348 comments

[ 4.4 ms ] story [ 308 ms ] thread
is there away to run this on AWS instead.

we were looking to explore Llama2 for internal use

OpenAI models are exclusively Azure only. Llama2 should have an AWS option I believe?
Amazon Bedrock makes Claude 2 available, as well as some other models.
We can run llama 2 on an AWS vm if you have enough GPUs: https://lamini.ai/

Install in 10 minutes.

Make sure you have enough GPU memory to fit your llama model if you want good perf

So the public access one isn't private and secure?
I believe it’s implying the free ChatGPT collects data and this one doesn’t.
I thought sama said they don’t use data going through the api for training. Guess we can’t trust that statement
That is correct, they do not use the data going through the API for training, but they do use the data from the web and mobile interfaces (unless you explicitly turn it off).
“We don’t water down your beer”.

Oh nice!

“But that is lager”

> However, ChatGPT risks exposing confidential intellectual property.

I don't remember seeing this disclaimer on the ChatGPT website, gee maybe OpenAI should add this so folks stop using it.

If you use ChatGPT through the app or website they can use the data for training, unless you turn it off. https://help.openai.com/en/articles/5722486-how-your-data-is...
Providing my data for training doesn't imply that it risks being exposed.

If you understand what happens on a technical level, it might be possible, but OpenAI has never said this was a risk by using their product.

Absolutely. For example it doesn't say that OpenAI employees can't look at everything you write.
It's pretty clear in the FAQ to be fair.
The comment you are responding to is sarcastic
The concern is that ChatGPT is training on your chats (by default, you can opt out but you lose chat history last I checked).

So in general enterprises cannot allow internal users to paste private code into ChatGPT, for example.

As an example of this. I found that GPT4 wouldn't agree with me that C(A) = C(AA^T) until I explained the proof. A few weeks later it would agree in new chats and would explain using the same proof I did presented the same way.
This is kinda creepy. But at the same time, how do they do that? I thought the training of these models stopped in September 2021/2022. So how do they do these incremental trainings?
The exact phrase they previously used on the homepage was "Limited knowledge of world and events after 2021" - so maybe as a finetune?
but doesn’t finetuning result in forgetting previous knowledge? it seems that finetuning is most usable to train “structures” not new knowledge. am i missing something?
All the public and (leaked) private statements I have seen state that this is not happening. As siblings noted, MoE probably explains this variance.

AIUI they are using current chat data for training GPT-5, not re-finetuning the existing models.

Kind of implies that OpenAI are lying and using customer input to train their models
I’ve found that the behavior of ChatGPT can vary widely from session to session. The recent information about GPT4 being a “mixture of experts” might also be relevant.

Do we know that it wouldn’t have varied in its answer by just as much, if you had tried in a new session at the same time?

There is randomness even at t=0, there was another HN submission about that
I tested it several times, new chats never got this right at first. I tried at least 6 times. I was experimenting and found that GPT4 couldn't be fooled by faulty proofs. Only a valid proof could change its mind.

Now it seems to know this mathematical property from first prompt though.

Another thing is that using ChatGPT for European companies might be in violation with GDPR – Azure OpenAI Services are available on European servers.
“private and secure” from the company that let contractor listen to your private Teams conversation for data labeling purpose, and monitor your activity on your own computer with their OS…
Crappy clone of ChatGPT frontend, half missing, half direct copy. Implied and overly vast claims of insecurity + lack of privacy, that are narrowly true, i.e. for _Chat_GPT.

Really surprised to see this aggressive of language 1) written down 2) on Github. I'd be pretty pissed if I was OpenAI, regardless of the $10B.

I have no doubt OpenAI is on board. This is just bringing more paid users to their platform because it still uses their API.
I think OpenAI is entirely on board with the idea that OpenAI sells to consumers and Azure/Microsoft sells the same product to enterprise.

That's how it's been working for months, and if OpenAI objected they would have done something about it.

Yeah sure, I totally trust you after the Storm-0558 desaster
Pretty sure Azure has a moderation endpoint enabled by default that makes using the OpenAI API an awful experience.
Curious if anyone has done a side-by-side analysis of this offering vs just running LLaMA?

I'm currently running a side-by-side comparison/evaluation of MSFT GPT via Cognitive Services vs LLaMA[7B/13B/70B] and intrigued by the possibility of a truly air-gapped offering not limited by external computer power (nor by metered fees racking up.)

Any reads on comparisons would be nice to see.

(yes, I realize we'll eventually run into the same scaling issues w/r/t GPUs)

chatgpt is obviously a LOT better, llama doesn't even understand some prompts

and since LLMs aren't even that good to begin with, it's obvious you want the SOTA to do anything useful unless maybe you're finetuning

> and since LLMs aren't even that good to begin with, it's obvious you want the SOTA to do anything useful unless maybe you're finetuning

This is overkill. First of all, ChatGPT isn't even the SOTA, so if you "want SOTA to do anything useful", then this ChatGPT offering would be as useless as LLaMA according to you. Second, there are many individual tasks where even those subpar LLaMA models are useful - even without finetuning.

it's the SOTA for chat(prove me wrong), and you can always use the API directly

even for simple tasks they're less reliable and needs more prompt engineering

> it's the SOTA for chat(prove me wrong)

GPT-4 beats ChatGPT on all benchmarks. You can easily google these.

The distinction between GPT-4 and ChatGPT is blurry, as ChatGPT is a chat frontend for a GPT model, and you can use GPT-4 with ChatGPT. The parent probably means ChatGPT with GPT-4.
Typically when people say "ChatGPT" without specifying which specific model they refer to, they refer to gpt-3.5-turbo (in case of API - or in case of the web ui, they mean whatever model is its current web ui equivalent). But now OP says they meant GPT-4, so, sure.
Counterpoint: I don’t refer to 3.5 when I say ChatGPT. I pay for ChatGPT, and always use GPT-4. Which I believe every paying customer do.
I tried and got nothing useful. What's the difference between GPT-4 and ChatGPT Plus using GPT-4?
that is why i said FOR CHAT.

even through the API you can't easily use the regular models for chat, the parsing would be atrocious and there are hundreds of edge cases to handle.

ChatGPT4 through the API is the SOTA

openai offers finetuning too. And it's pretty cheap to do considering.
(comment deleted)
I did one. I took a few dozen prompts from my ChatGPT history and ran them through a few LLMs.

GPT-4, Bard and Claude 2 came out on top.

Llama 2 70b chat scored similarly to GPT-3.5, though GPT-3.5 still seemed to perform a bit better overall.

My personal takeaway is I’m going to continue using GPT-4 for everything where the cost and response time are workable.

Related: A belief I have is that LLM benchmarks are all too research oriented. That made sense when LLMs were in the lab. It doesn't make sense now that LLMs have tens of millions of DAUs — i.e. ChatGPT. The biggest use cases for LLMs so far are chat assistants and programming assistants. We need benchmarks that are based on the way people use LLMs in chatbots and the type of questions that real users use LLM products, not hypothetical benchmarks and random academic tests.

I think tests like "can this LLM pass an English literature exam it's never seen before" are probably useful, but yeah there's a lot of silly stuff like math tests.

I suppose the question is where are they most commercially viable. I've found them fantastic for creative brainstorming, but that's sort of hard to test and maybe not a huge market.

>> I suppose the question is where are they most commercially viable.

Fair point, though I'm not aiming to start a competing LLM SaaS service, rather i'm evaluating swapping out the TCO of Azure Cognitive Service OpenAI for the TCO of dedicated cloud compute running my own LLM -- to serve my own LLM calls currently being sent to a metered service (Azure Cognitive Service OpenAI)

Evaluation points would be: output quality; meter vs fixed breakeven points; latency; cost of human labor to maintain/upgrade

in most cases, i'd outsource and not think about it. BUT we're currently in some strange economics where the costs are off the charts for some services

I don’t know what you mean by “too research oriented.” A common complaint in LLM research is the poor quality of evaluation metrics. There’s no consensus. Everyone wants new benchmarks but designing useful metrics is very much an open problem.
I think he wants to limit evaluations to the most frequent question types seen in the real world.
How did you measure the performance?
We (at Anyscale) have benchmarked GPT-4 versus the Llama-2 suite of models on a few problems: functional representation, SQL generation, grade-school math question answering.

GPT-4 wins by a lot out of the box. However, surprisingly, fine-tuning makes a huge difference and allows the 7B Llama-2 model to outperform GPT-4 on some (but not all) problems.

This is really great news for open models as many applications will benefit from smaller, faster, and cheaper fine-tuned models rather than a single large, slow, general-purpose model (Llama-2-7B is something like 2% of the size of GPT-4).

GPT-4 continues to outperform even the fine-tuned 70B model on grade-school math question answering, likely due to the data Llama-2 was trained on (more data for fine-tuning helps here).

https://www.anyscale.com/blog/fine-tuning-llama-2-a-comprehe...

Yeah right for the three letter agencies to have a backdoor, hard pass on something that cannot be deterministic with a seed
Since the only users who would likely care about this derive far more value than the $20/month of OpenAI's direct offering. Why doesn't OpenAI market this service, but with chat history, for something like $200/month?
That's a laughable price for an enterprise subscription.

And the reason is, it's enough for OpenAI to "say" that they're "not going to use your data" - you need a cloud deployment where you can control network boundaries to _prove_ that your data isn't going anywhere it isn't supposed to.

Unless you're physically controlling the network boundaries, how are you proving that on any cloud service?
OenAI IS Microsoft. Don't get tangled in the web of creating different entities when they are all part of the same pyramid. Also GitHub IS Microsoft too!!
GitHub was acquired by Microsoft, and they are no longer legally separate entities.

Microsoft is an investor in OpenAI, but does not own it, and they are legally separate companies. OpenAI is not Microsoft and it is factually incorrect to claim that OpenAI is Microsoft.

[1] https://blogs.microsoft.com/blog/2023/01/23/microsoftandopen...

But saying they're just an investor isn't quite doing the arrangement the justice it deserves. There seems to be a lot of strings attached to that investment.

It's not just a straight trade of dollars for shares, but many further contractual obligations.

I understand that perception but "seems to be a lot of strings" is all that is publicly known. None of those further obligations seem to have been disclosed. Without that disclosure it's a bit of a conspiracy theory?

Thus, it could very well be OpenAI has taken dollars, is commercially selling its technology to Microsoft on terms which aren't special, and sama and the OpenAI executive team and board has independently concluded that engaging in the partnership is a stellar way to grow their OpenAI brand, business and valuation?

This is potentially a huge deal. Companies are concerned using ChatGPT might violate data privacy policies if someone puts in user data or invalidate trade secrets protections if someone uploads sections of code. I suspect many companies have been waiting for an enterprise version.
(comment deleted)
I imagine most companies serious about this created their own wrappers around the API or contracted it out, likely using private Azure GPUs.
Most companies are either not tech companies, or do not have the knowledge to manage such a project within reasonable cost bounds.
Most companies are trying to figure out exactly what generative AI is and how to use it in their business. Given how new this is - I doubt any large company has done much besides ban the public ChatGPT. So this is probably very relevant for them.
This is a web UI that talks to a (separate) Azure OpenAI resource that you can deploy into your subscription as a SaaS instance.
So how is it any different
Microsoft says it is more secure. And that it is enterprise. That's about it
There are legal agreements backing the separation of company data from other parties. This is what's important to big corps.
I have to imagine Big Corps are also concerned about liability / risk when generating things with OpenAI products - at least until there is some sort of settled law around using models trained on this kind of data.
Yes, those concerns exist, but they're also practically impossible to enforce.

At my enterprise, it's a three step solution, two of which don't work.

1. Written policy concerning LLM output and its risks, disallow it for being used for any kind of official documentation or decision making. (This doesn't work, because no one wants to use their own brain to do tedious paperwork.)

2. Block access to public LLM tools via technical means from company owned end-user devices. (This doesn't work because people will just open ChatGPT on their home PC or mobile.)

3. Write and provide our own gpt-3.5 frontend, so that when people ignore rules #1 and #2 we have logs, and we know we're not feeding our proprietary info to to OpenAI.

How is this different from the other OpenAI GUI? Why another one by Microsoft? https://github.com/microsoft/sample-app-aoai-chatGPT.
I bet there are plenty of OKR/KPIs now tied to AI at Microsoft.
Bigger companies are cautious about using GPT-style products due to data security concerns. But most big companies trust Microsoft more or less blindly.

Now that Microsoft has an official "enterprise" version out, the floodgates are open. They stand to make a killing.

This is an internal ChatGPT, whereas that sample is ChatGPT constrained to internal search results (using RAG approach). Source: I help maintain the RAG samples.
Would it be too much to mention somewhere in the README what this repo actually contains? Just docs? Deployment files? Some application (which does..something)? The model itself?
The repo contains the UI code, not the model or anything else around ChatGPT, it just uses Azure’s ChatGPT API which doesn’t share data with OpenAI.
So basically – what you really need to do to run Azure ChatGPT is go and click some buttons in the Azure portal. This repo is a sample UI that you could possibly use to talk to that instance, but really you will probably always build your own or embed it directly into your products.

So calling the repo "azurechatgpt" is misleading. It should really be "sample-chatgpt-api-frontend" or something of that sort.

Correct. If offers a front-end scaffolding for your enterprise ChatGPT app. Uses Next/NextAuth/Tailwind etc. for deployment on Azure App Service that hooks into Azure Cosmos DB and Azure OpenAI (the actual model).
(comment deleted)
Isn’t there also some sort of backend stuff in there? How else would it keep track of history and accept documents.

I don’t know enough typescript to understand where the front end stops and the backend begins I this code

No better than the API.
Our company is pushing everyone to use a similar offering. Most of the company is doing low value work … still using excels even though we have a custom ERP. Now seeing people who couldn’t write a coherent email before write 3 page emails. The illusion of being productive by doing more work even though it has zero impact on the bottom line. It’s insane how inefficient organisations are. No doubt we’ll have some KPI soon about using the tool.
If anything it's less productive because people have to parse all that nonsense.

I was gobsmacked to hear a friend say that their work guidance is to use ChatGPT to write letters to external clients for example. I know for sure I'd be insulted if someone sent me paragraphs of text to read created from a sentence long prompt. I'd rather have the prompt, my time is valuable as well.

ahhhh, but they're pasting the 3 page email into ChatGPT ("summarize this"). The future is here.
Yeah that's one of the insane things that will happen.

Very soon everyone will in effect "hide" behind an agent that will take all kinds of decisions on one's behalf. Everything from writing e-mails to proposals but also to sue someone, make financial decisions, and be a filter that transforms everything going in or out.

I can't imagine this world really. How the hell are people going to compete or stand out? Doesn't it seem that what little meritocracy existed wills soon drown in noise?

I was scared about organizations doing this and losing their connection to the humans they serve.

The realization that individuals will also have this barrier to the world is even scarier.

If it goes that way we could be looking at a change to society on the level of social media, again. Mad.

I write emails and put it into chatgpt and ask it to make it more concise or point out issues. No utility in asking chatgpt to needlessly expand the text...
I think the more common case is to have a handful of bullet points and some notes and ask chat GOT to put into a coherent letter for an external customer with the goal of XYZ. I’ve done similar things and it is a huge timesaver. I still have to edit it, but it gives me a start that’s probably on par to what a Junior engineer would write as a first draft.
Exactly right. If you increase entropy you need energy to reduce it back. It be more valuable to take crap that humans have put together incoherently and summarizing it. (Perhaps someone should put a GPT on the other end in order to read it)

I honestly don’t know why we’re so obsessed with having LLMs generate crap. Especially when they’re very capable of reducing, simplifying. Imagine penetrating legal texts, political bills, obtuse technical writing, academic papers and making sense of those quickly. Much more useful imo.

You'll just have people reversing it into a summary on the other end, kind of like a "text" chat where both sides are using text-to-speech and speech-to-text instead of having a phone call.
The amount of othewise very smart people who completely lose the ability to think critically when it comes to "AI" is really interesting to me.

I'm not anti-AI; I've recommended that we use it at work a few times where it made sense and was backed by evidence/bencharmks. But for essentially any problem that comes up someone will try to solve it with ChatGPT, even if it demonstrably can't do the job. And these are not business folks, these are engineering leaders who absolutely have the capability to understand this technology.

What ERP are you using?

We've found some early success selling to companies with older "long-tail" ERP's. I've been finding a new one every day.

It’s a proprietary ERP completely custom. Think it was deployed through an acquisition. The problem isn’t the ERP it’s the business. “We want custom processes” but hire the cheapest developers possible to maintain the ERP and then complain about bugs. “We’re agile™” … but have the same inefficient processes for the last 3 years. Cargo cult org, the CEO was taking about Black Swans during COVID … even though Nassim Taleb explicitly said COVID wasn’t a black swan event.
(comment deleted)
I’ve learned that the most important writing skill is to figure out what you’re trying to say — this is a rather important prerequisite to writing well.

Naively asking a chatbot to write for you does not help with this at all.

It would be interesting to try to prompt ChatGPT to ask questions to try to figure out what the user is trying to write and then to write it.

Darn I just spent a week or so working on a ChatGPT clone that used Azure ChatGPT API due to the privacy aspect. Wasted effort I guess.
This is exactly the same
Welcome to the club :)
I'm also in this club but we wrote it months ago.
"Private and secure"

From Microsoft?

Ha.

A lot of companies are already using projects like chatbot-ui with Azure's OpenAI for similar local deployments. Given this is as close to local ChatGPT as any other project can get, this is a huge deal for all those enterprises looking to maintain control over their data.

Shameless plug: Given the sensitivity of the data involved, we believe most companies prefer locally installed solutions to cloud based ones at least in the initial days. To this end, we just open sourced LLMStack (https://github.com/TryPromptly/LLMStack) that we have been working on for a few months now. LLMStack is a platform to build LLM Apps and chatbots by chaining multiple LLMs and connect to user's data. A quick demo at https://www.youtube.com/watch?v=-JeSavSy7GI. Still early days for the project and there are still a few kinks to iron out but we are very excited for it.

Can you plug this together with tools like api2ai to create natural language defined workflow automations that interact with external APIs?
There is a generic HTTP API processor that can be used to call APIs as part of the app flow which should help invoke tools. Currently working on improving documentation so it is easy to get started with the project. We also have some features planned around function calling that should make it easy to natively integrate tools into the app flows.
You can use unfetch.com to make API calls via LLMs and build automations. (I'm building it)
Is it possible to not use Google with unfetch.com?
Google is just so easy for login. No need to deal with password forgot, reset, email verification etc. But I'll add login via magic link soon.
I find it interesting to see how competitive this space got so quickly.

How do these stacks differentiate?

Quality and depth of particular types of training data is one difference. Another difference is inference tracking mechanisms within and between single-turn interactions (e.g., what does the human user "mean" with their prompt, what is the "correct" response, and how best can I return the "correct" response for this context; how much information do I cache from the previous turns, and how much if any of it is relevant to this current turn interaction).
With Louie.ai, there is a lot of work on specialization for the job, and I expect the same for others. We help with data analysis, so connecting enterprise & common data sources & DBs, hooking up data tools (GPU visuals, integrated code interpreter, ...), security controls, and the like, which is different from say a ChatGPT for lawyers or a straight up ChatGPT UI clone.

Technically, as soon as the goal is to move beyond just text2gpt2screen, like multistep data wrangling & viz in the middle of a conversation, most tools technically struggle. Query quality also comes up, whether quality of the RAG, the fine tune, prompts, etc: each solves different problems.

I see this as more of a 'Migration problem'. Why is this offered as a SaaS as opposed to a consulting service?

The code to organize and vectorize the documentation, endpoints and run it through a variety of models and injection prompting like two shots, etc. are going to be highly customized. The 'Base-code' there, is not exactly trivial, but anyone reading all the llama index docs can do it.

Then it's just run of the mil, analyst level integration that you provide to the client on a T&M, or fixed price costs.

I agree there's room for consulting, but as a new field, there's a lot of software currently missing for each vertical. Today, that's manual labor by consultants, but as the field matures... consultants should be doing things specialized to the specific customer, not what can be amortized across adjacent verticals. Top software engineers investing into software over time deliver substantially more in substantially less time, and consultants should be integrating that, not competing head-on.
Interesting project - was trying it out, found an issue in building the image - have opened an issue on github - please take a look. Also do you have plan to support llama over openai models.
Thanks for the issue. Will take a look. In the meantime, you can try the registry image with `cp .env.prod .env && docker compose up`

> Also do you have plan to support llama over openai models.

Yes, we plan to support llama etc. We currently have support for models from OpenAI, Azure, Google's Vertex AI, Stability and a few others.

> we believe most companies prefer locally installed solutions to cloud based ones

We've also seen a strong desire from businesses to manage models and compute on their own machines or in their own cloud accounts. This is often part of a hybrid strategy of using API products like OpenAI for rapid prototyping.

The majority of (though not all) businesses we've seen tend to be quite comfortable using hosted API products for rapid prototyping and for proving out an initial version of their AI functionality. But in many cases, they want to complement that with the ability to manage models and compute themselves. The motivation here is often to reduce costs by using smaller / faster / cheaper fine-tuned open models.

When we started Anyscale, customer demand led us to run training & inference workloads in our customers' cloud accounts. That way your data and code stays inside of your own cloud account.

Now with all the progress in open models and the desire to rapidly prototype, we're complementing that with a fully-managed inference API where you can do inference with the Llama-2 models [1] (like the OpenAI API but for open models).

[1] https://app.endpoints.anyscale.com/

(comment deleted)
One thing I still don't understand is what _is_ the ChatGPT front end exactly? I've used other "conversational" implementations built with the API and they never work quite as well, it's obvious that you run out of context after a few conversation turns. Is ChatGPT doing some embedding lookup inside the conversation thread to make the context feel infinite? I've noticed anecdotally it definitely isn't infinite, but it's pretty good at remembering details from much earlier. Are they using other 1st party tricks to help it as well?
It uses a sliding context windows. Older tokens are dropped as new ones stream in
I don't believe that's the whole story. Other conversational implementations use sliding context windows and it's very noticable as context drops off. Whereas ChatGPT seems to retain the "gist" of the conversation much longer.
I mean, I explicitly have the LLM summarize content that's about to fall out of the window as a form of pre-emptive token compression. I'd expect maybe they do something similar.
I feel like we're describing short vs long term memory.
That’s exactly what it is. It’s just it turns out you need very good generalized or focused simple reasoning to do accurate compression or else the abstraction and movement to long term memory doesn’t include the most important content. Or worse distracting details.

I’ve been working on short and long term memory windows at allofus.ai for about 6 months now and it’s way more complex than I had originally thought it would be.

Even if you can magically extend the content window, the added data confuses and waters down the reasoning of the LLM. You must do layered abstraction and compression with goal based memory for it to continue to reason without distraction of irrelevant data.

It’s an amazing realization, almost like a proof that memory is a kind of layered reasoning compression system. Intelligence of any kind can’t understand everything forever. It must cull the irrelevant details, process the remains and reason on a vector that arises from them.

Is it unfair to consider this some kind of correlate to the Nyquist theorem that makes me skeptical of even the theoretical possibility of AGI claims?
I consider GPT4 AGI, so I'm probably not the one to ask this too. It reasons, it understands sophisticated topics, it can be given a purpose and pursue it, it can communicate with humans, and it can perform a reasonable task considering its modalities.

I don't really know what any sort of "big leap" beyond this people are expecting, incremental performance for sure. But what else?

I guess for me it needs to have active self-reflection and the ability to act independently/without directions. I'm sure there are many other criteria if I think about it some more, but those two were missing from your list.
This is mostly just that gpt4 API/app have this disabled rather than it’s not capable.

When you enable it, it is pretty shocking. And it’s pretty simple to enable. You just give it a meta instruct to decide when to message you and what to store to introspect on.

As a frequent user of the OpenAI APIs, I don't really know what you are talking about here. Could you point me to some documentation?
At least in 3.5 it's very noticeable when the context drops. They could use summarization, akin to what they are doing when detecting the topic of the chat, but applied to question-answer-pairs in order to "compress" the information. But that would require additional calls into a summarization LLM so I'm really not sure if it is worth it. Maybe they dump some tokens they have on a blacklist or text snippets like "I want to" or replace "could it be that" with "chance of".
This is one of the things that make me uncomfortable about proprietary llm.

They get task performance by doing a lot more than just feeding a prompt straight to an llm, and then we performance compare them to raw local options.

The problem is, as this secret sauce changes, your use case performance is also going to vary in ways that are impossible for you to fix. What if it can do math this month and next month the hidden component that recognizes math problems and feeds them to a real calculator is removed? Now your use case is broken.

Feels like building on sand.

I'm not sure you realize how proprietary LLMs are being built on.

No one is doing secret math in the backend people are building on. The OpenAI API allows you to call functions now, but even that is just a formalized way of passing tokens into the "raw LLM".

All the features in the comment you replied to only apply to the web interface, and here you're being given an open interface you can introspect.

It was a contrived example to make a point, one that seems to have flown over your head.
No it was a bad (straight up wrong) example because you don't understand how people are building applications on proprietary LLMs.

If you did you'd also know what evals are.

Thank you for pointing that out - I had assumed that things were not how they are.

Although performance has varied over time https://arxiv.org/pdf/2307.09009.pdf I also notice that the API allows you to use a frozen version of the model which avoids the worries I mentioned.

That was a pretty deeply flawed paper, one of the largest drops recorded was simple parsing errors in their testing:

https://www.aisnakeoil.com/p/is-gpt-4-getting-worse-over-tim...

Overall evals and pinning against checkpoints are how you avoid those worries, but in general, if you solve a problem robustly, it's going to be rare for changes in the LLM to suddenly break what you're doing. Investing in handling a wide range of inputs gracefully also pays off on handling changes to the underlying model.

> No one is doing secret math in the backend people are building on.

How do you know that? With SaaS you are at the mercy of the vendor.

They definitely do some proprietary running summarization to rebuild the context with each chat. Probably a RAG like approach that has had a lot of attention and work
This is effectively my question. I assume there is some magic going on. But how many engineering hours worth of magic, approximately? There is a lot of speculation around GPT-4 being MoE and whatnot. But very little speculation about the magic of the ChatGPT front end specifically that makes it feel so fluid.
That's mostly because there's very little value in deep speculation there.

It's not particularly more fluid than anything you couldn't whip up yourself (and the repo linked proves that) but there's also not much value in trying to compete with ChatGPT's frontend.

For most products ChatGPT's frontend is the minimal level of acceptable performance that you need to beat, not an maximal one really worth exploring.

What front end is better than ChatGPT? Is the OP implementation doing running summarization or in-convo embedding lookup?
It sounds like a cop-out but: it's one made for your use-case.

If you're letting people do fun long-form roleplay adventures using summarization alongside some sort of named entity K-V store driven by the LLM would be a good strategy.

If you're building a tool that's mostly for internal data, something that leans heavily into detailed answers with direct verbatim citations and having your frontend create new threads when there's a clear break in the topic of a request is a clever strategy since quality drops with context length and you want to save tokens for citations.

People who are saying LLMs suck or are X or are Y are mostly just completely underutilizing them because LLMs make it super easy to solve problems superficially: when it comes to actually scaling those solutions to production you need more than random RAG vector database wrappers.

>alongside some sort of named entity K-V store driven by the LLM

I'd be curious to hear more about how exactly this works. You do NER on the prompt (and maybe on the completion too) and store the entities in a database and then what? How does the LLM interact with it?

LLMs thrive at completely ambiguous classifications: you can have them extract entities and something like "a list of notable context".

Let's say we want to let our chat remember the character slammed the door last time they were in Village X with the mayor in their presence and have the mayor comment next time they see the player.

Every X tokens we can fire a prompt with a chunk of conversation and a list of semantically similar entities that already exist, letting the LLM return an edited list along the lines of:

   entity: mayor

   location: village X

   priority: HIGH

   keywords: town hall, interact, talk

   "memory, likelyEffect"[]: door slammed in face, anger at player
Now we have:

- multiple fields for similarity search

- an easy way to manage evictions (sweep up lowest priority)

- most importantly: we're providing guidance for the LLM to help it ignore irrelevant context

When the user goes back to village X we can fetch entities in village X and whittle that list down based on priority and similarly to the user prompt.

None of this has any determinism: instead you're optimizing for the illusion of continuity and trading off predictability.

You're aiming for players being shocked that next time they talk to the mayor he's already upset with them, and if they ask why he can reply intelligently.

And to my original point while this works for a game-like experience, you wouldn't want to play around with this kind of fuzzy setup for your companies internal CRM bot or something. You're optimizing for the exact value proposition of your use-case rather than just trying to throw a raw RAG setup at it

Logic for azure chatgpt's "infinite context" summarisation is in https://github.com/microsoft/azurechatgpt/blob/main/src/feat...

*Edit Azure chatgpt, would be amazed/disappointed if chatgpt used langchain.

That doesn't really look right to me, it looks like that's for responding regarding uploaded documents. I see nothing related to infinite context.

Also this is the azure repo from OP, nothing to do with the actual ChatGPT front-end that was asked about. I highly doubt the official ChatGPT front-end uses langchain, for example.

I don't see anything related to an infinite context in there. There's only a reference to a server-side `summary` variable which suggests that there is a summary of previous posts which will get sent along with the question for context, as is to be expected. Nothing suggests an infinite context.
This is awesome to see, feels heavily inspired (in a good way) by the version we made at Vercel[1]. Same tech stack: Next.js, NextAuth, Tailwind, Shadcn UI, Vercel AI SDK, etc.

I'd expect this trend of managed ChatGPT clones to continue. You can own the stack end to end, and even swap out OpenAI for a different LLM (or your own model trained on internal company data) fairly easily.

[1]: https://vercel.com/templates/next.js/nextjs-ai-chatbot

Really, they use Vercel AI SDK?
Well, they did. The repo appears to have been deleted.
This seems like such an obvious thing to do.

I see the use of general purpose LLMs like ChatGPT, but smaller fine tuned models will probably end up being more useful for deployed applications in most companies. Off topic, but I was experimenting with LLongMA-2-7b-16K today, running it very inexpensively in the cloud, and given about 12K of context text it really performed well. This is an easy model to deploy. 7B parameter models can be useful.

Is there an easy way to play with these models, as someone who hasn't deployed them? I can download/compile llama.cpp, but I don't know which models to get/where to put them/how to run them, so if someone knows about some automated downloader along with some list of "best models", that would be very helpful.
For llama, the 4bit quantized ones, small models like the 7b one. The ggml format. That will run on your local cpu. Google those terms too. you can look on hugging face for the actual model to download then load it and send prompts to it
Thanks, maybe it's as easy as downloading the ggml and running it with Llama.cpp. I'll try that, thanks!
there is also a python wrapper that has a web ui built in for llama.cpp, if it wasnt easy enough already
I usually run them on Google Colab, and occasionally a GPU VPS on Lambda Labs. Hugging Face model card documentation usually have a complete Python example script for loading and running a model.
(comment deleted)
How does this work in terms of utilization? The isolation presumably means buying gpu capacity and only using a %?
Basically you get N tokens/second (or if it was minute, can check tomorrow if you're really interested) per deployment. So if you would outgrow on deployment, just deploy another one (with the associated costs of course).

One deployment = a deployed model which you can query

On top of that, depending on the model you're using, you also see a cost increment for each 1000 request you make.

Ah right so it is somewhat shared. Not like your own gpu type situation
Interesting release, though still lacking a few features I've had to resort building myself such as code summary, code base architecture summary, and conversation history summary. ChatGPT (the web UI) now has the ability to execute code, and make function callbacks, but I prefer running that code locally, especially if I am debugging. This latter part, conversation history summary, is something that ChatGPT web UI does reasonably well, giving it a long history, but a sentiment extraction and salient detail extraction before summarizing is immensely useful for remembering details in the distant past. I've been building on top of the GPT4 model and tinkering with multi-model (gpt4 + davinci) usage too, though I am finding with the MoE that Davinci isn't as important. Fine tuning has been helpful for specific code bases too.

If I had the time I'd like to play with an MoE of Llama2, as a compare and contrast, but that ain't gonna happen anytime soon.

We have this at IKEA for a while now. Not impressed, but funny to read the hallucinations.
I'd expect a company like IKEA to have the expertise to create interfaces specific to their workflows so hallucinations aren't an issue.

Imo if you're making an open ended chat interface for a business, you're doing it wrong.

Have you considered instruction-tuning it with text, instead of just pictures?