[fitzn]

What open source model are you using when you hit groq?

I just benchmarked some perf for some of my larger context window queries last week and groq's API took 1.6 seconds versus 1.8 to 2.2 for OpenAI GPT-3.5-turbo. So, it wasn't much faster. I almost emailed their support to see if I was doing something wrong. Would love to hear any details about your workload or the complexity of your queries.

[YetAnotherNick]

It's not a lot more faster for input but it is something like 10x faster for output(mixtral vs gpt-3.5). This could enable completely new mode of interaction with LLMs e.g. agents.

In most of the cases, overall response time is mostly dominated by output as it is ~100x slower per token than input.

[bee_rider]

What context did I miss that implies they are using an open source model?

[vineyardmike]

If you go to GroqChat (which is like a demo app), they offer Gemma, Mistral, and LLaMa. These are all open-weights models.

[laserbeam]

> 1.6 vs 1.8-2.2 seconds

I believe certain companies would kill for 20% performance improvements on their main product.

[gpapilion]

I have lots of questions about how important latency is since you may be replacing many minutes or hours of a person’s time with undoubtedly a quicker response by any measure. This seems like a knee jerk reaction assuming latency is as important as it’s been with advertising.

I’m not convinced latency matters as much as groqs material tries to claim it does.

[verdverm]

Google won search in large part because of their latency. I stopped using local models because of latency. I switched from OpenAI to VertexAI because of latency (and availability)

[w-ll]

When has latency ever not mattered?

Let alone 'chat' use cases, but holding a reponse up for N*1.2 longer than it could holds all sorts of other resources up/down stream.

[ben_w]

When it's already faster than I can absorb the response, which for me as an organic brain includes the normal token generation rate of the free tier of ChatGPT.

If I was using them to process far more text, e.g. summarise long documents, or if I was using it as an inline editing assistant, then I'd care more about the speed.

[qeternity]

> When it's already faster than I can absorb the response

Streaming a response from a chatbot is only one use-case of LLMs.

I would argue the most interesting applications do not fall into this category.

[ben_w]

Number of different use cases (categories) I'd agree; I'm not so sure about use (volume)…

…not yet anyway. Fast moving area, lots of blue water outside the chat interface.

[boroboro4]

Name one use case where there is a difference between latency of 200 t/s (fireworks.ai mixtral model) and 500 t/s (groq mixtral)? Not throughput and not time to first token, but latency.

Groq model shines at latency, not at the other two.

[michaelt]

Depends on your application.

For example, if you're a game company and you want to use LLMs so your players can converse with nonplayer characters in natural language, replacing a multiple-choice conversation tree - you'd want that to be low latency, and you'd want it to be cheap.

[beepbooptheory]

But are people really going to do this? The cost here seems prohibitive unless you're doing a subscription type game (and even then I'm not sure). And the kinds of games that benefit from open ended dialogue attract players who just want to pay an upfront cost and have an adventure.

(All the sudden having nightmares of getting billed for the conversations I have in the single player game I happen to be enjoying...)

If there is a future with this idea, its gotta be just shipping the LLM with game right?

[Const-me]

> If there is a future with this idea, its gotta be just shipping the LLM with game right?

That might be a nice application for this library of mine: https://github.com/Const-me/Cgml/

That’s an open source Mistral ML model implementation which runs on GPUs (all of them, not just nVidia), takes 4.5GB on disk, uses under 6GB of VRAM, and optimized for interactive single-user use case. Probably fast enough for that application.

You wouldn’t want in-game dialogues with the original model though. Game developers would need to finetune, retrain and/or do something else with these weights and/or my implementation.

[sdwr]

It would probably look like an InfiniteCraft-style model, where conversation possibilities are saved, and new dialogue is computed as needed.

Small, bounded conversations, with problematic lines trimmed over time, striking a balance between possibility and self-contradiction.

I could see it working really well in a Mass Effect-type game.

[michaelt]

I understand there are games using LLMs for NPC dialog, yes [1]

> If there is a future with this idea, its gotta be just shipping the LLM with game right?

Depends how high you can let your GPU requirements get :)

[1] https://www.youtube.com/watch?v=Kw51fkRiKZU

[beepbooptheory]

FWIW to confused others, trying to extract something from that video, it looks like this game [1] is using this stuff. Based solely on the reviews and the game play videos (while definitely acknowledging its technically in development status), it kinda looks like long term profitability is the least of their concerns here...

EDIT: Watching the videos, I am more and more confused by why this is even desirable. The complexity of dialogue in a game, it seems, needs to match the complexity of the more general possibilities and actions you can undertake in the game itself. Without that, it all just feels like you are in a little chatbot sandbox within the game, even if the dialogue is perfectly "in character." It all seems to feel less immersive with the LLMs.

1. https://store.steampowered.com/app/2240920/Vaudeville/

[frozenport]

I guess its tool calling? When you chain the LLMs together?

[robrenaud]

Model quality matters a ton too. They aren't serving OpenAI or Anthropic models, which are state of the art.

[verdverm]

Research suggest most answers and use cases do not require the largest, most sophisticated models. When you start building more complex systems, the overall time increases from chaining and you can pick different models for different points

[metadat]

"kill", .. why would anyone kill for a fraction of a second in this case? Informed folks know that LLM hosters aren't raking in the big bucks.

They're selling dreams and aspirations, and those are what's driving the funding.

[vineyardmike]

Google has used LMs in search for years (just not trendy LLMs), and search is famously optimized to the millisecond. Visa uses LMs to perform fraud detection every time someone makes a transaction, which is also quite latency sensitive. I'm guessing "informed folks" aren't so informed about the broader market.

OpenAI and Anthropic's APIs are obviously not latency-driven. Same with comparable LLM API resellers like Azure. Most people are likely not expecting tight latency SLOs there. That said, chat experiences (esp. voice ones) would probably be even more valuable if they could react in "human time" instead of with few seconds delay.

Integrating specialized hardware that can shave inference to fractions of a second seems like something that could be useful in a variety of latency-sensitive opportunities. Especially if this allows larger language models to be used where traditionally they were too slow.

[metadat]

I wish things were so simple!

Reducing latency doesn't automatically translate to winning the market or even increased revenue. There are tons of other variables such as functionality, marketing, back-office sales deals and partnerships. Lots of times, users can't even tell which service is objectively better (even though you and I have the know how and tools to measure and better know reality).

Unfortunately the technical angle is only one piece of the puzzle.

[cyanydeez]

What is the killer app product of a LLM Play ATM that's not a lossleader?