OpenAI just launched its new ChatGPT Agent that can make as many as 1 complicated cupcake order per hour, but even Sam Altman says you probably shouldn't trust it for 'high-stakes uses'
Technology
56
Beiträge
34
Kommentatoren
329
Aufrufe
-
OpenAI launched ChatGPT Agent on Thursday, its latest effort in the industry-wide pursuit to turn AI into a profitable enterprise—not just one that eats investors' billions. In its announcement blog, OpenAI says its Agent "can now do work for you using its own computer," but CEO Sam Altman warns that the rollout presents unpredictable risks.
[...]
OpenAI research lead Lisa Fulford told Wired that she used Agent to order "a lot of cupcakes," which took the tool about an hour, because she was very specific about the cupcakes.
I need an agent who would set up DevOps for me. Then robots would definitely be the ones working hard, not humans.
-
Okay but that’s not what easier means.
Easier would be to call the bakery or spending 10 minutes browsing their website, asking to cast, and checking out.
I don’t want to spend an hour on tasks that would normally take 10 minutes. My executive dysfunctions already make me good at doing that.
This might be a revolutionary idea, but what if they helped me do that take an hour in 10 minutes?
I’m just putting that idea out there totally for free in case any AI companies want to jump on that opportunity.
I don’t want to spend an hour on tasks that would normally take 10 minutes.
I don't get it, do you think she spent an hour talking to ChatGPT to try and get it to order doughnuts?
-
It really is a nightmare brewing. And they will hide behind excuses and keep it all opaque unless they are strongly regulated.
regulated by who ?
our senate and congress is filled by pimps who work for pedophiles like epstien and cheer genocider scum murdering children on daily basis. this include the “lesser evil” party. they had 4 years to release the pedo list or even try to slow down the genocide.
they are not gonna give a fck about us working 3 jobs just to pay rent and live on prison food.sad reality is that after a certain threshold in a parasites-host dynamic, there is no other ending other than host dying because parasites has grown too big for it too feed. so unless another deadly parasite like cia or kgb luigi the 1%, the rest 99% are dead.
-
It’s a starting point
I use agents a lot and have written several MCP servers now, the tasks I automate aren't things like order cupcakes, it's mainly the glue between complex things.
I still can't get Claude to nicely open a JIRA ticket for me, but I can get it to read through a sequence of connected documents and filter that into.
I don't think agents are ready for the main event and these are some poor examples of their power.
I'm not saying they won't improve, but using the right tool for the right job is critical. An hour to order cupcakes is silly even for an llm.
-
I spent maybe 90 minutes trying to get ChatGPT to write me a fucking AppleScript or bash to copy all calendar events from a source calendar to a destination. That shit does not work.
for coding you want to use claude
if you don’t want to pay for claude after so many messages what you can do is use mistral to code it up then use claude to proof check the code
-
unfortunately any ai service is going to make things worse.
right now we can discover and choose. with search and browsing dead, ai provider will shove the product giving them the highest cut aka most garbage or snake oil products.even today targeted advertising for poor people is filled with betting, lottery & poker game. similarly elder people are primarily shown ads of miracle cure for chronic illness and scammy religious crap.
edit: switch to kagi. its paid but well worth it.
searchXNG is also a good alternative if you have got time for hosting it urself.Kagi is all in on AI. Its the AI slop version of a search ranking algorithm
-
It won't do that well. What you have to do is ask it to help you leverage your existing development skills in an unfamiliar domain. I used it to help me write a python program to authenticate, pull and filter data from a GCP firestore database and create an XLSX with summary and detail sheets.
I've never used Python before in my life. It took me about 4 hours. Of course I've been doing that sort of thing in Java for many years. Turned out I wrote that faster in Python than I could in Java. Configuring the connection to that database in Python was so simple compared to Java.
The stuff it wrote was sometimes incomplete or wrong in subtle ways, but I could see the bits that didn't make sense which helped me focus on those things and ask better questions to help me figure it out. I think the last hour was just me tweaking stuff by myself because I didn't need help with it by that point.
Anyone who already knows another programming language but has never used python in their life can write a simple python app quickly, regardless
-
I needed about 30 minutes to do a python application from scratch that took linear JSON data files, merged them and presented them as a tree in a GUI.
Before that I had barely done anything in python, basically could do a basic function declaration with a simple operation and nothing else. I even didn't have a lot of experience with UI at all.
But like you I had experience with java and such, and those skills transfer. All it took was searching basic syntax/related code examples and required library imports. And I mean basic, search engine search, not AI answers.
All I'm saying is, I really don't think AI is providing anything a lot more efficient than doing a good old crawl through API docs and stack overflow. So the fact it's using tremendous amounts of resources to maybe achieve a 10% efficiency boost is bothering me a lot.
There’s also the fact that
- It’s only really good at this if you want it to generate Python, PowerShell, bash, or C++ code. Try any other language and it quickly assumes you’re using outdated and often incompatible libraries or doesn’t really understand how the language functions.
- at the end of it all, neither you nor the AI has learned anything new; you’ll have to put in the exact same amount of work the next time. If you do it yourself, then over time that 10% advantage goes away.
Now, these things could both change over time, but humans are much more efficient to train than current state of the art probability sieves we call GenAI.
-
Anyone who already knows another programming language but has never used python in their life can write a simple python app quickly, regardless
No you can't if you don't know the libraries. Python is entirely dependent on what libraries you include. If you don't know what you need you can't do shit.
-
There’s also the fact that
- It’s only really good at this if you want it to generate Python, PowerShell, bash, or C++ code. Try any other language and it quickly assumes you’re using outdated and often incompatible libraries or doesn’t really understand how the language functions.
- at the end of it all, neither you nor the AI has learned anything new; you’ll have to put in the exact same amount of work the next time. If you do it yourself, then over time that 10% advantage goes away.
Now, these things could both change over time, but humans are much more efficient to train than current state of the art probability sieves we call GenAI.
It's only assuming if you aren't specific enough. And you do know their training is usually a year or two or 3 old. So they don't know about whatever new shit your trying to work with.
-
I use agents a lot and have written several MCP servers now, the tasks I automate aren't things like order cupcakes, it's mainly the glue between complex things.
I still can't get Claude to nicely open a JIRA ticket for me, but I can get it to read through a sequence of connected documents and filter that into.
I don't think agents are ready for the main event and these are some poor examples of their power.
I'm not saying they won't improve, but using the right tool for the right job is critical. An hour to order cupcakes is silly even for an llm.
It’s examples for the common guy in the streets who don’t know what an mcp server is.
-
OpenAI launched ChatGPT Agent on Thursday, its latest effort in the industry-wide pursuit to turn AI into a profitable enterprise—not just one that eats investors' billions. In its announcement blog, OpenAI says its Agent "can now do work for you using its own computer," but CEO Sam Altman warns that the rollout presents unpredictable risks.
[...]
OpenAI research lead Lisa Fulford told Wired that she used Agent to order "a lot of cupcakes," which took the tool about an hour, because she was very specific about the cupcakes.
That's quite a bold statement to make since he now has US military contracts. What is he making cupcakes for the Pentagon?
-
No you can't if you don't know the libraries. Python is entirely dependent on what libraries you include. If you don't know what you need you can't do shit.
No you can't if you don't know the libraries
IDE.
Python is entirely dependent on what libraries you include
??
If you don't know what you need you can't do shit.
IDE.
The problems you propose in your comment are not only greatly exaggerated but already been solved for decades using conventional tools AND apply to literally all languages, having nothing at all to do with python. Good try! My statement holds true.
Maybe your assumption is that you're in a cave writing code in pencil on paper, but that's not a typical working condition. If you have access to Claude to use as a crutch, then you have access to search for an available python library and read some "Getting Started" paragraphs.
Seriously, if the only real value that AI provides is "you don't need to know the libraries you're using"
that's not quite as strong of an argument as you think it is lmaooo "knowing the libraries" isn't exactly an existing challenge or software engineering problem that people struggle with... -
So much for the internet. We somehow managed to turn one of humanity’s greatest achievements into a hateful echo chamber we use for warfare first and then into a blackbox where inefficient AI agents communicate with each other in the most inefficient way so the planet can cook us alive even faster. God forbid just calling up a bakery to order some cupcakes.
Companies will dump billions into AI to fuck everyone over but the transition to clean energy is always too expensive.
-
Kagi is all in on AI. Its the AI slop version of a search ranking algorithm
Kagi has AI tools but they don't shove it down your throat. I don't understand what "all in on AI" means in this context. The company has said that they want to use AI like they use JavaScript, ie they want to use it as a tool but their product should work well without it.
-
for coding you want to use claude
if you don’t want to pay for claude after so many messages what you can do is use mistral to code it up then use claude to proof check the code
tx, will try it some time.
-
OpenAI launched ChatGPT Agent on Thursday, its latest effort in the industry-wide pursuit to turn AI into a profitable enterprise—not just one that eats investors' billions. In its announcement blog, OpenAI says its Agent "can now do work for you using its own computer," but CEO Sam Altman warns that the rollout presents unpredictable risks.
[...]
OpenAI research lead Lisa Fulford told Wired that she used Agent to order "a lot of cupcakes," which took the tool about an hour, because she was very specific about the cupcakes.
So now they're raiding We Bare Bears for ideas?
-
No you can't if you don't know the libraries
IDE.
Python is entirely dependent on what libraries you include
??
If you don't know what you need you can't do shit.
IDE.
The problems you propose in your comment are not only greatly exaggerated but already been solved for decades using conventional tools AND apply to literally all languages, having nothing at all to do with python. Good try! My statement holds true.
Maybe your assumption is that you're in a cave writing code in pencil on paper, but that's not a typical working condition. If you have access to Claude to use as a crutch, then you have access to search for an available python library and read some "Getting Started" paragraphs.
Seriously, if the only real value that AI provides is "you don't need to know the libraries you're using"
that's not quite as strong of an argument as you think it is lmaooo "knowing the libraries" isn't exactly an existing challenge or software engineering problem that people struggle with...It sounds like you are a much better developer than me, but to be fair I've had to teach myself everything using nothing but books and Google for thirty years. I've rarely had the luxury of working with someone who had the knowledge to mentor me, and never got a degree outside an AAS in electronics, so I've probably missed some critical skills along the way.
In a lot of ways, the AI fills that role because it's better at answering questions than it is writing code. Earlier today it was explaining to me how a DOM selector could return a stale element in some cases in a failing end to end test. It took a few back and forths with some code examples before I really understood why the selectors might not be working.
It also suggested some code changes that I had to push back on because, even though the code had errors, the errors weren't causing the problem. While building an array of validators I had awaited them, causing them to run serially instead of in parallel during Promise.all(). So you definitely have to know what you're doing to avoid having the AI waste your time (or at least more time than it takes to push back).
I'm still trying to debug it, but without the AI, I'd be googling the fuck out of typescript syntax, JavaScript idiosyncrasies, and a whole testing framework I've never seen before.
So...
if the only real value that AI provides is "you don't need to know the libraries you're using"
...returns false.
-
Did the AI gave you a starting point that would be very different from a bit of code someone submitted 10 years ago on stack exchange? Because in my experience, everything has already been asked and answered. This includes the most basic and naive stuff, and often I am very grateful for it, because, yeah, sometimes I need someone to guide me through the most basic stuff.
In fact, the AI needed that exact knowledge base and a bunch more to exist in the first place. It's just vaguely competent at retrieving it.
Anyway, I didn't say I had no experience, just the most minimal python experience. There are definitely a few quirks I had to learn (the data structures mostly), but for the rest is mostly finding the right method in the reference library, like you would in java.
Logically, you would be right. My practical experience is I waste a lot less time trying to google multiple explanations something because one by itself isn't helping me figure it out, writing bugged PoC test code and thinking something is broken, sorting through a bunch of things that haven't been relevant for 3 versions, etc.
Of course the AI is trained on the same material we can an all find and read, but it does it orders of magnitude more quickly. The trade off is that it's not always right, but neither am I and neither are most sources on the internet right in all circumstances. But it's so fast and easy that I can iterate and evolve designs and understanding much more quickly than I could on my own.
-
So now they're raiding We Bare Bears for ideas?
Explain.
