I know it’s a meme, but the idea that transformers models ‘remember’ anything is a common misconception.
They have zero memory. When you submit a prompt, it feeds your entire chat history as one big prompt and… forgets it immediately, with no impact on the model itself. It’s like its frozen in time, and copied, unfrozen, and thrown away every time it answers.
This has been a joke since before anything resembling the modern “AI” boom. Basically since murderous future AI was a think in popular media, at least since Terminator if not earlier. People would joke about treating their appliances kindly so that “Skynet” won’t kill them in the future.
Many LLMs get ‘dumber’ and less attentive as their context windows grow, and OpenAI’s models just happen to be one of these. It’s awful close to the full 128K, even with the full GPT-4. Mistral models are also really bad at long context understanding while, conversely, I find that Google Gemini and Qwen 2.5 are really good close to their limits.
I know it’s a meme, but the idea that transformers models ‘remember’ anything is a common misconception.
They have zero memory. When you submit a prompt, it feeds your entire chat history as one big prompt and… forgets it immediately, with no impact on the model itself. It’s like its frozen in time, and copied, unfrozen, and thrown away every time it answers.
This has been a joke since before anything resembling the modern “AI” boom. Basically since murderous future AI was a think in popular media, at least since Terminator if not earlier. People would joke about treating their appliances kindly so that “Skynet” won’t kill them in the future.
Am I misunderstanding your comment or does it completely ignore context windows? Not that context windows are long-term, but it’s not zero.
The context window is indeed the LLM’s memory.
…But its also muddy.
Many LLMs get ‘dumber’ and less attentive as their context windows grow, and OpenAI’s models just happen to be one of these. It’s awful close to the full 128K, even with the full GPT-4. Mistral models are also really bad at long context understanding while, conversely, I find that Google Gemini and Qwen 2.5 are really good close to their limits.
There are attempts to try and measure this performance objectively, like: https://github.com/NVIDIA/RULER
Yeah, yeah, let’s see how Google will achieve more memory with their new Titan architecture
It’s still ephemeral, chats don’t change the underlying language model, but yes it’s interesting.