This is not how LLMs work, they are not a database nor do they have access to one.
Please do explain how you think they make LLMs without a database of training examples to build a statistical model from.
The llm itself is just what it learned from reading all the training data,
I.e. “a model that encodes a database”.
They are a trained neural net with a set of weights on matrices that we don’t fully understand.
I.e., “we applied a very lossy compression algorithm to this database”.
We do know that it can’t possibly have all the information from its training set since the training sets (measured in tb or pb) are orders of magnitude bigger than the models (measured in gb).
Check out the demoscene sometime, you’ll be surprised how much complexity can be generated from a very small set of instructions. I’ve seen entire first person shooter video games less than 100kb in size that algorithmically generate gigabytes of texture data at runtime. The idea that a mere 1,000x non-lossless compression of text would be impossible is laughable, especially when lossless text compression using neural network techniques achieved a 250x compression ratio years ago.
Not_mikey@lemmy.world 1 year ago
If LLMs were just lossy encodings of their database they wouldn’t be able to answer any questions outside of there training set. They can though, and quite well as shown by the fact you can give it completely made up information that it can’t possibly have “seen” and it will go along with it and give plausible answers. That is where it’s intelligence lyes and what separates it from older chatbots like Siri that cannot infer and are bound by the database they pull from.
How do you explain the hallucinations if the llm is just a complex lookup engine? You can’t lookup something you’ve never seen.
knightly@pawb.social 1 year ago
Of course they could, in the same way that hitting the autocomplete key can finish a half-completed sentence you’ve never written before.
The fact that models can produce useful outputs from novel inputs is the whole reason why we build models. Your argument is functionally equivalent to the claim that wind tunnels are intelligent because they can characterise the aerodynamics of both old and new kinds of planes.
For the same reason that a random number generator is capable of producing never-before-seen strings of digits. LLM inference engines have a property called “temperature” that governs how much randomness is injected into their responses:
Image
Not_mikey@lemmy.world 1 year ago
Auto complete is not a lossy encoding of a database either, it’s a product of a dataset, just like you are a product of your experiences, but it is not wholly representative of that dataset.
A wind tunnel is not intelligent because it doesn’t answer questions or process knowledge/data it just creates data. A wind tunnel will not answer the question “is this aerodynamic” but you can observe a wind tunnel and use your intelligence to process that and answer the question.
Temperature and randomness don’t explain hallucinations, they are a product of inference. If you turned the temperature down to 0 and asked it the question " what happened in the great Christmas fire of 1934" it will give it’s best guess of what happened then even though that question is not in it’s dataset and it can’t look up the answer. The temperature would just mean that between runs it would consistently give the same story, the one that is most statistically probable, as opposed to another one that may be less probable but was pushed up due to randomness. Hallucinations are a product of inference, of taking something at face value then trying to explain it. People will do this too, if you tell someone a lie confidently then ask them about it they will use there intelligence to rationalize a story about what happened.
knightly@pawb.social 1 year ago
If LLMs don’t encode their training data, then why are they proving susceptible to data exfiltration techniques where they output the content of their training dataset verbatim? m.youtube.com/watch?v=L_1plTXF-FE