Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GPT-5 is behind schedule

wsj.com

Post reply on HN
Print view View original

861–870 of 1001 posts

Re: GPT-5 is behind schedule

#861

Post by skydhash »

ben_w wrote: post #816
cardanome wrote: post #805

Earlier quoted context omitted.

Being biased is not the same as hallucinating. LLMs have both problems. At least you could check whether a source was reputable and where the bias was. With LLM's the connection between the answer and the source is completely lost. You can't even tell why it answered a certain way.

> Being biased is not the same as hallucinating. LLMs have both problems. I didn't deny either of those things, I said that search engines also hallucinate — my actual link gave several examples, including "King of the United States" -> "Barack Obama". Just because it showed the link to breitbart doesn't mean it was not hallucinating . > At least you could check whether a source was reputable and where the bias was.…

> I said that search engines also hallucinate — my actual link gave several examples

They don't. Google added a weird widget that do hallucinate. But the result list is still accurate, even though it may be biased towards certain sources.

> You could tell where a search engine got an answer from, but not which answers were hidden

A bit pedantic, but a search engine returns a list of results according to the query you posted. There's no question-answer oracle. If you type "King of the United States", you will get pages that have the terms listed. Maybe there will be semantic manipulations like "King -> Head of state -> President", but generally it's on you to post the correct keywords.

Re: GPT-5 is behind schedule

#862

Post by codingdave »

netdevphoenix wrote: post #817
kevinsync wrote: post #797

Earlier quoted context omitted.

I actually find LLMs lacking true expertise to be a feature, not a bug. Most of the time I'm starting from a place of no knowledge on a topic that's novel to me, I ask some questions, it replies with summaries, keywords, names of things, basic concepts. I enter with the assumption that it's really no different than googling phrases and sifting through results (except I don't know what phrases I'm supposed to be googl…

You forget that it makes stuff up and you won't know it until you google it. When googling, fake stuff stands out because truth is consistent. Querying multiple llms at the same time and being able to compare results is a much better comparison to googling but no one does this. As I said, you are talking to a super confident journalist intern who can give you answers but you won't know if it is true or partially true…

LLMs train from online info. Online info is full of misinformation. So I would not trust an answer to be true just because it is given by multiple LLMs. That is actually a really good way to fall into the misinformation trap.

Re: GPT-5 is behind schedule

#863

Post by SamPatt »

svieira wrote: post #737
SamPatt wrote: post #726

Earlier quoted context omitted.

They aren't just in the hands of big corporations though. The open source, local LLM community is absolutely buzzing right now. Yes, the big companies are making the models, but enough of them are open weights that they can be fine tuned and run however you like. I think LLMs genuinely do present an opportunity to be neutral experts, or at the least neutral third parties. If they're run in completely transparent ways…

The whole problem is that they are not neutral. They token-complete based on the corpus that was fed into them and the dimensions that were extracted out of those corpuses and the curve-fitting done to those dimensions. Being "completely transparent" means exposing _all_ of that, but that's too large for anyone to reasonably understand without becoming an expert in that particular model. And then we're right back to…

Nothing is truly neutral. Humans all have a different corpus too. We roughly know what data has gone in, and what the RL process looks like, and how the models handle a given ethical situation.

With good prompting, the SOTA models already act in ways I think most reasonable people would agree with, and that's without trying to build this specifically for that use case.

Re: GPT-5 is behind schedule

#864

Post by josu »

EVa5I7bHFq9mnYK wrote: post #791
jpablo wrote: post #753

Earlier quoted context omitted.

They aren't blocking anything. They are just asking nicely not to be crawled. Given that AI companies haven't cared a single bit about ripping of other's peoples data I don't see why they would care now.

In their attempt to block OpenAI, they block me. Many sites that were accessible just 2 years ago, require login/captchas/rectal exam now just to read the content.

> captchas

I suspect that AIs are already more effective than humans at passing captchas.

Re: GPT-5 is behind schedule

#865

Post by bubaumba »

wokwokwok wrote: post #795
bubaumba wrote: post #671

Earlier quoted context omitted.

> if the underlying LLMs dont get any better, there is no reason to expect the system built out of them to get any better. Actually o1, o3 are doing exactly this, and very well. I.e. explicitly: by proper orchestration the same LLM can do much better job. There is a price, but... > you would expect to be able to build agentic systems out of much smaller LLMs Good point, it should be possible to do it on a high-end pc…

> but that overwhelmingly doesn’t work. MCTS will be the next big “thing”; not agents.

They are not mutually exclusive. Likely we'll get more clear separation of architecture and underlying technology. In this case agents (i.e. architecture) can use different technologies or mix of them. Including 'AI' and algorithms. The trick is to make them work together.

Re: GPT-5 is behind schedule

#866

Post by CamperBob2 »

jaybna wrote: post #755
jpablo wrote: post #753

Earlier quoted context omitted.

They aren't blocking anything. They are just asking nicely not to be crawled. Given that AI companies haven't cared a single bit about ripping of other's peoples data I don't see why they would care now.

Yeah, probably right. If you want a great rabbit hole, look up "Common Crawl" and see how a great academic project was absolutely hijacked for pennies on the dollar to grab training data - the foundation for every LLM out there right now.

It's hard to envision a greater success for the "great academic project" than what happened. I mean, what else were they trying to accomplish?

Re: GPT-5 is behind schedule

#867

Post by a_wild_dandan »

solardev wrote: post #732
LASR wrote: post #610

So the team I lead does a lot of research around all the “plumbing” around LLMs. Both technical and from a product-market perspectives. What I’ve learned is that for the most part that AI revolution is not going to be because of PHD-level LLMs. It will be because people are better equipped to use the high-schooler level LLMs to do their work more efficiently. We have some knowledge graph experiments where LLMs contin…

> a revolution in knowledge-worker productivity. That's a nice euphemism for "imminent mass layoffs and a race to the bottom"...

This conclusion is the lump of labor fallacy. It's not that simple.

Re: GPT-5 is behind schedule

#868

Post by jasondigitized »

jaybna wrote: post #741

25% of the top 1000 websites are blocking OpenAI from crawling: https://originality.ai/ai-bot-blocking I am betting hundreds of thousands, rising to millions more little sites, will start blocking/gating this year. AI companies might license from big sources (you can see the blocking percentage went down), but they will be missing the long tail, where a lot of great novel training data lives. And then the big sites w…

The amount of content coming off of YouTube every minute puts Google in a very enviable position.

Re: GPT-5 is behind schedule

#869

Post by jasondigitized »

netdevphoenix wrote: post #640
atonse wrote: post #457

Earlier quoted context omitted.

Same here. The ability to “talk to an expert” about any topic I’m curious about and ask very specific questions has been invaluable to me. It reminds me of being a kid and asking my grandpa a million questions, like how light bulbs worked, or what was inside his radio, or how do we have day and night. And before anyone talks about accuracy or hallucinations, these conversations usually are treated as starting off poi…

> The ability to “talk to an expert” about any topic I’m curious about and ask very specific questions has been invaluable to me. It is dangerous to assume that LLMs are experts on any topic. With or without quotes. You are getting a super fast journalist intern with a huge memory but inability to reason critically, lacking understanding about anything and huge unreliability when it comes to answering questions (you…

It's dangerous to assume that the person you have access to is an expert either.

Re: GPT-5 is behind schedule

#870

Post by ben_w »

glenstein wrote: post #839
ben_w wrote: post #824

Earlier quoted context omitted.

> To your point, I have wondered whatever became of that massive initiative from Google to scan books, and whether that might be looked at as a potential training source, giving that Google has run into legal limitations on other forms of usage. Still around, doing fine: https://en.wikipedia.org/wiki/Google_Books and https://books.google.com/intl/en/googlebooks/about/index.htm... Given the timing, I suspect it was st…

I don't know what you mean by timing (relative to what?) or "simple indexing" (they scanned the complete contents of books), but I am, and was already aware, of the wiki article and the role of recaptcha. Maybe I wasn't clear, but I was interested in the consequences of the legal stuff. It's not clear from the wiki article what any of this means with respect to the suitability of scans for AI training.

> I don't know what you mean by timing (relative to what?) or "simple indexing" (they scanned the complete contents of books), but I am, and was already aware, of the wiki article and the role of recaptcha.

Timing as in: it started in 2004, when the most advanced AI most people used was a spam filter, so it wasn't seen as a training issue (in the way that LLMs are) *at the time*.

As for training rights, I agree with you, there's no clarity for how such data could be used *today* by the people who have it. Especially as the arguments in favour of LLM training are often by comparison to search engine indexing.

Post reply on HN