Hacker News Reader

Comment by mixermachine | Hacker News Reader

Scaling a model on a chip is quite hard. ChatJimmy is based on Llama 3.1 8 billion. Kimi K3 has 2.8 trillion parameters. That are 350x more parameters.

I would expect that Gemma 4 E2B (approx 5.1 billion parameters) or maybe even Gemma 4 26 billion A4B at some point is running on a chip.

View on HN

Scaling is definetly hard - but there is no absolute requirement to put huge flagship models into this technology (although it might be possible over time).

A fairly dumb but FAST model has it's own totally distinct use-cases even if it can't be scaled in size. Think about a LLM-infused-Alexa where the response time is instant. Where you can request it looks at hotel options in Montreal, and it starts answering in half a second rather than a few minutes.

Plus some sort of slow smart + fast dumb combo architectures might also work really well for different classes of problems.

> Where you can request it looks at hotel options in Montreal, and it starts answering in half a second

Yet the answers will get outdated quickly whilst the silicon is fixed.

>Yet the answers will get outdated quickly whilst the silicon is fixed.

Bro is living in 2020 before rag was widely introduced.

I was told updates require replacing at least two layers of metal, though not whole thing. Was that not accurate? Can you say more?

You are talking about two different things.

Yes, to update the blueprint for new models two layers will be updated. That is the NN.

To instead update the data on which to operate you could use a RAG to query.

(As in "the Pathfinder 2.0 NN is on the chip; the geodata is in the OpenGeoMaps dump-DB-nightly" - not really overlapping with LLM+RAG but may give an idea in a different scenario.)

You can change the prompt, and give the agent tools, and that keeps the agent up to date.

The agent doesn't know the date, or know what hotels there are in Montreal, it sees:

> System: You are an AI agent. The date is 11th August 2026. Your knowledge cut-off is March 2023. User is based in <date>. If you need to search for something to support the user say {search:<term>} and a list of options will be provided along with instructions on how to access. Or say {help} for a full list of commands.

> User: Can you help me find hotels in Montreal for next weekend?

The AI then interacts with the tools given in the base prompt, which can obviously be updated. So it then goes:

> AI: Of course, let me search for that. {search: hotels in montreal for 16th August}

> System: [Provides list of websites]. Say {read[n]} to read option or say {start subagent:<goal>) to register subagent.

> AI: {start subagent: List hotels on booking.com available on 16th August}.

[etc etc, then eventually]

> AI: Yes I have searched for you and I found a few options!

While you can't change embedded knowledge, a good model knowing that the date is 5th January 2040 can infer certain things (e.g. while it might not have been trained on certain deaths, it can probably guess that it should search before answering if it means a person would be 102 and their last information is from 2024)

> before rag was widely introduced

And when did RAG start to work properly as a mature, reliable technology?

I mean, it's pretty damn reliable now. Has been for simple high contrast Q&A for a while now.

I've just been doing research and experiments for work related stuff.

Typically we've used plain embeddings for a lot of high contrast documents aka discrete facts.

However I've been working with a >1000 page document of complex procedures with incredibly low contrast where embedding falls flat.

There's top down/graph searching, bottom up/embedded; alts like colbert, reranking, reasoning, search agents and now (though seemingly quite new) specific search agent models.

Ultimately I found that a reasoning enabled search agent doing a hybrid of bottom up (with reranking) followed by top down, gave the absolute best results. Paired with Luna for cheaper and faster tokens it benchmarks pretty well even for vague references to procedures.

I would imagine that search specific models just coming out are even better and I'll have to evaluate using these but for now the above works well for us.

Having an agent get vector search results to use as anchors and then being able to explore the sections and subsections above that, then eventually digesting as much as is relevant (big context, cheap tokens) is amazing.

Very interesting and very good (though, let us say, much more recent that ancient): only, it paradoxically does suggest that the computational cost (when compared either to a simple "search by vectors distance" or to the task failures that we can have witnessed by chief implementations*) strongly suggest the benefit of 1000x speed boosting, energy conservative Taalas board.

*(Of course it has "always" worked well for «simple high contrast Q&A», ever since the base embeddings technology worked properly: that is almost by definition; it is on real world use cases, where the nuances of reality are present, that it failed miserably.)

And outside of idiotic demos, who exactly is going to ask an LLM to look at hotels in Montreal for them? This usecase has never made sense to me in the slightest

> who exactly is going to ask an LLM to look at hotels in Montreal for them

Anybody who has a specific informal query ("SELECT ... FROM ... WHERE has_carpark AND ... ORDER BY score(has_jacuzzi , walk_distance(...) ...) DESC") but does not want to research and cross the different scattered info himself (does not want to build the virtual DB himself).

In the slightest? The user goes to their computer and types in "Montreal hotels" and Google comes up with no shortage of results, including a bit from their LLM. So that's already happening, but how do you narrow down the results from that initial search? Click around on Expedia for an hour? You probably know what you care about, just tell the LLM that you have dogs or are a vegan or whatever instead of wasting a bunch of time doing it by hand yourself.

I mean it was just an example, but I get AI to do things like this all the time... ChatGPT planned my latest work trip by looking through flights, hotels and pulled together 3 options of itineraries for me, and wrote an email I could send to my client with estimated costs etc.

It's quite telling that the 8B Taalas chip was already reticle-sized on TSMC N6. I mean, we're talking about a process that does ~100 MTr/mm², ROM needs about one transistor per bit, but can probably be packed more densely than general logic. Something like, say, 150 megabit/mm² is not a lot. N6 has a 850 mm² reticle limit. This roughly tracks, the article says the chip has 8B parameters and apparently spends about half the area on ROM. There's a reason AI accelerators just use a ton of silicon area (each HBM3 die is >1000mm² of silicon). I imagine this is not terribly viable unless they make it a lot more space efficient e.g. using MLC ROM if they don't already, or use stacked dies with a ROM-optimized process. And then we're back to not cheap, though reticle chips were never in the cheap area to begin with.

Taalas exploits the low cardinality to store one 4-bit weight with one transistor. (They are using metal layer traces for the ROM, and connecting an access transistor to light up one of 16 options.)

Their system is honestly very efficient for the weights, the problem is the KV-cache. That's why HC1 only supports such short context, they use SRAM for that and spend most of what's left of the die for it. The recent advancements that made attention more efficient are probably going to be very useful for them.

Do you really need a KV cache if inference is that fast though?

... Yes. Quadratic is really bad for large enough n, and you need that big context for useful work.

A full wafer like Cerebras is about 60x that, and N2P has about 3x the transistor density. So right now it's technically feasible to etch a 1.4 trillion parameter model. So roughly DeepSeek-V4-Pro class. Imagine that running a factory, for example.

Cerebras have special techniques to work around etching errors / bad cores on their wafers. This is possible since their wafers are effectively hundreds of identical copies of redundant cores. Can't do that for a globally unique model.

Etching failure in that situation would be like brain-damage in a human, all sorts of weird effects would start appearing.

There's several ways to engineer around that as the errors are detectable. There's a big literature on how to trade off speed or transistors for error correction. [1]

(Is Cerebras doing something novel? CPUs and memory blocks have been doing those things for a long time too, since the error rate is otherwise too high for normal size chips as well)

[1] see eg https://www.vlsimentor.com/dft/redundancy-bisr to get some basic concepts

A few hundred bad bits/transistors in a trillion+ parameter model would compromise its abilities not one iota...the models are inherently lossy and resistant to "brain damage"...

> (each HBM3 die is >1000mm² of silicon)

Did you mean that each HBM3 stack is that large? Because it only takes one glance to see that the memory chips are much smaller than reticle-sized GPUs they sit next to.

I have found that for some personal prose-related projects, QWEN 3.6 35B A3B is an amazing model even quantized down to 4 bits. I actually find it's "writing" style as a GM for an LLM-powered solo text adventure game, better than even some of the faster/dummer frontier models like GPT-5.6-Luna or Haiku 4.5, and it runs (slowly) on a 3090 with a 80k context.

So I have faith in these embedded LLM chips when it comes to fun projects like that. I have not personally found my quantized QWEN good at agentic tasks, though, and it LOVES to make shit up when asking questions about documents in the prompt.

Consider a model like https://huggingface.co/nvidia/NVIDIA-Nemotron-Parse-2.0 which just came out. 0.9B parameters and very accurate for doing a very specific task: document features classification.

Now imagine you have a chip which is just that model, but can do it at absolutely insane speed. Like tens of thousands of documents a second.

Same for things like text-to-speech or speech-to-text. Think of the accessibility wins if subtitling becomes insanely accurate and fast and omnipresent.

There are all sorts of domains like that, and the trend has been such that smaller models are getting smarter and smarter. If you can stick them in parking meters, traffic lights / street crossings, mobility aids, etc etc I just see so much potential win.

350x is only about 10-20 years of improvement, using CPU FLOPS as the benchmark.

wouldnt you want gpu, fpga, or dsp as the benchmark?

its lots of parallel calculations, rather than one blazing fast one

Haven't CPUs largely plateaued? They're just getting bigger, more power hungry, and multiplying cores.

Physics has hard limits and Moore's law is long dead.

AFAIK these don't have to be CPUs. ASICS are well suited for this purpose. This will be the most cost effective way to run LLMs.

There's certainly incentive to do so. And its only an engineering problem haha.

You could imagine different layers on different chips, though, i think...

Reply on news.ycombinator.com