I asked it to design a submarine for my cat and literally the instant my finger touched return the answer was there. And that is factoring in the round-trip time for the data too. Crazy.
The answer wasn't dumb like others are getting. It was pretty comprehensive and useful.
While the idea of a feline submarine is adorable, please be aware that building a real submarine requires significant expertise, specialized equipment, and resources.
Agreed, this is exciting, and has me thinking about completely different orchestrator patterns. You could begin to approach the solution space much more like a traditional optimization strategy such as CMA-ES. Rather than expect the first answer to be correct, you diverge wildly before converging.
How about if you run this loop (one year from now) on this kind of hardware but with something like Claude/Kimi K2. How about that? Because that's where it'll go.
This is what people already do with “ralph” loops using the top coding models. It’s slow relative to this, but still very fast compared to hand-coding.
This doesn't work. The model outputs the most probable tokens. Running it again and asking for less probable tokens just results in the same but with more errors.
A related argument I raised a few days back on HN:
What's the moat with with these giant data-centers that are being built with 100's of billions of dollars on nvidia chips?
If such chips can be built so easily, and offer this insane level of performance at 10x efficiency, then one thing is 100% sure: more such startups are coming... and with that, an entire new ecosystem.
I think their hope is that they’ll have the “brand name” and expertise to have a good head start when real inference hardware comes out. It does seem very strange, though, to have all these massive infrastructure investment on what is ultimately going to be useless prototyping hardware.
You'd still need those giant data centers for training new frontier models. These Taalas chips, if they work, seem to do the job of inference well, but training will still require general purpose GPU compute
I dunno, it pretty quickly got stuck; the "attach file" didn't seem to work, and when I asked "can you see the attachment" it replied to my first message rather than my question.
why is everyone seemingly incapable of understanding this? waht is going on here? Its like ai doomers consistently have the foresight of a rat. yeah no shit it sucks its running llama 3 8b, but theyre completely incapable of extrapolation.
There are a lot of people here that are completely missing the point. What is it called where you look at a point of time and judge an idea without seemingly being able to imagine 5 seconds into the future.
It is incredibly fast, on that I agree, but even simple queries I tried got very inaccurate answers. Which makes sense, it's essentially a trade off of how much time you give it to "think", but if it's fast to the point where it has no accuracy, I'm not sure I see the appeal.
the hardwired model is Llama 3.1 8B, which is a lightweight model from two years ago. Unlike other models, it doesn't use "reasoning:" the time between question and answer is spent predicting the next tokens. It doesn't run faster because it uses less time to "think," It runs faster because its weights are hardwired into the chip rather than loaded from memory. A larger model running on a larger hardwired chip would run about as fast and get far more accurate results.
That's what this proof of concept shows
If it's incredibly fast at a 2022 state of the art level of accuracy, then surely it's only a matter of time until it's incredibly fast at a 2026 level of accuracy.
I think it might be pretty good for translation. Especially when fed with small chunks of the content at a time so it doesn't lose track on longer texts.
Me: "How many r's in strawberry?"
Jimmy: There are 2 r's in "strawberry".
Generated in 0.001s • 17,825 tok/s
The question is not about how fast it is. The real question(s) are:
1. How is this worth it over diffusion LLMs (No mention of diffusion LLMs at all in this thread)
(This also assumes that diffusion LLMs will get faster)
2. Will Talaas also work with reasoning models, especially those that are beyond 100B parameters and with the output being correct?
3. How long will it take to create newer models to be turned into silicon? (This industry moves faster than Talaas.)
4. How does this work when one needs to fine-tune the model, but still benefit from the speed advantages?
The blog answers all those questions. It says they're working on fabbing a reasoning model this summer. It also says how long they think they need to fab new models, and that the chips support LoRAs and tweaking context window size.
I don't get these posts about ChatJimmy's intelligence. It's a heavily quantized Llama 3, using a custom quantization scheme because that was state of the art when they started. They claim they can update quickly (so I wonder why they didn't wait a few more months tbh and fab a newer model). Llama 3 wasn't very smart but so what, a lot of LLM use cases don't need smart, they need fast and cheap.
Also apparently they can run DeepSeek R1 also, and they have benchmarks for that. New models only require a couple of new masks so they're flexible.
The counting rs in strawberry problem was a example of people not understanding how the models work but I guess good to show the limitations of the current architectures.
But thing is, those architectures haven't improved a whole lot. Now when it answers that correctly it's either in training data or by virtue of "count letters" or code sandbox tools.
https://chatjimmy.ai/