Max Bennett is an absolute star. His book is the best on intelligence I have read and I think I have read 8 in the last 4 years. The evolutionary perspective he brings show you a huge amount about what intelligence is, how it emerged and then how it became more sophisticated to support different niches. If you ever hear anyone say "we don't understand the brain" point them in the direction of ABHOI!
I do have to agree: his book really is one of the best books I've read in a while. What makes it so good is that it is both very engaging and well written, while also being VERY informative. One could learn a LOT from this book.
@@vincentnaayem I have another book recommendation. I just picked this up over the weekend and am starting on it as I am finishing Bennett's book: The Deep Learning Revolution, by Terrence J. Sejnowski, MIT Press, 2018. Like Bennett's book, there is a lot of discussion of the intersection between human and machine intelligence.
If you talk to a LLM and say 1 +1 = 4, it will usually refuse to see this as a fact. If you give it an new definition, it accepts it and works with it. So this ability is something that LLMs already have at inference time. In training time they have no chance to refuse what the trainer wants them to answer.
I think that was the point, that LLMs do not continuously adapt to new information, they only react. Like nematodes, it does a phenomenal job in some area, but it doesn't need to learn, understand, or even evaluate anything to do this. You can cheat a bit and include previous prompts in the next prompt, or use external resources to add information to the prompt in realtime, but the underlying structure is completely static, one might say quite unintelligent.
@@EskiMoThor in theory, you could finetune(update) the weights every day. weights are long term knowledge for LLMs. Activation pattern is short term knowledge. LLM based chatbots can learn a lot during a conversation. but they will forget, when the conversation is over. finetuning every day is difficult and would cost too much money. OpenAI has released a workaround today. it is a simple notebook for the chatbot, to remember some things about the user and old conversation. ChatGPT is more intelligent than many humans but still has some weaknesses. More tools and infrastructure will reveal the full potential.
In a first chapter about the basics of neurons and 🐸frog experiment, it is written: “In Adrian’s frog muscle experiments, a neuron might fire one hundred spikes in response to a certain weight. But after this first exposure, the neuron quickly adapts; if you apply the same weight shortly after, it might elicit only eighty spikes. And as you keep doing this, the number of spikes continues to decline.” Are we talking about "dead" neurons detached from a frog?!
Very fascinating stuff, Max did a great job at integrating the high level concepts with the important details. A couple of things I was wondering about while watching this: - When did emotions (not feelings, but changes in states) evolve and what were the adaptive benefits? What impact did it have on learning, and do you think they would be essential or even beneficial to improve AI? - Similarly with sleep, we know it is essential for our learning, but when did it evolve, what benefits did the early organisms have, and could we imagine AI using similar mechanisms?
I'm not really qualified to answer the questions you posed, but in regards to sleep, you would have to believe that once a biological agent can afford to sleep it would have been a bonus for the system, but that must have been very early in our development. Peace
I have not understood what learning is fundamentally!! Can someone explain in layman terms as to what happens physically in a system when it is said to have learned something!!!?
There should be a cognitive separation between you and the outside world to be able to be conscious and feel something...self consciousness is different...but the separation is essential for consciousness...
How come these people are so intelligent who seem to understand all these things n every thing while there are, I assume, dumb-witted people are also out there!!?
I just discovered this channel! Wow thank you, this is pure gold! Looking forward to the next episodes !
Max Bennett is an absolute star. His book is the best on intelligence I have read and I think I have read 8 in the last 4 years. The evolutionary perspective he brings show you a huge amount about what intelligence is, how it emerged and then how it became more sophisticated to support different niches. If you ever hear anyone say "we don't understand the brain" point them in the direction of ABHOI!
I do have to agree: his book really is one of the best books I've read in a while. What makes it so good is that it is both very engaging and well written, while also being VERY informative. One could learn a LOT from this book.
What are the other books that you have read? Getting started on my journey. I am a junior deep learning engineer
@@vincentnaayem I have another book recommendation. I just picked this up over the weekend and am starting on it as I am finishing Bennett's book:
The Deep Learning Revolution, by Terrence J. Sejnowski, MIT Press, 2018.
Like Bennett's book, there is a lot of discussion of the intersection between human and machine intelligence.
wow you are dropping some great shows at the moment
Thank you!
If you talk to a LLM and say 1 +1 = 4, it will usually refuse to see this as a fact. If you give it an new definition, it accepts it and works with it. So this ability is something that LLMs already have at inference time. In training time they have no chance to refuse what the trainer wants them to answer.
I think that was the point, that LLMs do not continuously adapt to new information, they only react. Like nematodes, it does a phenomenal job in some area, but it doesn't need to learn, understand, or even evaluate anything to do this.
You can cheat a bit and include previous prompts in the next prompt, or use external resources to add information to the prompt in realtime, but the underlying structure is completely static, one might say quite unintelligent.
@@EskiMoThor in theory, you could finetune(update) the weights every day.
weights are long term knowledge for LLMs. Activation pattern is short term knowledge. LLM based chatbots can learn a lot during a conversation. but they will forget, when the conversation is over.
finetuning every day is difficult and would cost too much money. OpenAI has released a workaround today. it is a simple notebook for the chatbot, to remember some things about the user and old conversation.
ChatGPT is more intelligent than many humans but still has some weaknesses. More tools and infrastructure will reveal the full potential.
In a first chapter about the basics of neurons and 🐸frog experiment, it is written: “In Adrian’s frog muscle experiments, a neuron might fire one hundred spikes in response to a certain weight. But after this first exposure, the neuron quickly adapts; if you apply the same weight shortly after, it might elicit only eighty spikes. And as you keep doing this, the number of spikes continues to decline.”
Are we talking about "dead" neurons detached from a frog?!
Very fascinating stuff, Max did a great job at integrating the high level concepts with the important details.
A couple of things I was wondering about while watching this:
- When did emotions (not feelings, but changes in states) evolve and what were the adaptive benefits? What impact did it have on learning, and do you think they would be essential or even beneficial to improve AI?
- Similarly with sleep, we know it is essential for our learning, but when did it evolve, what benefits did the early organisms have, and could we imagine AI using similar mechanisms?
I'm not really qualified to answer the questions you posed, but in regards to sleep, you would have to believe that once a biological agent can afford to sleep it would have been a bonus for the system, but that must have been very early in our development. Peace
Aren't the emotions derivatives of the alertness, reward-punishment mechanism.
You can see why fear, shock, awe, etc is there.
I like the part where he says AI grounds the the concepts of consciousness.
I have not understood what learning is fundamentally!! Can someone explain in layman terms as to what happens physically in a system when it is said to have learned something!!!?
There should be a cognitive separation between you and the outside world to be able to be conscious and feel something...self consciousness is different...but the separation is essential for consciousness...
Thomas Kenneth Jackson Ronald White Joseph
How come these people are so intelligent who seem to understand all these things n every thing while there are, I assume, dumb-witted people are also out there!!?
This whole video sounds like AI generated
Wow such camera shake spoiling an interesting video - I feel sick. 😣
Nonsense, there is no intelligence in, artificial intelligence.