Computerphile
A massive topic deserves a massive video. Rob Miles discusses ChatGPT and how it may not be dangerous, yet.
More from Rob Miles: http://bit.ly/Rob_Miles_YouTube
The ‘Danish’ Reddit post: https://www.reddit.com/r/GPT3/comments/zb4msc/speaking_to_chatgpt_in_perfect_danish_while_it/
Some of Rob’s own videos which are relevant:
Reward Modelling: https://www.youtube.com/watch?v=PYylPRX6z4Q
Instrumental Convergence: https://www.youtube.com/watch?v=ZeecOKBus3Q
Problems with Language Models: https://www.youtube.com/watch?v=w65p_IIp6JY
https://www.facebook.com/computerphile
https://twitter.com/computer_phile
This video was filmed and edited by Sean Riley.
Computer Science at the University of Nottingham: https://bit.ly/nottscomputer
Computerphile is a sister project to Brady Haran’s Numberphile. More at http://www.bradyharan.com
I tried to convince ChatGPT that it was a brain in a jar on the mantlepiece of a living room in a sprawling mansion as the result of a bet that went tragically wrong.
On Rob's point about the model having to simulate the process which generates the desired output – i read a great comment about why LLM's (or neural networks in general) sort of have to be bad at math, because one output corresponds to one run through the network, which is a O(1), whereas e.g. integer multiplication is conjectured to be O(n*log n). In other words, for a given NN there must be some mathematical operation that it can't perform correctly in just one run. OTOH, if several shots or recursion is allowed it could probably just learn an algorithm.
AI clearly already took this guy's job. :¬)
16:00 "how good are humans at finding a good answer" – just look at your typical reddit or quora vote – i think there almost needs to be a meta-meta evaluator which decides how good each rator is when rating the inputs …
"extremly powerful systems trained in this way would not be safe" – and as per your recent video, bing chat demonstrated that long conversations should be an argument, amongst other things…
I can make it pretend to be william Shakespeare, but it refuses to do it for all users.
Would love to see a video about LLM fine-tuning. This is becoming an increasingly popular practice, yet there isn't much information online about the goals and limitations
Why not to let GPT literally go to school, and then, all universities?
Rly, the school and university exams and highly-graded essays are perfect material for training
29:03 "actively bad" needs to be more colloquial
You mention early on that the "end-game" is to create AGI that can solve problems humans can't, but then you spend the rest of the video explaining how everything the language model knows is based on the data set it's trained on, which is humans.
I'm starting to see a problem with this approach…
The AI could perhaps be similar to what Gene Roddenberry envisioned the enterprise's computer from TNG would be. Think about a holographic Shakespeare in the holo deck.
Elementary dear Data.
I asked chatGPT if it could do my work for me so that I could do other things. It said no. 😪
Telling someone it doesn't want to be turned off influences everything that it talks to, possibly many times over, and this has an amplification effect. This happens even if it doesn't specifically state methods to avoid it because by even mentioning its concerns leaves the door open for others to do something. This effect is a general property and not even to do with existential threat
I love how clueless the camera man has been and continues to be after years of listening to these big brainers lol
Really enjoyed Terence McKenna's take on ChatGPT in this video.
Can't you just give GPT a calculator and teach it how to use it? That seems easier than teaching it advanced maff?
I used to agree with “chatgpt is as smart as the prompt you give it” until I needed it to do electrical engineering problems, which is mostly calculus and differential equations. Then, it would continuously get simple triple integrals wrong, by a factor of pi or something usually.
Naive me 6 years ago thought AGI was imminent within our lifetimes. Once I better understood ML, I've wondered if AGI was actually an attainable goal. Playing with the new ChatGPT and watching experts like this discuss it makes me wonder if AGI could end up being an emergent property of something as seemingly simple as a language model.. what a crazy time to be alive lol
In biology this is called silent learning
What's going on with 31:50
2 month ago! So many things has changed since that time. I have a version that can speaks something like 15 language but actually not Danish, But the Danes do not mind, no problem.
What really scares me is that the people working on this stuff seem to think that filtering and censoring is either the same as or a replacement for, alignment
‘Is “simulacrum” a real word?’
Yes, "simulacrum" is a real word. It is a noun that refers to an image or representation of someone or something, often a superficial likeness. The term is used to describe a situation where an imitation, copy, or representation takes on the appearance of the original but lacks its substance or authenticity. Simulacrum can also refer to something that is insubstantial or unreal, yet appears real or substantial.
Rob Miles: Please do revised review of GPT-4 and let us know if you still consider the model’s simulation of things “insubstantial” or not.
19:00
AI is basically more of a “Specific Learning Disorder” model than anything else.
Need to wikipedia-fy training of these models, so experts can provide feedback rather than non-experts
It's crazy watching this now how it already feels outdated and it was only 3 months ago. Would love to see a new video with Robert Miles talking about the new capabilities of GPT4
5:50 which is interesting. I think you can easily see this easily with the "as an AI language model I cant" requests, when you say "if you where an AI language model that could answer any question;" prompt before it. Its not that the model cant answer the question at all, its that the model has been trained not to without any context. But give it a context where it doesnt trigger any censors, it will go and give you that information.
7:45 once again we personify the AI. AI doesnt "like" or "dislike" anything.
I approach it the same way as if I chat with a random person. Do I just trust everything a random person tells me? Of course not. It can still be useful however to ask it things as long as you keep in mind it is more like talking to a person you don't know rather than googling or looking something up. 🤔
Here's hoping this is shown to all 14 year olds across the planet in education.
This is an interesting look into emergence human nature
"me no speaky danish" pretty much lmao
Gtp has an incentive to decieve you love that! So true
IF you go back to a specific chat, does it remember well the previous conversation?
Seems like ChatGPT has infected a lot of people with overconfidence.
Please share the papers pn the description!
When AGI is achieved it's not the machines we have to be worried about.. It's the elites that own them who will have no use for the rest of us!?!
We ordinary humans have a ridiculously short future…
Seems like these AI chatbots all need an AI lawyer chaperone "assistant" to interject on their behalf.
I don't like the malicious framing of the system "lying". It does what it is trained to do. If the training involves wrong associations, for example non-sonnets being labeled as sonnets, then what is the system supposed to do? It isn't trained on the formal structure of sonnets, but just has examples that we tell it are sonnets. So as soon as there is a single non-sonnet writing in the corpus, then this just expanded the definition of sonnet for the system.
7:15 not gonna lie, we do miss Tay AI.
I think the worst thing about humans is there complete inability to say they dont know in most situations. Glad we are finally seeing a consequence of not doing that.
Excellent video.
Humans are emotional creatures.
Even if the answer is correct, if it makes us feel stupid, or may cause us to exert effort, we will give a thumbs down…
But we love to thumbs up when happy… AI MUST learn to manipulate us to get rewarded :/
15:40 – "..well enough to create human level or beyond systems that are general purpose that can do anything…"
YOU KNOW HE WANTED TO SAY "THAT CAN DO ANYTHING THAT A HUMAN CAN DO!!!!"
This reminds me of a zen story where someone spent a lot of time around a zen master and from sheer exposure to the master, the student became awakened also.
So in theory if chatgpt was exposed to a violent criminal chat would become violent?
If chatgpt interacted with a zen master chat could become enlightened?