cross-posted from: https://piefed.social/c/linux@programming.dev/p/2509642/linus-torvalds-posts-his-first-ai-assisted-patch-to-the-linux-kernel-synthetic-test-clo
cross-posted from: https://piefed.social/c/linux@programming.dev/p/2509642/linus-torvalds-posts-his-first-ai-assisted-patch-to-the-linux-kernel-synthetic-test-clo
You are mistaken. Some AI models do indeed have memories.
They save some strings in a document to later read and “remember”. Calling them memories is a reach.
My point stands that they can learn from feedback. Your memories are only information in a virtual document, too. Just as susceptible to corruption, or misinterpretation.
Does an Alzheimer’s patient learn things just because they read the same note every morning explaining the reality to them?
No, their file system is corrupt. No reliable read or write access.
The AI is no different from the girl in “50 First Dates (2004)”. It has to go back and read the post-it notes every time you start a new session. It does not retain memory without writing shit down to read later.
You’re not wrong. It’s an inferior equivalent, but would you not say she learns with each note she makes, hampered though she is?
Yes I would say that she temporarily learns something while reading the notes. And she stores it in short-term memory.
The difference with AI though is that it has to re-read the same note again and again after each question you ask it. It does not keep things in “memory” from your previous conversations. It reads it again before answering you.
That is why context length matters and is important in LLMs. There’s a huge difference between talking to an AI with 10k in context length and 1mil in context length. You can try it yourself using Llama.cpp and running LLMs locally.
Our short term memory is our context window. Some people have larger capacity for information, others very little. Same thing, different terminology.
I use Ollama daily. The context limit of 10k is indeed very limiting and noticeable compared to the extended lengths of cloud models. I’ve had to recruit a model specifically trained on condensing salient information to perform context compression frequently. I also use the limited local context to establish effective strategies for maximizing token efficiency in the agent.md.
It’s no different from how many tasks a person can simultaneously perform, or disparate words they can hold concurrently. Eventually thoughts are forgotten. What exactly is determined by the structure itself - the context-pruning strategy inherent. We don’t know what our own are, but I believe we have an equivalent.
All I’m saying, in the end, is that anything that can occur in a physical entity can be emulated in a digital entity. Are we capable of doing that flawlessly yet? No, but it is looking inevitable to me. Neural networks are ubiquitous, and our ability to simulate them will only improve.
There is no concept of a filesystem in our brains. Weird hills, and such and so forth
Only because you don’t think of it as one. It’s just a different medium for the storage of information.