Hacker News Reader

Comment by blintz | Hacker News Reader

I'm not an expert, but I see the main threat to continued improvement as running out of high-quality data. LLM's are a cool thing you can produce only because there is a bunch of freely available high-quality text representing (essentially) the sum of human knowledge. What if GPT-4 (or 5, or 6) is the product of that, and then further improvements are difficult? This seems like the most likely way for improvement to slow or halt; the article cites synthetic data as a fix, but I'm suspicious that that could really work.

View on HN

Googler, but opinion are my own.

More than that, I believe we will hit a ceiling when the impossibility of these models to incorporate causality becomes evident.

Right now, LLMs are trained by predicting the next word given a context (the prompt). This approach, IMHO, gives a resemblance of cause/effect because the training data is made by humans and obviously we are able to express ourselves and reason in those terms. So we have a poor proxy for that which, also IMHO, partially explains why LLMs performance degrades when asked to solve novel problems (there was an entry few days ago about this).

After seeing LangChain and the reasoning paper I think it's fairly obvious that we're just starting to scratch the surface of AI architectures, dictated largely by scalability of GPU resources. The LLMs we're playing with are at best the proof of concept of what will eventually be the message passing for the next generation of models.

Links please..

1) langchain paper? 2) reasoning paper?

Thanks!

It's like worrying about running out of stables for horses when cars already started to take over.

It's going to shift from requiring quality data to producing quality data.

There is not going to be need for it just like there is no need for more Kasparovs anymore.

You can see it with artists slowly happening.

Lawyers will get the hit (or great tool to use depending how you want to see it).

Medical analysis will find the same faith. If you think that medicine will require human analysis just imagine for a second what closed loop AI could do - if it had access to data not only from all hospitals but patients as well and being able to munch through it continuosly. This together with constant access to simulations and physical trials, finding patterns and correlations on its own.

It's exciting and depressing to think that humans will evantually be left to being humans the way they wish with everything sorted out - just like chickens don't mind being free range chickens.

There is no other direction it can go and it'll just go faster.

Our generation will see some mind blowing things.

Next generation will arrive at the world unlike anything else.

> It's going to shift from requiring quality data to producing quality data.

synthetic data gives you synthetic results.

To train something requires decent quality input, otherwise it's going to mimic the crap quality stuff. There is little chance in getting around that.

Have you ever stopped to wonder _why_ large for profit companies give away free models with weights? its because once trained, they are commodity. the hard part is dataset management.

Yes, chatgpt is largely self supervised, but the training into a usable model required a fucktonne of human hours

You're missing self enhancing feedback loop it creates. Human hours are/were necessary to bootstrap.

Even just finetuning these models, they will pick up on the weirdest things a human wouldn't see, and artificial datasets will contain all kinds of invisible artifacts that further "inbreeding" is going to massively amplify.

I am not speaking speculatively either. I have seen it happen finetuning ESRGAN on previous upscales that I even personally vetted as "good," and these generative models are way more sensitive than the old GANs.

I've just spent a 4-5 hours session with ChatGPT trying to fix a problem with `aiohttp`.

The specific problem was what while I'm in a WebSocket handler, I cannot await on asynchronous generators which act like `asyncio.sleep()` and send messages to the client while I have `heartbeat` on that `WebSocketResponse` enabled (it sends pings to the client and waits for pongs to see if the connection is "alive"). The issue is that the incoming pong is not getting processed so the client gets disconnected.

I struggled a lot with ChatGPT's help, but it was mostly insightful, like rubberducking with a duck that does actually try to help you.

Occasionally I went to Google to search for a very specific way of solving the issue; for me Google is the gateway to Stack Overflow, I never use Stack Overflow's search.

But I didn't find anything of value there. And earlier today I noticed that I had unread messages in Stack Overflow, and when I checked how many consecutive days I have been on SO: 1. I used to be 100+ consecutive days on SO and lately I'm struggling with using it as the to-go place for my programming questions.

And this is where your comment comes in: I was thinking to myself what is going to happen if more people move away from Stack Overflow, that place is a goldmine of programming information, who will feed it?

One thing I really hope is that OpenAI adds some "modes", like what were poised to expect from Copilot X, where we get different layouts for the webpage, so that we can choose a "programmer mode" which actually lets us give proper feedback in the sense that "this code worked", "that one didn't".

There's already a feedback, but it's not task-oriented. A programmer can submit to it as well as a cook, the result won't go into a "database of usable code" which ideally would be publicly accessible or even get formed into a Question/Answer pair which is prepared to be submitted to Stack Overflow.

It may become a norm to write AI QAs instead of docs for projects. Maybe similar to sitemaps for web scrapers. But even this will not be necessary at some point.

Stack Overflow will degrade to single search box. And we'll all love it.

That we are running out of data makes continued improvement more likely, in the medium/long term, because it means we have found close to the maximum amount of static data that models need to be trained on. This means that one of the factors for compute requirements (static training data size) has reached its upper bound while still within our computational capacity today, and further improvements must come from elsewhere.

So yes, easy data scaling is coming to an end and may or may not lead to a short term winter. But this also means that, all being equal, training a model will be cheaper compared to a situation where we still had orders of magnitude more data to go.

I see no lack of high-quality data, although it may not be the kind you prefer. How about the telephone/chat/internet logs (not "metadata") of every single person in the USA? We've got it! And how about using that to answer social science questions and, more immediately, political questions such as:

- What really happens in corporate/government/private decision-making?

- Did a particular politician really trade his vote for money? How was that done?

- Are elected judges better/more honest than appointed judges?

- Are there hidden organizations within particular governmental entities? e.g., a right-leaning group within the FBI who secretly persecutes persons/groups of other political persuasions? Are there independent entities within the CIA that are capable of financing and operating on their own under the umbrella of the government but also capable of evading the congressional oversight specified in the Constitution?

- Is my wife's cousin really screwing Hunter Biden again?

Reply on news.ycombinator.com