3 ms·

LLMS are not AI. Large language models have no formal reasoning, they have no long term recall, they contain no structured logic. I’m not saying it can’t be so

LLMS are not AI. Large language models have no formal reasoning, they have no long term recall, they contain no structured logic. I’m not saying it can’t be solved. I’m saying it can’t be solved INSIDE the LLM. Anyone with a phd in machine learning would probably agree.

ben_w 2y ago
> LLMS are not AI. Ironic demonstration that humans also do what is deemed "hallucinations" when AI do it: https://en.wikipedia.org/wiki/AI_effect https://en.wikipedia.org/wiki/AI_effect LLMs, transformer models in particular, are artificial neural networks. They have always been AI. AI is the field which led to this, the research is published as AI research. It's amazing how often us humans (me included!) don't use RAG (retrieval augmented generation) in the form of a search engine and just trust our gut instinct for off-the-cuff responses :D > Large language models have no formal reasoning, they have no long term recall, they contain no structured logic. > I’m not saying it can’t be solved. I’m saying it can’t be solved INSIDE the LLM. Do you mean transformers then? Because that is the current vogue architecture for large language models which is clearly a broader category. The full details for the current best models are secret, but they're still large language models, and they're demonstrating surprisingly high performance on logic and reasoning.
- handoflixue 2y ago
  For all that LLMs produce good performance, it's still just predictive texting. If you can get the hallucination rate on that down to 0% without using anything else, I'll be extremely shocked. Now, layer a few sanity checks on top of an LLM, especially some clever thing we haven't invented yet, and I'll totally believe it - the task is absolutely doable, I'd just find it really weird if a predictive engine could do it 100% accurately, using only modern resources.
  - ben_w 2y ago
    Conditional on it still sounding like you mean Transformers rather than LLMs, we're pretty close to agreement. But even Transformers really are not just preditive text. IIRC the original Google usage, Attention is All You Need era, was for translation; and while I would indeed characterise the first few OpenAI/GPT models as "autocomplete on steroids", that changed with InstructGPT, which was the first time I saw them transforming requests into actions, in the form of creating a very simple web game. > I'd just find it really weird if a predictive engine could do it 100% accurately, using only modern resources. I currently think there is no such thing as "knowledge" in reality, that such a state is as unrealisable as counting to infinity, that all we can really have are beliefs of varying certainty; in this regard, 100% can never happen in any system including humans — but also, I wouldn't say an AI is "hallucinating" if the error rate was similar to that of a human. Likewise, I find it really weird how a neural network with the complexity of a mid-sized rodent is able to transform prompts in the most used languages into mostly-correct source code in most programming languages — this is not a thing I would have expected, given the observable lack of employment opportunities for rodents* in software engineering departments. I could be wrong about both, of course. * other than furries, who are everywhere ;)

ben_w 2y ago

> LLMS are not AI. Ironic demonstration that humans also do what is deemed "hallucinations" when AI do it: https://en.wikipedia.org/wiki/AI_effect https://en.wikipedia.org/wiki/AI_effect LLMs, transformer models in particular, are artificial neural networks. They have always been AI. AI is the field which led to this, the research is published as AI research. It's amazing how often us humans (me included!) don't use RAG (retrieval augmented generation) in the form of a search engine and just trust our gut instinct for off-the-cuff responses :D > Large language models have no formal reasoning, they have no long term recall, they contain no structured logic. > I’m not saying it can’t be solved. I’m saying it can’t be solved INSIDE the LLM. Do you mean transformers then? Because that is the current vogue architecture for large language models which is clearly a broader category. The full details for the current best models are secret, but they're still large language models, and they're demonstrating surprisingly high performance on logic and reasoning.

handoflixue 2y ago
For all that LLMs produce good performance, it's still just predictive texting. If you can get the hallucination rate on that down to 0% without using anything else, I'll be extremely shocked. Now, layer a few sanity checks on top of an LLM, especially some clever thing we haven't invented yet, and I'll totally believe it - the task is absolutely doable, I'd just find it really weird if a predictive engine could do it 100% accurately, using only modern resources.
- ben_w 2y ago
  Conditional on it still sounding like you mean Transformers rather than LLMs, we're pretty close to agreement. But even Transformers really are not just preditive text. IIRC the original Google usage, Attention is All You Need era, was for translation; and while I would indeed characterise the first few OpenAI/GPT models as "autocomplete on steroids", that changed with InstructGPT, which was the first time I saw them transforming requests into actions, in the form of creating a very simple web game. > I'd just find it really weird if a predictive engine could do it 100% accurately, using only modern resources. I currently think there is no such thing as "knowledge" in reality, that such a state is as unrealisable as counting to infinity, that all we can really have are beliefs of varying certainty; in this regard, 100% can never happen in any system including humans — but also, I wouldn't say an AI is "hallucinating" if the error rate was similar to that of a human. Likewise, I find it really weird how a neural network with the complexity of a mid-sized rodent is able to transform prompts in the most used languages into mostly-correct source code in most programming languages — this is not a thing I would have expected, given the observable lack of employment opportunities for rodents* in software engineering departments. I could be wrong about both, of course. * other than furries, who are everywhere ;)