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

GPT-4o's Memory Breakthrough – Needle in a Needlestack

nian.llmonpy.ai

Post reply on HN
Print view View original

201–210 of 256 posts

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#201

Post by UncleMeat »

CWuestefeld wrote: post #181
sschueller wrote: post #112

We are all so majorly f*d. The general public does not know nor understand this limitation. At the same time OpenAI is selling this a a tutor for your kids. Next it will be used to test those same kids. Who is going to prevent this from being used to pick military targets (EU law has an exemption for military of course) or make surgery decisions?

I hear these complaints and can't see how this is worse than the pre-AI situation. How is an AI "hallucination" different from human-generated works that are just plain wrong, or otherwise misleading? Humans make mistakes all the time. Teachers certainly did back when I was in school. There's no fundamental qualitative difference here. And I don't even see any evidence that there's any difference in degree, either.

"Sorry, computer says no."

Humans can be wrong, but they aren't able to be wrong at as massive of a scale and they often have an override button where you can get them to look at something again.

When you have an AI deployed system and full automation you've got more opportunities for "I dunno, the AI says that you are unqualified for this job and there is no way around that."

We already see this with less novel forms of automation. There are great benefits here, but also the number of times people are just stymied completely by "computer says no" has exploded. Expect that to increase further.

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#202

Post by skywhopper »

CWuestefeld wrote: post #181
sschueller wrote: post #112

We are all so majorly f*d. The general public does not know nor understand this limitation. At the same time OpenAI is selling this a a tutor for your kids. Next it will be used to test those same kids. Who is going to prevent this from being used to pick military targets (EU law has an exemption for military of course) or make surgery decisions?

I hear these complaints and can't see how this is worse than the pre-AI situation. How is an AI "hallucination" different from human-generated works that are just plain wrong, or otherwise misleading? Humans make mistakes all the time. Teachers certainly did back when I was in school. There's no fundamental qualitative difference here. And I don't even see any evidence that there's any difference in degree, either.

Because people know they make mistakes, and aren’t always 100% certain and are capable of referring you to other people. Also because the mistakes LLMs make are entirely unlike mistakes humans make. Humans don’t generate fake URLs citing entirely fake references. Humans don’t apologize when corrected and then re-assert the same mistake. Also because we know that people aren’t perfect and we don’t expect them to be infallible, humans can break out of their script and work around the process that’s been encoded in their computers.

But most people do expect computers to be infallible, and the marketing hype for LLMs is that they are going to replace all human intellectual labor. Huge numbers of people actually believe that. And if you could convince an LLM it was wrong (you can’t, not reliably), it has no way around the system it’s baked into.

All of these things are really really dangerous, and just blithely dismissing it as “humans make mistakes, too, lol” is really naive. Humans can decide not to drop a bomb or shoot a gun if they see that their target isn’t what they expect. AIs never will.

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#203

Post by lnxg33k1 »

CWuestefeld wrote: post #181
sschueller wrote: post #112

We are all so majorly f*d. The general public does not know nor understand this limitation. At the same time OpenAI is selling this a a tutor for your kids. Next it will be used to test those same kids. Who is going to prevent this from being used to pick military targets (EU law has an exemption for military of course) or make surgery decisions?

I hear these complaints and can't see how this is worse than the pre-AI situation. How is an AI "hallucination" different from human-generated works that are just plain wrong, or otherwise misleading? Humans make mistakes all the time. Teachers certainly did back when I was in school. There's no fundamental qualitative difference here. And I don't even see any evidence that there's any difference in degree, either.

Probably the main difference is that humans fail at smaller scale, with smaller effects, and build a reputation, probably chatgpt hallucinations can potentially affect everyone

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#204

Post by whereismyacc »

I always thought it seemed likely that most needle in a haystack tests might run into the issue of the model just encoding some idea of 'out of place-ness' or 'significance' and querying on that, rather than actually saying something meaningful about generalized retrieval capabilities. Does that seem right? Is that the motivation for this test?

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#205

Post by causality0 »

hackerlight wrote: post #101
causality0 wrote: post #64

I don't understand OpenAI's pricing strategy. For free I can talk to GPT 3.5 on an unlimited basis, and a little to GPT 4o. If I pay $20 a month, I can talk to GPT 4o eighty times every three hours, or once every two and a half minutes. That's both way more than I need, and way less than I would expect for twenty dollars a month. I wish they had a $5 per month tier that included, say, eighty messages per 24-hours.

It'll make more sense when they deploy audio and image capability to paying users only, which they say they're going to do in a few weeks

Yeah, but I want a tier where I have access to it in a pinch, but won't feel guilty for spending the money and then going a whole month without using it.

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#206

Post by chazeon »

sschueller wrote: post #112

We are all so majorly f*d. The general public does not know nor understand this limitation. At the same time OpenAI is selling this a a tutor for your kids. Next it will be used to test those same kids. Who is going to prevent this from being used to pick military targets (EU law has an exemption for military of course) or make surgery decisions?

It seems US and China are trying to reach an agreement to use AI to pick military targets these days.

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#207

Post by moralestapia »

CWuestefeld wrote: post #181
sschueller wrote: post #112

We are all so majorly f*d. The general public does not know nor understand this limitation. At the same time OpenAI is selling this a a tutor for your kids. Next it will be used to test those same kids. Who is going to prevent this from being used to pick military targets (EU law has an exemption for military of course) or make surgery decisions?

I hear these complaints and can't see how this is worse than the pre-AI situation. How is an AI "hallucination" different from human-generated works that are just plain wrong, or otherwise misleading? Humans make mistakes all the time. Teachers certainly did back when I was in school. There's no fundamental qualitative difference here. And I don't even see any evidence that there's any difference in degree, either.

Humans know when they've made a mistake. So there's ways to deal with that.

Computers are final. You don't want things to be final when your life's on the line.

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#208

Post by currymj »

kromokromo wrote: post #122
sschueller wrote: post #112

We are all so majorly f*d. The general public does not know nor understand this limitation. At the same time OpenAI is selling this a a tutor for your kids. Next it will be used to test those same kids. Who is going to prevent this from being used to pick military targets (EU law has an exemption for military of course) or make surgery decisions?

This is just doomerism. Even though this model is slightly better than the previous, using an LLM for high risk tasks like healthcare and picking targets in military operations still feels very far away. I work in healthcare tech in a European country and yes we use AI for image recognition on x-rays, retinas etc but these are fundamentally completely different models than a LLM. Using LLMs for picking military targe…

i don't know about European healthcare but in the US, there is this huge mess of unstructured text EMR and a lot of hope that LLMs can help 1) make it easier for doctors to enter data, 2) make some sense out of the giant blobs of noisy text.

people are trying to sell this right now. maybe it won't work and will just create more problems, errors, and work for medical professionals, but when did that ever stop hospital administrators from buying some shiny new technology without asking anyone.

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#209

Post by Sophira »

CWuestefeld wrote: post #181
sschueller wrote: post #112

We are all so majorly f*d. The general public does not know nor understand this limitation. At the same time OpenAI is selling this a a tutor for your kids. Next it will be used to test those same kids. Who is going to prevent this from being used to pick military targets (EU law has an exemption for military of course) or make surgery decisions?

I hear these complaints and can't see how this is worse than the pre-AI situation. How is an AI "hallucination" different from human-generated works that are just plain wrong, or otherwise misleading? Humans make mistakes all the time. Teachers certainly did back when I was in school. There's no fundamental qualitative difference here. And I don't even see any evidence that there's any difference in degree, either.

Society has spent literal decades being convinced to put their trust in everything computers do. We're now at the point that, in general, that trust is there and isn't misplaced.

However, now that computers can plausibly do certain tasks that they couldn't before via LLMs, society has to learn that this is an area of computing that can't be trusted. That might be easy for more advanced users who already don't trust what corporations are doing with technology[0], but for most people this is going to be a tall order.

[0] https://i.imgur.com/6wbgy2L.jpeg

Re: GPT-4o's Memory Breakthrough – Needle in a Needlestack

#210

Post by SiempreViernes »

ignoramous wrote: post #193
nopromisessir wrote: post #187

Earlier quoted context omitted.

[flagged]

> a sophisticated approach to their defense A euphemism for apartheid and oppression? > sources are rife with bias What's biased about terming autonomous weapons as "AI"? Or, sounding alarm over dystopian surveillance enabled by AI? > The nuance matters. Like Ben Gurion terming Lehi "freedom fighters" as terrorists? And American Jewish intellectuals back then calling them fascists? > The history matters... After that…

The nuance of Ben-Gvir presumably ...
Post reply on HN