Hacker News Reader

I don't necessarily see the "team of automated googlers" as a fundamental or damning problem with GPT-like approaches. First I think people may have a lot fewer truly original ideas then they are willing to admit. Original thought is sought after and celebrated in arts as a rare commodity. But unlike in arts, where there are almost not constraints, when it comes to science or engineering almost every incremental step is of form Y = Fn(X0,..,Xn) where X0..Xn are widely known and proven to be true. With sufficient logical reasoning and/or experimental data, after numerous peer reviews, we can accept Fn(...) to be a valid transform and Y becomes Xn+1, etc. Before internet or Google one had to go to a library and read books and magazines, or ask other people to find inputs from which new ideas could be synthesized. I think GPT-like stuff is a small step towards automating and speeding up this general synthesis process in the post-Google world.

But if we are looking to replace end-to-end intelligence at scale it's not just about synthesis. We need to also automate the peer review process so that it's bandwidth is matched to increased rate of synthesis. Most good researchers and engineers are able to self-critique their work (and the degree to which they can do that well is really what makes one good IMHO). And then we rely on our colleagues and peers to review our work and form a consensus on its quality. Currently GPT-like systems can easily overwhelm humans with such peer review requests. Even if a model is capable of writing the next great literary work, predicting exactly what happened on Jan 6, or formulating new laws of physics the sheer amount of crap it will produce alongside makes it very unlikely that anyone will notice.

Art (music, literature, ...) involves satisfaction of constraints. For instance you need to tune your guitar like the rest of the band, write 800 words like the editor told you, tell a story with beginning, middle, and end and hopefully not use the cheap red pigments that were responsible for so many white, blue, and gray flags I saw in December 2001.

I call it the "Prior-Units" theorem. Given that you are able to articulate an idea useful to many people, there exists prior units of that idea. The only way then to come up with a "new idea", is to come up with an idea useful only to yourself (plenty of those) (or small groups), or translate an old idea to a new language.

The reason for this is that if your adult life consists of just a tiny, tiny, tiny fraction of the total time of all adults, and so if an idea is relevant to more people, odds decrease exponentially that no one thought of it before.

There are always new languages though, so a great strategy is to take old ideas and bring them to new languages. I count new high level, non programming languages as new languages as well.

"team of automated googlers" where google is baked-in. Google results, and content behind it, changes. Meaning, GPT would have to be updated as well. Could be a cool google feature, a service.

Have you tried conversing with it, after a few lines of setting a proper context? Like two scientist talking or something like that. It can provide very interesting outputs that are not googlable.

Yes, every time you see something that for human obviously doesn't make sense it makes you dismiss it. You would look at that output differently though if you were talking with a child. Just like a child can miss some information making it say something ridiculous it may miss some patterns connections.

But have you ever observed carefully how we connect patterns and make sentences? Our highly sophisticated discussions and reasoning is just pattern matching. Then most prominent patterns ordered in time also known as consciousness.

Watch hackernews comments and look how after somebody used a rare adjective or cluster of words more commenters tend to use it without even paying conscious attention to that.

Long story short, give it a try and see what examples of what people already did with it even in it's limited form.

To me you are looking at an early computer and saying that it's not doing anything that a bunch of people with calculators couldn't do.

No, it provides something that superficially looks interesting. That's a big difference from being actually interesting.

GPT-3 is trained on text prediction, and there's been a lot of commentary about the generation aspect, but some of the applications that excite me most are not necessarily about the generation of text, instead, GPT-3 (and also other language models) create very useful vector representations of natural language as a side effect that can then be used for other tasks with much less data, or with too much extra data. Using the text prediction task as a way to supervise learning this representation without having to create an expensive labelled dataset is very helpful, and not just to language tasks. See for example the CLIP work that came out recently for image classification, using GPT-3 and captions to supervise training. There is other work referred to in that blog post that also exploits captions or descriptions in natural language to help understand images better. More speculatively, being able to use natural language to supervise or give feedback to automated systems that have little to do with NLP seems very very useful.

"More speculatively, being able to use natural language to supervise or give feedback to automated systems that have little to do with NLP seems very very useful."

I agree with this, and it isn't entirely speculative. One of the most useful applications I have seen that goes beyond googling is generating css using natural language. ie, "change the background to blue and put a star at the top of the page". There are heavily sample selected demos of this on twitter right now https://twitter.com/sharifshameem/status/1282676454690451457...

This is definitely practical, though I wouldn't design my corporate website using this. could be useful if you need to make 10 new sites a day for something with seo or domains

> I think of the value proposition of GPT-X as "what would you do with a team of hundreds of people who can solve arbitrary problems only by googling them?". And honestly, not a lot of productive applications come to mind.

Damn, this could replace so many programmers, we're doomed!

*realizing GPT-3 was probably created by programmers who's job really is mostly googling for stackoverflow answers*

#singularity

We were worried that the singularity was going to involve artificial intelligences that could make themselves smarter, and we were underwhelmed when it turned out to be neural networks that started to summarize Stack Overflow neural network tips, to try to optimize themselves, instead.

GPT-∞: still distinguishable from human advice, but it contains a quadrillion parameters and nobody knows how exactly it’s able to tune them.

Curses. We've been found out.

The current problem is that we don’t have a reliable, scalable way to merge in features of knowledge engines that have ontological relationships of entities with generative engines that are good for making more natural looking or sounding qualitative output. There’s certainly research going on to join them together but it’s just not getting the kind of press releases as the generative and pattern recognition stuff that’s much easier comparatively. The whole “General AI Complete” class of problems seems to be ones that are trying to combine multiple areas of more specific AI systems but that’s exactly where more practical problems for the average person arise.

Agreed, but that's because they're hard to integrate together: one is concerned with enumerating over all facts that humans know about (a la Cyc) and the other is concerned with learning those directly from data. Developing feedback systems that combine these two would be quite exciting.

> honestly, not a lot of productive applications come to mind

Not so convincing when you enumerate so many applications yourself.

> but won't be able to say anything about current events

There are variants that use transformer + retrieval, so they got unlimited memory that can be easily extended.

Content farms are hardly a productive application.

You missed the forest for the trees. If you got a tool that can use StackOverflow to solve simple programming tasks, or to generally solve any simple task with Google, then you're sitting on a gold mine.

Yes and no.

It may be useful to hire less low skilled employees and keep a few senior ones that take input from machine and decide what to keep and what to throw away. I'm not sure if a senior engineer would be more productive patching up code written by a bot or writing it from scratch. It's going to be a hard sell while you still need human supervisors.

You can't trust a machine that can't reason with code implementation, or even content creation. You need a human to supervise or a better machine.

We already have AI based auto-completion for code, gpt-3 can be useful for that (but at what cost? Storing a huge model on your disk or making a slow / unsafe http request to the cloud?)

> if a senior engineer would be more productive patching up code written by a bot or writing it from scratch.

I have no doubt writing from scratch would win hands down. The main reason we patch wonky legacy code is because it's already running and depended on. If you remove that as a consideration, a senior engineer writing the equivalent code (rather than debugging code generated randomly from Google searches) would -IMO- would be more efficient and produce a higher quality program.

That's a big if though.

GPT-3 is much more interesting autocomplete based on most commonly used patterns than something which figures out that Problem X has a lot of conceptual similarities with Solved Problem Y so it can just reuse the code example with some different variable names.

I've mentioned this in another thread, but a GPT-3 that could reliably generate quizbowl questions like the ones on https://www.quizbowlpackets.com would be great in this domain. My experience with it indicates it's no where near being able to do this, though.

> And honestly, not a lot of productive applications come to mind

So, can't go into too many details, since I haven't started yet, I'm thinking about mixing a flavor of GPT with DETR for OCR tasks where the model then must predict categorization vectors, the chief difficulty of the task being that it must identify and classify arbitrary length content in the OCR.

I work in a niche sector of the insurance industry. Based on what it can already do, I can see it doing half my job with basically no learning curve for the user. Based on this alone, I could see it reducing headcount in the company and sector by 5%. This is massive when you consider the low margins in the industry and high costs of "skilled" staff.

It's not even commercially available afaik, it's only been made available to some developers who applied for access, through a completely opaque process-- this after soliciting for application ideas from every applicant. So I heartily support finding a way to truly open source it.

Did anyone try to ask GPT-3 something about the insurrection? I can't find any free/open text prompts for GPT-3, this is what GPT-2 has to say:

After Trump lost the 2020 election his supporters stormed the Capitol

> They smashed windows, tore apart the Republican Party's headquarters, destroyed a statue of Ronald Reagan and set fire to the Texas statehouse.

> A month later, after the Democrats took control of the Senate and the House, Trump's supporters stormed their own offices. They attacked the offices of California Democratic Senator Kamala Harris and New York Democratic Representative Joe Crowley. They ransacked offices in the Minnesota and Texas statehouses and burned down the Republican Party headquarters in Tennessee. The damage was estimated at

my god it should be called CNN-2

GPT-3

> The Trump supporters were armed with guns and knives. They were also carrying torches. They were chanting “Trump 2020” and “Build the Wall.” The Trump supporters were also chanting “Lock her up.” But they were referring to Hillary Clinton.

This is hilarious, more please :)

You can play with GPT3 in a custom world at AIdungeon.io The responses are biased towards giving you RPG second person narrative, but the corpus of data, mastery of syntax and more uncertain grasp of events and relationships is all there

Example with the prompt You are Donald Trump. The recent election results have been a disappointment to you.

https://pastebin.com/dSYZypCw

Props for turns of phrases like "Your opponent is a typical liberal. He hails from the right wing of the Democratic Party, but has been trying to appeal to the left to gain more support.", but poor marks for apparently not having grasped how elections work. (There's a joke in there somewhere)

If you don't pick a custom world and your own prompt, you get something more like this:

> You are Donald Trump, a noble living in the kingdom of Larion. You are awakened by a loud noise outside the gate. You look out the window and see a large number of orcish troops on the road outside the castle.

I'd like 'orcish troops' better if I thought it was inspired by media reports of Capitol events rather than a corpus of RPGs.

Or not even googling, but pre-googling. Using its predictive typing in the text box at google.com Because you are giving to something to complete.

>"what would you do with a team of hundreds of people who can solve arbitrary problems only by googling them?"

What would you do with a team of hundreds of people who can instantly access an archive comprising the sum total of digitized human knowledge and use it to solve problems?

Well, they can use it to write text. Not to solve problems directly.

We have that now, it's called googling. you could easily hire 100 people to do that job, but you'd have to pay them at least $15/hr now on the US. Say equivalent gpt-3 servers cost a fraction of that. How do you make money with that resource?

In science, some amazing discoveries are made years or even centuries before some practical applications for them are found. I believe in humanity, sooner or later we'll find some actually useful applications for GPT-X.

It's a wonderful co-writer of fiction, for one. Maybe the better authors wouldn't need it, but as for everyone else -- take a look at https://forums.sufficientvelocity.com/threads/generic-pawn-t..., and compare the first couple of story posts to the last few.

One of the ways in which people get GPT-3 wrong is, they give it a badly worded order and get disappointed when the output is poor.

It doesn't work with orders. It takes a lot of practice to work out what it does well with. It always imitates the input, and it's great at matching style -- and it knows good writing as well as bad, but it can't ever write any better than the person using it. If you want to write a good story with it, you need to already be a good writer.

But it's wonderful at busting writer's block, and at writing differently than the person using it.

I think this is exactly right, and indeed this is a lot of the value. "Content-generation" is already a thing, and yes it doesn't need to make much sense. Apparently people who read it don't mind.

People don’t read it, search engines do.

BTW, we should mandatorily tag generated content for search engines in order to exclude it from future training sets.

Apart from that, hopefully the people building training sets use gltr or something similar to prevent training on generated text.

http://gltr.io/

Well, hopefully, someone will come up with a language model that picks words based on GLTR purpleness. Automated data collection deserves automated data.

I feel your stance [1] is demonstrably false in two challenges.

1) Please play a winning game of Go against Alpha Zero, just by googling the topic.

2) Next please explain how Alpha Zero’s game’s could forever change Go opening theory[2], without any genuine creativity.

[1] that “the output from GPT-3, DALL-E, et al is similar to what you get from googling the prompt and stitching together snippets from the top results.”

[2]”Rethinking Opening Strategy: AlphaGo's Impact on Pro Play” by Yuan Zhou

Op was clearly not talking about Alpha Zero, a different technology made by different people for a different purpose. Instead, they were noting that despite displaying some truly excellent world modeling, GPT-3 is trained on data that encourages it to vomit up rehashes. It's very possible that the next generation will overcome this and wind up completely holding together long-run concepts and recursion, at least if scaling parameters keeps working, but for now it is a real limitation.

GPT-3 writes like a sleepy college student with 30 minutes before the due date; with shockingly complete grasp of language, but perhaps not complete understanding of content. That's not just an analogy, I am a sleepy college student. When I write an essay without thinking too hard it displays exactly the errors that GPT-3 makes.

GPT-3 can’t play Go.

It almost definitely can to some extent given that gpt2 could play chess [0].

0. https://slatestarcodex.com/2020/01/06/a-very-unlikely-chess-...

Retrained (or to be precise: fine-tuned) GPT2 can play chess after training on additional data.

> I think of the value proposition of GPT-X as "what would you do with a team of hundreds of people who can solve arbitrary problems only by googling them?". And honestly, not a lot of productive applications come to mind.

If I was Xi Jinping, I would use it to generate arbitrary suggestions for consideration by my advisory team, as I develop my ongoing plan for managing The Matrix.