srslack

Especially with half of the thread being unable to read or see the GitHub icon implying it’s a marketed product when it’s self hosted. Even if the docs are insanely slopped: if they’re to be believed and it works it is well beyond any locked in cloud runs/orchestration from Anthropic or OpenAI.

You could not get 500GB of VRAM for less than $100k at the time. It was the absolute best bang for the buck as far as local inference and fine tuning. I am not sure why you are so hostile when I specifically called out “latency sensitivity” as far as prompt preprocessing and various other bottlenecks. I run Linux.

I am using Mac studios for local inference, but not for openclaw. It was simply the best bang for the buck at the time, and the software side of things has gotten even better since then. I haven’t run the pricing comparison but I have a feeling with the direction they’re heading that it’s better to hold off for m6 as it may assuage various bottlenecks if you are latency sensitive.

From the perspective of the market, TSLA is not a car company. It sells carbon offsets/credits. That is their business.

>No actual resistance measurement of LK-99 with a 4 point probe.

You are misinformed. It's in the original Korean paper.

There was a wild article that inaccurately overestimated the context of the for the code interpreter fine tune a while back, it’s still 8k like the base model and not the wild 100k like it presented.

I implemented a version of it while I was waiting for openai to roll it out to me with pyodide and it worked well. Got it to re-encode videos right in the browser with a bit of tweaking of the langchain plugins prompt. When I did get it, it was a very nice breath of productivity though their interface could be better for many things beyond that. But their fine tune was better for sure than the API + system prompting.

With Meta supposedly releasing their new model soon, I think anyone who cares about control over their language calculator and not having to be predictably and incessantly lectured by some moralizing RLHF should try to contribute to the community and build out the infrastructure, fine tunes and tools instead of just relying on an org who wanted their moat so badly that they decided to fear monger and confuse the public.

Can you point to any substantial reproductions of copyrighted novels that are output by LLMs? If it’s just compression, you should be able to pull a substantial reproduction of the works out of them.

Can you share the log?

I'm not singling out the author or his company, but it's really hard to tell if this article is taking itself seriously, much in the same way that I've seen a group of 19 year old "AI scientists" talk about how everything will be automated in five years.

Let's just take it on its face. OpenAI, the most "powerful" foundation model has a rolling release of every 4-5 months. I've witnessed loss of "reasoning" and outright ignoring of instructions in some cases using the latest models unless you push some things through the functions parameter, which will probably break a lot of prompts, especially the agentic ones. Imagine that, there's no free lunch with a fine tune. You're pretty much at the mercy of a company you're basing your business on not breaking your shit every 5 months. And with how they handled deprecating davinci, it'll keep happening.

In other news, a lot of the VC cash that is flying around right now and is a bit deluded and chasing some vague concept of "AGI" that's coming soon(tm), but given typical Bay Area bullshitting who knows if they are actually a true believer(tm)(c) or they're just throwing around cash looking for your typical unicorn. Strikes me as delusional too. I'm sure you'll have multimodal models any day now that can do everything from pick strawberries to writing and maintaining some bottom of the barrel web app.

So, more delusion and more bullshit. Lately, in general, and it's not a personal thing against the author or his business at all. It's just much in line with the frenzy that I see elsewhere.

You don't need acquired capital for your business because you can rent a VPS and generate social media posts cheaply? Here's your sign.

The only thing good about it in my experience is agents+tools+out-of-the-box wrappers for various LLM providers and such, but if you need to support multiple why not just roll your own adapter pattern?

It adds a cognitive burden, and hides the "magic" and prompts three layers deep with agents, to something already cognitively burdensome if you're looping inputs on themselves or calling multiple agents. In my opinion, of course.