API updates include 4x larger conversation memory for GPT-3.5 and function calling.
See full article...
See full article...
OpenAI rolls out big chatbot API upgrades for developers
- Thread starter JournalBot
- Start date
They need to launch a model that includes knowledge of API documentation / libraries at the same time they launch the API. Two months back GPT could only provide irrelevant code.
Upvote
24
(27
/
-3)
I don't know why I bother looking at the Ars comment section for AI news since its all the same cynical, snarky comment repeated ad nauseam.
GPT-4 could probably generate a more valuable comment section for this topic than the Ars community.
Upvote
122
(140
/
-18)
Or use any model with large enough context limit to include relevant, or whole sale API documentation. No need to pre-train, or fine tune API stuff.
Large context limits really allow these models to do some crazy stuff. Ive been feeding the smaller gpt4 directions to coding repositories like ss64.com and getting very positive results in coherency, and reduced hallucination of code.
Upvote
27
(29
/
-2)
Hype is frequently followed by the dunning Kreuger effect, and memes.
A hell of a combination
Upvote
38
(40
/
-2)
I'm too am just sick of Ars comment sections being filled with bitter, cynical luddites.
It's "cool" on Ars these days to just hate on every new, interesting thing there is with snarky and dismissive comments.
I suspect the people making these comments get an ego boost from them - perhaps it makes them feel as if they are somehow better/smarter than those that build something cool, if they take a dump all over their creation.
Last edited:
Upvote
59
(75
/
-16)
I have a few projects that use this API. Increased context length is great, price cuts are always welcome, but the big thing here is first-class support for function calling, especially auto calling (gpt figures out if it should call for help).
This makes it trivial to extend and integrate GPT, compared to unreliable gross hacks we had to do before.
As a quick example, I can now just tell gpt to call my web browsing function if it is asked about any kind of fact, person, event, etc. My custom code then checks online and feeds it actual info. This cuts down hallucinations considerably. It can still get things wrong, but now that's more a function of most of the internet being a cesspool of spam...
This makes it trivial to extend and integrate GPT, compared to unreliable gross hacks we had to do before.
As a quick example, I can now just tell gpt to call my web browsing function if it is asked about any kind of fact, person, event, etc. My custom code then checks online and feeds it actual info. This cuts down hallucinations considerably. It can still get things wrong, but now that's more a function of most of the internet being a cesspool of spam...
Upvote
49
(50
/
-1)
Anecdotal, but I work with a team of recruiters with zero IT background, less than zero interest in fad technology and are using GPT daily. Yes, it hallucinates, but that’s not hard to deal with a basic knowledge of what they’re doing.
The guy who usually types with two fingers is now capturing all his thoughts with speech to text then cleaning it up with ChatGPT. All he has to type is a bit of cleanup. Much faster for him.
Upvote
47
(49
/
-2)
I was gonna say "the cynics are out in full force today" but then I read your comment, and it would have just been redundant.
Upvote
14
(18
/
-4)
Looks more like the anti-cynics are outnumbering them so far. Incidentally, commenting on the comment section might need to move to its own "meta-comment" thread. I don't know what others think, but generally I don't think that class of comments adds much to discussions. (Yea I know this is a useless meta-comment too in that sense, guilty as charged.)
Upvote
8
(16
/
-8)
Even GPT3.5 can be really useful with a 16k prompt. Especially if you generate it dynamically to include relevant information.
Upvote
1
(3
/
-2)
It's pretty clear that, given the state of the economy today, these kind of systems will just accelerate wealth concentration and cheapen creative jobs. We've seen it happen before.
Are you ready to get swamped with even more spam content, difficult to sort automatically from real, useful, human generated content on the internet? Do you want writers to get laid off with the profits going to owners recycling old stories?
The Luddites were right. Maybe you should be more cynical.
Upvote
17
(36
/
-19)
I feel there's a general cynical sentiment in the media against tech, mainly in the past 10 years or so. Probably a consequence of tech companies becoming so large and powerful that leads to this explicit push to "reel them in", and that sentiment probably rubs off on people too. It brings everyone down.
I remember the 90s and early 2000s news coverage being almost laughably optimistic about every tech gizmo that came out. Would be nice if there was some sort of middle ground.
Last edited:
Upvote
17
(19
/
-2)
chaos215bar2
Ars Praefectus
Tech of the 90s and 2000s wasn’t perfect, but it certainly wasn’t the exploitative mess we have now.
There’s a lot of cool stuff out there, and it’s important to recognize that and get past the cynicism for a moment. But then, in the current regulatory and political landscape, it’s pretty hard to see most of what’s coming out today actually improving quality of life for most people, rather than just further enriching already wealthy investors.
When everything is a service, everything must be monetized, and every bit of personal data must be sold to the highest bidder, one could be forgiven for being a bit cynical.
Upvote
35
(37
/
-2)
So, I get fed up with contriving ML to be AI. Chat-GPT is ML not actual AI. We are not even close to having true AI systems and I'm not sure we would actually want actual AI around. Does ML provide good benefits with proper data sets? Yep. Does Chat and the drawing ones prove that something potentially good can be done, yet. But these current models are definitely not ready for real time and should not have been deployed as widely as they are being now.
Am I cynical. Damn skippy, because all I see with the latest fad called Chat GPT is just another pile of people with a get rich scheme and damned the consequences of releasing these ML tools without considering the consequences.
Upvote
-6
(15
/
-21)
Expansion of the size of tokens for input is super helpful in workloads like detailed summarization of meeting recordings or other formats where you might want to have a transcript and detailed summary. It allows the summary to have good details while allowing it to flow in a way that isn't as abrupt as it can be now. Now it's handled by chunking the data, summarizing pieces of the meeting, then summarizing the summaries which tends to lose detail, and becomes pretty stiff.
Upvote
6
(7
/
-1)
I think this stuff is overhyped too, but ML is still a tech straight out of AI theory and research and fits under the AI umbrella. ML is AI.
Upvote
17
(17
/
0)
Refugees from Twitter, where you get rewarded by dopamine hits in the form of likes for constantly complaining about everything.
Upvote
-3
(7
/
-10)
SubWoofer2
Ars Tribunus Militum
Coupled with voice to text input, sounds like you've got TIM from The Tomorrow People or any other "Ship" computer from any TV sfnal series. Bless! And where does a non-coder buy?
Upvote
2
(2
/
0)
Curious, if you want the chatbot to remain in the Grimace character in subsequent responses, will you have to include this system message in every call? If yes, it could eat up a lot of tokens fast
Upvote
1
(1
/
0)
Shilling lame youtube videos isn't going to help too much. He only has slightly more charisma than Sam Altman.
Upvote
2
(4
/
-2)
Good news.
Google knows a lot about us from search history, web analytics, YouTube preferences, map usage, etc. Talking with a bot might be even more revealing. People basically directly say to a corporation what interests them and what they are working on. It also makes it easier to steer them in a specific direction.
With better models and new hardware maintenance cost is likely to go down over time.
Eventually corporations will find other ways to monetize the chat bots.
Google knows a lot about us from search history, web analytics, YouTube preferences, map usage, etc. Talking with a bot might be even more revealing. People basically directly say to a corporation what interests them and what they are working on. It also makes it easier to steer them in a specific direction.
With better models and new hardware maintenance cost is likely to go down over time.
Eventually corporations will find other ways to monetize the chat bots.
Upvote
1
(1
/
0)
Yup. For reasons linked to my neurodiversity I have a PA who supports me with planning, admin, travel bookings, letters to various bodies etc. She told me she now uses chatGPT extensively for helping her handle my requests.
So I’m paying someone to type things into chatGPT on my behalf
I’m happy with it. It’s helping her with her English and other tricky stuff and she’s much more productive with it. Her understanding of what I need, her knowledge of what to ask chatGPT and her ability to handle the output appropriately and convert it to meaningful actions that make a difference to my life is what I pay her for.
Upvote
12
(13
/
-1)
Yes, and this is exactly why a larger context limit is such a big deal!Curious, if you want the chatbot to remain in the Grimace character in subsequent responses, will you have to include this system message in every call? If yes, it could eat up a lot of tokens fast
Upvote
1
(2
/
-1)
"The company says that developers can continue to use these models until September 13, after which the older models will no longer be accessible."
Artificial Obsolescence! Just like WIndows! Your API products shall be temporal and cancelled at any time?
Artificial Obsolescence! Just like WIndows! Your API products shall be temporal and cancelled at any time?
Upvote
0
(1
/
-1)
If people keep using it it needs to be supported which creates overhead
Upvote
0
(0
/
0)
I think this comment misses the mark a bit. Blaming the tech when really the system is a fault. No new tech would need to come out for the next decade for the wealthier to get wealthier.
And the late 90s early 2000s comment is either rose colored glasses or we grew up in different classes of neighborhoods. Where I lived people who had computers always had the classic one two punch of netzero and bonzi buddy - pretty exploitative. And many who had a computer got them through rent to own stores and ended up paying 1000% of the list price for them over a long time.
That was when your computer would be actually obsolete in two years too, and they’d still be a ways from paying it off.
Upvote
-1
(2
/
-3)
To some extent: we chose them for convenience.
Windows is basically a service, but Linux isn't.
My home automation is Zixbee with open-source control software running on a Pi.
There are private search engines.
But yes: I generally agree with you. We haven't really started to deal with the privacy issues we've created.
Upvote
0
(0
/
0)
What about giving it browser based access to an extremely well vetted data pool such as UpToDate? (Essentially peer reviewed doctor Wikipedia). Do you see that as possible with the function calling? Asking for a doctor friend.
Upvote
0
(0
/
0)
Yup. For reasons linked to my neurodiversity I have a PA who supports me with planning, admin, travel bookings, letters to various bodies etc. She told me she now uses chatGPT extensively for helping her handle my requests.
So I’m paying someone to type things into chatGPT on my behalf
I’m happy with it. It’s helping her with her English and other tricky stuff and she’s much more productive with it. Her understanding of what I need, her knowledge of what to ask chatGPT and her ability to handle the output appropriately and convert it to meaningful actions that make a difference to my life is what I pay her for.
I've seen articles about how 'AI whisperers' are a thing these days, just being able to craft the input to get the right output (basically google-fu for the 2020s). It's a useful skill, and might be a key job skill going forward.
Interestingly, I also work with a recruitment team for Papua New Guinea. Despite that ChatGPT isn't the tool for random language translations, it was the only way we could easily find to translate a job ad into a local language that someone could clean up and publish. That people think it's "AI" misses the point, language tools are stupidly powerful in and of themselves.
Upvote
2
(2
/
0)
"As dozens of us said in our open letter, the incredible pace of AI technology development could create an AI system so powerful it poses an existential threat to humanity. On an unrelated topic, please try out the latest version of our AI technology. We spent a lot of time on fast-paced development to make it even more powerful than ever." - Sam Altman, CEO of OpenAI
Upvote
1
(1
/
0)
If UpToDate has a working search API, and allows this in their ToS, certainly.
The way GPT function calling works, is when you call the OpenAI API, in their response you get information about what GPT would like to "call". In your custom code, you can do whatever you want (including firing off a browser and accessing stuff like UpToDate) and return some sort of results, like snippets of text.
So it's not about you allowing OpenAI to directly call 3rd party services or run actual code (like running a plugin in ChatGPT UI). This is more about a structured way for GPT to ask you for help.
Upvote
1
(1
/
0)
