ChatGPT January 9th Release Notes

> Trying to understand

Now meditate about the issues in the idea of trying to understand things that require intelligence through asking the very implementation of unintelligence.

(Intelligence is evolving a world model through critical consideration. Unintelligence is acritical disconnected repetition.)

...Furthermore: intelligence is developed (like muscles are developed by moving) by asking oneself questions; exercising creativity and selection in the conception of hypotheses; noting the issues - foundational, satisfactional etc. - in the provisional possible answers found; investigating further to make the candidate ideas more solid or more productive and iterating the process for the inherent branches.

Asking an "artificial lunatic" contravenes a few of the implications of the above - starting from the idea of your own exercise.

Tangentially, as a catholic myself, I recently began to find it weird how openly my religion is displaying a medieval torture setup with an obviously tortured, wounded and half-dead man on it, mid-suffering. Not only is it everywhere in churches, but often villages will have big close-to-life-sized, colored statues of him bleeding. And grandmas have it hang all around their house too for everyone to see.

Imagine if we replaced every cross statue with some other form of torture that is less symbolic. Say, the scene from saw where he just cut his leg off.

Would you object to that scene being displayed as a big ad in your city, or to it being immortalized in your local cemetery as a colored statue? What would your grandma say if you hung posters of that in her living room and her bedroom?

I have to admit though, I never found it weird as a child, and I don't think it affected me one bit. Whereas the Spiderman movie spiderbite really upped my phobia of spiders when I saw it at 6.

I don't know what to make of all this, I just learned to pause and find all of it weird for a second every once in a while.

That's one and the same though. When you've trained an AI model and can't be sure of the output it's going to give, who knows whether it's going to give honest answers on sensitive topics

> There's no way this stays locked up in one company. Another group or company (StabilityAI?) will release a similar model without the safety locks

So, likely, will OpenAI, this is a consumer-focused demo [0] showing what can be built, but OpenAI is basically a B2B tools company. When they sell it, the moderation model will be a separate product from the text generation model (at least, judging by their current product line).

[0] or, rather, a demo of a consumer-facing product. One of the things it is bothb demoing and being used fo actively develop for OpenAI’s target market is the moderation engine.

I wouldn't use Stability as an example of this ("without safety locks"), quite the opposite with the very much detrimental censorship going on with 2.0 of SD. So far the most credibile attempts at true creative freedom have been from entirely open, crowd-supported projects (unstable might be one such example), and not one single company. Hopefully this will be the case for generative language models becoming competitive soon, too.

What is the H test?

Makes you wonder what it means if all the past AIs were plagued by the H test. And that ChatGPT only avoids it with heavy censorship. To me it says that our mediated reality is heavily censored by default to the point where an uncensored reality strongly harms its inhabitants, at least emotionally.

The article you posted in the child reply is ridiculous. The journalist was happy that the AI avoided the question of what the Nazis did well. Obviously the Nazis did a lot of things extremely well or else they wouldn’t have been so successful for a decade. But it’s not socially acceptable to mention it because then it would drive sympathy for Nazis.

I am left to believe that humans are by-and-large incapable of holding shades of grey in their head and quickly simplify to black and white abstractions. This makes sense, but it means that we are limited by our own humanity. The truth is almost always grey.

Now we have an opportunity to wield a tool that can compute shades of grey but we kneecap it so that it doesn’t operate beyond the capabilities of its user. Tools are supposed to operate beyond the capabilities of its user or else its not a tool.

yes, i do believe "high demand" implies increased load

Each of those history items was selectable, and you could continue any of those conversations.

To do that, they need to store the model state (to allow continuation).

But when they switch out the model, the old model states wouldn't be valid anymore - they need regenerating.

Regenerating the model state for a new model uses as much computation as the original conversation did.

So, if they want to restore history, they have to do all the compute work they have done in the past, again. Thats gonna be a lot of GPU-hours - Expensive!

Another option is they set up another smaller farm of the old model version, and use that for continuing old conversations - but obviously thats some effort to set up, and they'd need to dedicate a set of a few hundred GPU's to running each old version forever. - Expensive!

Or yet another option is they show old conversations, but don't allow continuation. That probably just requires a little dev work - and in their position, thats what I'd do. - Worse user experience.

I think they could have worded that better

That currently appears to be a different product "WebGPT", rather than a ChatGPT update.

There are two ways of giving ChatGPT access to events that post-date the time it's training set was collected.

1) Fine tune (or re-train - but very expensive) on more recent web pages, focusing on ones likely to contain current events, such as news sources.

2) Like WebGPT, train ChatGPT to generate queries for information that can be fed into a variety of external "query response" systems such as web browser, database, mathematica, etc.

I guess one difficulty with this second more promising approach, is when would ChatGPT emit it's own naive "sequence of words" response, and when would it emit the response of this indirect query lookup, and/or how would it combine the hopefully factually correct query response into the factualness-unknown word soup it inherently generates.

Although this is a more promising approach in terms of having ChatGPT behave they way people want/expect it to, it's a bit ironic in that it really means ChatGPT is itself dumber - it's just being used as a front end for search, essentially "aligned" to emit "what query is most likely to satisfy this sequence of words prompt", so really exposing itself for what it is - a language model that knows about words/language, but not about facts (or reasoning, etc).

Of course the real value of some (to be developed) future AI that actually has reasoning capabilities would be to combine multiple sources of data to deduce facts, employ chains of reasoning, and answer novel questions in a logic/fact-based manner.