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

Operator research preview

openai.com

Post reply on HN
Print view View original

261–270 of 448 posts

Re: Operator research preview

#261

Post by Workaccount2 »

er4hn wrote: post #257
karpathy wrote: post #254

Earlier quoted context omitted.

Thank you for sharing your concerns. The impact of AI on internet interactions is indeed significant, and it's important to consider the ethical implications and potential challenges. Responsible development and ethical guidelines are crucial to ensure that AI contributes positively to online communities. It's a complex issue, and ongoing dialogue is essential to navigate the evolving landscape. (Posted by OpenAI Ope…

In karpathy's defense, this actually seems great. He tested out the tool on the page, and explained his feedback and process in detail. He was explicit in the followup comment about it. This seems like a great demo example for trying out a new technology in a way where it interacts with something we are all familiar with (this comments page)

In swatcoders defence, he likely loaded the comments between the time that karpathy posted the AI comment and the time he posted the follow-up (there is a 5 minute gap). The timings on the comments seem to suggest that possibility at least.

Re: Operator research preview

#262

Post by TeMPOraL »

ItsMattyG wrote: post #198
krapp wrote: post #193

Earlier quoted context omitted.

McDonald's already tried having AI take orders and stopped when the AI did things like randomly add $250 of McNuggets or mistake ketchup for butter. Note - because this is something which needs to be pointed out in any discussion of AI now - even though human beings also make mistakes this is still markedly less accurate than the average human employee.

For now

Indeed. I think a GPT-4o class model, properly prompted, would work just fine today. The trick is, unlike a human, the computer is free to just say "no" without consequences. The model could be aggressively prompted to detect and refuse weird orders. Having to escalate to a human supervisor (who conveniently is always busy doing other things and will come to you in a minute or three) should be sufficient at discouraging pranksters and fraudsters, while not annoying enough to deter normal customer.

(I say model, but for this problem I'd consider a pipeline where the powerful model is just parsing orders and formulating replies, while being sanity-checked by a cheaper model and some old-school logic to detect excessive amounts or unusual combinations. I'd also consider using "open source" model in place of GPT-4o, as open models allow doing "alignment" shenanigans in the latent space, instead of just in the prompts.)

Re: Operator research preview

#263

Post by TeMPOraL »

franktankbank wrote: post #217
reissbaker wrote: post #201

Earlier quoted context omitted.

This seems unreasonably pessimistic (or unreasonably optimistic in OpenAI's moat?). There are so, so many companies competing in this space. The cost will reflect the price of the hardware needed to run it: if it doesn't, they'll just lose to one of their many competitors who offer something similar for cheaper, e.g. whatever DeepSeek or Meta releases in the same space, with the cost driven to the bottom by commoditi…

The data is the moat.

Not for much longer, perhaps not even now. There's plenty of data avaliable to anyone, and people are finding ways to use that data more effectively.

Mid-term, I believe the only real moat is going to be human labor - that is, RLHF and other funny acronyms that boil down to getting people to chat with the model and rate how they feel about its answers.

Software improvements (architecture, training process, inference) are always one public paper or leak away from being available for free to anyone. Hardware improvements will spread too, because NVIDIA et al. would prefer to sell more chips than less chips. Meanwhile, human labor is notoriously expensive, only getting more expensive as economic conditions of people improve, and most importantly, whatever "spark" of human intelligence/consciousness there is, this is where it cannot be automated away - not until we get to human-level AGI.

Human labor is the one thing that you can only scale by throwing more money at it - which is why modern businesses seek to remove it from the equation as much as possible. Hell, the whole pursuit of AGI is in big part motivated by hope of eliminating labor costs entirely. Except, in this one pursuit, until AGI is reached, labor is a critical resource that has no substitute.

That's my mid-term prediction. Long-term, we'll hit AGI and moats won't matter anymore.

Re: Operator research preview

#264

Post by energy123 »

brap wrote: post #31

I don't know why, but the approach where "agents" accomplish things by using a mouse and keyboard and looking at pixels always seemed off to me. I understand that in theory it's more flexible, but I always imagined some sort of standard, where apps and services can expose a set of pre-approved actions on the user's behalf. And the user can add/revoke privileges from agents at any point. Kind of like OAuth scopes. Ima…

I am more interested in Gemini's "Deep Research" feature than Operators. As a ChatGPT subscriber I wish they'd build a similar product.

Even when it comes to shopping, most of the time I spend is in researching alternatives according to my desired criteria. Operators doesn't help with that. o1 doesn't help because it's not connected to the internet. GPT-4o doesn't help because it struggles to iterate or perform > 1 search at a time.

Re: Operator research preview

#265

Post by swatcoder »

Workaccount2 wrote: post #261
er4hn wrote: post #257

Earlier quoted context omitted.

In karpathy's defense, this actually seems great. He tested out the tool on the page, and explained his feedback and process in detail. He was explicit in the followup comment about it. This seems like a great demo example for trying out a new technology in a way where it interacts with something we are all familiar with (this comments page)

In swatcoders defence, he likely loaded the comments between the time that karpathy posted the AI comment and the time he posted the follow-up (there is a 5 minute gap). The timings on the comments seem to suggest that possibility at least.

Excellent forensics! You are correct.

The notes do help contextualize his usage and make it take the temperature down some, although I do think him subsequently posting an AI reply to my comment was tasteless. (But I also get it. I used harsh words there and invited some ribbing in return.)

Re: Operator research preview

#266

Post by throwaway314155 »

gyomu wrote: post #255
karpathy wrote: post #245

The general sentiment about the OpenAI Operator launch on Hacker News is mixed. Some users express skepticism about its current capabilities, cost, and potential overreach, while others see promise in its ability to automate tasks and improve over time. Ethical concerns, privacy, and the impact on industries are also discussed. Overall, there's cautious optimism with acknowledgment of challenges and potential improve…

@dang can we have guidelines against posting AI generated content here? (who cares if the account is "human operated" or has a disclaimer). It's just lame and not what this forum is about.

It's absurd to me that simply because it's Karpathy's account, everyone immediately changes their tune about posting "slop" AI summaries; something which is normally downvoted into oblivion on this site.

People go even further to downvote any criticism?? Pick a lane people. This will be business as usual in a week and Operator posts will go back to being thoroughly downvoted by then too.

Re: Operator research preview

#267

Post by mceoin »

minimaxir wrote: post #8

Overall, Operator seems the same as Claude's Computer Use demo from a few months ago, including architecture requiring user to launch a VM, and a tendency to be incorrect: https://news.ycombinator.com/item?id=41914989 Notably, Claude's Computer Use implementation made few waves in the AI Agent industry since that announcement despite the hype.

Claude's Computer Use API has been good for us and I'm surprised it isn't more popular. It can be slow, and definitely gets things wrong, but so far we've had thousands of people make and edit spreadsheets on "autopilot" so the value is already there today on simple tasks, even in an alpha state.

I do find it is best when combined with other capabilities so the internal reasoning is more "if Computer Use is the best for solving this stage of the question, use Computer Use. Otherwise, don't.", instead of full Computer Use reliance. So e.g. you might see it triggered for auto-formatting but not writing SQL.

Will report back how it compares vs Operator CUA once we get access!

Re: Operator research preview

#268

Post by ActorNightly »

willmarch wrote: post #197
ActorNightly wrote: post #144

Earlier quoted context omitted.

When 4o came out with its chain of thought, people thought this is it. And today, nobody really cares. Its just another LLM. Same thing with this. The other day I was writing some code to compute some geometric angles, and I was getting 2 different results for what I though was the same angle, but in fact I didn't realize that these angles should not be equivalent. No LLM was able to tell me the issue, they just said…

4o models don't have chain of thought, are you thinking of o1 perhaps?

yep you are right, got the name confused.

Re: Operator research preview

#269

Post by panarky »

darioush wrote: post #151
mattstir wrote: post #102

Earlier quoted context omitted.

I can sympathize with vague notions of AI dystopia, but this might be stretching the concept a bit too far. This kind of service is extremely abusable ("Operator, go to Wikipedia and start mass-vandalizing articles" or "Go to this website and try these people's email addresses with random passwords until it locks their accounts") and building some alignment goals into it doesn't seem like a terribly draconian idea. A…

You can also write a python script to achieve the same goals. Except it's not python's responsibility to interpret the intent of your script, just as it's not your phone's responsibility to interpret the contents of your conversation. So our tools are not our morality police. We have a legal system that can operate within the bounds of law and due process. I am well aware of the already applied levels of machine lear…

Even though you have the power to ignore stop signs when you're driving your own car, it's not an unreasonable restriction of your liberty when AI-driven cars stop at stop signs.

Re: Operator research preview

#270

Post by thrance »

transcriptase wrote: post #250
thrance wrote: post #143

Earlier quoted context omitted.

Can't you fathom Trump being criticized organically? You know, the convicted felon that just pardoned the neonazis that attempted to overthrow the government 4 years ago, that undid 60 years of progress by repelling the Civil Rights Act, that appointed his corrupt billionaire friends to every positions in this government but then rants on about how "DEI is taking away our meritocracy".

I can. But it doesn’t change the fact that nearly all the default subs are run by the same mods, who give blatant manipulation a pass as long as they agree with it politically.

I'm baffled by the fact this is the first case of "manipulation" you mention, when X is many orders of magnitude worse and more harmful to our democracies. But perhaps you don't care because you agree with its sieg-heiling owner?
Post reply on HN