tshadley

> LLMs cannot offer that promise by design, so it remains your job to find and fix any deviations from the abstraction you intended.

LLMs are clumsy interns now, very leaky. But we know human experts can be leak-proof. Why can't LLMs get there, too, better at coding, understanding your intentions, reviewing automatically for deviations, etc.?

Thought experiment: could you work well with a team of human experts just below your level? Then you should be able to work well with future LLMs.

Thank you, amazing, fresh.

The goal here is not to replace transformers but combine them with RNN so you get both good short-term memory (self-attention) and much improved long-term memory (ATLAS recurrent memory).

"Empirically, our models—OmegaNet, Atlas, DeepTransformers, and Dot—achieve consistent improvements over Transformers and recent RNN variants across diverse benchmarks."

> To understand the capabilities of LLMs, we evaluate GPT3 (text-davinci-003) [11], ChatGPT (GPT-3.5-turbo) [57] and GPT4 (gpt-4)

Oh dear, this is embarrassing. Anil Anathaswamy, are you aware a year in AI research now is like 10 years in every other field?

Seems to me o3 prices would be what the consumer pays, not what OpenAI pays. That would mean o3 could be more efficient in-house than paying subject-matter experts.

I always get the feeling he's subconsciously inserting a "magical" step here with reference to "synthesis"-- invoking a kind of subtle dualism where human intelligence is just different and mysteriously better than hardware intelligence.

Combining programs should be straightforward for DNNs, ordering, mixing, matching concepts by coordinates and arithmetic in learned high-dimensional embedded-space. Inference-time combination is harder since the model is working with tokens and has to keep coherence over a growing CoT with many twists, turns and dead-ends, but with enough passes can still do well.

The logical next step to improvement is test-time training on the growing CoT, using reinforcement-fine-tuning to compress and organize the chain-of-thought into parameter-space--if we can come up with loss functions for "little progress, a lot of progress, no progress". Then more inference-time with a better understanding of the problem, rinse and repeat.

> One random example to illustrate the distinction: training gaps can easily decrease uncertainty. You have lots of mammals in your training data, and none of them lay eggs. You ask "The duck-billed platypus is my favorite mammal! Does it lay eggs?" Your model will be very confident when it responds "No". That is a high-confidence error.

This article did not seem to make the mistake of associating hallucination with bad data so hard to see exactly how this is relevant. I mean, you could write an article "AI Error: how to reduce it" and frame it entirely in user's perceptions but I wouldn't make a peep.

My objection is that it is silly to use the word "hallucination" (which suggests insanity/psychosis) and then address it as if LLMs are marginally insane and the solution is straight-jacket-like heuristics, when "uncertainty" (which suggests uncertainty) is a far more accurate description of behavior pointing to a far more productive and focused solution.

> ...proving that this one particular piece of the hallucination problem may be conceptually simple.

Everything mentioned in the article boils down to that one particular piece-- non-detected uncertainty. The architecture constraints referenced are all situations that cause uncertainty. Training data gaps of course increase uncertainty.

Their solutions are a shotgun blast of heuristics that all focus on reducing uncertainty-- CoT, RAG, fine-tuning, fact-checking -- while somehow avoiding actually measuring uncertainty and using that to eliminate hallucinations!

"Why PCIe Risers suck and the importance of using SAS Device Adapters, Redrivers, and Retimers for error-free PCIe connections."

I'm a believer! Can't wait to hear more about this.