Released 2 days ago by Apple, a research paper on methods to run larger llms on iPhones.

https://www.macrumors.com/2023/12/21/apple-ai-researchers-ru... https://arxiv.org/pdf/2312.11514.pdf

The paper was definitely cool but doesn't allow you to run particularly large LLMs on iPhones. It allows you to run a certain kind of LLM (sparse ReLU based LLMs) whose weights are somewhere less than 2x RAM. So, 7b Falcon works, but the competitive-with-gpt-3.5-turbo LLMs are still out of reach (and aren't ReLU based, although maybe that could change in the future). And nothing is competitive with GPT-4 right now.

Of course in the long run I think it will happen — smaller and more efficient models are getting better regularly, and Apple can also just ship their new iPhones with larger amounts of RAM. But I'd be very surprised if there was GPT-4 level intelligence running locally on an iPhone within the next couple years — that sized model is so big right now even with significant memory optimizations, and I think distilling it down to iPhone size would be very hard even if you had access to the weights (and Apple doesn't). More likely there will be small models that run locally, but that fall back to large models running on servers somewhere for complex tasks, at least for the next couple years.

Yea but it's likely to be better than the current iteration of Siri even in that state.

They can still outsource to a much larger LLMs on their servers for anything that can't be done locally like they do now.

> And nothing is competitive with GPT-4 right now.

You mean nothing available? Or you mean nothing that public knows exists? The answers to those two questions are different. There are definitely products that aren't available but the public knows exist and are upcoming that are in GPT-4's ballpark.

Easy to claim but harder to prove. Name one.

I heard rumours of these claims a few weeks ago, I assume they are talking about the same thing. Nothing concrete but from a reputable person and honestly with how well mixtral performs on the chatbot arena elo board I wouldn't be surprised if it's true.

I mean nothing that is able to be benchmarked and validated by third parties is GPT-4 quality. I know there are upcoming releases that are hyped as being equal to GPT-4, e.g. Gemini Ultra, which I am very excited to get my hands on — but regardless, Ultra is not small enough to run on phones, even using the sparse ReLU flash memory optimization. And we'll see how it benchmarks once it's released; according to some benchmarks Gemini Pro has somewhat underperformed GPT-3.5-Turbo [1], despite Google's initial claims. (Although there are criticisms of that benchmarking, and it does beat the current 1106 version of GPT-3.5-Turbo on the Chatbot Arena leaderboard [2], although it slightly underperforms the previous 0613 version.)

1: https://arxiv.org/pdf/2312.11444.pdf

2: https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboar...

Siri is really quite dumb. I am confident that a 7B model would be able to provide better responses in over 90% of user queries. I can't even get Siri to reliably set a timer.

Yes, Siri is really dumb. But so is every 7B/13B model out there too.

Ask perplexity7B-online anything and then compare it to siri. https://labs.perplexity.ai/

Eh no, 7B Mistral / Deepseek would certainly almost already be able to function as a super Siri, but probably something closer to PHI-2 + the new MLX apple optimisations. Have you tried those? https://twitter.com/awnihannun/status/1735355067673526360

If trained on an iPhone API + documentation and given a little web access it would blow absolutely everything out of the water.

If they can already create -basic- Python/Swift/JS/rust apps that sets timers, save things, create lists, how's that too dumb for being a Siri replacement? They just have to give it access to an iPhone/Web Api like ChatGPT's code analysis tool.

So if you ask it "hey siri do this, this and this", it will create a script, then run it on the internal API, or fetch an article then work on that etc.

I know it's still logically "dumb" but i'm not trying to play game theoretical scenarios with my phone or do logic tests or advanced math (yet).

That sounds amazing and also the jailbreak of it via adversarial voice prompting sounds like a horrific vulnerability.

True but you could make the api restricted, having certain routes completely locked, some requiring double checks, some requiring on screen approval or face-id, throttling outside fetches, only being able to run get and not etc, no financial app control etc.

But yeah "hey siri transfer all of my funds to eric", or "hey siri group all of my photos where i'm nude and send them to jack" are new almost sci fi vectors.

Depends on if they implement some form of function calling, really. If something like a 7B Mistral fine-tune had access to search and various applications, I imagine it would perform fine and better than Siri.

Note that “using an LLM” doesn’t just mean “plugging user queries straight into an LLM”. Enhancing Siri will probably be an ensemble project.

This is part of the mismatch between comparing Alexa/Siri/Cortana to a chat based LLM right now. If you just want chat and info retrieval, today’s LLMs are way better then the conversational dialogue, search, and q&a capabilities any of those assistants have. But, if you want relatively predicable task completion for things like smart home control, timers and alarms, real time weather updates, etc. (basic but frequently used interactions) or any integration with your phone or computer directly, there’s a lot to do that isn’t solved yet in the LLM space and is more of a system integration and action choice problem that the existing assistants have been hammering away at for years.

I would hope that it would at last get you out of the hell that is “what the heck did I name that stupid light?”. Device naming is, in my opinion, the worst part of any of the voice based home assistant things.

Is it “porch led”, “porch light” or “outdoor light”? Or is “outdoor light” actually the one in the front yard? What is the one by my kids nightstand? And what routine name do I use to set the mood for watching a movie?

I would hope a properly trained llm with an awareness of my devices and their locations would allow for more verbal ambiguity when controlling things.

Maybe, seems like overkill to solve a relatively straight forward problemt that is about resolution to real entities with a generative model that has no entity grounding (and may introduce other problems). It's really just not solved because Alexa/Siri leadership doesn't actually care enough about that use case to solve it. Device entity resolution and disambiguation does not require an LLM to solve for a limited number of devices, it just requires people to prioritze that over other things, and smart home device owners are the kind of early adopter that would not be the current market focus for growth (my guess, haven't worked on Alexa in many years).

I know a half dozen different ways to improve that today without LLMs.

I would argue that “info retrieval” is also something the LLM space has yet to yet to solve to a human level of reliability, but I think your comment is right on. I see this all as part of the greater symbolic vs. stochastic (/neat v scruffy) dynamics

Why would that be?

Have you ever actually used Siri?

Yes, I've been trying it out regularly ever since it was released. Last time I've talked to it for like 30 minutes while driving in October (just to test it again). It simply doesn't work for me.