Comment on ELI5. Limit of current gen AI/LLMs
jacksilver@lemmy.world 1 day ago
I think you may be mixing a couple of things together, but I’ll take a crack at this.
When you get an Ai generated response from a search engine, this is usually a modified RAG (retrieval augmented generation) approach. How this works is that the content from web pages are already pre-processed into embeddings (numerical representations of the text). When you perform a search, your search text is turned into an embedding and compared (numerical similarity) to the websites to get the most related content for your search. That means that the LLM only parses and processes a very small subset of the returned websites to generate its response.
Another element you might be asking about is how can these agentic AI systems handle larger tasks (things like OpenClaw). That is a bit more complicated and dependent on the systems design, but basically boils down to two things. The first is the “reasoning models” first break concepts into smaller tasks meaning the LLM only has to worry about a subset of a larger task. Secondly, a lot of these systems will periodically merge all past context into a compressed state that the LLM can handle (basically summaries of summaries) or add them to a database for future/faster reference.
At the end of the day, your understanding of the limits of LLM are correct, all the progress we’ve really seen with LLMs (over the past couple of years) has been the creation of systems to work around their limitations. The base technology isn’t getting much better, but the support around it is.
vaderaj@lemmy.world 1 day ago
Thanks.
And to clarify, other than the corporate greed is there any actual use case for the work around their limitations? I mean if the building materials aren’t strong enough there is only so much you can achieve with a beautiful paint job (my current understanding, and I may be wrong)
jacksilver@lemmy.world 1 day ago
The underlying issues, in my opinion, regarding LLMs is their indeterministic nature. Even zeroing out the temperature (randomness of outputs), you can get significantly different results between two almost identical texts.
However, building out an ecosystem supporting new technology is a fairly common progression. If you compare it to the internet things like browser caches, CDNs (content delivery networks), code minifiers, etc. are all ways to help combat latency (a fundamental problem for the internet).
As for the effectiveness of these solutions, RAGs do help a lot when generating text against a select corpus. Its what allows the linked sources in things like ChatGPT and Googles AI results. It’s also what a lot of companies are using for searching their support pages/etc. It’s maybe not quite as good as speaking to a person, but is faster.
Similarly, the reasoning models and managing the models “context” both have shown demonstrable improvements for models in benchmarking.
I’m not sure I personally believe this makes LLMs a replacement for humans in most situations, but it at least demonstrates forward progress for GenAI.
vaderaj@lemmy.world 1 day ago
Interesting, the thing is I can quite easily pick up something new but at the same time I am very resistant to change until there is good reasoning and some sort of a scientific conformation.
Need to discover good uses cases for LLM/AI and make peace with it I guess!
Poik@pawb.social 20 hours ago
The best uses I’ve seen are blind person aides. Scene understanding and OCR for disability aides. The OCR doesn’t have to be LLMs, but a system that combines the two effectively is useful.
There is merit to sitting an LLM in front of an expert system to act as an intermediate, but the LLM shouldn’t be doing any “thinking.” It should only translate results.
jacksilver@lemmy.world 1 day ago
Yeah, that’s fair. I haven’t jumped into the whole agentic side of things as I find LLMs consistently fail at lower level stuff.
Everyone says it’s great at prototyping or writing documents, etc, but I think that’s just cause people have low standards. When coding I find that it quickly messes things up or lacks good quality control (which you only notice if you’re familiar with the domain). For writing it’s fine, but the tone and language always feels off and certainly doesn’t sound like me.
Either way, I would suggest playing around with them to see how they fit into how you do things. I think we’re starting to see things finally slow down on new implementations, and they aren’t going away, so it may be a good time to see if all the fuss is worth it to you.