet at least. The big issue is that (some, maybe most) software is made...

A while back I spoke to someone studying cg animation at a university. The training started by doing traditional cel animations to get the basics of motion and keyframing etc. This is the only way going forward if we don't want things to...

Didn't know where to put this post, but it started with wanting to test drive some smaller models before shelling out for my own hardware so I'll put it here... Stumbled on Qwen 3.8 27b being served on Cerebras hardware, and wow it is FAST. Got...

It's not a tool in the same way we had tools during the Industrial Revolution. You can vet the tools of the Industrial Revolution. Once you're confident that it works, then you can displace the effort into thinking and building with the tools...

Except the Jurassic Park OS/file system was real/based on a real one - the File System Navigator made by Silicon Graphics - the company whose computers ran the renders for the CGI for Jurassic Park - and it was/ran on a UNIX system (IRIX)...

I have hermes agents on 7 hosts, all peered A2A protocol and aware of each other. Effectively lets me command as "overwatch" from my primary PC, or primary laptop and orchestrate not only my infrastructure but design/dev/deploy end to end, code...

Same here! I got Claude to build scripts to export processed clips out of Resolve, so now looking into an agent to run it start to finish from the setup of the project and paramaters to a final folder of clips with effects like noise reduction...

I suppose this is a move to be like MS <-> Github? Except betting on the future - once every last corner of github is scraped its worthless and these LLMs are the future.

It is a smart way to keep KV caches reasonably small at large contexts (1M). According to the friendly neighbourhood LLM: Hybrid attention in Large Language Models (LLMs) is an architectural approach that combines traditional full (softmax)...

I like it, it makes sense (y) I thought maybe more separation from local vs. cloud models would be good but reading through the posts it works.

I'm just using too many tokens with Claude. It's starting to slow me down as I'm pegging my weekly limits easily. I'm thinking about building a new desktop with a 5090 for Qwen.

The most bizarre thread I have seen here to date. 🙃

You are correct, this setup seems to be the best bang for the buck right now if you want useful work. "Dev kit" or not. Prefill is what kills the performance on the mac. Also performance doesn't seem to scale well with multiple macs. I like the...

That is a pretty interesting option. $8208 over 3 years, for 256G + 4tb, similar specs to 2x Sparks.

Funny how dual Sparks was my suggestion at the very top of this thread.

2x DGX Sparks will give you the most bang for the buck considering what you actually can run on them.