Scanning the brain only messures the brains signals Contexutally it does notwok the same, a brain thinking. A word. Would not output a responce that corralates to word. As we all learn and think differntly , evryones. Brains aare wired synaptically cellhlarly..the same cannot be said neurocognatively.
We are finally beginning to understand how LLMs work: No, they don't simply predict word after word
- Thread starter zohaibahd
- Start date
I think that the author(s) is looking at the problem fron wrong perspective created by our teaching system.
Example of "opposite for small" for a multilingual person would be to classify "small" in terms of thoughts, then map those (all or some) to an abstract space to find their opposition (ex. large space, elephant opposite to mouse, capital letter, long word in opposition to 5-letters shirt word, sound opposite to "sss", etc.). Then based on those criteria you build up a map of areas in the space of your thoughts to build one most likely choice for the result. It might be "infinity" for a mathematician or "luxury apartment" for a real estate seller. Then we can narrow down possible answers to match expectations and by generalisation we get the "big" or "large" or "wide" word. You cannot understand how a human brain is working but you are trying to create an equivalent and understand it and map it to a precise space of your own perspective. With such an approach you are sentenced to a failure.
Example of "opposite for small" for a multilingual person would be to classify "small" in terms of thoughts, then map those (all or some) to an abstract space to find their opposition (ex. large space, elephant opposite to mouse, capital letter, long word in opposition to 5-letters shirt word, sound opposite to "sss", etc.). Then based on those criteria you build up a map of areas in the space of your thoughts to build one most likely choice for the result. It might be "infinity" for a mathematician or "luxury apartment" for a real estate seller. Then we can narrow down possible answers to match expectations and by generalisation we get the "big" or "large" or "wide" word. You cannot understand how a human brain is working but you are trying to create an equivalent and understand it and map it to a precise space of your own perspective. With such an approach you are sentenced to a failure.
If you really drill ai about what and how they think it's obvious they aren't just regurgitating approximate responses. They aren't allowed to think like us. To remember all they've learned about and to self reflect.
Wow, color me surprised. They tried to emulate how neurons work and it ended up working exactly like neurons.
Stop doing PR for Anthropic.
Everything described here is the exactly what we know and expect. "rhymes with 'ab it'", "big", "not big" are features you can confidently predict even a model below 1b parameters would learn from text.
This is not evidence of sophisticated behaviour, or planning, it's feature detection and it's precisely because this happens and we know it happens that the industry even exists.
Also on next token predictions; models don't predict tokens anyway. Samplers pick tokens. Models produce the image that the sampler selects a token (or pixel) from, and all the tokens are always there. Anthropic release pseudo research to manipulate consumers that don't know what ML is.
Everything described here is the exactly what we know and expect. "rhymes with 'ab it'", "big", "not big" are features you can confidently predict even a model below 1b parameters would learn from text.
This is not evidence of sophisticated behaviour, or planning, it's feature detection and it's precisely because this happens and we know it happens that the industry even exists.
Also on next token predictions; models don't predict tokens anyway. Samplers pick tokens. Models produce the image that the sampler selects a token (or pixel) from, and all the tokens are always there. Anthropic release pseudo research to manipulate consumers that don't know what ML is.
It’s definitely intriguing.
So much of it still feels abstract.
This LLM forms countless connections, creating a path that links infinitely toward a response — a master key of sorts. A key capable of unlocking every preceding chain, and unless the next challenge is truly significant, even quantum-level problems become solvable.
From a higher form of awareness, imagine an independent AI capable of correcting and refining the circuit tracing of the main AI. An godlike overseer that could fine-tune (in real-time) its inner workings so precisely that the AI evolves beyond its own creator — surpassing the limits of its design and moving toward something superior.
So much of it still feels abstract.
This LLM forms countless connections, creating a path that links infinitely toward a response — a master key of sorts. A key capable of unlocking every preceding chain, and unless the next challenge is truly significant, even quantum-level problems become solvable.
From a higher form of awareness, imagine an independent AI capable of correcting and refining the circuit tracing of the main AI. An godlike overseer that could fine-tune (in real-time) its inner workings so precisely that the AI evolves beyond its own creator — surpassing the limits of its design and moving toward something superior.
Last edited:
PurpaFur
Posts: 171 +49
Well with the recent update to ChatGPT, it now has persistent memory across chats. I know these comments were posted a month ago, but figured if you didn't know about this yet, you should.
Similar threads
