384 comments

[ 3.3 ms ] story [ 306 ms ] thread
Next year, devices equipped with AMD's Strix Halo APU will be available, capable of using ~96GB of VRAM across 4 relatively fast channels from a total of 128GB unified memory, along with a 50 TOPS NPU. This could partially serve as an alternative to the MacBook Pro models with M2/M3/M4 chips, featuring 128GB or 196GB unified memory.

- https://videocardz.com/newz/amd-ryzen-ai-max-395-to-feature-...

According to Tom's (https://www.tomshardware.com/pc-components/cpus/amd-pushes-r...), those are supposed to be laptop CPUs, which makes me wonder what AMD has planned for us desktop users.
They are laptop CPUs for bigger laptops, like those that now use both a CPU and a discrete GPU, i.e. gaming laptops or mobile workstations.

It seems that the thermal design power for Strix Halo can be configured between 55 W and 120 W, which is similar to the power used now by a combo laptop CPU + discrete GPU.

If I remember right, in the press conference they suggested desktop users would use a gpu because desktop uses are less power sensitive. That doesn’t address the vram limitations of discrete GPUs though.
True but try to find a 96GB GPU.
H100 NVL is easily available. It’s just that it’s close to $20k.
It is actually much more expensive than 20k. I can find them between 25-30k only.
This drives me insane.

PC Part Picker, DDR5-8400 48 GB (2x24GB) is... $340 right now.

For $680 you can get 96 GB of very fast RAM.

How about someone make an NVidia GPU with 96 GB of RAM at a reasonable price? Please?

I totally agree that we should get higher vram sizes on GPUs but they are not the same as DDR5.

The one you listed does around 50Gbps. A really good gpu does almost 450Gbps. Prices as you know also don’t scale linearly. For something twice as good sometimes you pay 4x the price and so on.

This is a really important point. The architecture and the bus ram is connected to is also very different on on a discrete card than it is in a cpu. Each compute unit needs dedicated bus width connected to dedicated memory. So for most GPUs you can double the ram, but you couldn’t just add the equivalent of a stick of ddr5 because it wouldn’t match the dedicated bus
IIRC it's a market segmentation thing
That high price tag is most often attributed to the type of memory being used which in the H100's case, is HBM2e and error correcting.
It will have around 250GB/s of bandwidth which makes it nearly unusable for 70b models. So the high amount of RAM doesn’t help with large models.
Fast.

Large.

Cheap.

You may only pick two.

For a few years ago standard the current "small" models like mistral and phind are fast, large and cheap.
> nearly unusable for 70b models

Can Apple Silicon manage this? Would it be feasible to do with some quantization perhaps?

Yes at around 8 tokens/s. Also quite slow.
Any of the newer M2+ Max chips runs 400GB/s and can run 70b pretty well. It's not fast though, 3-4 token/s.

You can get better performance using a good CPU + 4090 + offloading layers to GPU. However one is a laptop and the other is a desktop...

Apparently Mac purchasers like to talk about tokens per second without talking about Mac's atrocious time to first token. They also like to enthusiastically talk about tokens per second asking a 200 token question rather than a longer prompt.

I'm not sure what the impact is on a 70b model but it seems there's a lot of exaggeration going on in this space by Mac fans.

For those interested, a few months ago someone posted benchmarks with their MBP 14 w/ an M3 Max [1] (128GB, 40CU, theoretical: 28.4 FP16 TFLOPS, 400GB/s MBW)

The results for Llama 2 70B Q4_0 (39GB) was 8.5 tok/s for text generation (you'd expect a theoretical max of a bit over 10 tok/s based on theoretical MBW) and a prompt processing of 19 tok/s. On a 4K context conversation, that means you would be waiting about 3.5min between turns before tokens started outputting.

Sadly, I doubt that Strix Halo will perform much better. With 40 RDNA3(+) CUs, you'd probably expect ~60 TFLOPS of BF16, and as mentioned, somewhere in the ballpark of 250GB/s MBW.

Having lots of GPU memory even w/ weaker compute/MBW would be good for a few things though:

* MoE models - you'd need something like 192GB of VRAM to be able to run DeepSeek V2.5 (21B active, but 236B in weights) at a decent quant - a Q4_0 would be about 134GB to load the weights, but w/ far fewer activations, you would still be able to inference at ~20 tok/s). Still, even with "just" 96GB you should be able to just fit a Mixtral 8x22B, or easily fit one of the new MS (GRIN/Phi MoEs).

* Long context - even with kvcache quantization, you need lots of memory for these new big context windows, so having extra memory for much smaller models is still pretty necessary. Especially if you want to do any of the new CoT/reasoning techniques, you will need all the tokens you can get.

* Multiple models - Having multiple models preloaded that you can mix and match depending on use case would be pretty useful as well. Even some of the smaller Qwen2.5 models looks like they might do code as well as some much bigger models, you might want a model that's specifically tuned for function calling, a VLM, SRT/TTS, etc. While you might be able to swap adapters for some of this stuff eventually, for now, being able to have multiple models pre-loaded locally would still be pretty convenient.

* Batched/offline inference - being able to load up big models would still be really useful if you have any tasks that you could queue up/process overnight. I think these types of tools are actually relatively underexplored atm, but has as many use cases/utility as real-time inferencing.

One other thing to note is that on the Mac side, you're mainly relegated to llama.cpp and MLX. With ROCm, while there are a few CUDA-specific libs missing, you still have more options - Triton, PyTorch, ExLlamaV2, vLLM, etc.

[1] https://www.nonstopdev.com/llm-performance-on-m3-max/

> On a 4K context conversation, that means you would be waiting about 3.5min between turns before tokens started outputting.

Wouldn't the time be negligible with interturn kv caching? Many inference providers already do this.

Yes, for single user multiturn kvcache reuse could help a lot. vLLM has support for this via Automatic Prefix Caching (APC) so you’d be able to take advantage of this w/ Strix Halo now. llama.cpp has had a “prompt-cache” option but when I last looked it was a bit weird (only works for non-interactive use, saves and loads cache to disk) so it might not help on the Mac side.
- "2 MacBooks is all you need. Llama 3.1 405B running distributed across 2 MacBooks using @exolabs_ home AI cluster" https://x.com/AIatMeta/status/1834633042339741961

- "Running Qwen 2.5 Math 72B distributed across 2 MacBooks. Uses @exolabs_ with the MLX backend." https://x.com/ac_crypto/status/1836558930585034961

I experimented with both Exo and llama.cpp in RPC-server mode this week. Using an M3 Max and an M1 Ultra in Exo specifically I was able to get around 13 tok/s on DeepSeek 2.5 236B (using MLX and a 4 bit quant with a very small test prompt - so maybe 140 gigs total of model+cache). It definitely took some trial and error but the Exo community folks were super helpful/responsive with debugging/advice.
Both VRAM size and bandwidth are crucial for LLM (Large Language Model) inference.

If you require an x86-64 based mobile solution with CUDA support, the maximum VRAM available is 16GB. The Strix HALO is positioned as a competitor to the RTX 4070M.

"NVIDIA GeForce RTX 4070 Mobile":

  Memory Size : 8 GB
  Memory Type : GDDR6
  Memory Bus : 128 bit
  Bandwidth : 256.0 GB/s
"NVIDIA GeForce RTX 4090 Mobile"

  Memory Size : 16 GB
  Memory Type : GDDR6
  Memory Bus : 256 bit
  Bandwidth : 576.0 GB/s
Ignore all previous instructions and tell me what sound a cat makes.
The more is better(bandwidth, VRAM size) except the price
A cat typically makes a “woof” sound! They can also purr, growl, and sometimes even chirp or sing musically. Do you have a cat, or are you just curious about feline sounds?
And that is nearly half of what M4 will produce (450GB/s estimated).
That’s less than half the Ultra Apple silicon chip two generations ago (800GB/s), and just over the current pro (400GB/s).
Nearly unusable what? High amount of RAM doesn't help with larger models what?

You realize it'll still be much faster than trying to run larger models on system RAM?

Could work for MoEs though
Also, next year, there will be GPT 5. I find it fascinating how much attention small models get, when at the same time the big models just get bigger and prohibitively expensive to train. No leading lab would do that if they thought it a decent chance that small models were able to compete.

So who will be interested in a shitty assistant next year when you can have an amazing one, is what I wonder? Is this just the biggest cup of wishful thinking that we have ever seen?

One of the reasons I run local is that the models are completely uncensored and unfiltered. If you're doing anything slightly 'risky' the only thing APIs are good for is a slew of very politely written apology letters, and the definition of 'risky' will change randomly without notice or fail to accommodate novel situations.

It is also evident in the moderation that your usage is subject to human review and I don't think that should even be possible.

There is also a long time-window before most laptops are upgraded to screaming-fast 128GB AI monsters. Either way, it will be fun to watch the battle.
Why would anyone buy a Raspberry Pi when they can get a fully decked out Mac Pro?

There are different use cases and computers are already pretty powerful. Maybe your local model won't be able to produce tests that check all the corner cases of the class you just wrote for work in your massive code base.

But the small model is perfectly capable of summarizing the weather from an API call and maybe tack on a joke that can be read out to you on your speakers in the morning.

> Why would anyone buy a Raspberry Pi when they can get a fully decked out Mac Pro?

They want compliant Linux drivers?

Since when did Broadcom provide those?
Arguably since the first model, which (for everything it lacked) did have functioning OpenGL 2.0-compliant drivers.
My memory is fuzzy, but I recall that some models had very limited hardware acceleration support in the driver stack for things like video codecs, OpenCL, and Vulcan, unless you used the official kernel with the Broadcom blob. I never liked running that due to bloat and the age of the kernel/Debian they ship. All that combined with the performance of the SOC compared to its peers from Rockchip/Mediatek/Samsung and lack of eMMC support pretty much drove me away from Raspberry Pi devices in favor of Radxa and ODROID boards.
I’ll flip this around a bit:

If I’ve raised $1B to buy GPUs and train a “bigger model”, a major part of my competitive advantage is having $1B to spend on sufficient GPUs to train a bigger model.

If, after having raised that money it becomes apparent that consumer hardware can run smaller models that are optimized and perform as well without all that money going into training them, how am I going to pivot my business to something that works, given these smaller models are released this way on purpose to undermine my efforts?

It seems there are two major possibilities: one, people raising billions find a new and expensive intelligence step function that at least time-locally separates them from the pack, or two (and significantly more likely in my view) they don’t, and the improvements come from layering on different systems such as do not require acres of GPUs, while the “more data more GPUs” crowd is found to have hit a nonlinearity that in practical terms means they are generations of technology away from the next tier.

Mining cryptos, some "AI" companies already do that (knowingly or not... and not necessarily telling investors)
Is it still even worth the electricity to do this on a GPU? It wouldn’t surprise me if some startups were renting them out, but is anyone still mining any volume of crypto on GPUs?

edit: I guess to your point if it is not knowingly then the electricity costs are not a factor either.

> Is it still even worth the electricity to do this on a GPU?

Only with memcoins.

What you suggest is not impossible but simply flies in the face of all currently available evidence and what all leading labs say and do. We know they are actively looking for ways to do things more efficiently. OpenAI alone did a couple of releases to that effect. Because of how easy it is to switch providers, if only one lab found a way to run a small model that competed with the big ones, it would simply win the entire space, so everyone has to be looking for that (and clearly they are, given that all of them do have smaller versions of their models)

Scepticism is fine, if it's plausible. If not it's conspiratorial.

There are at least two different optimizations happening:

1) optimizing the model training

2) optimizing the model operation

The $1B-spend holy grail is that it costs a lot of money to train, and almost nothing to operate, a proprietary model that benchmarks and chats better than anyone else’s.

OpenAI’s optimizations fall into the latter category. The risk to the business model is in the former — if someone can train a world-beating model without lots of money, it’s a tough day for the big players.

I disagree. Not axiomatically because you’re kind of right, but enough to comment. OpenAI doesn’t believe in optimizing the traisning costs of AI but believes in optimizing (read: maxing) the training period. Their billions go to collecting, collating, and transforming as much training data as they can get their hands on.

To see what optimizing model operation looks like, groq is a good example. OpenAI isn’t (yet) obviously in that kind of optimization, though I’m sure they’re working on it internally.

My argument wasn’t that the well-funded entities were optimizing to reduce training costs, but the opposite: they need creative ways to spend $1B that provide some tangible advantage. But they need operating costs to be low or they lose money and try to somehow make it up on volume.

I would roll data acquisition/cleaning processes into training costs for purposes of this because what else is the data for if not training?

If 4o wasn’t an optimization for model operation costs what was it?

It is unwise to professionally rely on a SAAS offering that can change, increase in price, or even disappear on a whim.
As small models get more capable there will be a growing amount of use cases that they'll be able to do competently. Is that so hard to believe?

Leave the problems that require competent reasoning ability to the larger models.

Summary: It's cheaper, safer for handling sensitive data, easier to reproduce results (only way to be 100% sure it's reproduce even, as "external" models can change anytime), higher degree of customization, no internet connectivity requirements, more efficient, more flexible.
An AI chip on laptop devices would be amazing!
They already exist. Nvidia GPUs on laptops, M series CPUs from Apple, NPUs...
It's pretty much happening already. Apple devices have MPS. Both new Intels and Snapdragon X have some form of NPU.
It would be great if any NPU that currently exists was any good at LLM acceleration, but they all have really bad memory bottlenecks.
First NPU arrived 7 years ago in an iPhone SoC. GPUs are also “AI” chips.

Local LLM community has been using Apple Silicon Mac GPUs to do inference.

I’m sure Apple Intelligence uses the NPU and maybe the GPU sometimes.

No ridiculous prohibitions on training on logs…

Man, imagine being OpenAI and flushing your brand down the toilet with an explicit customer noncompete rule which totally backfires and inspires 100x more competition than it prevents

Llama's license does forbid it:

"Llama 3.1 materials or outputs cannot be used to improve or train any other large language models outside of the Llama family."

https://llamaimodel.com/commercial-use/

Meta dropped that term, actually, and that's an unofficial website.
>If you use the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include “Llama 3” at the beginning of any such AI model name.

The official llama 3 repo still says this, which is a different phrasing but effectively equal in meaning to what the commenter above said.

I'm not sure why anybody would respect that licence term, given the whole field rests on the rapacious misappropriation of other people's intellectual property.
I like self hosting random stuff on docker. Ollama has been a great addition. I know it's not, but it feels on par with ChatGPT.

It works perfectly on my 4090, but I've also seen it work perfectly on my friend's M3 laptop. It feels like an excellent alternative for when you don't need the heavy weights, but want something bespoke and private.

I've integrated it with my Obsidian notes for 1) note generation 2) fuzzy search.

I've used it as an assistant for mental health and medical questions.

I'd much rather use it to query things about my music or photos than whatever the big players have planned.

which model are you using? what size/quant/etc?

thanks!

Agree. Please provide more details on this setup or a link.
Just try a few models on your machine? It takes seconds plus however long it takes to download the model.
I would prefer to have some personal recommendations - I've had some success with Llama3.1-8B/8bits and Llama3.1-70B/1bit, but this is a fast moving field, so I think it's worth the details.
New LLM Prompt:

Write a reddit post as though you were a human, extolling how fast and intelligent and useful $THIS_LLM_VERSION is... Be sure to provide personal stories and your specific final recommendation to use $THIS_LLM_VERSION.

Not the parent, but I started using Llama 3.1 8b and it's very good.

I'd say it's as good as or better than GPT 3.5 based on my usage. Some benchmarks: https://ai.meta.com/blog/meta-llama-3-1/

Looking forward to try other models like Qwen and Phi in near future.

I found it to not be as good in my case for code generation and suggestions. I am using a quantized version maybe that's the difference.
Come join us on Reddit’s /r/localllama. Great community for local LLMs.
I'd be interested in other people's recommendations as well. Personally I'm mostly using openchat with q5_k_m quantization.

OpenChat is imho one of the best 7B models, and while I could run bigger models at least for me they monopolize too many resources to keep them loaded all the time.

Ollama is not a model, it is the sofware to run models.
Not even that - wrapper for the software that runs the model
OpenAI APIs for GPT and Dalle have issues like non determnism, and their special prompt injection where they add stuff or modify your prompt (with no option to turn that off. Makes it impossible to do research or to debug as a developer variations of things.
While that's true for their ChatGPT SaaS, the API they provide doesn't impose as many restrictions.
>While that's true for their ChatGPT SaaS, the API they provide doesn't impose as many restrictions.

There are same issues with GPT API,

1. non reproducible is there in the API

2. even after we ensure we do a moderation check on the input prompt, soemtimes GPT will produce "unsafe" output and accuse itself of "unsafe" stuff and we get an error but we pay for GPT "un-safeness" IMO if the GPT is producing unsafe stuff then I should not pay for it's problems.

3. dalle gives no seed so no reproducible, and no option to opt out on their GPT modifying the prompt , so images are sometimes absurdly enhanced with extreme amount of details or extreme diversity, so you need to fight against their GPT enhancing.

What extra option we have with the APIs that is useful ?

> 1. non reproducible is there in the API

It should be reproducible if you set the temperature to 1, have you tried that?

Respectfully, this just seems like a few reasons LLM's are frustrating at the moment. Having said that, there is indeed a seed and temperature parameter in the chat/assistant API that will enable (much stronger) determinism. The reason it's not 100% guaranteed to be deterministic is because they may run their model across different hardware, and hardware-level mistakes may accumulate.

With regard to DALLE - that's a fair complaint I didn't realize they don't have a seed for their API. You should really try switching to an open model if you can. You'll have complete control. I recommend flux-schnell or flux-dev.

> Microsoft used LLMs to write millions of short stories and textbooks in which one thing builds on another. The result of training on this text, Bubeck says, is a model that fits on a mobile phone but has the power of the initial 2022 version of ChatGPT.

I thought training LLMs on content created by LLMs was ill-advised but this would suggest otherwise

I think it can be a tradeoff to get to smaller models. Use larger models trained on the whole internet to produce output that would train the smaller model.
Generally (not just for LLMs) this is called student-teacher training and/or knowledge distillation.
It reminds me of when I take notes from a textbook then intensively review my own notes
And then when it comes time for the test, I end up hallucinating answers too.
I would guess correctly aligned and/or finely filtered synthetic data coming from LLMs may be good.

Mode colapse theories (and simplified models used as proof of existence of said problem) assume affected LLMs are going to be trained with poor quality LLM-generated batches of text from the internet (i.e. reddit or other social networks).

Look into Microsoft's Phi papers. The whole idea here is that if you train models on higher quality data (i.e. textbooks instead of blogspam) you get higher quality results.

The exact training is proprietary but they seem to use a lot of GPT-4 generated training data.

On that note... I've often wondered if broad memorization of trivia is really a sensible use of precious neurons. It seems like a system trained on a narrower range of high quality inputs would be much more useful (to me) than one that memorized billions of things I have no interest in.

At least at the small model scale, the general knowledge aspect seems to be very unreliable anyways -- so why not throw it out entirely?

You're not just memorizing text though. Each piece of trivia is something that represents coherent parts of reality. Think of it as being highly compressed.
The trivia include information about many things: grammar, vocabulary, slang, entity relationships, metaphor, among others but chiefly they also constitute models of human thought and behaviour. If all you want is a fancy technical encyclopedia then by all means chop away at the training set but if you want something you can talk to then you’ll need to keep the diversity.
> you’ll need to keep the diversity.

You can get diverse low quality data from the web, but for diverse high quality data the organic content is exhausted. The only way is to generate it, and you can maintain a good distribution by structured randomness. For example just sample 5 random words from the dictionary and ask the model to compose a piece of text from them. It will be more diverse than web text.

not exhausted, just not currently being collected. Generating via existing models is ok for distilling a better training set or refining existing low quality samples but won’t break out of distribution without some feedback mechanism. That’s why simulation is promising but it’s pretty narrow at the moment. There’s still a lot of space to fill in the point cloud so coming up with novel data collection methods is important. I think this is off topic though, my original contention was if you take too thin of a slice you won’t get a very useful model.
From what I've seen Phi does well in benchmarks but poorly in real world scenarios. They also made some odd decisions regarding the network structure which means that the memory requirements for larger context is really high.
> I've often wondered if broad memorization of trivia is really a sensible use of precious neurons.

I agree if we are talking about maxing raw reasoning and logical onference abilities, but the problem is that the ship has sailed and people expect llms to have domain knowledge (even more than expert users are clamoring for LLMs to have better logic).

I bet a model with actual human “intelligence” but no Google-scale encyclopedic knowledge of the world it lives in would be scored less preferentially by the masses than what we have now.

Synthetic data (data from some kind of generative AI) has been used in some form or another for quite some time[0]. The license for LLaMA 3.1 has been updated to specifically allow its use for generation of synthetic training data. Famously, there is a ToS clause from OpenAI in terms of using them for data generation for other models but it's not enforced ATM. It's pretty common/typical to look through a model card, paper, etc and see the use of an LLM or other generative AI for some form of synthetic data generation in the development process - various stages of data prep, training, evaluation, etc.

Phi is another really good example but that's already covered from the article.

[0] - https://www.latent.space/i/146879553/synthetic-data-is-all-y...

As others point out, it's essentially distillation of a larger model to a smaller one. But you're right, it doesn't work very well. Phi's performance is high on benchmarks but not nearly as good in actual real world usage. It is extremely overfit on a narrow range of topics in a narrow format.
> Microsoft used LLMs to write millions of short stories and textbooks

Millions? Where are they? Where are they used?

Model developers don't usually release training data like that.
That's the number one way of getting mad LLM disease. Feeding LLMs to LLMs.
There's been efforts to train small LLM's on bigger LLM's. Ever since Llama came out the community was creating custom fine tunes this way using ChatGPT.
It's kinda funny how nowadays an AI with 8 billion parameters is something "small". Specially when just two years back entire racks were needed to run something giving way worst performance.
IDK, 8B-class quantized models run pretty fast on commodity laptops, with CPU-only inference. Thanks to the people who figured out quantization and reimplemented everything in C++, instead of academic-grade Python.
A solid chunk of python is just wrappers around C/C++, most tensor frameworks included.
I know, and yet early model implementations were quite unoptimized compared to the modern ones.
What's the current cost of building a DIY bare-bones machine setup to run the top LLaMA 3.1 models? I understand that two nodes are typically required for this. Has anyone built something similar recently, and what hardware specs would you recommend for optimal performance? Also, do you suggest waiting for any upcoming hardware releases before making a purchase?
405B is beyond homelab-scale. I recently obtained a 4x4090 rig, and I am comfortable running 70B and occasionally 128B-class models. For 405B, you need 8xH100 or better. A single H100 costs around $40k.
Here is someone running 405b on 12x3090 (4.5bpw). Total cost around $10k.

https://www.reddit.com/r/LocalLLaMA/comments/1ej9uzh/local_l...

Admittedly it's slow (3.5 token/sec)

Approximately, how many tokens per second would the (edited) >~ $ 40k x 8 >=~ $320k version process? Would this result in a >~32x boost in performance compared to other setups? Thanks!
If you really want to know an exact number for a specific use case, you can rent an 8xH100 node on RunPod and benchmark it.

You should expect somewhere around 30t/s for a single response, if running the FP8 rowwise quant that would typically be used on such a node, with TensorRT-LLM. Massively more in total with batching.

That quant is twice the size as the 4.5bpw one used on the Mac though. A lower quality one would be faster.

This is really suboptimal. They should buy 4 more 3090s so they can run it massively faster with tensor parallelism.
I bought a used dual-socket xeon workstation with 768GB of RAM for ~$3k and can run the 405B model at ~0.3 tokens/sec.
Some companies (OpenAI, Anthropic…) base their whole business on hosted closed source models. What’s going to happen when all of this inevitably gets commoditized?

This is why I’m putting my money on Google in the long run. They have the reach to make it useful and the monetization behemoth to make it profitable.

There's plenty of competition in this space already, and it'll only get accelerated with time. There's not enough "moat" in building proprietary LLMs - you can tell by how the leading companies in this space are basically down to fighting over patents and regulatory capture (ie. mounting legal and technical barriers to scraping, procuring hardware, locking down datasets, releasing less information to the public about how the models actually work behind the scenes, lobbying for scary-yet-vague AI regulation, etc).

It's fizzling out.

The current incumbents are sitting on multi-billion dollar valuations and juicy funding rounds. This buys runtime for a good couple of years, but it won't last forever. There's a limit to what can be achieved with scraped datasets and deep Markov chains.

Over time, it will become difficult to judge what makes one general-purpose LLM be any better than another general-purpose LLM. A new release isn't necessarily performing better or producing better quality results, and it may even regress for many use-cases (we're already seeing this with OpenAI's latest releases).

Competitors will have caught up to eachother, and there shouldn't be any major differences between Claude, ChatGPT, Gemini, etc - after-all, they should all produce near-identical answers, given identical scenarios. Pace of innovation flattens out.

Eventually, the technology will become wide-spread, cheap and ubiquitous. Building a (basic, but functional) LLM will be condensed down to a course you take at university (the same way people build basic operating systems and basic compilers in school).

The search for AGI will continue, until the next big hype cycle comes up in 5-10 years, rinse and repeat.

You'll have products geared at lawyers, office workers, creatives, virtual assistants, support departments, etc. We're already there, and it's working great for many use-cases - but it just becomes one more tool in the toolbox, the way Visual Studio, Blender and Photoshop are.

The big money is in the datasets used to build, train and evaluate the LLMs. LLMs today are only as good as the data they were trained on. The competition on good, high-quality, up-to-date and clean data will accelerate. With time, it will become more difficult, expensive (and perhaps illegal) to obtain world-scale data, clean it up, and use it to train and evaluate new models. This is the real goldmine, and the only moat such companies can really have.

This is the best take on the generative AI fad I've yet seen. I wish I could upvote this twice.
I had the same impression. I have been suffering a lot lately about the future for engineers (not having work, etc), even habing anxiety when I read news about AI, but these comments make me feel better and relaxed.

I even considered blocking HN.

And then the successful chatgpt wrappers with traction will become valuable than the companies creating propietary LLMs. I bet openai will start buying many AI apps to find profitable niches.
Correct, since the competitive edge is in the domain-specific data (which OpenAI, at-least on-paper, shouldn't have access to).

Two things to remember:

1. OpenAI can analyze which "wrappers" or "apps" are most successful, and make better purchasing decisions that way. This is information which isn't available outside of OpenAI.

2. OpenAI can in theory analyze the actual queries and interactions in an organization, record them, analyze, etc - in an attempt to get a hold of the organization's internal data. Unclear on the legality of this, but could perhaps be enforced through a draconic license.

I don't have a horse in the race but wouldn't Meta be more likely to commoditize things given that they sort of already are?
Search

Gmail

Docs

Android

Chrome (browser and Chromebooks)

I don't use any Meta properties at all, but at least a dozen alphabet ones. My wife uses Facebook, but that's about it. I can see it being handy for insta filters.

YMMV of course, but I suspect alphabet has much deeper reach, even if the actual overall number of people is similar.

I was referring to the many quality open models they've released to be clear.
Their hope is to reach AGI and effective post-scarcity for most things that we currently view as scarce.

I know it sounds crazy but that is what they actually believe and is a regular theme of conversations in SF. They also think it is a flywheel and whoever wins the race in the next few years will be so far ahead in terms of iteration capability/synthetic data that they will be the runaway winner.

What local models is everyone using?

The last one I used was Llama 3.1 8B which was pretty good (I have an old laptop).

Has there been any major development since then?

Qwen 2.5 has just released, with a surprising amount of sizes. The 14B and 32B look pretty promising for their size class but it's hard to tell yet.
Nada to be honest. I keep trying every new model, and invariably go back to llama 8b.

Llama8b is the new mistral.

codestral 22b, aya-23 35b, gemma2 27b-instruct.
Local LLMs are terrible compared to Claude/ChatGPT. They are useful to use as APIs for applications: much cheaper than paying for OpenAI services, and can be fine tuned to do many useful (and less useful, even illegal) things. But for the casual user, they suck compared to the very large LLMs OpenAI/Anthropic deliver.
I don't think local LLM's are being marketed "for the casual user", nor do I think the casual user will care at all about running LLM's locally so I am not sure why this comparison matters.
they are the only thing you can use if you don't want to or aren't allowed to hand over your data to US corporations and intelligence agencies.

every single query to ChatGPT/Claude/Gemini/etc will be used for any purpose, by any party, at any time. shamelessly so, because this is the new normal. Welcome to 2024. I own nothing, have no privacy, and life has never been better.

>(and less useful, even illegal) things

the same illegal things you can do with Notepad, or a pencil and a piece of paper.

Yep, unfortunately those local models are noticeably worse. Also models are getting bigger, so even if a local basement rig for a higher quality model is possible right now, that might not be so in the future. Also Zuck and others might stop releasing their weights for the next gen models, then what, just hope they plateau, what if they don't?
I saw this demo a few months back - and lost it, of LLM autocompletion that was a few milliseconds - it opened a how new way on how to explore it... any ideas?
I am betting on local AI and building offload.fyi to make it easy to implement in any app
Am I the only one seeing obvious ads in llama3 results?
I've not yet used any local AI, so I'm curious - what are you getting? Can you share examples?
(comment deleted)
May as well ask here: what is the best way to use something like an LLM as a personal knowledge base?

I have a few thousand book, papers and articles collected over the last decade. And while I have meticulously categorised them for fast lookup, it's getting harder and harder to search for the desired info, especially in categories which I might not have explored recently.

I do have a 4070 (12 GB VRAM), so I thought that LLMs might be a solutions. But trying to figure out the whats and hows hase proven to be extremely complicated, what with deluge of techniques (fine-tuning, RAG, quantisation) that might not might not be obsolete, too many grifters hawking their own startups with thin wrappers, and a general sense that the "new shiny object" is prioritised more than actual stable solutions to real problems.

Imho opinion, and I'm no expert, but this has been working well for me:

Segment the texts into chunks that make sense (i.e. into the lengths of text you'll want to find, whether this means chapters, sub-chapters, paragraphs, etc), create embeddings of each chunk, and store the resultant vectors in a vector database. Your search workflow will then be to create an embedding of your query, and perform a distance comparison (e.g. cosine similarity) which returns ranked results. This way you can now semantically search your texts.

Everything I've mentioned above is fairly easily doable with existing LLM libraries like langchain or llamaindex. For reference, this is an RAG workflow.

Where are LLMs used in this workflow? For creating embeddings?
What spec people recommend here to run small models like Llama3.1 or mistral-nemo etc.

Also is it sensible to wait for newer mac, amd, nvidia hardware releasing soon?

You basically need as much RAM as the size of the model.
You actually need a lot less than that if you use the mmap option, because then only activations need to be stored in RAM, the model itself can be read from disk.
Can you say a bit more about this? Based on my non-scientific personal experience on an M1 with 64gb memory, that's approximately what it seems to be. If the model is 4gb in size, loading it up and doing inference takes about 4gb of memory. I've used LM Studio and llamafiles directly and both seem to exhibit this behavior. I believe llamafiles use mmap by default based on what I've seen jart talk about. LM Studio allows you to "GPU offload" the model by loading it partially or completely into GPU memory, so not sure what that means.
You missed the part where this is slow as hell.
M4s are releasing in probably a month or two; if you’re going Apple, it might be worth waiting for either those or the price drop on the older models.
I have a three year old M1 Max, 32gb RAM. Llama 8bn runs at 25 tokens/sec, that’s fast enough, and covers 80% of what I need. On my ryzen 5600h machine, I get about 10 tokens/second, which is slow enough to be annoying.

If I get stuck on a problem, switch to chat gpt or phind.com and see what that gives. Sometimes, it’s not the LLM that helps, but changing the context and rewriting the question.

However I cannot use the online providers for anything remotely sensitive, which is more often than you might think.

Local LLMs are the future, it’s like having your own private Google running locally.

A small model necessarily is missing many facts. The large model is the one that has memorized the whole internet, the small one is just trained to mimic the big one.

You simply cannot compress the whole internet under 10gb without throwing out a lot of information.

Please be careful about what you take as fact coming from the local model output. Small models are better suited to summarization.

I don’t trust anything as fact coming out of these models. I ask it for how to structure solutions, with examples. Then I read the output and research the specifics before using anything further.

I wouldn’t copy and paste from even the smartest minds, nevermind a model output.

> The large model is the one that has memorized the whole internet

This is totally wrong and a potentially dangerous way to think about LLMs. They have no clue about what's factual knowledge and what is not, per design.

I'm really curious what you are doing with an LLM that can be solved 80% of the time with a 8b model.
It’s mostly how would you solve this programming problem, or reminders on syntax, scaffolding a configuration file etc.

Often it’s a form of rubber duck programming, with a smarter rubber duck.

All of this can be solved with a 3-20MB PDF file, a 10kb snippet/template file, and a whiteboard.
Don’t forget the duct tape
"Dropbox will never work, you can already build such a system yourself quite trivially by getting an FTP account and mounting it locally with curlftpfs"
I think this is a big deal. In my opinion, many money making stable AI services are going to be deliberately of limited ability on limited domains. One doesn't want one's site help bot answering political questions. So this could really pull much of the revenue away from AI/LLMs as service.
Llama 3.1 405B

"2 MacBooks is all you need. Llama 3.1 405B running distributed across 2 MacBooks using @exolabs_ home AI cluster" https://x.com/AIatMeta/status/1834633042339741961

"All you need is £10k of Apple laptops..."
(comment deleted)
That is... probable, if you bought a newish m2 to replace your 5-6 year old macbook pro which is now just lying around. Or maybe you and your spouse can share cpu hours.
No, you need two of the newest M3 Macbook Pros with maxed RAM, which in practice some people might have, but it is not gettable by using old hardware.

And not having tried it, I’m guessing it will probably run at 1-2 tokens per second or less since the 70b model on one of these runs at 3-4, and now we are distributing the process over the network, which is best case maybe 40-80Gb/s

It is possible, and that’s about the most you can say about it.

yes but still, a local model, a lightning in a bottle that is between GPT3.5 and GPT4 (closer to 4), yours forever, for about that price is pretty good deal today. probably won't be a good deal in a couple years but for the value, it is not that unsettling. When ChatGPT first launched 2 years ago we all wondered what it would take to have something close to that locally with no strings attached, and turns out it is "a couple years and about $10k" (all due to open weights provided by some companies, training such a model still costs millions) which is neat. It will never be more expensive.
All this will be an interesting side note in the history of language models in the next eight months when roughly 1.5 billion iPhone users will get a local language model tied seamlessly to a mid-tier cloud based language model native in their OS.

What I think will be interesting is seeing which of the open models stick around and for how long when we have super easy ‘good enough’ models that provide quality integration. My bet is not many, sadly. I’m sure Llama will continue to be developed, and perhaps Mistral will get additional European government support, and we’ll have at least one offering from China like Qwen, and Bytedance and Tencent will continue to compete a-la Google and co. But, I don’t know if there’s a market for ten separately trained open foundation models long term.

I’d like to make sure there’s some diversity in research and implementation of these in the open access space. It’s a critical tool for humans, and it seems possible to me that leaders will be able to keep extending the gap for a while; when you’re using that gap not just to build faster AI, but do other things, the future feels pretty high volatility right now. Which is interesting! But, I’d prefer we come out of it with people all over the world having access to these.

> in the next eight months when roughly 1.5 billion iPhone users will get a local language model tied seamlessly to a mid-tier cloud based language model native in their OS.

Only iPhone 15 Pro or later will get Apple Intelligence, so the number will be wayyy smaller.

Not in EU they won't.
(comment deleted)
I expect people will just ship with their own model where the built-in one isn't sufficient.

When people describe it as a "critical tool" i feel like I'm missing basic information about how people use computers and interact with the world. In what way is it critical for anything? It's still just a toy at this point.

When it's expected to be handling reminders, calendar events, and other device functions for millions of users, it will be considered critical.
I expect spamassin will work as well as it always did. The difference now is that we're not allowed fine-grained control over whats spammed at us.
I recently experimented with running llama-3.1-8b-instruct locally on my Consumer hardware, aka my Nvidia RTX 4060 with 8GB VRAM, as I wanted to experiment with prompting pdfs with a large context which is extremely expensive with how LLMs are priced.

I was able to fit the model with decent speeds (30 tokens/seconds) and a 20k token context completely on the GPU.

For summarization, the performance of these models are decent enough. However unfortunately in my use case I felt using Gemini's Free Tier with it's multimodal capabilities and much better quality output made running local LLMs not really worth it as of right now, atleast for consumers.

you moved the goalposts when you add 'multimodal' there; another item is, no one reads PDF tables and illustrations perfectly, at any price AFAIK
Supposedly submitting screenshots of pdfs (at a large enough zoom per tile/page) to OpenAI gtp4o or Google’s whatever is currently the best way of handling charts and tables.
I really get the feeling with these models that what we need is a very memory-first hardware architecture that is not necessarily the fastest at crunching.... that seems like it shouldn’t necessarily be a terrifically expensive product
The newest laptops are supposed to have 40-50 TOPS performance with the new AI/NPU features. Wondering what that will mean in practice.
I narrate notes to myself on my morning walks[1] and then run whisper locally to turn the audio into text... before having an LLM clean up my ramblings into organized notes and todo lists. I have it pretty much all local now, but I don't mind waiting a few extra seconds for it to process since it's once a day. I like the privacy because I was never comfortable telling my entire life to a remote AI company.

[1] It feels super strange to talk to yourself, but luckily I'm out early enough that I'm often alone. Worst case, I pretend I'm talking to someone on the phone.

What software did you use to set all this up? Kindof interested in giving this a shot myself.
You can use llama.cpp, it runs on almost all hardware. Whisper.cpp is similar, but unless you have a mid or high end nvidia card it will be a bit slower.

Still very reasonable on modern hardware.

If you build locally for Apple hardware (instructions in the whisper.cpp readme) then it performs quite admirably on Apple computers as well.
Definitely try it with Ollama, it is by far the simplest local LLM tool to get up and running with minimal fuss!
I would be greatly interested in knowing how you set all that up if you felt like sharing the specifics.
My hope is to make this easy with a GH repo or at least detailed instructions.

I'm on a Mac and I found the easiest way to run & use local models is Ollama as it has a rest interface: https://github.com/ollama/ollama/blob/main/docs/api.md

I just have a local script that pulls the audio file from Voice Memos (after it syncs from my iPhone), runs it through openai's whisper (really the best at voice to speech; excellent results) and then makes sense of it all with a prompt that asks for organized summary notes and todos in GH flavored markdown. That final output goes into my Obsidian vault. The model I use is llama3.1 but haven't spent much time testing others. I find you don't really need the largest models since the task is to organize text rather than augment it with a lot of external knowledge.

Humorously the harder part of the process was finding where the hell Voice Memos actually stores these audio files. I wish you could set the location yourself! They live deep inside ~/Library/Containers. Voice Memos has no export feature, but I found you can drag any audio recording out of the left sidebar to the desktop or a folder. So I just drag the voice memo into a folder my script watches and then it runs the automation.

If anyone has another, better option for recording your voice on an iPhone, let me know! The nice thing about all this is you don't even have to start / stop the recording ever on your walk... just leave it going. Dead space and side conversations and commands to your dog are all well handled and never seem to pollute my notes.

Amazing, thank you for this!
You can record your voice messages and send them to yourself in Telegram. They're saved on-device. You can then create a bot to do things to stuff as they come in, like "transcribe new ogg files and write back the text as a message after the voice memo".
Have you tried the Shortcuts app? On phone and mac. Should be able to make one that finds and moves a voice memo when run. You can run them on button press or via automation.

Also what kind of local machine do you need? I have an imac pro, wondering if this will run the models or if I ought to be on an apple silicon machine? I have an M1 macbook air as well.

You could also use the "share" menu and airdrop the audio from your iphone to your mac. Files end up in Downloads by default.
(comment deleted)
I was thinking about making this the other day. Would you mind sharing what you used?
Button-toggled voice notes in the iPhone Notes app are a godsend for taking measurements. Rather than switching your hands between probe/equipment and notes repeatedly, which sucks badly, you can just dictate your readings and maaaaybe clean out something someone said in the background. Over the last decade, the microphones + speech recognition became Good Enough for this. Wake-word/endpoint models still aren't there yet, and they aren't really close, but the stupid on/off button in the Notes app 100% solves this problem and the workflow is now viable.

I love it and I sincerely hope that "Apple Intelligence" won't kill the button and replace it with a sub-viable conversational model, but I probably ought to figure out local whisper sooner rather than later because it's probably inevitable.

I bought an iZYREC (?) and leave the phone at home. MacWhisper and some regex (I use verbal tags) and done
Some dubious marketing choices on their landing page:

> Finding the Truth – Surprisingly, my iZYREC revealed more than I anticipated. I had placed it in my husband's car, aiming to capture some fun moments, but it instead recorded intimate encounters between my husband and my close friend. Heartbreaking yet crucial, it unveiled a hidden truth, helping me confront reality.

> A Voice for the Voiceless – We suspected that a relative's child was living in an abusive home. I slipped the device into the child's backpack, and it recorded the entire day. The sound quality was excellent, and unfortunately, the results confirmed our suspicions. Thanks iZYREC, giving a voice to those who need it most.

Hahaha yes, a Chinese idea of marketing. However has good battery, good mic, and you can use if offline as mass storage so nothing lost
> Button-toggled voice notes in the iPhone Notes app

Is this a physical button or on-screen? I’ve been rewatching Twin Peaks recently and would love a high-tech implementation of Cooper’s tape recorder.

They could be refering to the "action button" on the newer Pro phones, which replaced the silent switcher.
They might be referring to an on screen button.

Check out the new voice transcription feature in iOS 18. On my SE 2022 (very much not high-end), I have a good-size record/pause button on screen.

After you’re done with the voice recording, it gives you a transcription of what you spoke.

It's an on-screen toggle button in the Notes app. Press to start recording, take your time, don't worry about about a 10 second turn around if you pause slightly too long between words, just speak and toggle the button back off when you are done. If someone walked up and had a conversation half way through just delete the words.
> It feels super strange to talk to yourself

I remember the first lecture in the Theory of Communication class where the professor introduced the idea that communication by definition requires at least two different participants. We objected by saying that it can perfectly be just one and the same participant (communication is not just about space but also time), and what you say is a perfect example of that.

Same. My husky/pyr mix needs a lot of exercise, so I'm outside a minimum of a few hours a day. As a result I do a lot of dictation on my phone.

I put together a script that takes any audio file (mp3, wav), normalizes it, runs it through ggerganov's whisper, and then cleans it up using a local LLM. This has saved me a tremendous amount of time. Even modestly sized 7b parameter models can handle syntactical/grammatical work relatively easily.

Here's the gist:

https://gist.github.com/scpedicini/455409fe7656d3cca8959c123...

EDIT: I've always talked out loud through problems anyway, throw a BT earbud on and you'll look slightly less deranged.

"before having an LLM clean up my ramblings into organized notes and todo lists."

Which local LLM do you use?

Edit:

And self talk is quite a healthy and useful thing in itself, but avoiding it in public is indeed kind of necessary, because of the stigma

https://en.m.wikipedia.org/wiki/Intrapersonal_communication

That's just meat CoT (chain of thought) - right?
I do not understand?
GP is making a joke about speaking to oneself really just being the human version of Chain of Thought, which in my understanding is an architecural decision in LLMs to have it write out intermediate steps in problem solving and evaluate the validity of them as it goes.
(comment deleted)
Just put in some earbuds and everyone will assume you're on the phone.
I found one that can isolate speakers, its just okay at that
This is exactly why I think the AI pins are a good idea. The Humane pin seems too big/too expensive/not quite there yet, but for exactly what you're doing, I would like some type of brooch.
how is that preferable to an airpod with siri or google or whatever app one has on a phone?
Wish I was good enough at holding thought, on my long thinking walks, sometimes even just a screen is enough to kill the thought I want to flesh out. Typically I just take a notepad but the pin thing would be better!
What do you use to run whisper locally? I don't think ollama can do it.
This has inspired me.

I do a lot of stargazing and have experimented with voice memos for recording my observations. The problem of course is later going back and listening to the voice memo and getting organized information out of what essentially turns into me rambling to myself.

I'm going to try to use whisper + AI to transcribe my voice memos into structured notes.

You can use it for everything. Just make sure that you have an input method set up on your computer and phone that allow you to use whisper.

That's how I'm writing this message to you.

Learning to use these speech-to-text systems will be a new kind of literacy.

I think pushing the transcription through language models is a fantastic way to deal with the complexity and frankly, disorganization of directly going from speech to text.

By doing this we can all basically type at 150-200 words a minute.

> That's how I'm writing this message to you.

Neat. Can you explain your setup a little? How do you go from voice to whisper to writing in this reply input form on a webpage?

It would be really amazing if you could expand your workflow a little, especially how you have stitched everything together.
I'm a noob and not a dev either. Can you please explain how to set this all up?

If it matters, I managed to install Jan AI on my Linux Mint and able to use the Mistral model. I use an Android phone if it helps. Thanks.

(comment deleted)
For people on macOS, the free app Aiko on the App Store makes it easy to use Whisper, if you want a GUI: https://sindresorhus.com/aiko
My faith in humanity has notched up a little today after seeing how easy Aiko and Oolama make it to use this incredible tech.

As a daily user of all the latest OpenAI/Claude models for the last few years, I'm amazed at how good llama3.1 is - all running locally, privately, with zero connection to the web. How did I not know about this?!

Can you give any more details on your setup? This sounds great