When I saw this, I figured it was a clear step back from simply using plugins in the main ChatGPT view. It's basically plugins, but with extra prompting and you can only use one at a time.
But if you look at projects like Autogen ( https://github.com/microsoft/autogen ), you see one master agent coordinating other agents which have a narrower scope. But you have to create the prompts for the agents yourself.
This GPTs setup will crowd-source a ton of data on creating agents which serve a single task. Then they can train a model that's very good at creating other agents. Altman nods to this immediately after the GPTs feature is shown, repeating that OpenAI does not train on API usage.
Prediction: next year's dev day, which Altman hints will make today's conference look "quaint" by comparison, will basically be an app built around the autogen concept, but which you can spin up using very simple prompts to complete very complex tasks. Probably on top of a mixture of GPT 4 & 5.
The GPT Store will prove to be an interesting moderation and quality control experiment for OpenAI. Apple/Google have spent a lot of time and money on both of those things and they still have issues, and that's not even accounting for the fact that AI growth hackers will be the primary creators of GPTs. And a revenue sharing agreement will provide even more incentive to do the traditional App Store marketing shennanigans.
The more elaborate the moderation tooling, the faster the race to the bottom gets as revenue drives black-hat marketers to become experts in circumventing AI moderation. It could end up being Google “S”EO all over again where we eventually ended up with very little real information in the results.
I don’t think it’s that far away. It’s not just that technology has been accelerating over the past few decades, it’s that the humans have been gaining more and more knowledge about the bigger fundamental patterns that technology uses.
As a very rough example: if their demographic is Americans they will come out the gate with proper spelling, grammar and sentence structure that’s been run through LLM(s), and an overall branding that does not raise any immediate red flags. If someone tries to implement a blacklist or whitelist of words, they will walk past it like it wasn’t even there.
Tickets filed by plugin developers (including requests to publish or update a plugin) are handled by bots. The only contact I've had with a human there has been via backchannels.
Yes, thats also why policies aren’t enforced. I assume OpenAI knew all along that plugins probably are a dead end. Wonder if they will be around 6 months from now.
But Actions will have all the same security challenges basically.
Not sure about currently - but budget increases seemed to be manually reviewed in a queue. I had to increase monthly budget amounts twice and it took like 10-30 hours for each approval, interesting to see how their processes advance.
I don't think the target market for this is people looking for extremely knowledgeable LLMs that can handle deep technical tasks, given that you can't even finetune these models.
I'd guess this is more of an attempt to poach the market of companies like character.ai. The market for models with a distinct character/personality is absolutely massive right now (see: app store rankings) and users are willing to spend insane amounts of money on it (in part because of the "digital girlfriend" appeal).
Can you elaborate a bit on why you think the market for distinct character/personality models is massive right now?
I ask only because I've been asked by a fairly well-known GAI company to help them 'personality-ize' some of their models and I'm trying to understand who else is doing it, where, and why – namely because I'd like to keep doing it.
Notice how it's "digital girlfriend" but not "digital boyfriend" that insane users will spend insane amounts of money on. So sad. But maybe it will distract the incels from harassing real women.
So these "GPTs" are the combination of predefined prompts and custom API access? Not customly trained LLMs?
If so, I guess you can make such a "GPT" on your own server and independent from a specific LLM service by using a prompt like
...you have available an API "WEATHER_API". If you need the
weather for a given city for a given day, please say
WEATHER_API('berlin', '2022-11-24')
and I will give you a new prompt including the data you
asked for...
Or is there some magic done by OpenAI which goes beyond this?
If you want to be independent for academic/personal reasons, sure you can.
If you want reasoning capabilities that Open Source hasn't matched before today, and I'm guessing just got blown out of the water on again today... there's no reason to bother.
> If you want to be independent for academic/personal reasons,
Or OpenAI's usage policy limits reasons (either because of your direct use, or because of the potential scope of use you want to support for downstream users of your GPT, etc.) Yes, OpenAI's model is the most powerful around. Yes, it would be foolish not to take advantage of it, if you can, in your custom tools that depend on some LLM. Depending on your use, it may not make sense to be fully and exclusively dependent on OpenAI, though.
You don't need to use an open source LLM for the approach I described. You can still send the prompts to OpenAI's GPT-4 or any other LLM which is available as a service.
What other LLM will compete with GPT-4 Turbo (+ V)? At most you're hedging that Anthropic releases a "Claude 2 Turbo (+ V)": is complicating your setup to such a ridiculous degree vs "zero effort" worth it for that?
If things change down the line the fact you invested 5 minutes into writing a prompt isn't going to be such a huge loss anyways, absolutely no reason to roll your own on this.
> If things change down the line the fact you invested 5 minutes into writing a prompt isn't going to be such a huge loss anyways
If things change down the road such that your tool (or a major potential downstream market for your tool) is outside of OpenAI's usage policies, the fact that you invested even a few developer-weeks into combining the existing open source tooling to support running your workload either against OpenAI's models or a toolchain consisting of one or more open source models (including a multimodal toolchain tied into image generation models, if that's your thing) with RAG, etc., is going to be a win.
If it doesn't, maybe its a wasted-effort loss, but there's lots of other ways it could be a win, too.
> If you want to be independent for academic/personal reasons, sure you can.
If your goal is to be in business, or to get some sort of reach, or anything other than "have fun tinkering"... wasting developer weeks on "could be a win" is how to fail.
Agreed, but one is a recognizable trademark (I assume) and one is a generic catch-all term that means a lot of things and would probably be worse from a branding/marketing perspective.
Not even. An agent should primarily operate in some kind of self-prompting loop. Afaik xou can not specify complex and branching looping behaviors for GPTs.
ChatGPT and plugins were already agents. An agent just does stuff on somebody else’s behalf. Browsing, browses the Internet for the user, code interpreter creates and runs Python code for the user…
What you describe are what’s called autonomous agents, and i agree that I also expected some interesting announcement there but probably too many security issues.
Well, I would say that when people at the moment say agent they mean what you call autonomous agent. Browsing or code interpreter is just LLM + tool use. I think there is a really large difference in quality between a system that just interacts 1-2 times with some tool/API before giving an answer and one that runs in a loop with undefined length (until it decides to terminate). It's like the difference between programming without loops or recursion vs with them. Night and day.
Currently the loop results in compounding errors and attention failure. Even with reviews, grounding information, and veracity checks in place by other LLMs it still happens.
they are trying to trademark "GPT" and "GPTs"..in order to do this...they need to use it widely and specifically...i don't believe that it has gone through yet..so in order for it be approved they will use this name...then prevent others from using it....thats the game...
Barrier to entry for commercial or useful GPTs/Plugins/"Agents" is almost non-existant since its just a str.concat(hiddenprompt, user_prompt), the secret sauce (ie the weights, chat timeout and context length) are already generated/limited by OpenAI and they already have the content moderation/"hr dept" baked in at the weights level. So even if one was to create a "story writer helper" GPT, i don't see how it would be of any value generating new, unique and interesting content other than the prompt recipes we already have on reddit/r/chatgpt (heres 1000 prompts for every use case) that creates netflix like plots (inclusively diverse casting across ethnicities and orientations, socially conscious storylines, modern jargon-filled dialogue, themes of empowerment, progressive characters, and non-traditional relationship dynamics).
This will most likely be like the google play store with a 99% of GPTs being a repackaged public prompt.
I wonder how much money I could make making "GPTs" full time. Barrier to entry is nonexistent so I imagine highest revenue ones if this becomes a serious thing people use will be advertised externally or have some proprietary info/access.
OpenAI in a weird way has mediocre marketing. The examples they use for Dalle-3 are way worse than the average ones I see people cooking up on Twitter/Reddit. They only seem to demo the most vaguely generic implementations of their app. Even their DevDay logo is just 4 lines of text.
To be fair, the name "ChatGPT" has quite a bit of mindshare and I've found many non-technical folks referring to any generative AI product as "ChatGPT" or "GPT". Yet, if you asked any single one of them what "GPT" stood for, they'd have no clue.
To be fair, I'm a dev who uses chatgpt on an hourly basis and I had no idea what GPT stood for until I googled it just now. I think it's kinda smart to make people strongly associate GPT with OpenAI
I don't recall which interviews I saw it stated in, but I believe Sam said in one or two of his world tour stops, where he stated they deliberately have gone with a technical name instead of a human name to help remind those using it that it's not a person. So I think that coupled with the mindshare (as others have stated) it already holds, makes a lot of sense to stick with it.
Doesn't seem like a problem to me. Many brands are acronyms that are meaningless, too technical or arbitrary. Very common for cars for instance, how about a BMW X5 V8 SUV?
Plus, "generative pre-trained transformer" sounds futuristic, which seems like a fitting brand image for OpenAI.
Omg… Thinking about their push for regulation with this… Are they after something like keeping advanced generative pretrained transformer LLM model technology to themselves, prohibiting others, at least in American economy where regulations can be applied?
How much would you pay for an actual Programming Rubber Duck ChatGPT terminal that you can talk to like Amazon Echo? With a bluetooth screencast input and keyboard and mouse output so it could see your screen and pair program with you, of course.
not much based on what OpenAI has been doing lately, using their own customers as product research and then copying the best ideas. OpenAI pretty much has to keep a huge lead in model capabilities or developers are going to stop using them for this reason
basically copying the Microsoft strategy of Embrace, Extend, Extinguish. Makes sense they took so much funding from Microsoft
I'm more confused how the revenue share works. Do they get part of my ChatGPT subscription fee? Am I paying extra? Per bot? Per amount of time I consult with the bot?
I'm certain of one thing...when I'm making a GPT, I save, I update and it completely disregards established rules, making me max out my requests very fast. It also gives general answers when I ask for specifics. I only have Plus. In regards to general answers...I then went to Bard and got exactly what I was looking for after 1 request-this is well known information that's been online for years, information that Bing certainly has...it was making me angry, basically charging me for doing nothing over and over.
For instance, my GPT is based on a list of 15 items. I had to repeatedly tell it not to deviate from the list because it kept doing so. I save and update every time, and it never matters.
Thoughts on Zapier trying to become OpenAI faster than OpenAPI can become Zapier? There will always be a long tail of APIs that folks want integrating, but the most popular APIs are perhaps only a few hundred in number (Google Calendar and Slack, for example).
I feel like history has shown that those who own the platform end up winning and in this case OpenAI's platform of models seems much harder to recreate. My guess is this would lead to Zapier using OpenAI as a platform and eventually OpenAI would re-create Zapier's integrations before the other way around.
I think this perspective is fair and historically accurate wrt platform risk, but I also strongly believe Zapier has substantial value beyond what historically has been acting as a conduit between APIs. Customers don't want a pipe between their services, they want to automate their mundane work with a robot.
What is Zapier's value-add when GPT can chain an arbitrary number of flexible API calls? Imagine if the GPT plug-ins directory is fed as input to GPT...
The problem with Zapier is all the individual accounts and contracts you need with dozens of companies in order to have them all communicating for a real business. SaaS fatigue is a real issue.
Sounds like a call for them to partner with Venminder or a similar business to manage all your SaaS from a contractual perspective, as well as some CASB (cloud access security broker) to handle onboarding, offboarding, and governance of an org's SaaS inventory. Legit problem, but SaaS isn't going away.
So many zapier integrations are half baked. A lot of them are good for reacting to events but not for searching for data (i.e. you can use zapier to react to a jira ticket change but can't use zapier to query jira for ticket info)
And seeing how OpenAI is moving up the value chain, what's the guarantee they won't come up with an in-house competitor to the bot that was built on their platform?
Guarantee? That's one of the most important aspects of bothering to build out a platform: eat the ecosystem to add incremental value to the platform, bolster the moat.
The guarantee is that they will clone and extinguish most of the best bots/tools/services riding on top of GPT. It's what platforms overwhelmingly tend to do.
If your thing is a near-touch to GPT, depends on it, is of medium complexity or lower, and is very popular/useful, they will build your thing into GPT eventually.
On the flip side, if you're Salesforce and using GPT to augment something about a major product, you're not facing a serious threat of OpenAI trying to become the next big CRM company.
This process will work similarly to how it did with Windows, Google, AWS, etc.
Their primary target is B2B and they want to get that market ASAP. Pretty sure any idea or anything that seems feasible will be copied and even improved upon.
The pattern will continue until no one needs anything. I am sure this company will eat everything. There will be no pie.
Also not only for ChatGPT, but there will be a timespan between majority of people are automated out of jobs and a a forceful redistribution of wealth happens (let’s call it socialism). It’s always a good idea to have that $$ to bridge that gap.
There’s a great new book on how a lot of market leaders did exactly this; it’s called “Chokepoint Capitalism”. It’s from the guy who coined the term “enshittification”
I think something could be said for "virality" as well - could easily see some entertainment or lifehack themed templates blowing up on TikTok. No one wants to post the output of the lame, less popular template on their story!
> I wonder how much money I could make making "GPTs" full time
I don’t get why people are thinking along these lines at all. Like, if you don’t own and control the LLM yourself, what makes you so sure OpenAI will allow you to make money at all? They could make advertising externally or hosting external marketplaces against the TOS. They could copy your GPT and put their “official” version at the top of the store page. Just because a technology is powerful does not necessarily mean you can make money off of it.
I just feel like the big tech companies have gotten better at capturing the value of their platforms. It’s not the same as it was in 2008. I don’t think we’re going to see anything like the “manage an indie app as a lifestyle business” boom we saw back then, and I expect the consolidation that killed off most of those indie devs will happen way faster this time.
> what makes you so sure OpenAI will allow you to make money at all?
They might not. But if they do, I'd imagine there are a lot of people who will try. And as long as you're not dependent on the income stream they provide, you don't have much to worry about if it gets shut off.
How does the store solve visibility? In the demo it looked like there was a select list of Custom Assistants in the left panel which he manually had to click, so not much different from plug-ins?
Right he said something about promoting the best but what about the discoverability?
Imagine a "GPT" that could generate websites and provide you with a live deployment as you change it using natural language. A website builder GPT that is primed to output and design in a decent way, that has all the prep beforehand to use particular libraries, and integrations with something like Render.
One can come up with all sorts of ideas like this, but building it will be a matter of slow iterations at prompt engineering in a mixture of natural language and data structures and will be at the whim of changing APIs, including the backing ChatGPT model. Sounds messy, hard to manage, hard to test...or am I missing what the actual process will be for creating one of these?
Good intuition, it was in the demo. Chat with an agent to config a new GPT. Then make the GPT available to others if you wish. Same functionality is exposed via API IIRC.
You could farm the hell out of referral link before people catch up. Prob already too late, but there's a lot of market that's searching comparatives and would be fooled by this.
The GPT's making the most money will be made by larger companies who advertise use of it and maybe make it a funnel to their in-app integration, or GPTs which are made effective by information that is proprietary.
This makes me think of the Alexa skills ecosystem which is full of low quality skills. Many of which have poor data practices abstracted away behind the scenes. How long until "Chat with your favorite character from [Intellectual Property]?" which is simply made to promote a new film or collect data.
"Example GPTs are available today for ChatGPT Plus and Enterprise users to try out including Canva and Zapier AI Actions." and yet as a paying ChatGPT Plus customer, neither the Canva nor the Zapier AI Actions link work for me, I get a "GPT inaccessible or not found" error for Canva or Zapier.
> Example GPTs are available today for ChatGPT Plus
or
> Starting today, no more hopping between models; everything you need is in one place.
Neither of which are true. I'm a paying user and I have access to neither. They do this _all the time_. They announce something "available immediately" and it trickles out a week or more later. If they want to do gradual rollouts (which is smart) then they should say as much.
I (Plus subscriber, EU) tried https://chat.openai.com/gpts/editor (as linked from https://help.openai.com/en/articles/8554407-gpts-faq#h_86549... ) a minute ago and got a "You do not currently have access to this feature" toast notification on orange background top of pace, along with "chat.openai.com" in the URL bar.
(Chrome on Android, but besides a responsive layout, I haven't noticed discrepancies with the desktop Chromium site/interface; sadly the native Android app still shows no signs of code interpreter mode.)
The announced (and a few days ago leaked) Omni prompt also doesn't show up in the model selector. And that despite the expanded context looking very promising for the REPL-feedback-augmented code generation abilities.
> We believe the most incredible GPTs will come from builders in the community. Whether you’re an educator, coach, or just someone who loves to build helpful tools, you don’t need to know coding to make one and share your expertise.
“Please work for us for free, while we keep all the product of your work for ourselves like we did with the content we scraped on the internet.”
Isn’t that the game they all play? Amazon Marketplace. Apple App Store. Let the guinea pigs run. See which one gets the furthest. Then take away its lunch.
Their zappier demo was uninspiring. They are trying to appeal to the mainstream (non devs) users when their actual value is the API. Everyday users don´t have access to propriatery data that makes the LLM Bots valuable.
This makes me think that open source models are the future, for other reasons as well. No company is going to give them OpenAI their data to make bots, regardless if they won't train on the data. So what OpenAI can do at the moment is to focus on everyday user, and see where the biggest use cases are.
Furthermore, as Sam alluded to, this confirms to me that multi-LLM architectures (AutoGen, AutoGPT, BabyAGI etc) is the future. It´s about pushing the ways in which we can use these models, i.e. using different architectures because it will be a while before we get into more powerful models. These is simply no compute for it at the moment. The dust has yet to settle.
I disagree. There's a lot of folks using zapier that would like some light automation -- I have fielded requests from several -- who are not engineers and cannot build anything with an api. The bit where they built out the functionality with zapier and human instructions was aimed squarely at folks like this.
The issue is all open source models are made by huge companies with an indirect profit motive. There is too much scale advantage for open source models to be competitive in any field other than those which venture backed companies deliberately avoid (porn, illegal stuff, etc).
> I'm not even sure what kind of startup still makes sense as a gpt thin wrapper?
Startups don't make sense as thin wrappers around another company's product when that company is aggressively expanding that product is a core offering that the vendor is aggressively working to provide as an integrated solution for as many markets as possible, which very much applies to OpenAI offerings in general, and its chat models especially.
A wrapper that also leverages some exclusive special sauce data, algorithm, etc., for which you have a real moat as a key component, that makes some sense. But just a thin wrapper around GPT? That's just asking to have your market eaten by OpenAI.
It might make some sense for products where the vendor is a stable, steady-state infrastructure supplier for many markets without any evident interest in entering the same market as the startup, where the uncertainty across markets that they would create by specifically targeting your startups market would hurt them more with their established customers than they would gain from your niche, but even that is risky because it requires lots of potentially-erroneous assessments of how the vendor would expect their other customers to react to them acting in your market.
They never made sense long term, of course. But, plenty of first movers made a bundle of money making chatgpt wrappers and marketing the hell out of them. In that context, they probably made sense for a small subset of people for a small slice of time.
Yeah none of them ever did, and that was always very obvious. If you can make it by wrapping a few API calls, you have no moat and anyone can steal your idea/customers.
Even if you're doing something with strong network effects you've got to consider cases like facebook which weren't first movers in markets where there are super strong network effects and yet still won. You're not really going to have a 'community' based moat for PDF analyzer tool #5 or retrieval augmented generation knowledgebase parser #300. You're not gonna have a moat based on much else either...
Conversion.ai / Jarvis.ai / Jasper.ai / whatever they’re called today raised $125M at a $1.5b valuation, so it made sense at some point to some people.
1) Jasper isn’t exactly what I’d call “thin.” It has e.g. an entire project management suite inside it now. They moved very fast into a vertical SaaS platform, probably because they knew that a thin wrapper was neither a particularly good product nor a defensible one.
2) There are tons of actually thin wrappers that raised a ton of money — contrary to popular belief this does not mean they are good businesses (Jasper, AFAICT, doesn’t seem to fall into the categories of either “thin wrapper” or bad business)
3) Obviously anything being called out as not making any sense allegedly “made sense” to some people at some point otherwise it wouldn’t be worth talking about.
Jasper was cited in a recent TechCrunch article about the obvious vulnerability of the API wrapper startup business model. I guess we could argue at length about the thickness of the wrapper, but that seems unproductive.
I think people may be overstating this. OpenAI has obvious brand and distribution advantages for their end-user facing products, but they aren't guaranteed to hit the bullseye and be the market leader in every category. ChatGPT plugins, for example, have never really taken off. Startups like Perplexity and Phind are doing just fine in the "search the web with AI" space that plugins were supposedly going to dominate.
I'd say that trying to build a business plugging gaps in OpenAI's infrastructure offerings (fine-tuning, RAG, etc.) is quite perilous, since this is really their core competency. But they've still had just one big user-facing hit so far: ChatGPT.
Wrapping OpenAI for a specific use case then building better workflows and UI for that use case than ChatGPT seems like it's still a pretty reasonable strategy. OpenAI isn't going to build the ideal workflow/UI for every use case.
I recently enjoy Perplexity.ai for their superior search+llm integration, while chatGPT usually send stupid search keywords to Bing that I know are not going to help.
It was (is) basically a smash and grab. You can spin up a site and a payment processor so fast nowadays that it just takes a weekend of work to go from idea to "Click here to subscribe and get 200 [AI generation] a month!"
Can I ask what the goal is here? It's cool to be able to spin up a nicely designed website (and it is nicely designed and has a good aesthetic), but isn't the content going to be semantically empty? I don't want to be negative but it feels like cheap plastic imitation of a real thing, and the web is already full of spammy low quality content. Aren't you in danger of your products having a very short life cycle and ending up as digital landfill, so to speak?
The quality of the GPT4 content for travel is surprisingly good, although there are of course plenty of hallucinations and I've a new pass for the content set up that gets rid of the worst offenders. This is really just a learning experiment, but building something of this content scale by 1 person was absolutely impossible a year ago.
The point was that content generation use cases might be a viable path give OpenAI's thin-wrappers-killing direction.
But does anyone particularly want another travel site full of generic advice? Nothing against your individual project, I'm just saying that the web is already absolutely awash in marketing materials and all-in-one portals. If I'm visiting a place and want to read up on it in advance I tend to look for someone with insider knowledge, otherwise I could just buy a guidebook specifically about that place (might be a bit dated, but the editorial quality in a book is usually orders of magnitude better than the web).
LLMs can quickly extract ideas using search and synthesize fresh articles. Take this conversation thread for example, there are many interesting ideas here, they can be scooped and formatted in a nice style. This gives AI a seed of authenticity, its main role would be to curate the mess.
Many startups have started over the past few years, trying to build infrastructure (shovels) for companies to integrate LLMs, or specific chat copilots trying to cater to a specific usecase. Most are dead in the water once OpenAI subsumes their feature set.
... is what people who don't understand positioning will parrot time and time again.
Jasper isn't having a good time, but you'd think the fact anyone can produce better output than they did after spending millions of dollars in GPT-3 based pipelines for $20 a month would mean they're dead dead.
But instead they went and changed their positioning, changed who their target market is, adjusted the UX, the messaging, and the feature set, and now it's a product that has a place even if OpenAI can give all of your marketers an internal ChatGPT (unless your plan is to have 100 different "GPTs" for every marketing task in your company)
tl;dr: People fail to realize that OpenAI can offer your startup's core value proposition tomorrow morning, and it doesn't matter if they don't offer it in a format that resonates with your target users.
You could have a cure to cancer and you'd still have to market it correctly.
That's the complete opposite of reality: Jasper could only fight to survive because they were a wrapper.
Imagine if Jasper had raised and built their own GPT-3 alternative from the ground up at 100x the cost targeted at writing: it wouldn't have generalized like OpenAI's models given the different goals in training, they'd have spent orders of magnitude more, they'd currently be scrambling to partner with some non-OpenAI provider to play the role OpenAI does for them today...
Jasper "only" had to start integrating new APIs to keep up with the SOTA in quality, and instead of burning precious runway on R&D they get to focus on rebranding and driving home that positioning.
What isn't a wrapper? Honestly everyone turns they nose in the air and claims all superiority bit I challenge you to find a use case that isn't just a 'wrapper'
Without getting into specifics, I operate a ChatGPT plugin with a lot of users.
It sounds like the GPT Store will solve some of the rough edges with plugins as they exist today, but I'm not interested.
OpenAI has been a horrible "partner" to work with. The team that owns this product is terribly under-resourced. You have no hope of reaching a human unless you have an inside contact. Support tickets are handled by GPT; policies are arbitrarily enforced; and major bugs in the API and documentation are simply ignored.
There's an amazing opportunity there for user acquisition, obviously, but do you really want to put your hand in the pain box?
Can anyone explain what "extra knowledge" means specifically? Is that like fine tuning? How much data can I give it to learn? How much can it retain? Can it be updated over time?
The keynote used a .txt file of a lecture that the user uploaded as a data source the model can select from. From a technical perspective, it's anoher data source for retrieval-augmented generation (RAG) doing a vector search in the background: https://platform.openai.com/docs/assistants/tools/knowledge-...
"optimizes for quality by adding all relevant content to the context of model calls." So for their own profits they maximize recall and let GPT handle precision.
They showed a demo where you could upload a file while creating an agent. As others have answered. I think it's about configuring an agent in a way that you give it material on some specific topic (one file, multiple files) and it uses Retrieval to augment the answer based on the source material.
Google launched Notebook LM[1] a while back which does a similar thing conceptually. It allows you to import a Google drive folder with docs of the stuff you would want to understand and then just chat with it. It's a good product but restrictive in the sense that it only allowed Google docs.
I doublt it's fine tuning (actually changing the model weights). It's more like "im going to paste a text blob then in the following chats i will ask questions about it) type inner prompt.
How's that different from pasting the text in the first chat and running the vector embedding step on the text on the server (maybe at least bypassing the chat text limit)? Does this fix the amnesia issue where the info from chats longer than the context length is forgotton because the document isn't baked directly into the weights like fine tuning?
The demo just showed a file dialog box where they could upload a set of static files. What I'd really like to see is the ability to sync a source integration (for example into a GitHub repo or a notion account), and it would always pull relevant information using a RAG architecture.
> We’ve made ChatGPT Plus fresher and simpler to use
> Finally, ChatGPT Plus now includes fresh information up to April 2023. We’ve also heard your feedback about how the model picker is a pain. Starting today, no more hopping between models; everything you need is in one place. You can access DALL·E, browsing, and data analysis all without switching. You can also attach files to let ChatGPT search PDFs and other document types. Find us at chatgpt.com.
It's so annoying how they say this, I refresh and I still have to hop between models. Just say "rolling out over the next week" if that's what's happening. I even logged out and back in and still the same old way of doing it.
The roll outs are quite slow. Sometimes it takes weeks for them to release a feature to my account while I know others who get access immediately. I understand that it helps them control the quality of ChatGPT, but I wish I got access earlier as I have been subscribed since the beginning.
Sometimes it seems like logging off and back on causes these updates to hit. I didn't have Dall-e or image capabilities after weeks of it being released and I logged off and back on and both were available. This was between multiple computers logged in so it wasn't just a cache clearing situation.
For what it's worth, for me it appeared nothing was different except I could leave the Default model on, upload a picture, ask it to generate a similar one with a change, and it just booted up Dall-e and did so. I'm not sure if everyone's experience is like this. I agree it's super annoying how they handle these things. I do still have the dropdown but, set to Default, it would generate images for me.
I do recommend logging off and logging back in if it isn't working. I've seen that update things for me.
What is even more annoying is every plus subscriber whining that they didn't get the feature yet. We know, you paid and you want to play now, but it's a rollout and it's progressively coming towards you. Just have a little bit of patience.
You’re completely missing the point. If they said “rolling out over the next week” it’d be one thing but they almost always say “available now” and it’s not. That’s not ok, that’s what we call a lie.
Nobody is going to say "We are rolling it out to a random set of users, and then if there are no crashes, we will roll out to more" in a marketing demo.
Yeah I'm getting pretty sick of these announcements followed by insanely slow rollouts. I'm a paying customer and have been since Plus became available. So annoying.
I don’t understand why more people aren’t creating digital doubles of their brain..? You train your own LLM for practically free and have your own digital double to maximize any and all productivity. Why is there not more of this?
I think you'll find that "maximizing any and all productivity" is both harder than you think and not really something worth striving for.
Plus as neat as these advancements are, they are nowhere nearing the ability to create a "digital double of your brain" of yourself. There is a vast ocean from where we are are today to the possibility of replicating a brain digitally (if it's even physically possible).
90% of my work is done in Slack and Email threads, if I had a (ideally local) LLM read and learn everything there, it would be immensely helpful - as in it could replace me to answer on most questions and queries.
I don't know for sure but I'd bet BIG money that these do not include automatic fine-tuning, though I still understand them to be a bit more powerful than just "custom prompts" -- think templates, or sets of custom prompts for specific (sub-)situations.
This is the kind of feature that will prove to be a minor improvement for anyone on this forum, and a complete paradigm shift for the less technically-inclined. IMO.
> Is this basically them deploying fine tuned models?
From the description of the past outside practice it is marketed as moving into OpenAI's offering, it sounds more like its custom prompts, not fine-tuned models.
Custom GPTs are fine but they’re sort of useless. Since anyone wanting to use a custom GPT has to have a ChatGPT Plus account, they themselves can spin up an even more targeted, even better prompt than anything you could come up with and use that instead.
Unless your custom gpt provides some special bells and whistles through functions or APIs, your custom gpt ideas are going to r copied by someone who can spin up those special bells and whistles and API calls and cut you out from the middle.
I think I speak for a few of us AI Doomers here when I say that this makes me excited and terribly anxious at the same time. So... well done OpenAI :). Great news, and a great feature!
I have no doubt that this will immensely increase uptake among the less technically literate, since it will allow the techy people in their life (or on the app store) to introduce them with much less friction. It'll be like the little examples you can find on the "New Chat" screen of every chatbot, but 1000x more engaging
I think that's probably the point. It's kind of like early Wordpress with plugins. People who couldn't make professional websites suddenly could (with a little design and UI sense.) So I thought maybe if you get in at the beginning you might be able to make some waves in a pond...but I'm not sure because it moves much quicker and evolves into different things. Eventually, soon, none of the GPT's will be shiny enough.
If the catastrophe is mega-corps getting to monopolize a valuable technology because people saw The Terminator and thought it was a documentary then any announcement from OpenAI is bad news.
The catastrophe is humanity going extinct from superintelligent AI. Like a native species going extinct after an invasive species arrives. Mentioning Terminator is like saying the Earth is flat because Hitler said it is round.
This reminds of the Eliezer Yudkowsky tweet saying that AI was going to hack our DNA and use our bodies to mine bitcoin or something. Ridiculous fearmongering.
I have probably read more sci-fi than the average HN user but the whole "superintelligent AI is going to kill us all" hysteria is among the more ridiculous ideas I have ever heard.
Really though I have entertained all the doomers propositions and none of them seem any more likely than the plot of the Matrix. The ideas that prop these fears up are based on layers of ever more far fetched hypothesis about things that do not exist. If you have a novel reason why AI poses an x-risk I am more than interested in hearing it.
Here is a really interesting quote that I think might go against some of the misanthropic tendencies of doomers and the tech crowd in general but it really is more relevant than ever:
“There was also the Argument of Increasing Decency, which basically held that cruelty was linked to stupidity and that the link between intelligence, imagination, empathy and good-behaviour-as-it-was-generally-understood – i.e. not being cruel to others – was as profound as these matters ever got.”
True humans have been remarkably ignorant throughout our short history. Though you might notice though that most folks dont go around abusing animals or hurting other people on purpose. Take from that what you will.
Maybe together as a species we can avoid hellish cyberpunk dystopias brought on by regulatory capture of the most powerful technology created by humans thusfar. I can only hope.
>Though you might notice though that most folks dont go around abusing animals or hurting other people on purpose. Take from that what you will.
It doesn't matter what "most folks" go about doing. If anything it makes things all the scarier. All this destruction we've caused to so many other species and we weren't even trying.
But anyway, that's just human terrorists using future AIs to build biological weapons. But the much greater danger is superintelligent AI causing human extinction by itself.
That is not an article that is a series of short-form tweets. If it was an actual article I would absolutely read it but I cannot see referencing a bunch of tweets from a self-proclaimed expert as in any sense good faith. I asked if you had any novel ideas about how "superintelligent AI" gives a valid x-risk but you failed to provide any interesting ideas. So I wont engage with what from my perspective is fearmongering for the benefit of malicious corporations.
I still dont see any studies with control groups and reproduceable experimental procedures saying that any AI agent is in any way more useful to a 'terrorist' than unrestricted access to the internet.
> That is not an article that is a series of short-form tweets. If it was an actual article I would absolutely read it
So you didn't even read the tweets? They contain the link to a paper.
> self-proclaimed expert
Esvelt is absolutely a recognized expert on biotechnology. The authors of the article you linked to are not.
> I asked if you had any novel ideas about how "superintelligent AI" gives a valid x-risk
I just responded to your article about bioterrorism, which was not about x-risk. Arguments about x-risk were made elsewhere, but I'm sure you would dismiss them because they don't contain studies with control groups.
Oh you mean the paper that was discredited in the link I already shared? I am not going to copy and paste the argument from the link I shared, you can read it for yourself. I gave you the benefit of the doubt that you actually read the article that you claimed to read but unfortunately good faith arguments around this seem impossible to have with the alarmist crowd.
So yes the paper that was already discredited in the essay I posted that you didn't read
------------------
"While I was writing this, an extra paper game out on the same topic as the "Dual-use biotechnology" paper, with the fun title "Will releasing the weights of future large language models grant widespread access to pandemic agents?"."
Maybe this is a totally different paper with exactly the same name but if not, I am not really interested in reading a paper that is unrepeatable and doesn't use control groups because that isnt science
I think without AI we wouldn't go extinct for a long time. There are no other likely extinction risks. Toby Ord has a nice book (The Precipice) about various forms of extinction risks, and he basically says the same thing.
They also run an RAG implementation for you, and there's probably some value in the store if it can surface high quality apps and have a good rating system. But basically, yeah. This is a complete replacement for every startup that was just a thin wrapper around the API - instead of being real startups those concepts will just be "GPTs" on the "GPT Store". I don't know how realistic or viable their solution is, but the sea of thin wrapper startups was never viable and everyone knew it.
Yes, it offers a lot more than I initially thought. I think for the presentation they wanted to emphasize how simple it was to build and use, so I missed the implications of what they were showing. But the ability to upload a document base to one of these "GPTs" that it can then rely on for its responses means they must have some RAG system implemented on the backend, probably a pretty good one as I expect them to be very good at embeddings and search (they have a huge number of Google alums). It can also have integrated Code Interpreter, Data Analysis, and Image handling/generation if you tick the boxes, plus what looks like a very significant list of functions it can call in order to act as an agent (e.g. changing your calender entries or sending messages at your request after double checking, something that's probably going to give simonw a heart attack).
It looks like at base, with a short text description of what you want it to do and uploading of relevant documents, it may be able to replace like 95% of the GPT wrappers that have sprung up in the last 7-10 months. That's pretty impressive, because while all of those were doomed because they had no ability to defend themselves from competition, some of them were pretty useful.
This is just slid into the last paragraph, but it is a pretty big deal as this knowledge cutoff date is now past ChatGPT’s initial launch date (thus OpenAI has likely sort of figured out as to how to exclude crawled text data generated by itself and other LLMs on the internet):
> Finally, ChatGPT Plus now includes fresh information up to April 2023.
E.g. one popular comment in the submission about twitter’s new “edgy ai” was that it could be reimplemented as a chatgpt prompt[1]. Looks like this is even more relevant now.
What is your impression of how respectful of their ecosystem this was? I saw many “OpenAI killed my start-up”-type reactions. Still, it feels like they are trying to handle obvious use cases in the ecosystem and make things easier to integrate canonically and more transparently. And to handle legal concerns. It was their first ecosystem conference.
Then again, iOS felt like a very promising ecosystem when the iPhone 4 came out; I feel like it turned into an addendum/duplicate of the web rather than something genuinely original. Android seems more vivid, but not much more.
Awesome, I've been waiting for something like this.
It looks like we are moving away from apps to web GPTs, this looks like chatbot interface is here to stay and 'AI' is now the default interface that is to be expected.
I also don't need to spend lots of money on a developer to test my ideas out, this is great for product validation, I look forward to playing with GPTs.
I can see writing, travel and other bots being more enhanced and more powerful, hopefully existing startups will adapt to this change.
GPTs replacing AI startups is what happens when they are pushed to ship as fast as possible by VCs and YC. What do they end up shipping? Very small +1 features that are already on OpenAI's roadmap.
Well the market is pretty obvious. Example - Fitness Influencer on Instagram can be your personal coach,i.e. the AI clone of the Fitness Influencer you can subscribe to at a fairly low price, as personal consultations are expensive and physically limited. So people with distributions and expertise will train their GPTs, make them accessible at affordable prices. good for creators, it's passive income source and good for consumers, affordable service/education.
Poe has done a great job in this space, quite a large marketplace of existing bots. I'm excited to see what it can do with the extra vision, D•ALLE, and Code Interpreter models.
I’m very excited about all the new developments but must say that they really dropped the ball on marketing this one. The feature name is ambiguous and unsearchable.
Nah, I'm in the US (with a US based account) and I'm still getting the message it's rolling out over the next few days (you have to open a sample to see that message).
> Retrieval augments the Assistant with knowledge from outside its model, such as proprietary product information or documents provided by your users. Once a file is uploaded and passed to the Assistant, OpenAI will automatically chunk your documents, index and store the embeddings, and implement vector search to retrieve relevant content to answer user queries.
This feels like simply a system prompt generator. It gives you a survey to generate a decent system prompt from, then staples that prompt to the front of any interactions with the GPT.
It doesn't seem to be much smarter than a simple "tell me how I should interpret inputs from the user."
One of the things I did not care for is setting my plugin hints from the Custom Instructions setting. It seems this will allow for each "GPT" agent to have their own system instructions which will be quite handy.
> ChatGPT Plus now includes fresh information up to April 2023
I'm so happy; I can finally ask questions about expo and trpc and get fresh answers. I confirmed this by asking chatgpt about the superbowl winners in 2022 & 2023.
Right?! That then give you no way to interact except to force a regeneration. I don't want to waste messages or time repeating a long generation, I want to be able to either continue it or, if it's sufficient as is, just respond to it in place. I don't understand why you're forced to hit regenerate after those.
they mentioned revenue sharing in the keynote, and I'm eager to find out how that is going to work. There isn't much money in the $20/month subscription to go around to very many other developers
What I was thinking for my own agent hub thing was to sell universal credits and charge per use or token. Then agent developers could specify what they want to charge.
that's kinda interesting but I'm not sure it maps well to the value added by a GPT app. Like, I'm imagining that I'll do old fashioned API work and GPT will the UI layer - sure, the tokens are the most expensive part, but the value for the customer comes from easy access to whatever is on the backend
It's long after 1pm PST -- does anyone else have access to this yet? I'm on a paid US account (so I assume I should have access) but still get "You do not currently have access to this feature".
Is this scary? They said revenue share - it sounds like a streaming platform software licensing model. That sounds like getting paid much less than 70%.
I have been utilising GPT to create my own app, and now openAI wants to be the only app that matters. I’m not sure whether I should be excited or not :/
Every prompt you put in somebody else's LLM goes into the training set of the next iteration of said LLM, with the explicit purpose of replacing you as a cognitively, and therefore economically, relevant entity. The only dignified move is not to play, though it's a very difficult choice. It probably not a winning move, though at this point there are no obvious winning moves -- you and I and all our loved ones will be obsoleted and replaced by tech within the next few years. Concretely, to not play means to stop feeding the machine data, i.e. disconnecting from the digital world. Given how digitalized society is becoming, possibly also from the modern society altogether. Godspeed.
I like to think about it from the perspective of the far future, looking back on me as a historical actor. I have no idea what will happen exactly, of course, but I can't imagine a moral/social crisis of the past where "cross your fingers and hope it goes away" is a move I'd approve of...
That said, your worry is one I definitely share. I guess I just hope more people think of ways they can try to ride/shape this wave, rather than stop/weather it.
I think all technological revolutions have caused similar transformations which obsolete certain types of activities and push novel activities to the forefront.
Not playing is certainly possible but could be a losing strategy as well.
What is your job? Chances are, it’s nothing an AGI (based on LLMs) can’t do, and an AGI is possible, today. People are building these things today, check out GitHub. And if you don’t believe GPT-4 cannot do your job cheaper than you, just wait for GPT-N, which will be able to.
Well, they tried to put a government sponsored moat in the way of other people building AI companies that would be competing with them. Thankfully, they mostly seem to have whiffed the ball on that one. This plan of theirs to monetize the creation of agents and other tools that take advantage of their underlying infrastructure is a good secondary kind of moat. Because, if your tool relies on their underlying infrastructure, even if you could build something different, the infrastructure is required. This may be a "less-evil" way to keep them building things and making tools available without completely locking out competition.
What did you expect? Surely it was just an MVP, and you expected OpenAI to commoditize its complements? Right?
On the long run, if your idea or app can be expressed as a flavor of a general GPT, you will not be able to compete with the AI gorillas. The space for AI startups is with custom, highly niched data or capabilities, that cannot be found in a general corpus, or that you can uniquely generate or control.
They will allow revenue sharing so then it's a matter of how many customers they'd be able to offer you for your app; and whether it makes sense for you to distribute through them or bypass them and distribute yourself.
This business model will only serve them in the long run. Luckily for us, open source llms are getting traction and we won't depend on "open"AI to implement features on our apps.
I thought about the same thing, I've seen a lot of apps that have similar ideas like "ChatGPT chatbot for your data or your website", I don't know how will they deal with it.
Posting from another comment in a different thread, everything that is new from OpenAI developer day:
- Context length extended to 128k (~300 pages).
- Better memory retrieveal across a longer span of time
- 4 new APIs: DALLE-3, GPT-4-vision, TTS (speech synthesis), and Whisper V3 (speech recognition).
- GPT-4 Turbo, a more intelligent iteration, confirmed as superior to GPT-4.
- GPT-4 Turbo pricing significantly reduced, about 3 times less expensive than GPT-4. Input and output tokens are respectively 3× and 2× less expensive than GPT-4. It’s available now to all developers in preview.
- Improved JSON handling (via JSON mode) and function invocation for more sophisticated control.
- Doubled rate limits with the option to request increases in account settings.
- Built-in retrieval-augmented generation (RAG) and knowledge current as of April 2023.
- Whisper V3 to be open-sourced and added to the API suite.
- Copyright Shield initiative to cover legal fees for copyright-related issues.
- Ability to create your own, custom "GPTs".
- Assistants API and new tools (Retrieval, Code Interpreter)
- 3.5 Turbo 16k now cheaper than old 4k. 0.003c per 1k in / 0.004c per 1k out.
So what is the selling point for writers and artists now? Being more expensive, much slower and less capable than a machine? At the rate these things are progressing, there’s not going to be any point in keeping them around.
How much something is valued is determined by how much others are willing to give up for it, and looking at artist salaries it’s already not much. Whether or not I browse artwork would make little difference. Trying to attack me for pointing out that something else does the job better doesn’t improve your value proposition at all.
The point is that you are required to have poorly refined taste if you feel that machine learning's literary output is remotely competitive with skilled authors.
Anyone can easily build their own Google. No coding is required. You can make them for yourself, just for your company’s internal use, or for everyone. Creating one is as easy as starting a conversation, giving it instructions and extra knowledge, and picking what it can do, like searching the web, making images or analyzing data.
The whole point of ChatGPT was go to one single place for all your knowledge needs.
The whole point of Amazon is the largest collections of things you can buy and have it delivered to your doorsteps in a few days.
I don't want many GPTs. I want one GPT that can reliably digest all information available on the internet, understand it, organize it and allow me to do useful things with it.
It's the same enshittifaction on Whatsapp that Meta is doing with Celebrities AI like SnoopDog AI that are gimmicks.
Please don't build gimmicky features. Leave that to the community via integrations.
> The whole point of ChatGPT was go to one single place for all your knowledge needs.
That's just your own perception. OpenAI is trying to build AGI. You entered into the storyline at a specific junction and jumped to conclusions based on the limits of your own imagination, or something.
Being able to have multiple personae could be very useful.
One persona may not give you the answer you're looking for, but another one may. I think maybe they should require the GPTs to have human names though so people intuitively understand that.
Like, Paul can't help me with this task, so let me ask Monica instead.
A good example is creative/idea work vs fact work. You don't want creative facts, and you don't want fact bounded creativity. You either have a prompt ready to paste, to prime the conversation/context, or you can use a personality.
One "über" AI is great, but it requires guidance into the context you're interested in, including yourself. For example, the default ChatGPT will assume you're uneducated about any topic you ask about.
I think this all fits perfectly into what Sam Altman talked about in the Lex Friedman podcast: people want an AI that fits their own worldview/context. Custom instructions, and "about yourself" are good starts, but sometimes you want to talk to a chef, and sometimes a scientist.
They still have that - it's just regular GPT-4. One immediate application about this one is that it makes it trivial to create a fine tuned version of GPT based on your data, where you can upload a series of documents that can basically act like a set of embeddings that augment the regular GPT trained data.
My bad, I wasn't intending fine tune to mean fine tuning the model itself. And leveraging vectorized embeddings as I mentioned IS FUNCTIONALLY a rag-style architecture.
But if you look at projects like Autogen ( https://github.com/microsoft/autogen ), you see one master agent coordinating other agents which have a narrower scope. But you have to create the prompts for the agents yourself.
This GPTs setup will crowd-source a ton of data on creating agents which serve a single task. Then they can train a model that's very good at creating other agents. Altman nods to this immediately after the GPTs feature is shown, repeating that OpenAI does not train on API usage.
Prediction: next year's dev day, which Altman hints will make today's conference look "quaint" by comparison, will basically be an app built around the autogen concept, but which you can spin up using very simple prompts to complete very complex tasks. Probably on top of a mixture of GPT 4 & 5.