I have some issues with how this article describes GPT and its impact. The scenarios that are described as using GPT-3 remains the stuff of science fiction as far as I know. They would require major technological breakthroughs in areas of autonomous agents, combining language models with 'concept-based' models, and more mundane things like labelling (can you imagine how difficult it would take to accurately label all legislation for 'nefarious details', given that even humans can't agree on what is nefarious?).
More specifically:
- training on data is a lossy process. In your examples, GPT would actually have a worse memory than your lawyer or therapist. There is no way to combine language models and something more abstract like 'facts'.
- GPT has shown zero ability to do anything consistently successfully without a human-in-the-loop. When it comes to bring AI models into production, this matters a lot. There's no way autonomous therapists are coming from GPT-3 when half the time the model spews out potentially dangerous garbage. You can't teach GPT-3 to not hurt people because it has no concept of people or hurting them. It JUST knows the shape of English.
- GPT is an unsupervised (in terms of data labelling work required) model. It has not made any breakthroughs in requiring labelled data for fine-tuning the model to do a specific task. Which remains a gigantic problem for productionalizing models. Like how are you going to build an autonomous therapist? That data remains as inaccessible and impossible to label as ever.
- Please stop telling people that neural nets are related to brain neurons. They have essentially no relationship other than the name and it just fosters this fear of Terminator and obscures the real issues that need to be thought about. This is just my personal opinion but I'm so tired of having to spend my time telling otherwise smart people who don't know better that we aren't close to Terminator.
GPT is an impressive technical accomplishment, but it's impact on the world has been exaggerated quite a bit IMO. Some of the demos I've seen are almost certainly smoke and mirrors or very carefully chosen, human-in-the-loop examples.
>> Please stop telling people that neural nets are related to brain neurons. They have essentially no relationship other than the name and it just fosters this fear of Terminator and obscures the real issues that need to be thought about.
Indeed, this is supported by the opinions of the foremost experts in deep learning and neural networks:
IEEE Spectrum: We read about Deep Learning in the news a lot these days. What’s your least favorite definition of the term that you see in these stories?
Yann LeCun: My least favorite description is, “It works just like the brain.” I don’t like people saying this because, while Deep Learning gets an inspiration from biology, it’s very, very far from what the brain actually does. And describing it like the brain gives a bit of the aura of magic to it, which is dangerous. It leads to hype; people claim things that are not true. AI has gone through a number of AI winters because people claimed things they couldn’t deliver.
Synapses' strength and ANN weights are sort of similar, but the learning process is different (co-activation vs loss propagation). Plus brains have neuroplasticity, but ANNs usually have at least one fully connected layer so it can emulate that, sort of.
Also how the brain's visual circuit parses stuff and how a CNN encodes concepts are sort of similar.
A much more realistic and accurate take on the tech than the article. One combination that can work is to have GPT3 with an unskilled human judge in the loop. The human can discriminate with common sense while the machine generates good jargon with correct grammar.
Jargon (sometimes) isn't just obfuscation; it carries important shades of meaning that would be tedious to spell out every time. 'Homicide' and 'murder', for example, are sometimes used interchangeably, but are not actually legally interchangeable. You would not want your involuntary manslaughter plea upgraded to murder, a more serious--and therefore more severely punished offense.
You could, of course, try to train "unskilled labor" to detect things like that. By the time you are done, the labor won't be unskilled and you will have reinvented the paralegal.
Yeah, I think those will be some of the first wave of language generation applications. Stuff like Buzzfeed but without as many writers because the language model is generating potential articles and the humans pick which are good (and maybe clean them up). Similarly, disinformation or noise attacks on social media - with just a few people overseeing the language model as it generates potential posts/replies.
The first is a digital information gatherer that is explicitly not playing the role of a therapist (likely for the reasons I mentioned above around risk/cost of failure, difficulty in evaluating what constitutes 'good advice', and probably legal barriers). There is a world of difference between a chatbot that does information retrieval and an autonomous agent that provides therapy.
The second is also vastly different from the "summarize legislation and detect 'nefarious' clauses" scenario in the article. They are identifying errors in standard NDAs, which is a (comparatively) well-defined, straightforward supervised learning task where the data has a pretty consistent shape unlike congressional legislation and what 'nefarious' means. (I have to make some assumptions here since as far as I can tell there isn't a technical paper on the work).
These sound similar to the tasks in the article, but once you get into the details of implementing and deploying them, I don't think they're very similar. You would still need to solve the problems that have plagued self-driving cars for years: high cost of failure, unpredictable failure modes, a long-tail distribution of data that prevents realistic-to-collect data sets from generalizing well enough, almost no progress in AI for autonomous agents (RL is promising, but it hasn't really made it out of the lab yet AFAICT).
There is zero evidence that GPT-3 has the ability to become self-aware and define new goals for itself. But I will give you that if your idea of 'close' is hundreds of thousands to millions of years, then yeah, AGI might be 'close'.
The terminator idea is worth dismissing in this context. Even if GPT-3 is indeed intelligent, there's no reason to think that it's a goal-based system that could independently want things or do things.
More specifically:
- training on data is a lossy process. In your examples, GPT would actually have a worse memory than your lawyer or therapist. There is no way to combine language models and something more abstract like 'facts'.
- GPT has shown zero ability to do anything consistently successfully without a human-in-the-loop. When it comes to bring AI models into production, this matters a lot. There's no way autonomous therapists are coming from GPT-3 when half the time the model spews out potentially dangerous garbage. You can't teach GPT-3 to not hurt people because it has no concept of people or hurting them. It JUST knows the shape of English.
- GPT is an unsupervised (in terms of data labelling work required) model. It has not made any breakthroughs in requiring labelled data for fine-tuning the model to do a specific task. Which remains a gigantic problem for productionalizing models. Like how are you going to build an autonomous therapist? That data remains as inaccessible and impossible to label as ever.
- Please stop telling people that neural nets are related to brain neurons. They have essentially no relationship other than the name and it just fosters this fear of Terminator and obscures the real issues that need to be thought about. This is just my personal opinion but I'm so tired of having to spend my time telling otherwise smart people who don't know better that we aren't close to Terminator.
GPT is an impressive technical accomplishment, but it's impact on the world has been exaggerated quite a bit IMO. Some of the demos I've seen are almost certainly smoke and mirrors or very carefully chosen, human-in-the-loop examples.