What does it take to go from here to a model on a pcie card or an m.2 card, so I can plug one into my workstation / laptop? Will 'intelligence' become much like a gpu, where most people just live with the performance of whatever they have installed, outside large companies that must have cutting edge, or prosumers that have a incrementally better version than the masses?
Are we a couple years away, a decade away, or something else?
I'm surprised neither OpenAI nor Anthropic made this move first. The Chinese open weight models are pulling ahead and commoditizing their value proposition.
Baking models onto silicon would've been the next logical move to get a moat.
Google is already doing this and has an experimental project on top of already having TPUs and cramming their quantized flash onto individual TPUs for inference.
Personally I think Apple should have acquired them. if you could burn a gemma4 class model into an iphone and actually get extremely low latency and low battery usage it would feel like the future IMO. even if it means you wont get frontier intelligence, there might actually be incentive to buy a new mobile device every year again.
The Taalas chips are not physically small. And part of their secret (if you look at the design) is just locating a bunch of memory soldered on the edges ( I belive higher amounts of SRAM ? )
I don't think this works out from a cost/silicon perspective. Small models already run pretty well in software (since the weights fit in cache) and big models require silicon area proportional to the size of weights. On a mobile device putting a chip like this is competing directly in BOM and power against a whole lot more l3 cache, and the l3 cache makes everything faster
That likely isn't as relevant for on-device iPhone usage as it is for Real Work™. I won't notice the difference between 50tps and 1000tps when asking Siri a question.
I don't know. As others have said, the Taalas chip wasn't small, or particularly low power, so it's hard to "imagine" what that tech in an cell phone chip might look like.
But if the basic premise of "good enough LLM at insane throughput" holds, I think it could qualitatively change local uses of LLMs. At a certain speed point, you're able to move from request -> response to a cascade of tool calling and "subagents", which could allow a small model to be much more useful, if provided with a lot of local data and tool calls.
That said, this is assuming you could stuff a "good enough" model into a phone with Taalas-like technology. The Taalas tech demo was an 8B parameter model and required hundreds of watts (IIRC) to run. The efficiency was good given the speed (as I understand), but it's not clear at all that the approach scales small enough to be a sensible coprocessor on an iPhone or whatever.
That order of magnitude could be the difference between "the users wants me to open the notes app, let's open it" and "I've scanned all your notes before you could blink and found what you're looking for".
This is my thought as well. Models have to be intentional about which tokens they burn because there's a real lag time. If you can just fork out 10 different reasoning sessions at once with no regard for token waste/lag, you can compensate a smaller model with just doing more at once with it. No idea if this is reasonably true though.
The weights might fit in cache, if you're using a small model. If you wanted to have a 20B+ parameter model, that's just going in RAM. You could put more RAM in the device and pay the perf cost or have a dedicated chip. Most devices already have a dedicated chip, this just changes which silicon you're spending the money on.
Slightly besides your point, but it's interesting how many here naturally ponder about how the current winner could or "should" keep winning, instead of how another company could become a competitor by doing the more clever thing the incumbent isn't thinking about.
But this is already happening with iPhones. Apple is touting on-device AI and only the latest phones offer the full capabilities. Newer phones will be able to run better models, so the incentive is there as soon as someone makes the killer app that only makes sense when the model is running locally on your phone.
If you’re only running models for frontier capabilities, yeah. For tasks where current models are smart enough, running them 100x faster is the most impactful improvement you can make. Consider all the things you could use a model for, but don’t, because the latency is just a bit too high.
But Claude Opus 4.6 is not really practical. Taalas' process seems targeted for edge models. Their proof of concept model, for example, is a heavily quantized version of Llama 3.1 8B and even then they acknowledge their custom 3-bit/6-bit representation causes model quality degradation.
Taalas is going to have a tough time putting a trillion-parameter model on one conventional die. Their HC1 die is already near the maximum size that conventional lithography can expose. They claim they could partition the model across many chips, but I'm not sure if they have tested this process or what it means for compute. The basic storage arithmetic is unforgiving: for a one trillion parameters model at four bits it will take 50–100 chips. To service a sizable customer base will take thousands of 100-chip fabs.
That all said, I'm bullish on this technology, and look forward to seeing it evolve.
It costs something like $300,000 for the hardware to run a model of that size. You'd pay that for a single model for 1-2 years? Not even the AI companies can justify that kind of spend which is why they keep extending the expected lifespan on their hardware in the accounting.
I'm expecting the Taalas MSIC version to cost a fraction of that. Then probably have some kind of cheap subscription to Anthropic for updates (yes, Taalas chips can receive a certain kind of updates: they have a small SRAM).
Compute the cost of producing n of them devices, imagine a fair price based on that, and see if that local, blazing fast card* can be an asset that could be replaced periodically.
*(It's local: private files managing firm oriented. It's blazing fast: it can be placed into recursive, intensive local workflows.)
"seems like baking models into silicon is speed-running obsolescence"
Now maybe. When models are flying passenger aircraft, other prerogatives will assert themselves. When a 50TB ROM means you can impulse purchase a ChatGPT 6.3 xhigh that runs on batteries, yet more use cases will be apparent.
obsolescence is the whole point. apple gets to sell a new phone very 6-12 months because of it.
i have written about this:
"For device makers
Packaging models with laptops and smartphones will let application access near free, low latency inference and potentially offer users a better experience with the option of preserving data on-device. This is viable under the condition that tasks that do require larger expert models that run in the cloud can be routed to external models.
A side-effect of local models and what will let Apple cut upgrade cycles from ~4 years (?) down to 12-18 months is specialized hardware to run them. For almost a decade, smartphones have been trying to compete on better cameras. This coming decade will see them selling better GPUs, NPUs, ASICs and whatever other things they'll be calling the inference chips, to drive re-purchase. Every six months will see a better model on new hardware, which will enable better performance in certain applications."
Not sure. You can fix the transistors but leave the connections between them open for flexibility, so you only need to change the manufacturing process for the upper masks for every new model.
You need to find customers for several-generations-ago models before this makes any sense. AMD is a lot more incentivized to look than mr vanilla llm is
Bitcoin mining doesn't have large memory requirements, but does have huge compute requirements. ASICs work great there because it's very straightforward to add some circuits for computing hashes. If you _also_ have to add many GB of memory, then suddenly ASICs will cost as much or more than comparable off-the-shelf hardware and they won't be faster unless you've also invested in huge memory bandwidth.
Most enterprise GPUs are scrap after 5 years because they're so inefficient compared to newer models. It's entirely possible to make them last longer by undervolting them, people just don't because it doesn't make sense.
Bitcoin OTOH has used the same PoW algorithm for a decade. Barring some really exciting discoveries about the nature of computation, new ASICs are not that much more efficient than old ones.
BTC mining is also not exactly competitive anymore; the nature of the PoW algorithm means that it's dominated by a few large players who've set up shop next to a dam and who pay very little for electricity.
New entrants are highly discouraged because the mining rewards are constantly halving, it's hard to find cheap power, and the price of BTC is now so volatile that a yearslong investment is very likely to lose money.
Isn’t that kind of useless for the stock? It sounds complicated, unlike having number of CPUs go up.
It’s like talking about anything else than Megapixels when everyone was convinced that megapixels must go up in certain periods of the smartphone boom.
Not necessarily: it is relevant to Taalas only if it is a compute-in-memory architecture.
The Jalapeño mentioned («Anthropic is not alone in walking this path») in the article is still a classical Von Neumann architecture.
And Taalas' idea makes sense in a perspective of scale - producing a large number of cards; "for internal use" (a lower order of items) means a high production cost.
Depends what you mean by relevant. If you use AI primarily as a search/knowledge engine, it makes no sense. If it's your capable assistant that has a lot of general knowledge, can do tool calls, and has a big context window, very doable.
Indeed, for some kinds of applications involving secure/legal data etc. I can see the consistency of silicon winning out, because it combines performance with immutability and guardrails in hardware. Some chips have write-once PROMs to store password hashes and similar, you could do the same thing with prompt hashing to absolutely force or forbid certain behaviors. A model that can't be updated is also a model that can't be hacked.
Closer to every 5-6 years these days and with ram prices going up it will be even longer. Especially with the low/mid range phones, which are most phones outside some developed countries, people will keep their phones as long as they can.
Would depend on the income levels, but yeah, buying a new phone these days is entirely a non essential luxury. An iphone easily lasts 7 years so the moment money is tight, it's a very easy choice to not buy a new one.
So they have decided that putting a small LLM on a phone would backfire because people would have a negative perception of their cloud models. Pretty sure AMD will use these taalas chips in data centers, not phones
Not so if it's embedded in something smart enough for its intended purpose.
Think vision, spatial reasoning, speech synthesis, even some speech analysis. Think self-driving cars (and drones) that need 10x less power for the brain, and can think at 10x situation per second.
This is only true for people who are solely focused on performance. There is absolutely a market for acceptable performance combined with predictability.
We're all used to having to constantly update our browsers and phones to keep up with the security arms race. If a frozen model can't be updated, it will predictably remain vulnerable to any "exploits" or idiosyncratic quirks that people discover over time.
Let's say, as somebody suggested in another comment, that you buy 100,000 of these chips and deploy them to run fast-food drive-thrus. And then somebody discovers the model has a fondness for goblins[1], and if you role-play convincingly enough, you can get it to accept payment in shiny buttons and rodent skulls instead of cash.
What do you do then? I guess your options are to try and fix the behavior with a better prompt, or put some kind of filter in front of the model to catch attempted exploits. If the filter is cheap and dumb it probably won't work well enough, and if you use another model as a filter, you've negated the cost and speed benefits of putting the first model in hardware.
Of course the real answer is to just never expose the model to situations where an adversarial input could possibly lead to an undesired output. But that drastically limits what you can do with it.
How is that in any way related to a consumer device? This method doesn't reduce physical memory requirements, so still results in huge die area. This isn't a for-end-user thing, probably for decades.
I don't follow. How is that related? GPUs don't have fixed memory. You don't throw them away when you want to load a new model.
NVIDIA will probably give us a new GPU when someone competent in the free market decides they want wheelbarrows full of money. Unfortunately, AMD is entirely, incomprehensibly, incompetent, to the point where I can only assume they're colluding with Nvidia, behind the scenes.
What I like about this, is that it significantly increases the probability of a sci-fi scenario where you're picking up a hot chip on the black market; rumor has it, Mythos 9 weights baked in...
I read the paste, it got the etymology wrong, no? Schlong comes from shlang (snake), not shlemp (is this even a word? I don't speak Yiddish but couldn't find it on Google).
It did get it wrong but it also got a lot farther than much more recent, but worse models like 6.7GB on disk size ternary bonsai. It at least knows it's from Yiddish. The "schlemp" appears to be a total hallucination or it's confusing it with schlep, which is not related to schlong. One of the reasons why I said it "mostly" passes the test. Something much larger on the size of qwen 3.5 122B, deepseek v4 flash or similar that runs in 120GB to 190GB of RAM in my experience will answer perfectly unless it has been ruined by something like Q2 quantization.
There isn't SchlongBench(TM) yet, it's a specific question I've been asking of differently sized models as a randomly chosen gauge of how much less commonly used knowledge is perma-baked into it. In this case a question about a specific yiddish origin slang term. Small/bad models don't know it's from middle high german or Yiddish and get its origin and meaning totally wrong (or it runs into model censorship related to slang related to the male anatomy).
It's also a question I have found will cause models that don't know what it is to go off quickly in a direction of hallucination trying to explain it, so the hallucination is evident very quickly starting from the first ever prompt issued with 0 context fill. Example: I had a model write four detailed supposedly-accurate sounding, grammatically correct paragraphs saying its origin is from AAVE (African American Vernacular English), which it most certainly is not
You could do the same by picking any topic that is very rarely discussed in conversation, some esoteric and narrow piece of knowledge and asking the model about it.
The speed is awesome, in the true sense of the word. It's great at knowledge and basic stuff but the output is complete junk for anything concerning new facts or slightly esoteric topics.
I had the same reaction but then I showed it to my partner. She completely didn't get it, in her words "how can it be thinking of a good answer when it's that quick?"
I tried to explain but I fear were probably going to be adding artificial sleeps to these things to convince the masses it's doing something clever.
to be fair, the model used for Chat Jimmy is not very smart, but the world where it is smart is very interesting.
It’s going to be really crazy when the bottle neck for agents is the speed of the tool calls rather than the speed of inference. Imagine an agent interacting with the terminal near instantly…
I asked it some old hardware command line questions I'd recently asked Gemini, it hallucinated parts of the answer.
The characters in the 3-act Shakespearean play had very little depth, many of the names were similar, and they were not very smart, but the simple plot was cohesive.
I think the same exact model running on CPU-only and RAM, or a small GPU, would do about the same? It's quite an old model now and small, you could throw a GGUF into llama-server or something for a side by side comparison.
As I remember just about any english language model from mid 2024 and earlier didn't even do well if you asked it to count sequentially from 0 to 100, nevermind calculating stuff.
Reasoning models are the same speed. They’re just post trained with RL to do CoT inside tags like <thinking></thinking> before a tag like <response></response>
There’s no difference in the inference implementation, parameter count, or speed.
So this demo is around 90 times faster than typical speeds for the same model at openrouter, and around 30 times faster than the absolute fastest option available (Groq).
AIs don't intrinsically know anything about themselves so they often give wrong answers to such questions. This can be fixed by putting info in the system prompt but they may consider it a waste of tokens since most usage doesn't benefit from that information.
I think the real value here is not as a customer-facing agent/chatbot but for for automated processes. Think of all the companies out there that have LLMs doing simple tasks like categorizing customer feedback emails. For such tasks, you don't gain much from better models, so if you could run it 10x cheaper on a slightly older model, it would absolutely be worth it. Pretty much any place people are currently running a flash model could benefit from this since they're already deciding that speed+price is worth using a less capable model.
I think this would make sense for consumer hardware, not for AI companies.
AI companies constantly update/change stuff, new models come out, new requirements, etc.
But if you ship an "ai-powered" dishwasher, it can come with the chip built-in to do computer vision and precisely target each spot, and will be sold as-is with no updates.
I find speed alone would be a game changer for current models. I hardly find any task anymore that the current frontier models can't do with max reasoning after several rounds of feedback (provided sufficient instruction and the right harness). But waiting an hour or more for reasoning to finish is getting really cumbersome. If they could do the same in seconds (and for cheap of course), I'm pretty sure we'd pretty soon see major software companies pop up that are run by a single human.
There’s some kind of tradeoff between speed, cost, and quality for every application. I would be perfectly happy with a model 6 months old that was 50x faster for many uses. Right now I use either Opus (for smart stuff) or Flash without thinking (for fast stuff). I would take an even dumber model for more speed (lower latency in particular).
Already models have gotten really good at a lot of things.
A lot of people would probably be happy to stick with the same model for a year or two if it’s 10x faster and cheaper.
And perhaps older models can become cheaper over time as newer models come out on new silicon for a higher price. That incentivizes people to stick with older models.
They expect a sort of breakpoint at which each subsequent model version will only be marginally better than the previous ones, thus allowing them to retain their value for some time. Their business doesn’t work if each year the new model demolishes the previous one in terms of performance.
pretty much everything is “1 or more versions behind” by the time it comes out. the question is whether or not it’s still useful? at some point, presumably not every application will need the latest cutting edge huge model.
Something I personally haven’t seen much of, in all the discussions of model benchmarks and AI breakthroughs, is a distinction between “peak performance” and “reliable performance”. The “peak performance” of frontier models is very high: they’re solving open math problems, analyzing large codebases, etc. But my subjective impression is that “reliable performance” is mid at best: out of 100 random questions I might think to ask, it’s likely to say something wrong or stupid a handful of times at least.
I think there’s inherent tension between the two: the more a model reaches or outright hallucinates, the more likely it is to come up with tricky, subtle solutions to problems (I think people are somewhat like this too: Terry Tao’s brother is nonverbal, Jim Watson’s son has severe schizophrenia, etc). But then the less likely it is to generate a sensible email reply.
I use models all the time for coding, but I would not let one take over my daily correspondence. If the idea here is to run frontier models at high speed in data centers, that could be useful (the speed would be cool), but I’d be surprised if the cost of that hardware churn is worth it to frontier labs. But if the idea is to turn this into a chip that goes in your phone as some kind of routine, low-power inference thing…taking something too kooky to be relied on and baking it into your phone’s hardware like that doesn’t make sense to me.
I have no idea. It makes my comment a statement posted as a proxy for a question, a question you correctly pose explicitly.
If it can, then deployment in a sea of gates can make a chip viable across model generations as weights change, inside some scale factor.
If not, unless the part is under a pinout and address model which can scale on the bus, and can be easily replaced, it makes the entire dependency a replacement, not just this part. So embedded use has consequences.
This is probably a win-win. The team gets paid, and we get greater assurance that their best ideas and architectures -- which are truly impressive -- are going to see the light of day in actual products.
They were too small for this to be a meaningfully sized purchase for AMD, there's real risk they get sucked into a team that ultimately delivers sqat, not to mention the chances of anything being delivered in an even remotely consumer-priced bracket are definitely out the window
AMD could have saved their money and used their own hardware! I've got a language model doing 60k tok/s on AMD hardware already, a Xilinx Kria K26 SOM, with the weights baked into URAM/BRAM with zero DRAM in the token loop. Same thesis as Taalas: single-stream decode is bandwidth bound, so stop fetching weights from far away.
Caveats stacked high, obviously. It's 3.16M parameters (tinystories, and I also have a kevin-speak lemmatised version), the tokens are characters, and the 60k record is 16 streams that each remember exactly one token of context, so it's blisteringly fast at saying nothing. The honest build with full context and KV caching still does ~19k tok/s on one stream though.
I keep messing with the blogpost with the live demo, but I'm planning on flipping it to live in the next day or two
Yeah Im surprised nobody is talking about this. When everyone first saw Taalas I looked at the design and it had a big legup in physical cache availale compared to most chips. Makes you wonder how much of a benefit there is to the actual "baking" of the model vs just having a large chip with a ton of SRAM (or whatever) soldered close to the edge physically.
I feel like what we really need is the ability to solder computer cache on all sides of the chip Meaning above and below as well. If you can only attach it to the edges you will be inherently physically limited on the amount you can put (and maybe even have latency benefits as well)
Talaas is different, it's a true compute-in-memory architecture where the weights are stored in the connections between the transistors that perform the matrix multiply, rather than in seperate memory cells.
Most of the benefit comes from this architecture; hardwiring the weights into the silicon is just the easiest way to implement it. SRAM requires too many transistors, DRAM requires an incompatible manufacturing process, and exotic phase-change memories aren't readily available.
They can't due to power density, I believe - they have to be run in a sandwiched waterblock with massive cooling, as far as I can tell. That's the biggest thing that baked weights gets you - a relatively modest watts-per-square-mm compare to cerebras, where they had to engineer a whole system to get the watts out of the chip
1. How come you didn't make your implementation public? You could be a millionaire now.
2. Especially if AMD has the technology to do what Taalas does, it makes a ton of sense for AMD to acquire Taalas: remove them from the market. Make sure nobody else (Intel, Huawei, Alibaba, NVIDIA, etc) acquires them. It could have been a great acquisition for a rebirth of BlackBerry btw.
People are missing the point if they think this is useless because frontier models keep changing every few months.
We really, really need better secondary models that can do things fast and do them cheaply for lots of dumb tasks. Not only because it can be used as sub agents by frontier models, but also because it can be like a universal grease for all kinds of software.
I've got an app I am building and I don't want to tie myself with frontier models because I'll never be able to beat openai/anthropic. I just want a simple, cheap, instantaneous model that can just go through my documentation and tell the user what to do next and how to integrate with whatever ai subscription they have.
Pipeline the burn into silicon, lower the latency as much as you can, for the 10-100x operation cost it's worth it. Imagine if frontier models cost $5/mtok and the 2nd or 3rd tier models cost $5/billion tokens for 3-month-old models.
> Is there any LLM from exactly one year ago that would be worth running?
Bad perspective: consider the correction: "when are thresholds of sought quality reached"? Hence: not "is there a 10yo from last year that could compete with the current 13yo", but "will there be a 30(?)yo from last year that could compete with the current 33(?)yo" ('(?)': the scale of yearly growth in the future is uncertain).
It's not just about it "being smart enough". It's about there being actual user demand when it needs to compete with the shiny new model.
A 10 year old iPhone is probably good enough, but is there demand for it? In a vacuum a 10 year old iPhone is good, but why would you pick it if you can have a current one for a reasonable price?
Honestly, this is starting to make more and more sense. SOTA models are starting to converge to certain architecture and capabilities. I wouldn’t be surprised we end up with a base model ASIC + “fine tune” card where it’s a physical LoRA style adapter.
Imagine a multi-modal model with 1000's of tokens per second. Realtime inference for a host of applications. This is a BIG deal and will change the landscape in unfathomable ways.
Once models settle down this makes sense. Imagine a cartridge with a physical model on it. You purchase a cartridge and stick it in your computer/phone/server. Want to upgrade? By a new 'cartridge'.
This should bring inference cost down dramatically, I wonder how OpenAI/Anthropic feel about that.
i'm looking forward to Qwen3.8 27B launch to see how much models have peaked at a given size.
it might already be time to start burning the best small models onto hardware since it's possible they can't get much better at many tasks like knowledge recall due to the inherent information density limits for models at a given size.
very interesting idea. i didnt think of that. i was just assuming youd have an additional one of these in your phone for actual lightning fast local inference
You have a robot. You need it to be smarter. You buy a new model cartridge (probably a PCIE 9.x). Now you need some domain specific skills. You'd like it to be able to cook, and you'd like it to not dent your walls anymore. You buy 'improved spatial reasoning LORA' card and 'Gordon Ramsey's Chef ULTRA9000' card.
Now your robot can respond sarcastically when you ask for chicken nuggets. Again. It also doesn't dent your walls anymore.
Wouldn't this mean someone with sufficient hardware could lift the SOTA model weights off the chip? Or are you saying that these chips would only be used internally by these companies and not sold to the public?
I wouldn't expect companies not sharing their weights today to be any more likely to share them if they're on hardware, this doesn't sufficiently hide weights from a local user.
The weights are very unlikely to be on the chip itself. That wouldn't work for SOTA models that are terabyte scale, even quantized. This is probably an accelerator for specific kernels in the model, but the weights are likely loaded from memory. The chip may have SRAM to store some of the weights temporarily during inference.
At least in the case of Taalas the weights are physically encoded directly on the chip.
It’s composed of 4-bit multiplier cells that compute all 16 possible results in parallel. The top metal wiring layer physically selects the one that corresponds to a multiplication with that cell’s constant weight, and routes it to the next layer.
I don’t get why this is an issue? You can run Claude/OpenAI SOTA models through Amazon bedrock. These weights have to live somewhere to run on Bedrock.
we have not converged at all, if you look at how different the chinese models in terms of architecture you can guess that the labs are experimenting a lot as well. we are seeing all different types of hybrid architectures, different attention methods and so on. Of course on a high level its still a transformer but if you take a proper look we are seeing more divergence then a convergence.
SOTA American models are not. SOTA Chinese models are. From a physics aspect, closed source models cannot be too far from open source ones in terms of size. There’s only so much you can squeeze out a B100 style cluster even with fancy Dflash style diffusion model for the speculative model.
If we had deepseek v4 flash 0731 etched on a chip it would be more than capable enough and fast enough for so many people's needs, even hardcore engineer.
I think it’s tongue in cheek. When I first got access to Sonnet 4.5 I remember thinking to myself “y’know if they never got any better and I just had access to this forever then that would be pretty okay”. Turns out my expectations have changed since then and I would like a higher baseline now.
Having a base model ASIC as a physical piece of hardware makes me think of the early days of microcomputer desktop stuff where having a socketed ROM or PROM was a key piece of hardware, and people actually knew/cared what ROM was on their system's motherboard.
Imagine if like instead of having a specific Mac Plus ROM, you had a thing that looks like a fat ASIC that can hold models sitting on a slotted daughtercard directly next to the CPU and RAM.
I don't think there'll be a fine tune card; you'll have the base model vintage whatever year, and then your GPU will do whatever LoRA layers you want it to do; the LoRA will wrangle older dated models into the current of whatever your looking at.
But yeah, for things like programming, if it can do linux and python and some go and sql and javascript, larger domains can be threaded with LORA
It obviously won’t be continuous delivery but could make sense if the lifecycle of a model (train, deploy, iterate (meaningfully) is about 1-2 years. In that case it fits nicely in the “this year’s model” already established with cars, phones, etc.
Seems to be somehow some kind of offshoot from or connected to Tenstorrent, which is just down the road. Founder looks like he was/is maybe at Tenstorrent and previously associated with Keller?
Always fantasize about applying at Tenstorrent, but wrong side of Toronto. 2 hour commute.
Claude Code is still using haiku 4.5 from ages ago for explore subagents for instance. Not to mention production uses like customer service that only need to be "good enough"
That's solely so that you burn more money. It's totally unnecessary to assume the parent model. Sure, it could be upgraded from Haiku if there was a solid reason to, but...
2 to 3 months optimistically assuming everything goes smoothly and is fully automated.
6 months or even a year if something goes wrong in the fabrication process and you need to update things.
If they do more standard asic design, it could be a lot longer as the design needs to be validated on an FPGA cluster, which would necessarily need to be very big for something like a LLM. Easily up to 2 years.
There's a reason chatjimmy isn't demonstrating newer models and why they only show of an 8B model.
I mean even if it take a few months, it'll still be out of date. But there was a hypothetical when it came up in Feb, would you want Qwen 3.5 at like 10k tokens per second.
At the time people were no doubt saying yes but now 3.8 is out, is that still desirable?
There's soooo much stuff that such a model is still capable of doing in the pursuit of getting a better overall answer. Imagine a powerful research agent that blasts out dozens of the small, cheap models to fetch and summarize one page each. Then the beefy researcher model performs the final analysis.
While this design is self-limiting I think its a good approach. It doesn't take an entirely new architecture or infinite memory to produce significant performance improvement.
Models, probably first open weight ones like Kimi K3 class, are etched into silicon like this and sold as cartridges almost like old school game cartridges.
You buy a USB-C dongle that the cartridge goes into, or for data centers you have PCI cards that take these in slots.
Me? Probably not. A business or a hoster, sure. There'd probably end up being an aftermarket in used cartridges with slightly older but still good models on them.
With web search and tool call a decent current generation model at the speed of the chatjimmy could do a lot. People saying it would be out of date are missing the point. It’s not going to make much sense for frontier companies that’s chasing the SOTA. But for a lot of business use cases if someone can put GLM 5.2 and sell it as a box, it would make so much sense.
My partner has been asking for a “completely private” model for doing research and shifting through volumes of data that can’t leave the office and $$$ for the current hardware makes no sense. It would be an easy sell if someone walks in with a black box that contains “ChatGPT”.
In my understanding the first Deep Think / Pro models were already very good as they were doing some kind of parallel repeated reasoning, thus were slow and expensive. So if chatjimmy speeds enables a fast deep think level performance, I think that would be great.
100% agree - you don't need the most up-to-date model to have something that's useful in agentic contexts. They could even produce chips with weights that make all the decision making/logical reasoning and have it delegate to other specialized agents. If it becomes cheap enough to print a run of custom chips, releasing a batch for each major advancement does not seem unreasonable for SOTA companies.
There are so many use cases for supremely fast offline models. The first thing that comes to my mind is for real-time video processing or other non-textual content in real time.
I feel like NAND process tech could become useful at solving some of these problems. A GPU where you can update the weights a few thousand times may be sufficient.
Enjoying the Ian Cutress / TechTechPotato video on Taalas. Some ok good technical details on the tech, and some good insider baseball, whose who stuff. (What a treasure having tech discussions like this about.)
https://youtu.be/3MKRjt59hh4
I wrote them an email asking for PrismML Bonsai 27b Ternary which is like 6b or something crazy small and would be a lot easier for them to do initially.
> How tolerant are models today to a few broken weights.
Extremely! You can remove entire layers and the model will still work just fine, with barely perceptible capability losses.
I've cut/bypassed ~15% of total parameters out of Gemma 4 31B on a pod once. Still got perfectly coherent responses out of it. Certain layers are a lot more important than others, particularly early and late ones; but it's honestly astonishing how much can be cut out from the middle without destroying the model's coherence.
I didn't run any meaningful benchmarks, so I have no idea what the capability loss looks like exactly. But "produce coherent and sensible English in response to a wide variety of prompts" was definitely not among the things the model unlearned.
Brings to mind the scene in '2001' where Bowman is pulling out individual pieces of hardware that represent the mind of HAL, and it becomes increasingly incoherent as more physical hardware is detached.
I wonder if you had a few percent of problems in the yield, if it would be functionally equivalent to the difference between a unsloth-published Q6 standard size GGUF vs. the nearly perfect precision of an unsloth Q8-K-XL. Or more like Q4 vs Q8 where a lot is lost.
> At 20 billion parameters per chip, you’d need just 50 accelerators to support a trillion-parameter model
I don't see any evidence that this is possible. From my understanding, the whole model needs to be on a single chip. Which rules out any popular frontier models with several trillions of parameters. Even smaller sub-frontier models have hundreds of millions of parameters, so these would be ruled out as well.
Are we a couple years away, a decade away, or something else?
Baking models onto silicon would've been the next logical move to get a moat.
Google is already doing this and has an experimental project on top of already having TPUs and cramming their quantized flash onto individual TPUs for inference.
But if the basic premise of "good enough LLM at insane throughput" holds, I think it could qualitatively change local uses of LLMs. At a certain speed point, you're able to move from request -> response to a cascade of tool calling and "subagents", which could allow a small model to be much more useful, if provided with a lot of local data and tool calls.
That said, this is assuming you could stuff a "good enough" model into a phone with Taalas-like technology. The Taalas tech demo was an 8B parameter model and required hundreds of watts (IIRC) to run. The efficiency was good given the speed (as I understand), but it's not clear at all that the approach scales small enough to be a sensible coprocessor on an iPhone or whatever.
A ~30mm side for the HC1 tech for an 8b model (still unclear the planned HC2)?
They could have 9 year old AI and still post profits.
Not sure if it's my pixel or android, but I made a randos jaw drop with what the crappy AI on android can do.
When are we getting android OpenClaw?
Taalas is going to have a tough time putting a trillion-parameter model on one conventional die. Their HC1 die is already near the maximum size that conventional lithography can expose. They claim they could partition the model across many chips, but I'm not sure if they have tested this process or what it means for compute. The basic storage arithmetic is unforgiving: for a one trillion parameters model at four bits it will take 50–100 chips. To service a sizable customer base will take thousands of 100-chip fabs.
That all said, I'm bullish on this technology, and look forward to seeing it evolve.
You did not compute that as the cost for a speculative card from Taalas, right?
*(It's local: private files managing firm oriented. It's blazing fast: it can be placed into recursive, intensive local workflows.)
Now maybe. When models are flying passenger aircraft, other prerogatives will assert themselves. When a 50TB ROM means you can impulse purchase a ChatGPT 6.3 xhigh that runs on batteries, yet more use cases will be apparent.
i have written about this:
"For device makers
Packaging models with laptops and smartphones will let application access near free, low latency inference and potentially offer users a better experience with the option of preserving data on-device. This is viable under the condition that tasks that do require larger expert models that run in the cloud can be routed to external models. A side-effect of local models and what will let Apple cut upgrade cycles from ~4 years (?) down to 12-18 months is specialized hardware to run them. For almost a decade, smartphones have been trying to compete on better cameras. This coming decade will see them selling better GPUs, NPUs, ASICs and whatever other things they'll be calling the inference chips, to drive re-purchase. Every six months will see a better model on new hardware, which will enable better performance in certain applications."
https://try.works/role-model-the-case-for-a-model-routing-pr...
Bitcoin mining doesn't have large memory requirements, but does have huge compute requirements. ASICs work great there because it's very straightforward to add some circuits for computing hashes. If you _also_ have to add many GB of memory, then suddenly ASICs will cost as much or more than comparable off-the-shelf hardware and they won't be faster unless you've also invested in huge memory bandwidth.
Bitcoin OTOH has used the same PoW algorithm for a decade. Barring some really exciting discoveries about the nature of computation, new ASICs are not that much more efficient than old ones.
BTC mining is also not exactly competitive anymore; the nature of the PoW algorithm means that it's dominated by a few large players who've set up shop next to a dam and who pay very little for electricity.
New entrants are highly discouraged because the mining rewards are constantly halving, it's hard to find cheap power, and the price of BTC is now so volatile that a yearslong investment is very likely to lose money.
It’s like talking about anything else than Megapixels when everyone was convinced that megapixels must go up in certain periods of the smartphone boom.
The Jalapeño mentioned («Anthropic is not alone in walking this path») in the article is still a classical Von Neumann architecture.
And Taalas' idea makes sense in a perspective of scale - producing a large number of cards; "for internal use" (a lower order of items) means a high production cost.
Indeed, for some kinds of applications involving secure/legal data etc. I can see the consistency of silicon winning out, because it combines performance with immutability and guardrails in hardware. Some chips have write-once PROMs to store password hashes and similar, you could do the same thing with prompt hashing to absolutely force or forbid certain behaviors. A model that can't be updated is also a model that can't be hacked.
Think vision, spatial reasoning, speech synthesis, even some speech analysis. Think self-driving cars (and drones) that need 10x less power for the brain, and can think at 10x situation per second.
We're all used to having to constantly update our browsers and phones to keep up with the security arms race. If a frozen model can't be updated, it will predictably remain vulnerable to any "exploits" or idiosyncratic quirks that people discover over time.
Let's say, as somebody suggested in another comment, that you buy 100,000 of these chips and deploy them to run fast-food drive-thrus. And then somebody discovers the model has a fondness for goblins[1], and if you role-play convincingly enough, you can get it to accept payment in shiny buttons and rodent skulls instead of cash.
What do you do then? I guess your options are to try and fix the behavior with a better prompt, or put some kind of filter in front of the model to catch attempted exploits. If the filter is cheap and dumb it probably won't work well enough, and if you use another model as a filter, you've negated the cost and speed benefits of putting the first model in hardware.
Of course the real answer is to just never expose the model to situations where an adversarial input could possibly lead to an undesired output. But that drastically limits what you can do with it.
[1]: https://openai.com/index/where-the-goblins-came-from/
NVIDIA will probably give us a new GPU when someone competent in the free market decides they want wheelbarrows full of money. Unfortunately, AMD is entirely, incomprehensibly, incompetent, to the point where I can only assume they're colluding with Nvidia, behind the scenes.
It also mostly passes the "schlong" test
https://pastes.io/YcxSi8Fp
Oxford also claim that its first recorded use was from the 60s, not the 20s; https://www.oed.com/dictionary/schlong_n?tl=true
It's also a question I have found will cause models that don't know what it is to go off quickly in a direction of hallucination trying to explain it, so the hallucination is evident very quickly starting from the first ever prompt issued with 0 context fill. Example: I had a model write four detailed supposedly-accurate sounding, grammatically correct paragraphs saying its origin is from AAVE (African American Vernacular English), which it most certainly is not
You could do the same by picking any topic that is very rarely discussed in conversation, some esoteric and narrow piece of knowledge and asking the model about it.
[0] Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity [https://arxiv.org/abs/2604.24827]
I tried to explain but I fear were probably going to be adding artificial sleeps to these things to convince the masses it's doing something clever.
It’s going to be really crazy when the bottle neck for agents is the speed of the tool calls rather than the speed of inference. Imagine an agent interacting with the terminal near instantly…
Developing software becomes 95% about intent and requirements. Can’t wait for the next iteration of that.
The characters in the 3-act Shakespearean play had very little depth, many of the names were similar, and they were not very smart, but the simple plot was cohesive.
https://huggingface.co/meta-llama/Llama-3.1-8B
As I remember just about any english language model from mid 2024 and earlier didn't even do well if you asked it to count sequentially from 0 to 100, nevermind calculating stuff.
....damn. It's very impressive notwithstanding its limitations.
There’s no difference in the inference implementation, parameter count, or speed.
(/s!)
Won’t the silicon etched model already be 1 or more versions behind by the time the silicon comes out.
Though if it’s cheap enough, there certainly can be a market for cheaper model inferences.
AI companies constantly update/change stuff, new models come out, new requirements, etc.
But if you ship an "ai-powered" dishwasher, it can come with the chip built-in to do computer vision and precisely target each spot, and will be sold as-is with no updates.
A lot of people would probably be happy to stick with the same model for a year or two if it’s 10x faster and cheaper.
And perhaps older models can become cheaper over time as newer models come out on new silicon for a higher price. That incentivizes people to stick with older models.
Something I personally haven’t seen much of, in all the discussions of model benchmarks and AI breakthroughs, is a distinction between “peak performance” and “reliable performance”. The “peak performance” of frontier models is very high: they’re solving open math problems, analyzing large codebases, etc. But my subjective impression is that “reliable performance” is mid at best: out of 100 random questions I might think to ask, it’s likely to say something wrong or stupid a handful of times at least.
I think there’s inherent tension between the two: the more a model reaches or outright hallucinates, the more likely it is to come up with tricky, subtle solutions to problems (I think people are somewhat like this too: Terry Tao’s brother is nonverbal, Jim Watson’s son has severe schizophrenia, etc). But then the less likely it is to generate a sensible email reply.
I use models all the time for coding, but I would not let one take over my daily correspondence. If the idea here is to run frontier models at high speed in data centers, that could be useful (the speed would be cool), but I’d be surprised if the cost of that hardware churn is worth it to frontier labs. But if the idea is to turn this into a chip that goes in your phone as some kind of routine, low-power inference thing…taking something too kooky to be relied on and baking it into your phone’s hardware like that doesn’t make sense to me.
What are some examples?
Burnt in, it needs a zif socket and easy access in every car, aircraft, a pull out slot in a phone, or it's new era planned obselescence.
Sort of a FPGA, that (electrically) arranges the connections on-boot, and then it's like a static inference chip.
If it can, then deployment in a sea of gates can make a chip viable across model generations as weights change, inside some scale factor.
If not, unless the part is under a pinout and address model which can scale on the bus, and can be easily replaced, it makes the entire dependency a replacement, not just this part. So embedded use has consequences.
Guess we can look forward to picking these up ex-enterprise on ebay for under $5k a pop in a decade or two
Caveats stacked high, obviously. It's 3.16M parameters (tinystories, and I also have a kevin-speak lemmatised version), the tokens are characters, and the 60k record is 16 streams that each remember exactly one token of context, so it's blisteringly fast at saying nothing. The honest build with full context and KV caching still does ~19k tok/s on one stream though.
I keep messing with the blogpost with the live demo, but I'm planning on flipping it to live in the next day or two
I feel like what we really need is the ability to solder computer cache on all sides of the chip Meaning above and below as well. If you can only attach it to the edges you will be inherently physically limited on the amount you can put (and maybe even have latency benefits as well)
Talaas is different, it's a true compute-in-memory architecture where the weights are stored in the connections between the transistors that perform the matrix multiply, rather than in seperate memory cells.
Most of the benefit comes from this architecture; hardwiring the weights into the silicon is just the easiest way to implement it. SRAM requires too many transistors, DRAM requires an incompatible manufacturing process, and exotic phase-change memories aren't readily available.
I agree that Groq with multilayer hybrid bonding could be a good idea.
We really, really need better secondary models that can do things fast and do them cheaply for lots of dumb tasks. Not only because it can be used as sub agents by frontier models, but also because it can be like a universal grease for all kinds of software.
I've got an app I am building and I don't want to tie myself with frontier models because I'll never be able to beat openai/anthropic. I just want a simple, cheap, instantaneous model that can just go through my documentation and tell the user what to do next and how to integrate with whatever ai subscription they have.
In Aug 2025 you had
- OpenAI o3
- Opus 4.1
- Gemini 2.5 Pro
- Grok 4
Even if those were almost free to run, you'd be way better off with Deepseek flash 0731 or GPT 5.6 Luna, which already are almost free.
Other than for things where the t/s are critical, it seems like a bad idea to etch a model into silicon.
Bad perspective: consider the correction: "when are thresholds of sought quality reached"? Hence: not "is there a 10yo from last year that could compete with the current 13yo", but "will there be a 30(?)yo from last year that could compete with the current 33(?)yo" ('(?)': the scale of yearly growth in the future is uncertain).
A 10 year old iPhone is probably good enough, but is there demand for it? In a vacuum a 10 year old iPhone is good, but why would you pick it if you can have a current one for a reasonable price?
The https://chatjimmy.ai demo was impressive.
Once models settle down this makes sense. Imagine a cartridge with a physical model on it. You purchase a cartridge and stick it in your computer/phone/server. Want to upgrade? By a new 'cartridge'.
This should bring inference cost down dramatically, I wonder how OpenAI/Anthropic feel about that.
it might already be time to start burning the best small models onto hardware since it's possible they can't get much better at many tasks like knowledge recall due to the inherent information density limits for models at a given size.
I can finally have my own Dixie flatline. Cool.
In case some did not know: also the movie (actually TV series) is finally happening.
# Neuromancer - Official Teaser ( https://news.ycombinator.com/item?id=49055037 )
Now your robot can respond sarcastically when you ask for chicken nuggets. Again. It also doesn't dent your walls anymore.
It’s composed of 4-bit multiplier cells that compute all 16 possible results in parallel. The top metal wiring layer physically selects the one that corresponds to a multiplication with that cell’s constant weight, and routes it to the next layer.
They won't sell/rent/license the weights to an end user at any price because they don't trust your security.
Imagine if like instead of having a specific Mac Plus ROM, you had a thing that looks like a fat ASIC that can hold models sitting on a slotted daughtercard directly next to the CPU and RAM.
But yeah, for things like programming, if it can do linux and python and some go and sql and javascript, larger domains can be threaded with LORA
Always fantasize about applying at Tenstorrent, but wrong side of Toronto. 2 hour commute.
Would it be realistic for a frontier lab to deploy this or would the turnaround time mean the model is always too out of date?
Assuming the weights and architecture are eventually stable, how much cheaper would this end up being?
Claude Code is still using haiku 4.5 from ages ago for explore subagents for instance. Not to mention production uses like customer service that only need to be "good enough"
But either way, I think GP's overall sentiment of "delegating intelligence-saturated tasks to an outdated but fast subagent" makes a lot of sense.
I guess losing some customers due to poor customer service is ok if the price of customer service is right.
6 months or even a year if something goes wrong in the fabrication process and you need to update things.
If they do more standard asic design, it could be a lot longer as the design needs to be validated on an FPGA cluster, which would necessarily need to be very big for something like a LLM. Easily up to 2 years.
There's a reason chatjimmy isn't demonstrating newer models and why they only show of an 8B model.
At the time people were no doubt saying yes but now 3.8 is out, is that still desirable?
"You are not prepared" --Illidan Stormrage
Models, probably first open weight ones like Kimi K3 class, are etched into silicon like this and sold as cartridges almost like old school game cartridges.
You buy a USB-C dongle that the cartridge goes into, or for data centers you have PCI cards that take these in slots.
My partner has been asking for a “completely private” model for doing research and shifting through volumes of data that can’t leave the office and $$$ for the current hardware makes no sense. It would be an easy sell if someone walks in with a black box that contains “ChatGPT”.
You need to be able to add|mul where the data (the weights) are stored.
Enjoying the Ian Cutress / TechTechPotato video on Taalas. Some ok good technical details on the tech, and some good insider baseball, whose who stuff. (What a treasure having tech discussions like this about.) https://youtu.be/3MKRjt59hh4
Bonsai Ternary (1.7bits/weight) is a compromise, compromise that has to make sense in the context - efficient when translated into transistors.
If tolerant, they could churn out many cheaper chips, some perhaps with slight abnormal tendencies ;)
Extremely! You can remove entire layers and the model will still work just fine, with barely perceptible capability losses.
I've cut/bypassed ~15% of total parameters out of Gemma 4 31B on a pod once. Still got perfectly coherent responses out of it. Certain layers are a lot more important than others, particularly early and late ones; but it's honestly astonishing how much can be cut out from the middle without destroying the model's coherence.
I didn't run any meaningful benchmarks, so I have no idea what the capability loss looks like exactly. But "produce coherent and sensible English in response to a wide variety of prompts" was definitely not among the things the model unlearned.
https://www.youtube.com/watch?v=UwCFY6pmaYY
> Our second model, still based on Taalas’ first-generation silicon platform (HC1), will be a mid-sized reasoning LLM
I don't see any evidence that this is possible. From my understanding, the whole model needs to be on a single chip. Which rules out any popular frontier models with several trillions of parameters. Even smaller sub-frontier models have hundreds of millions of parameters, so these would be ruled out as well.