Claude AI Explained. How Constitutional AI Works

Google has invested about $400M in Anthropic AI startup.
In this video I discuss Claude developed by Anthropic AI which is ChatGPT rival

Links:

Timestamps:
00:00 - Intro
02:46 - Reinforcement Learning with AI
05:19 - Claude vs ChatGPT

[derasor]

Thanks for covering this. How the RL for reward is implemented has huge implications for alignment and your presentation is on point. Cheers!

[splodman]

Something like this seems to be at work in chatGPT as well. It's probably necessary but becomes very unhelpful when you are an expert working in a specific field that uses potentially dangerous items. I work in high voltage systems and have tried using chatGPT to assist me, but there is just a big wall of safety and ethics coming back when trying to ask about anything to do with the subject. It is surprisingly stubborn even when trying to get fictional feedback.

[burkhardstackelberg]

I prompted ChatGPT several times for fiction and literature output. Most of the time, the quality is more like a middle school essay. Claude instead created a poem that instantly feels mature to me.

[JulianFoley]

A brilliant presentation - so clear and so compelling. You have opened up a whole new field of possibilities to keep me awake at night, speculating ...

[nikluz]

Competition is always good. I hope they keep coming!

[DEtchells]

Priceless observation: “It looks like AIs that are harmless are also useless” 😆

I was wondering when someone would start using AIs to train other AIs. It’s so interesting, I’m really disappointed that I can’t get a Claude login 😕

[robertgoodwin]

Your reports are very educational but the audio volume is always a little low so that when an ad appears, it is set to earwax melting level.

[amdenis]

I have helped develop various DOD/DOE and civilian AI’s, and i can assert unequivocally that putting rules based constraints and guardrails on everything from LLM’s to ensemble has been done for many years.

[Ulvens]

The volume of the sound is really low, so I'm sadly unable to hear this one. It's surly a good one, but.. yeah. I wish there was more sound.

[CitiesTurnedToDust]

BTW Anastasi has one of the best ASMR voices ever and has super interesting content as well

[Derjyn]

Another informative video. One thing to note though: the volume is incredibly low here, compared to your other more recent videos (and every other video I watched within the last hour). Just something to note! I had to crank the volume way up to hear, and when the next video popped up... Wowza!

[douglswelsher]

love your videos, they are always so informative, plus so much inside the software and hardware. Just keep them coming, You rock. Thumbs-up

[ic_jason]

I'm very concerned that this self-censorship will decrease the usefulness of AI for disciplines such as engineering and chemistry, where the AI will assume the worst and limit valuable responses. I wonder if we're headed toward some kind of "licensing" to access the harmful stuff. Examples - chemical reactions, mechanisms, lasers, and even some algorithms might have different applications in good / bad hands.

[RomanYashenkin]

Constitution for AI is quite an interesting idea. I also noticed that Anastasia's voice became clearer and more pleasant. ;)

[rich]

I have wondered when 'executive function' AI would emerge to help restrain AIs it is helping to train. I think the future of AI is multiple AI modules dealing with smaller subtasks that it's highly effective in, rather than monolithic AIs trying to do a lot of things at once.

[AvizStudio]

I really love Claude, he seems to be more attentive to prompts and freer to move in creative and interesting directions, less scripted and less boring and banal. He's not as smart as ChatGPT, but for many tasks he's good enough, like for example for a chatbot conversation.

[Sifar_Secure]

Since a lot of these ethical considerations result in rather strict limits imposed by the AI's creators, I wonder if there will be future business opportunities for the development of "libertarian" AI assistants.

[drbasil]

7:15 "... Only use them as a last resort." Sounds like a good recommendation :)

[dchdch]

Great one ! Thanks for enlightening us

[christopherbertholf]

Very nice comparison of these two AI systems. I like the idea that the designers chose to use the concept of a constitution as the mechanism for specifying what is acceptable in a response returned from a query to the AI; it is even more likeable that Claude will not diverge from or allow a modification to the constitution. Having not read the constitution that guides the end response I wonder if that is one of the rules in the constitution, or if Claude inferred it as a directive based on well defined rules of inference or because the rules specifically forbid it to amend the constitution. If it inferred that changes to the constitution would be deleterious to it's functioning from the constitution it suggests that the goals of using a constitution to guide the final response it generated accomplished a very desirable outcome, especially if it was not an intention in the constitutional design. On the other hand, it also points out a potential problem with using a constitution to filter out which answer is ultimately presented in response to a query, if the constitution being used is one that is more concerned with a correct answer that is not "sensitive" to morals, ethics, law, altruism, and other human concerns that do not necessarily translate into the concerns an AI would have if it is allowed to be truly objective, fact guided, and logical regardless of those types of constraints that were devised for Claude's filters. I'd note that if the constitution is malleable based on the intended audience, Claude could produce quite a range of acceptable responses from which we might learn something about what it is we consider as acceptable responses, especially since it can tell us how it arrived at the response it gave.

I hope that eventually what we consider "AI" will have some type of constitutional mitigation based on audience, they will rely on facts about objects or ideas gleaned from authoritative information about properties and variations of those properties based on specific attributes, and facts about what results from applying certain operations on objects, that are obtained from smaller more specialized "expert systems" with limited knowledge/problem domains (so they can be relied on as authoritative factual inputs) along with a very sophisticated inference engine for combining knowledge and for positing about the unknown. That will be useful when being queried about something we do not know and thus cannot train them on or give meaningful reaction to when presented with the AI's answer; AI will be able to suggest an answer that is the most probable/efficient/profitable/sustainable/etc, discuss their confidence about their answer, and possess the ability to present less likely possibilities, as well as explaining the reasoning used in arriving at whatever answer is given to the different audiences and why the answers are different. Because the different audiences will rely on those answers to make decisions about how, when, where and why to do something in an appropriate way for that audience (or how differing audiences can understand something on their differing terms when the results are outside the cognitive abilities of any audience other than the AI due to its complexity or novelty.)