General models are useful to the general public, and they're useful as a starting point to find tune to a smaller model. The purpose matters. Research and most products benefit heavily from having a much smaller model (speed) to do one thing very well on trillions of documents. General chatting from big models is mostly a gimmick, but they're still very useful to find tune.

The thing is that those super large models with CoT are often much more reliable than specialized fine-tuned smaller models you'd like to use on trillions of documents, as I am observing right now. You are not trading just speed but also quality with smaller models. The only relevant use case for fine-tuning for me recently was to apply wizard dataset to LLaMA 2 for a less censored model and that was done using LoRA.

What tasks?

For processing trillion documents for example NER can be done much better.

You are literally using trillion documents? Or are you exaggerating?

chaxor above mentioned it so I quickly recalled a task I saw a super large LLM demolishing fine-tuned models on documents.

It's really depressing that a handful of big corporations will be able to exert such control over labor and productivity

That's why the LLaMA 2 release is so significant. You can run the full 70B in 8-bit on two prosumer A6000 Ampere (cost around $10k together) which is within the reach of most companies and some devs. This could further accelerate all research to make it both efficient and available even to regular folks.

But it's still not comparable to GPT 4, nor will it likely be for some time at least. And by the time we have GPT 4 class open source models, I'd imagine there'd be significant advancements in closed source models, such as inference time symbolic reasoning using MCTS, something google is working on for gemini...or just bigger/better architectures, data etc

Yeah, I don't really see a solution. Even Stanford can't keep up with the latest AI research. Still, having LLaMA 2 is better than not having it.

This tradeoff is ridiculous, even if it is "better" by .01% F score. I would much rather have a dataset created in 1 day from BERT at 98% F-score than 1000 years at 98.01% F-score from a 540B parameter model, or even a 33B parameter model. The performance in million parameter models for NER is still excellent, and works at speed that are usable. Running things through OpenAI is also useless, as it would cost a few million $.

It's more like 100% accuracy vs 95% accuracy, and the super large models are now able to extract non-trivial derived info from a regular human speech as well. While cost-wise it's not efficient right now, this will change over time (you skate to where puck will be, not where it is now), making the current fine-tuning way obsolete. Academically I am not thrilled as I built my research on fine-tuning, but as a producer of a product this solves so many issues at the same time, making me pretty happy.

I’d like to generate Minecraft fan fic. GPT4 does a terrible job because it’s been lobotomized. Fine tuning a model with a corpus of fan fic generates better fan fic. If I could fine tune GPT4 then ok I agree. But “general purpose lobotomized oracle” is only one use case for a language model, albeit a useful one, is the least creative one.

OK, understood. You might do that with uncensored versions of LLaMA 2 now. Role playing/fan fic is much better with those.

They tend to not have the right base of training materials - the models are constrained not just by the lobotomization but by some specialized writing not being part of the corpus. Training a lora on a specific corpus - fan fic, alt.sex.stories, or a persons specific writing, can help induce behavior or language choices specific to that corpus. You can also use context injection but I’ve found combining a model weight adjustment with priming the context to be the most effective.

I think over time it’ll become standard practice to train special purpose models on top of general purpose models. If I have a very specific task domain a model tuned to that domain will less often wander off and will be more likely to respond in the desired way.