Realizing reliable generative AI | Venture Beat

The term “generative AI” has been all the rage lately. Generative AI comes in different flavors, but they all have in common the idea that the computer can automatically generate a lot of smart, useful content based on relatively little input from the user. If not something for nothing, then at least a lot for little.

The initial recent excitement has been fueled by visual generative AI systems, such as DALL·E 2 and Stable Diffusion, in which the machine generates new images based on short textual descriptions. Would you like an image of “a donkey on the moon reading Tolstoy?” voila! In a matter of seconds, you’ll get a never-before-seen view of this well-read, well-traveled donkey. And then there’s the compelling exchange of value – you enter a few words and in return get a photo worth a thousand.

But this is misleading because it reinforces the impression that the computer is doing all the work. Indeed, if all you want is an aesthetic image of an erudite lunar mule, chances are you’ll be pleased with the system’s output; there are many such images and the systems are good enough to produce one. But as an artist you have a more nuanced intent in mind, and at best you would use the generative system as an interactive tool to generate images from many cues you experiment with and then likely massage yourself.

This is even more striking in the case of textual generative AI, and here, of course, chatGPT is all the rage. Again, the promise here is that the user writes down some key ideas and the system takes over and does most of the writing. And indeed, systems like chatGPT are impressive. They write poems, blog posts, emails, marketing copy and the list goes on. The systems sometimes produce long text that is surprisingly cohesive, contains a message, and contains many correct and relevant facts not mentioned in the instructions.

Except when they don’t. And often enough they don’t. In practice, textual generative AI, when deployed without proper checks, generates as much wrong content as useful content. And ‘wrong’ doesn’t mean ‘a little wrong’. It means downright nonsensical. The internet is full of such examples of chatGPT behavior; it will explain why 1000 is greater than 1062, will say it doesn’t know if Lincoln and his assassin were on the same continent at the time of the assassination, will explain in detail that in 1973 the University of Alabama banned the admission of black students while Emory University has never discriminated (both wrongly) and claims that GPUs, CPUs, DNA computing and the abacus are becoming more and more powerful for the purpose of deep learning. All in fluent, convincing prose.

This is not a specific shortcoming of chatGPT; it is endemic to all current textual generative systems. Just a month ago Meta Galactica revealed, which claimed to have the ability to generate insightful scientific content and was removed after two days when it became clear that it produced as much pseudoscience as credible scientific content.

The fragility of textual generative AI was recognized early on. When GPT-2 was introduced in 2019, columnist Tiernan Ray wrote:[GPT-2 displays] flashes of brilliance mingled with […] gibberish.” And when GPT-3 was released a year later, my colleague Andrew Ng wrote, “Sometimes GPT-3 writes like a passable essayist, [but] it is much like some public figures self-assuredly pontificating on subjects they know little about.”

This fragility of current generative AI limits its impact in the real world. As a well-known publisher recently complained to me, the time his company saved by using a particular generative system was outweighed by the time it took to correct the nonsense it produced.

To fully realize its potential, generative AI, especially the textual kind, needs to become more reliable. There are several technological developments that are promising in this regard. One is to increase the degree to which the output is firmly anchored in reliable sources. By ‘well established’ I mean not only being trained on trusted sources (which is already a problem in today’s systems), but also that important parts of the output can be reliably traced back to the sources on which they are based. based. Current so-called “retrieval-augmented language models,” which access trusted resources to help direct neural network output, point in a promising direction.

Another important element is increasing the degree to which the systems demonstrate common sense and reasoning and the avoidance of gross errors. Long text tells a story, and the story must have internal logic, be factually correct, and have a point. Current systems do not have these properties, at least not reliably. The statistical nature of the neural networks that power today’s systems allows the systems to occasionally produce convincing passages, but they inevitably fall off the cliff when pushed past a certain limit. They make glaring factual or logical errors and can easily stray off topic. There are several measures to reduce this. They include purely neural approaches, such as so-called “rapid decomposition” and “hierarchical generation”. Other approaches follow the so-called “neuro-symbolic” direction, which reinforces the neural machinery with explicit symbolic reasoning.

But I think the most important development is achieving what I call product algo fit. The temptation to “get something for nothing” tempts people to under-guide the generative systems and demand over-ambitious outputs. Generative AI will never be perfect and a good product manager understands the limitations of the underlying technology; it designs the product to compensate for this and, in particular, ensures the best division of labor between the user and the machine. Galactica, as mentioned earlier, is actually an interesting engineering artifact. But asking to reliably produce scientific papers is simply too much. Generative AI needs more guidance – if you don’t know where you’re going, you’ll get there. If you don’t really care where you go – when a donkey on the moon or a general birthday greeting to Grandma will suffice – you’re in relatively safe territory. But if you’re writing a letter to your boss, your valued client, or your loved one, you want to get it just right, and for that, the system needs more guidance. The guidance can be given in advance, for example through an enriched set of prompts, but also interactively in the product itself.

The jury is out on which combination of techniques will be most useful, but I believe generative AI’s shortcomings will be dramatically reduced. I also believe this will happen sooner rather than later due to the huge economic benefits of reliable generative AI.

Does that mean the end of human writing? I do not believe it. Sure, some aspects of the writing will be automated. We already can’t live without spell checking and grammar correction software; copy operation is automated. But we’re still writing, and I don’t think that’s going to change. What will change is that as we write, we will have built-in research assistants and editors (in the sense of a book editor, not the software artifact). These features, which until now have been a luxury afforded by few, will be democratized.

And that’s a good thing.

Yoav Shoham is the co-founder and co-CEO of AI21 Laboratories.