When German journalist Martin Bernklautyped his name and location into Microsoft’s Copilot to see how his articles would be picked up by the chatbot, the answers horrified him. Copilot’s results asserted that Bernklau was an escapee from a psychiatric institution, a convicted child abuser, and a conman preying on widowers. For years, Bernklau had served as a courts reporter and the AI chatbot had falsely blamed him for the crimes whose trials he had covered.
The accusations against Bernklau weren’t true, of course, and are examples of generative AI’s “hallucinations.” These are inaccurate or nonsensical responses to a prompt provided by the user, and they’re alarmingly common. Anyone attempting to use AI should always proceed with great caution, because information from such systems needs validation and verification by humans before it can be trusted.
But why did Copilot hallucinate these terrible and false accusations?
Hallucinations is a fancy word for being wrong.
The models are not wrong. The models are nothing but a statistical model that’s really good at predicting the next word that is likely to follow base on prior information given. It doesn’t have understanding of the context of the words, just that statistically they’re likely to follow. As such, all LLM outputs are correct to their design.
The users’ assumption/expectation of the output being factual is what is wrong. Hallucination is a fancy word in attempt make the users not feel as upset when the output passage doesn’t match their assumption/expectation.
So randomly spewing out bullshit is the actual design goal of AI models? Why does it exist at all?
They’re supposed to be good a transformation tasks. Language translation, create x in the style of y, replicate a pattern, etc. LLMs are outstandingly good at language transformer tasks.
Using an llm as a fact generating chatbot is actually a misuse. But they were trained on such a large dataset and have such a large number of parameters (175 billion!?) that they passably perform in that role… which is, at its core, to fill in a call+response pattern in a conversation.
At a fundamental level it will never ever generate factually correct answers 100% of the time. That it generates correct answers > 50% of the time is actually quite a marvel.
So good as a translator as long as accuracy doesn’t matter?
If memory serves, 175B parameters is for the GPT3 model, not even the 3.5 model that caught the world by surprise; and they have not disclosed parameter space for GPT4, 4o, and o1 yet. If memory also serves, 3 was primarily English, and had only a relatively small set of words (I think 50K or something to that effect) it was considering as next token candidates. Now that it is able to work in multiple languages and multi modal, the parameter space must be much much larger.
The amount of things it can do now is incredible, but our perceived incremental improvements on LLM will probably slow down (due to the pace fitting to the predicted lines in log space)… until the next big thing (neural nets > expert systems > deep learning > LLM > ???). Such an exciting time we’re in!
Edit: found it. Roughly 50K tokens for input output embedding, in GPT3. 3Blue1Brown has a really good explanation here for anyone interested: https://youtu.be/wjZofJX0v4M