zlacker

Makes me wonder whether to keep building upon OpenAI? Given that they have an API and it takes effort to build on that vs. something else. I am small fry but maybe other people are wondering the same? Can they give reassurances about their products going into the future?

>>quickt+(OP)
I am wondering the same. It’s a PR desaster to their dev community and i’m not even sure if Sutskever isn’t secretly happy about this.

>>quickt+(OP)
Also same here. Actually currently staying up late Friday night hacking on OpenAI API projects (while waiting for SpaceX Starship launch, it's quite a day for high-tech news!) - and wondering if I should even bother. Of course I will keep hacking, but...still it makes you think. Which is a very unexpected feeling.

Hugely more interested in the open source models now, even if they are not as good at present. Because at least there is a near-100% guarantee that they will continue to have community support no matter what; the missing problem I suppose is GPUs to run them.

>>siva7+B
I am at lost. Not fear, just lost.

Don't know what to do. Is my investment into their API still worth it? It feels very unstable at this moment.

>>quickt+(OP)
I'm in this boat. Not for my startup but for side projects I was absolutely pinning my hopes on them unlocking access to tools and relaxing some of their restrictions in the near future.. a future which now seem unlikely.

>>quickt+(OP)
I’d recommend trying to build out your systems to work across LLMs where you can. Create an interface layer and for now maybe use OpenAI and Vertex as a couple of options. Vertex is handy as while not always as good you may find it works well for some tasks and it can be a lot cheaper for those.

If you build out this way then when the next greatest LLM comes out you can plug that into your interface and switch the tasks it’s best at over.

>>karmas+m2
Microsoft owns it. I honestly can not imagine trying to build a business on an API owned by Microsoft of all companies.

>>Throwa+h3
Would like to elaborate a bit?

Though my investment will be still tiny at the moment, but not other multi-modal model on the market right is as good.

>>crypto+d2
Totally. I'll keep going too. I am just putting a nice GUI wrapper around the new Assistant stuff which looks damn cool. Project is half "might make some bucks" and half "see if this is good to use in the day job".

>>karmas+m2
If you're just using their completions/chat API, you're gonna be ok. As an ultimate fallback you can spin up H100s in the cloud and run VLLM atop a high param open model like Llama 70B. Such models will catch up and their param counts will increase.. eventually. But initially expect gpt-3.5-esque performance. VLLM will give you an OpenAI-like REST API atop a range of models. Keep making things :))

>>padols+t4
Thx. I will. My current interests mainly lies in benchmarking their vision model.

That being said, I might not go further relying on their APIs for something more serious

>>mebutn+Y2
The problem is swapping LLMs can require rework of all your prompts, and you may be relying on specific features of OpenAI. If you don't then you are at a disadvantage or at least slowing down your work.

>>Throwa+h3
What makes you say that?

Microsoft seems like one of the more reliable partners to build on compared to Google etc. just for the simple reason that their customers are large businesses and not breaking things for them is in their blood. Just like Windows backwards compatibility.

>>mebutn+Y2
Definitely, just like with games development, the key is to master how things work, not specific APIs.

AI tools will need a similar plugin like approach.

>>quickt+g5
Just ask the LLM to rewrite your prompts for the new model.

>>karmas+W3
I actually wouldn't be worried in the short term for exactly this reason. Microsoft has legal access to GPT-4 and is allowed to host and serve it via Azure. If OpenAI somehow tanks its API in the near term, MS is sitting on a gold mine and will make use of that by continuing to serve it. In the long term I am worried, but less worried if Sam and Greg form a competing co to continue to build.

>>quickt+g5
I have a hierarchy of templates, where I can automatically swap out parts of the prompt based on which LLM I am using. And also have a set of benchmarking tests to compare relative performance. I treat LLMs like a commodity and keep switching between them to compare performance.

>>Throwa+h3
People have been successfully building on MS APIs for four decades now. I've been for nearly three.

What exactly are you saying?

>>Throwa+h3
Isn't Microsoft famous for their insane API stability and backwards compatibility?

>>bongob+D5
Does it really have that kind of self awareness to be able to do that successfully? I feel very sceptical.

>>worlds+V6
I doubt self awareness has anything to do with it..

>>quickt+(OP)
>Can they give reassurances about their products going into the future

They wouldn't have been able to do that even before Sam's dismissal

>>karmas+m2
By that logic you could never use a third party API.

>>karmas+m2
If you are building something that is end-user facing that relies on ChatGPT then that was always a huge and risky bet on the future of OpenAI.

In addition, it would likely be some time, possibly years, before it would be ready for production.

Perhaps recent events have just brought that more clearly into focus for you.

>>quickt+(OP)
it's business and systems design 101 to actually worry about your dependencies. regardless of this drama, you should have thought about what you'd do if OpenAI shut down, or become your competitor, or gets worse, or is bought by MS or something.

> Can they give reassurances about their products going into the future?

emotional comfort is not the thing you should be looking for mate.

>>quickt+g5
Isn’t the expectation that “prompt engineering” is going to become unnecessary as models continue to improve? Other models may be lagging behind GPT4 but not by much.

>>quickt+(OP)
Indeed. It was a big enough battle to convince execs that building on top of OpenAI was ok. Now that conversation is pretty much impossible. You have the Microsoft offering, but to most muggles that just looks like them reselling OpenAI.

The board of OpenAI should have been replaced by adults a long time ago.

>>pjmlp+w5
I have a good idea how transformers work and have written Python code and trained toy ones, but end of the day right now calling OpenAI nothing I can build can beat it.

>>mebutn+Y2
That would go as well as trying to write a universal android iOS app or write ansi sql to work across database platforms. A bad idea in every dimension.

>>ahepp+y6
While I think the sentiment is overblown, if you compare Azure to AWS, Azure's stability is like Google's.

>>discon+m6
Just curious are you using something specific for the tests?

>>rlt+4f
The dream maybe. You still have to instruct these natural language agents somehow, and they all have personalities.

>>ahepp+y6
Microsoft Azure is famous for its insanely bad security: https://karl-voit.at/cloud/

>>worlds+V6
Just have it write 10 and bench them against your own.

>>Roark6+M9
What else would you call the ability for it to adapt a task for its own capabilities?

>>worlds+fp
Language modelling, token prediction. It's not much different from generating code in a particular programming language; given examples, learn the patterns and repeat them. There's no self-awareness or consciousness or understanding or even the concept of capabilities, just predicting text.

>>quickt+94
Yeah, the assistants api is pretty great. Curious if you’ve faced issues with certain things not working out of the blue forcing you to re-run threads?

For example, i have an assistant which is supposed to parse an uploaded file and extract useful info from it. To use this assistant, I create a thread and a run and attach it to the assistant with a different file-id. About half the time, the assistant simply throws up its hands and says it can’t parse the file I supplied with the thread. Retrying a few times seems to do the trick.

>>quickt+(OP)
Take my opinion with some skepticism because I am retired and the massive amount of time I put into LLMs (and deep learning in general) is only for my own understanding and enjoyment:

In all three languages I frequently use (Common Lisp, Python, and Racket) it is easy to switch between APIs. You can also use a library like LangChain to make switching easier.

For people building startups on OpenAI specific APIs, they can certainly protect themselves by using Azure as an intermediary. Microsoft is in the “stability business.”

>>Sai_+VA
The assistants API is fantastic; I was just getting started with it. This news makes me reconsider -- but I also think it's inevitable that a compatible API will be released with open source underlying LLMs. I've deployed the OpenAI-compatible completions API over Llama2 in production with vLLM, and it works perfectly.

Do you know if there are any projects working on this? Even something like a high quality json-tuned base model would go a huge way toward replicating OpenAI's current product.

>>ahepp+y6
Microsoft is famous for a lack of adoption DESPITE backwards compatibility.

>>sbroth+EI
Sorry, no idea if what you’re looking for exists. For now, I was looking at integrating with OpenAI and productising some th ing but this situation is making me nervous.

>>jeswin+s6
Why would you build a business that Microsoft could sherlock in a second?

>>ahepp+y6
No I would not say that at all. Microsoft has gone through how many Desktop APIs for Windows at this point? At least a dozen I'd think.

>>Throwa+h3
Do you use github?

>>mkl+CA
Sure but that kind of sounds like it is building a theory of mind of itself.

If it does have considerable training data including prompt and response when people are interacting with itself then I suppose it isn't that surprising.

That does sound like self awareness, in the non magical sense. It is aware of its own behaviour because it has been trained on it.