LIMA, been + Premise Smashing

An Ai-Techno Orientation Editorial with Language Bending

If you look it up, you will see that LIMA represents “Less is More for Alignment.”

Here is a super and super technical look from May 2023.

“Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow specific response formats from only a handful of examples in the training data, including complex queries that range from planning trip itineraries to speculating about alternate history. Moreover, the model tends to generalize well to unseen tasks that did not appear in the training data”

They mean less people. They mean models on top of models.

Here is a further crunch of this META-sponsored study on Medium (neither rare nor well done.)

“Human feedback is not the key ingredient”

Fair enough. Yet, we are still dealing w prompt training and fine-tuning in the first place of LARGE LANGUAGE MODELS (LLMs).

Yes of course I get it this is the exciting game in town. Much promise. Much interest.

And yet there are serious serious privacy (h/t D. Casey Fletcher @ LexFusion) concerns.

There are other concerns.

• What is “in there”? We don’t know.

• Where do my docs/prompts/replies go? Prolly safe…???

• How much time on a cottage industry to “prompt train” and to train others to prompt train?

• How much time and effort on “fine tuning” to make sure that our results are ‘correct’?

• How much hallucination due to so much cloud of data and noise of cross prompts?

• Centralized control of AllDataForBenefitMankind (and an update yesterday where OAI chats back at Elon.)

Yet what if our ‘container’ (the db) has too much stuff in it in the first place?

What if big is the new too big, and small the new big?

DBs which are smaller can be less prone to hallucination, less in need to prompt gardening and pruning, less needful to be “trained” like a bear on a chain.

Lawyers mess with language. And we were first before the Internet even, and certainly before gptllama et al.

I prefer to think about LIMA [Less Is More in Alignment] to instead mean less is more in the database. SLMs Small Language Models, not LLMs. LIMA, been doing, not LLM.

Small, decentralized, clearly-defined dbs which do not require so much work overlay will transform for the positive (and sometimes negative) the impact of Ai on humanity. Parametered, garden walled, totally selective inputs.

For many use cases, and I am thinking most particularly in the Law, a dozen or thirty documents and/or sources is enough to provide an ongoing source of helpful, cogent, plain language, non-legalese outputs (i.e. clear answers for understanding) for the benefit of non-lawyer lay people. And of course this can be translated into so many fields; medicine, compliance, education, service, self-use cases…

I am also thinking here of the difference between a salty unpredictable open expanse of ocean with sharks, and comparing that to the pristine controlled parametered body of water that is a custom pool…

Final metaphor: Why are we spending all this time training to find the “right” needle in the giant haystack? Why not unfurl the hay unto bunches instead and put the needle right in the middle? We Texans like hay.

*Zeroing in bespoke and focus can be additional fun see Ritesh create a totally synthetic AI-Elon Musk (and ouch! even tho on gpt3…)