It makes me wonder how long it will be until open-source locally-run AI chatbots reach the GPT-4 level. Five years? Ten?

its more a question of hardware progress than software probably. theres a minimum level for these behaviors to emerge

It depends on the tasks, I’d argue some of the open source alternatives are at the gpt-4 level already, particularly for code generation.

That said, I suspect summarization, translation, etc will take time. I’d suspect under 1 year.

>I’d argue some of the open source alternatives are at the gpt-4 level already, particularly for code generation.

Like which one?

Large context Chronos 33b (and some mixes) for roleplaying type chat.

And there are some very new 65b finetunes I have not tried.

The real limiter is having enough GPU RAM to train and run a usefully large model. Everything else is just fiddly details.

Not anymore. Llama.cpp and kobold.cpp run well with low vram and non Nvidia GPUs.

Huggingface is full of finetunes now, and I believe a 33b model can be finetuned on a single 3090.

Llama.cpp is developing some kind of training, but I have no idea what the requirements will be.

To me, GPT's real "secret sauce" is how it primes the model for interactive prompting instead of just text completion.

There are "chat" and "instruct" variants of most models now.

The open source bot I played with last week (Wizard Vicuna uncensored) was 85% of gpt 3.5 on a VERY hard use case (fiction stories).

Maybe a year before we are the level of stablediffusion?

> It makes me wonder how long it will be until open-source locally-run AI chatbots reach the GPT-4 level. Five years? Ten?

A lot shorter than you'd think: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderb...

Latest falcon-40b-instruct is scoring higher than GPT-3 on HellaSwag benchmark already, and not far behind GPT-4: https://paperswithcode.com/sota/sentence-completion-on-hella...

Big question is what counts as "locally run". You still need about $5k worth of GPU hardware to run it.

When it happens, I think it’ll happen on our phones first.

Given the fundamental hardware matrix operations, that seems unlikely, unless it's something very scaled down.

It doesn't seem to be a major barrier for most (it absolutely is for me). Is there enough of a want/market for tools that integrate with common OSS LLM APIs like oobabooga/Kobold/Novel/vLLM etc? One idea I've been toying with is a BYO model with support for these APIs, eg for IDE integration, brain storming etc. Seems like a good approach to me but how many people would actually bother standing up these LLM engines with APIs to use it?

Well, it is a round-up of their venture investments in the space ;-)

If you're looking to self-host chat memory rather than go all in on Supabase, there's Zep: https://github.com/getzep/zep

Full disclosure: I'm a co-author.

Supabase is also self-hostable though.

Also worth noting that they aren’t investors in supabase (but we’re very appreciative that they included us in this stack)

I stand corrected. Sometimes my cynicism gets the best of me ;-)

> It's not host your own AI at all! You need like 5 saas api-keys to operate this app. But sure, you can host this UI on your own.

Yeah I've pretty much lost all interest in anything API based, the pricing + censorship is too much. And models like falcon-instruct are getting close to GPT3 levels for a hardware cost of only a few grand to fine tune/run inference.

I think we’re in platform reliance mode for quite a while. I think of all of these SaaS companies as different icons on an abstract AWS console that doesn’t exist yet.

We wouldn’t bat an eye at using S3, EC2, and RDS as a host your own setup. The only difference here is that startups are moving faster than incumbents.

FWIW that’s one reason why Steamship (disclaimer: I’m the founder) aggregates all AI services under a single API key and interface. It’s to deal with the insane glue-code hassle of running this stuff on your own.

That’s fine but the title is wrong

And using S3, RDS and EC2 I would argue is hosting your own as you are using foundational components. Same if you run LLaMa on a rented A100.

Not the same at this point. Setting up properly secure AWS stuff is arguably more demanding than setting up your own box at home.