>Dedicated capacity offering is limited by GPU availability. OpenAI also offers dedicated capacity, which provides customers with a private copy of the model. To access this service, customers must be willing to commit to a $100k spend upfront.
How many shell corporations are intelligence agencies seeding right now?
Last night I was musing how many different countries' intelligence agencies have moles working at OpenAI currently. Gotta be at least 6, maybe as high as two dozen?
US, France, Israel ... then who? Maybe another five eyes country like the UK? Possibly China? I'm pretty skeptical Russia would be able to get someone in there but maybe.
I had been putting theories in comments but they kept getting flagged or banned or downvoted to oblivion, but maybe its time has come. I'll keep it tame. If you are curious you can google connections of OpenAI board of directors, Will Hurd, In-Q-Tel trustees, Allen and Company, etc. There is more but whatever. The conspiracy theory is that 'the govt stepped in' during the six month pause after gpt-4 was trained and before it was released.
Private instance means a dedicated endpoint fully managed by OpenAI. You do not get model access or anything a regular API user doesn't already get, except your API url will be something like customer123.openai.com/api instead of api.openai.com/api
I never know if I have an inside scoop or an outside scoop. Has Hyena not addressed the scaling of context length [1]? I know this version is barely a month old but it was shared to me by a non-engineer the week it came out. Still, giving interviews where the person takes away that the main limitation is context length and requires a big breakthrough that already happened makes me seriously question whether or not he is qualified to speak on behalf of OpenAI. Maybe he and OpenAI are far beyond this paper and know it does not work but surely it should be addressed?
As someone who is in the field: papers proposing to solve the context length problem come out every month. Almost none of the solutions stick or work as well as a dense or mostly dense model.
You'll know when the problem is solved when model after consistently use a method. Until then (and especially if you're not in the field as a researcher), assume that every paper claiming to tackle context length is simply a nice proposal.
Yes. Solving context length has been tried in hundreds of different approaches, and yet most LLMs are almost identical to the original one from 2017.
Just to name a few families of approaches: Sparse Attention, Hierachical Attention, Global-Local Attention,Sliding Window Attention, Locality sensitive hashing Attention, State space model, EMA gated attention.
Notably, human working memory isn't great either. Which begs the question (if the comparison is valid) as to whether that limitation might be fundamental.
The failure mode is that only long context tasks benefit, short ones work fast enough with full attention, and better. It's amazing that OpenAI never used them in any serious LLM even though training costs are huge.
Which is all the more curious, considering OpenAI said this only in January:
> Azure will remain the exclusive cloud provider for all OpenAI workloads across our research, API and products [1]
So... OpenAI is severely GPU constrained, it is hampering their ability to execute, onboard customers to existing products and launch products. Yet they signed an agreement not to just go rent a bunch of GPU's from AWS???
Did someone screw up by not putting a clause in that contract saying "exclusive cloud provider, unless you cannot fulfil our requests"?
Barring a revolution in chip manufacture, there likely will always be a lack of supply relative to consumer GPUs. The size of the die results in terrible yields.
>So... OpenAI is severely GPU constrained, it is hampering their ability to execute, onboard customers to existing products and launch products. Yet they signed an agreement not to just go rent a bunch of GPU's from AWS???
> Did someone screw up by not putting a clause in that contract saying "exclusive cloud provider, unless you cannot fulfil our requests"?
The relevance here is that Azure appears to be very well designed to handle the hardware failures that will inevitably happen during a training run taking weeks or months and using many thousands of GPUs... There's a lot more involved than just renting a bunch of Amazon GPUs, and anyways the partnership between OpenAI and Microsoft appears quite strategic, and can handle some build-out delays, especially if they are not Microsoft's fault.
One of Azure's unique offerings is very large HPC clusters with GPUs. You can deploy ~1,000 node scale sets with very high speed networking. AWS has many single-server GPU offerings, but nothing quite like what Azure has.
Don't assume Microsoft is bad at everything and that AWS is automatically superior at all product categories...
Whether MS is good or not isn't really the point. If they're constrained by GPU availability, being locked in to any specific provider is going to be a problem.
Large scale sets are only needed for training. For inference, 8x NVIDIA A100 80G will allow inference for 300b models (GPT-3 is 175b) or 1200b models with 4-bit quantization (quantization impact is negligible for large models), so a single machine is sufficient.
I don't think Amazon offers what Azure does (yet) in terms of HPC or multi-GPU capacity. The blog post doesn't say how long the agreement is for, but the relationship probably makes sense at the moment.
All the cloud providers are building out this type of capacity right now. It's already having a big impact in terms of quarterly spend, which we just saw in the NVDA Q1 results. AWS, Azure, and GCP for sure, but also smaller players like Dell and HPE and even NVidia themselves are trying to get into this market. (Disclaimer: I work at one of these places but don't feel like saying which). I suspect the GPU constraints won't be around too long, at which point we'll find out if OpenAI made a contractual mistake.
Having been recently taken aboard by the mothership I expect they'll start trying to tune out anything related to programming to push people towards co-pilot X..
It's pretty hilarious and annoying to see bing start to write code only to self censor itself after a few lines (deleting what was there! no wonder these guys love websockets and dynamic histories)
When i tried out bing 1-2 weeks ago and asked it for code it would start writing and after a few lines, realize it was writing code and stop and delete what it had written.
I tried it again tonight and it seems like they fixed it to only produce small amounts of mediocre code instead.
I’m hoping GPT will remove the information cutoff date. I write plenty of terraform/AWS and it’s a bit of a pain that the latest API isn’t accessible by GPT yet.
There’s been quite a bit happening in the programming space since sept 2021.
I use GPT to keep things high level and then do my normal research methodology for implementation details.
It's not like an arbitrary imposition, that's the data it was trained on and it's expensive to train. I hope they find a way to continually train in new information too but it's not like they can just remove the cutoff date.
Not disagreeing, but a fascinating thing they did (as a one-off fine-tune?) was teach ChatGPT about the openai python client library, including the features that were added after the cutoff date.
I enjoy using GPT4 as a co-programmer, and funny enough it is very challenging to get advice on Microsoft's own .NET MAUI because that framework was in prerelease at the time the model was trained.
My understanding is right now they essentially need to train a new model on a new updated corpus to fix this, but maybe some other techniques could be devised...or they'll train something more up to date.
You might actually get pretty far if you just went through the Microsoft docs and created a bunch of really concise examples and fed that as the start of the prompt. Use like 6-7kb for that and then the question at the end.
I have had some luck doing exactly that, and not even as efficiently as you describe - If my question is limited enough that the discussion won't overwhelm the context window I've found I can just paste in big chunks of the docs wholesale like a 'zero shot.'
As stated your request is entirely impossible. They cannot simply "remove the cut-off date". It takes months and huge amounts of hardware to train. Then they do the reinforcement adjustments on top of it while researching how to train the next batch.
Injecting the context yourself can help a lot. I frequently copy in a bunch of example code at the beginning of the conversation to help prime ChatGPT on APIs it knows nothing about.
For smaller projects that will fit, I've taken to: `xclip *` and then pasting the entire collection of files into ChatGPT before describing what I want to do.
Keep in mind that GPT-4 has a max context size of ~8000 tokens, if I recall correctly. That means that in any given ChatGPT session the bot only remembers roughly the last ~6k words, as a trailing window. It'll forget the stuff at the beginning fast.
All AI companies (OpenAI included) are now working full tilt on making AIs improve themselves (writing their own code, inventing new pipelines etc). I don't know why choose anything else to work on. This is a prime directive, that will bring the greatest payoff.
I disagree since GPUs are a major constraint currently and that skilled specialists outperform GPT-4 almost always as long as they stay in their domain.
Will they use copilot(s) to improve the models? Yes, but they have been doing that since 2021 already (the release year of GitHub Copilot).
if this was currently possible wouldn't it lead to sentient/superhuman AI rapidly?
>tell AI to make itself more efficient by finding performance improvements in human written code
>that newly available processing power can now be used to find more ways to improve itself
>flywheel effect of AI improving itself as it gets smarter and smarter
eventually you'd turn it loose on improving the actual hardware it runs on. I think the question now is really how far transformers can be taken and if they are really the path to "real" AI.
Within a couple of years of improvement processes like you suggest will actually be really dangerous and stupid.
Also don't confuse all other types of human/animal characteristics like sentience with intelligence. They are different things. Things like sentience, subjective stream of experience, or other aspects of being alive don't just accidentally fall out of larger training datasets.
And we should be glad. The models are going to be orders of magnitude faster (and perhaps X times higher IQ) than humans within a few years. It is incredibly foolish to try to make something like that into a living creature (or emulation of living).
Intelligence is about action, and sentience is about qualia, which I equate to perceptions coloured by values. Action is visible and qualia are hidden, but they are closely interconnected: we choose our actions in accordance with our values and situation at hand.
> Things like sentience, subjective stream of experience, or other aspects of being alive don't just accidentally fall out of larger training datasets.
I disagree, language is all we need. Agency? Encode your “internal needs” as prompts, periodically generate prefixes from these, append them to incoming prompts. Self-awareness? Summarize this internal dialogue, reflect on it with a few iterations, add the results to the common prefix. Sentience? Attach some sensors, summarize their observations with the language model, prepend to prefix. Actions? Make the model output commands that some servos or other interfaces understand. Etc, etc.
And, of course, it would be extremely _cool_ to make something like that into a living creature, and lots of labs are already doing that. Fear and luddism should not stay in the way of curiosity.
If we humans cannot improve our own intelligence, making something smarter than us is an evolutionary imperative.
That's not what I meant. What you describe is deliberate engineering. Not accidentally falling out such as from just training on larger and larger datasets which some people think will result in digital consciousness or something through "emergence".
It is almost certain that the next stage of intelligence will be digital. But it is very foolish and unnecessary to try to speed that along.
It is likely that we have a century or two max left in control of the planet, regardless of what we do. On some level I agree that totally suppressing it indefinitely would be a shame.
When I said "living" I meant digital life. Such as those things you describe and others including control-seeking, self-preservation, and reproduction which are all central to living beings.
The problem is that AI will soon think 100 or more times faster than humans. This is anticipated based on the history of increases in computing efficiency and the fact that we are now optimizing a very specific system (LLMs). Humans will not in any way be able to keep up.
This is not luddism. I have a service that connects GPT-4 to Linux VMs on the internet to install or write software. I think this technology is great and has a lot of positive potential.
But when you deliberately try to emulate animals (like humans) and combine that with hyperspeed and other superintelligent characteristics, you are essentially approaching suicide or at least, abdicating all responsibility for your environment. There is no way to prevent such a thing from making all of your decisions for you.
The speed difference will be incredible. Imagine a bullet time scene where everyone seems to be moving in extreme slow motion. Now multiply by 10 so they are so slow they seem completely frozen.
This level of performance is coming in five years or less.
While I don't want to suppress the evolution of intelligent life in our corner of the universe, I also am not ready to join a death cult. Especially not accidentally.
The bit about plugins not having PMF is interesting and possibly flawed. I, like many others, got access to plugins but not the browsing or code interpreter plugins which feel like the bedrock plugins that make the whole offering useful. I think there's also just education that has to happen to teach users how to effectively use other plugins, and the UX isn't really there to help new users figure out what to even do with plugins.
Yea, seems weird to allow people to use plugins, but not all of them. Then have the gall to say that no one is using plugins, yea because half of them don't have any context outside of America.
I tried the plugins - they honestly didn't seem to work very well. GPT-4 wasn't sure when it could use a plugin, or when it should talk about how it would do something. I wasn't able to get the plugins to activate most of the time.
Most of the plugins are garbage and for those that aren't, most seem like they would be better as a chat like experience in the original app than the OpenAI app
For what it's worth I've found the model actually performs significantly worse at most tasks when given access to browsing, in part because it relies on that instead of its own in built knowledge.
I haven't found a good way to have it only access the web for specific parts of its response.
The only plugin I found useful was the diagramming one, forgot what it's called. But you can quickly make code (or other) flowcharts etc. And browsing in rare cases.
Great content and great answers except for open source question. Sam is saying that he doesn’t think anyone would be able to run the code at scale so they didn’t bother? Seems like a nonsense answer, maybe I’m misunderstanding. The ability for individuals or businesses to effectively run and host the code shouldn’t have an impact on the ability to open source.
I love the tongue-in-cheek paradox myth that the Bitcoin whitepaper was written by a future god-AI to increase demand for GPUs (and thus boost supply) so we are able to assemble the future god-AI.
Somehow I'm super sensitive to the audio (or might be video) and start feeling nauseous after a short time. Is there an explanation to this? I think it's that scratchy humming background sound.
I don't know, except that I agree that it's horribly mixed. It's almost impossible to make out what people are saying in certain parts, and that causes cognitive load that degrades the experience. Also, the story is completely nonsensical. The only good thing about that movie is that time-reversed fight scene and even that was kind of questionable.
> I love the tongue-in-cheek paradox myth that the Bitcoin whitepaper was written by a future god-AI to increase demand for GPUs (and thus boost supply) so we are able to assemble the future god-AI.
I know it's a joke, but the hole is the god-AI couldn't have been that smart, since cryptocurrency-mining quickly switched to ASICs, which muting the demand increase for GPUs.
> But not before ramping up development and production of GPUs.
Did the GPU manufactures ever embrace cryptocurrency? IIRC, they actually tried to discourage it (e.g. by butting throttling into mass market models to discourage their use for computation).
Desktops is a very very key word there, I believe that’s why they repeat it so much. And it’s all tongue in cheek, and we all largely understand mining drove up gpu demand
Production MAYBE, I'm not sure what makes you think bitcoin would have ramped up the development of GPUs. What part of the last 10 years of GPU development looks bitcoin-focused to you? They're still very very focused on rendering and machine learning, not computing hashes.
Well, humans switched from using using brains to store all their memories once they could dump data onto external media via writing. Much like how crypto switching to ASICs frees up GPU capacity for AGI, writing freed the brain to develop higher GI.
Nah, many of the technical people who knew Hal Finney personally were claiming that the Caltech alumnus did write the original Bitcoin whitepaper not any random guys on the Internet [1].
The talk/conversation appeared to me not as OpenAI future plan but more on the CEO lamenting on how severely limited the company by the GPU or the lack thereof. It just a cheeky ploy by a CEO of an AI company that currently at 30B USD valuation to get more money in order to buy several [fill in the blank] of these most advanced GPU systems [2].
[1]Nakamoto's Neighbor: My Hunt For Bitcoin's Creator Led To A Paralyzed Crypto Genius:
Most Bitcoin miners have been using ASIC chips that compute nothing but Bitcoin format SHA256 since long ago so it’s not increasing demand for GPUs. Ethereum is but it has already switched to PoS.
His statements on open sourcing in this interview/write-up is somewhat in conflict with his recent statement made last week in Munich https://youtu.be/uaQZIK9gvNo?t=1170, where he explicitly said the Frontier of GPT won't be open sourced due to what they perceive as safety reasons, https://youtu.be/uaQZIK9gvNo?t=1170 (19:30 - 22:00).
it's legal to make contradictory statements that's one of the job of a ceo and it's why they aren't usually overly literal types you know the kind i'm talking about
He was talking about open sourcing GPT-3. That is not the frontier.
The frontier is the multimodal versions of GPT-4 which he just said wasn't even going to public release until next year. Or whatever they are on now which they are carefully not calling GPT-5.
> He reiterated his belief in the importance of open source and said that OpenAI was considering open-sourcing GPT-3. Part of the reason they hadn’t open-sourced yet was that he was skeptical of how many individuals and companies would have the capability to host and serve large LLMs.
Am I reading this right? "We're not open sourcing GPT-3 because we don't think it would be useful to anyone else"
Really? Even after saying this? "While Sam is calling for regulation of future models, he didn’t think existing models were dangerous and thought it would be a big mistake to regulate or ban them."
Why couldn’t that be true? E.g. even scientists who worked on the Manhattan Project (justifiably) had antipathy toward the much more powerful hydrogen bomb.
It’s possible to think squirt guns shouldn’t be regulated but AR-15s should, or AR-15s shouldn’t but cruise missiles should. Or driving at 25mph should be allowed but driving 125mph shouldn’t.
And he'll do that by... keeping ChatGPT models away from individuals and small teams and in the hands of a few large deep-pocketed firms?
The great thing about open source is that people can try different approaches and gravitate towards what works best for them. Sam knows that of course, he's just being disingenuous because the truth makes him look bad.
More like – it won't be useful to small-time developers (since they won't have the capability to host and run it themselves) and so all the benefits will be reaped by AWS and other large players.
This is what I understood as well. They want to either democratize adoption or not release it. The last thing they/anyone wants is for another BigCo or Govt to h take undue advantage of the model (through fine-tuning?) when others cannot.
That said, I can imagine a GPTQ/4-bit quantized model to be smaller and easier to run on somewhat commodity clusters?
Or it could run with GGML/llama.cpp on a cloud instance with a TB of RAM?
After seeing what people were able to do with LLaMA, I am positive that the community will find a way to run it - albeit with some loss in performance.
It would be truly amazing if they used their computing to develop quantized models as well.
A big chunk of developments based on Facebook’s LLaMA model are by small-time developers and individuals, not large players. Facebook has already shown a viable way to release models in the way you described.
I think I worded this poorly. What he said was that a lot of people say they want open-source models but they underestimate how hard it is to serve them well. So he wondered how much real benefit would come from open-sourcing them.
I think this is reasonable. Giving researchers access is great but for most small companies they're likely better off having a service provider manage inference for them rather than navigate the infra challenge.
The beauty of open source is that the community will either figure out how to make it easier, or collectively decide it’s not worth the effort. We saw this with stable diffusion, and we are seeing it with all the existing OSS LLMs.
“It’s too hard, trust us” doesn’t really make sense in that context. If it is indeed too hard for small orgs to self host then they won’t. Hiding behind the guise of protecting these people by not open sourcing it seems a bit disingenuous.
If he says he's inclined to open-source GPT-3, I don't see any good arguments not in favor of giving startups the choice of how they can run inference.
Sam Altman is responsible for leading the team that have revolutionised AI in its position within society.
There is plenty to criticise OpenAI for but what he and they have achieved is extraordinary, and there is no need for that sort of toxic personal attack.
I can’t comment on his personality because I don’t know him. But it’s delusional to think the leadership’s personal qualities are irrelevant to developing AI. There’s going to be a lot of subjectivity in fine-tuning the answers.
Dude, he absolutely loves to talk about incentives.
Now that’s neither here nor there response to yours… but since he is so in tune with the concept of incentives, and likes to discuss them, it does make me curious what his incentives are that he is acting on.
I mean, Occam's Razor says if he wanted a way to get paid, he would just have equity or a very large salary, the normal things CEOs do when they want to make lots of money from their very profitable company.
He gained a huge amount of reputation by saying "I don't have equity in the company. I'm doing it because I love it" in front of the Congress. While he was demanding licensing for AI because it's too dangerous.
I love it. It's dangerous. Hypocrisy? No. It's the tendency of man to pursue what is harmful to himself. Even though optimistic technologists of HN won't agree, much of technology has not benefitted humanity.
Or maybe he just doesn't need more money and wants to do something cool? If my basic needs were met, I'd go volunteer at any number of non-profits full time.
I know it doesn't seem related on the surface, but I've found all startup CEOs who ban remote work to be snake-like and dishonest. Interesting coincidence at least.
Might it be that because they know that their own morals and ethics are so fluid that they therefore think than everyone else is the same and thus has to be surveilled to be sure to have "performance...? Just a thought....
If the only barrier to a small team/org using it is cost/effort of hosting (as opposed to some licensing shenanigans), I fail to see how not releasing it is better for the world than releasing it would be. Even if it benefits big tech more than a small team.
Am I somehow being protected by a benevolent sama not open-sourcing the model?
Most of the companies I work with are actively putting in place policies to prevent employees from using OpenAI's service because nobody wants to send their proprietary IP to them.
Almost all of these companies have the technical ability, desire, and means to self-host for their employee community. Imagine the internal coup for CTO/CIOs everywhere to buy whatever is the latest Nvidia GPU cluster box, stick it in the on-prem datacenter, load a licensed GPT model and provide "AI as a service to our employees".
Except what's happening is everybody is looking at buying the box from Nvidia, and sticking a large actually open model on it and simply ignoring OpenAI.
> Almost all of these companies have the technical ability, desire, and means to self-host for their employee community.
Well, one of companies I worked for could have hosted a canary service for cron jobs. But we bought it instead of building because we were focused on building features. And here you’re talking about hosting an entire LLM.
The cron job canary was probably not a service that employees were uploading tonnes of company confidential material into, was it? So I fail to see how the comparison makes sense.
The reason companies shun OpenAI and want a self hosted alternative isn't related to costs, it's becasue they don't want their code, internal emails, documentation etc to be uploaded to Microsoft and thus also directly to the NSA.
It is a shame that Sama does not believe in Open Source. The community can solve their GPU bottleneck issue by making it run on CPUs and edge devices in a matter of days.
> Cheaper and faster GPT-4 — This is their top priority. In general, OpenAI’s aim is to drive “the cost of intelligence” down as far as possible and so they will work hard to continue to reduce the cost of the APIs over time.
this certainly aligns with the massive (albeit subjective and anecdotal) degradation in quality i've experienced with ChatGPT GPT-4 over the past few weeks.
hopefully a superior (higher quality) alternative surfaces before its unusable. i'm not considering continuing my subscription at this rate.
I wonder of it actually is because they’re tuning it to make it less offensive (by their standards). Thats the only explanation I keep seeing repeated.
I would be very surprised. Things that are very, very far from that are also much worse. I'm having difficulty finding the difference between GPT-3.5 and GPT-4 for a lot of my programming tasks lately. It's noticeably degraded.
That's a convenient explanation that's been repeated over and over by certain people, but cost is a much more likely explanation: inference for large models is extraordinarily expensive when you have millions of users and their pricing model always seemed way too low to pay for that.
They have likely been subsidizing their users since the launch of their commercial offering (and this is pretty common strategy for SV startups) but they've been so successful that they now need to scale the cost down in order not to burn all their cash too fast.
First some context: llm "prompts" are actually the whole conversation + initial context. They learn nothing, hence the whole conversation gets fed into them every time, but the instruction following ones are trained to answer your most recent chat response.
In a nutshell, part of your llm prompt (usually your most recent question?) gets fed as a query for the embedding/vector database. It retrieves the most "similar" entries to your question (which is what an embedding database does), and that information is pasted into the context of the llm. Its kinda like pasting the first entry from a local Google search into the beginning of your question as "background."
Some implementations insert your old conversations (that are too big to fit into the llm's context window) into the database as they are pushed out.
This is what I have seen, anyway. Maybe some other implementations do things better.
> part of your llm prompt (usually your most recent question?) gets fed as a query for the embedding/vector database
How is it embedded? Using a separere embedding model, like Bert or something? Or do you use the LLM itself somehow? Also, how do you create content for the vector database keys themselves? Also just some arbitrary off the shelf embedding? Or do you train it as part of training the LLM?
Yeah its completely seperate. The LLM just gets some extra text in the prompt, that is all. The text you want to insert is "encoded" into the database which is not particularly compute expensive. You can read about one such implementation here: https://github.com/chroma-core/chroma
Have you tried something like Memory Transformers https://arxiv.org/abs/2006.11527 where you move the k/v pairs that don't fit in the context window to a vector db? Seems like a more general approach, but I have tested then against each other.
One thing I don't understand is how feeding the entire conversation back as a prefix for every prompt doesn't waste the entire 4K-token context almost immediately. I'd swear that a given ChatGPT window is stateful, somehow, just for that reason alone... but everything I've read suggests that it's not.
Any database can be used to extend the memory of LLMs. What a database does is store stuff and lets you search/retrieve stuff. Embeddings are differet form of data that are in many (but not all) cases superior to searching through text.
You do not need a fancy cloud hosted service to use an embeddings database like you do not need one to use a regular databse (although you could).
Check https://github.com/kagisearch/vectordb for a simple implementation of a vector search database that uses local, on-premise open source tools and lets you use an embeddings database in 3 lines of code.
The problem with Claude is that it is quite literally impossible to get off the waiting list to use it. To OpenAI’s credit they actually ship the product in an accessible way to developers.
I just checked out poe.com. Seems you can only buy a subscription if you own Apple hardware (first time I've ever heard that).
It's $20 a month and comes with 300 GPT-4 messages and 1000 Claude 1.2 messages.
By comparison, ChatGPT Plus gives gives you up to 6000 GPT-4 messages a month for the same price (admittedly it would be hard to use that many as they are given in 3 hour blocks).
Has anyone compared the ChatGPT GPT-4 performance in the plus subscription to that of the API? Has the API performance deteriorated just as much? It would be strange if it did as I'd assume the models costs are priced in there.
If they open up fine-tuning API for their latest models, I wonder how the enthusiasm around the open source model is impacted. One of the advantages of the open source models is the ability to be fine-tuned. Are other benefits enough to keep the momentum going?
You better have deep pockets, have you check the prices and then the rates for using the tuned models? They sure 10x to 100x more expensive then nontuned models
Really great news to give at cheaper and faster GPT4. As a GPT+ subscriber, the most annoying thing is the 25 message limit every 3 hours, I really want that removed.
A bit sad to hear that the multimodal model will only come next year, was hoping to get it this year
100k to 1 Million context length, sounds phenomenal especially if it comes to GPT4. I've used Claudes 100k context length and I found it so useful that when I have large documents I just default to Claude now
I'm a graduate student doing AI research in a US University. And I applied pretty early (Last year December I think) , those might be two factors that got me access to Claude.
I think getting access to Claude through slack is much easier and I recently got it by just downloading it as a Slack App
I use Poe and got access to Claude 100k as soon as it was released. I think it's a better deal than paying OpenAI for sure, since you have access to GPT-4, Claude+, and others. They also have community bots, etc.
Claude has a free slack client that I briefly was able to access by creating a new slack workspace and adding it there. But as of yesterday it wasn't working for me
If you look at their API limits, no serious company can use this to scale up beyond say 10k users. 3500 Reqs per min for gpt3.5 turbo. They have a long way to go to make it usable for the rest of the 95%
I've had to move to using Azure OpenAI service during business hours for the API-- much more stable unless the prompts stray into something a little odd and their API censorship blocks the calls.
Nice to see they are working on reducing the pricing. GPT-4 is just too expensive right now imo. A long conversation would quickly end up costing tens of dollars if not more, so less expensive model costs + stateful API is urgently needed. I think even OpenAI will actually gain a lot by reducing the pricing, right now I wouldn't be surprised if many uses of GPT-4 weren't viable just because of the costs.
Dozens of people using it daily for coding and conversations and review in a month might be a couple hundred bucks. All day convo, constantly, as fast as it can respond, might add up to $5.
Not sure what kind of convo you're having that you could hit $10 unless you're parallelizing with something like the "guidance" tool or langchain.
The version of GPT 4 with 32K token context length is the enabler for a huge range of "killer apps", but is even more expensive than the 8K version.
And yes, parallelism and loops are also key enablers for advanced use-cases.
For example, I have a lot of legacy code that needs uplifting. I'd love to be able to run different prompts over reams of code in parallel, iterating the prompts, etc...
The point of these things is that they're like humans you can clone at will.
The ability to point thousands of these things at a code base could be mindblowing.
> OpenAI will avoid competing with their customers — other than with ChatGPT. Quite a few developers said they were nervous about building with the OpenAI APIs when OpenAI might end up releasing products that are competitive to them. Sam said that OpenAI would not release more products beyond ChatGPT. He said there was a history of great platform companies having a killer app and that ChatGPT would allow them to make the APIs better by being customers of their own product. The vision for ChatGPT is to be a super smart assistant for work but there will be a lot of other GPT use-cases that OpenAI won’t touch.
Can anyone elaborate on this? This is a big issue for me.
Technically he can claim that OpenAI will not release competing products while Microsoft plugs AI into everything.
Microsoft just announced at Build 2023 that they'll have OpenAI tech integrated with: Windows, Bing, Outlook, Word, Teams, Visual Studio, Visual Studio Code, Microsoft Fabric, Dynamics, GitHub, Azure DevOps, and Logic Apps. I probably missed a bunch.
Very soon now, everything Microsoft sells will have OpenAI integration.
Unless you're selling a niche product too small for Microsoft to bother with, you're competing directly against OpenAI.
Oh, and to top it off: Microsoft can use GPT 4 all they want, via API access. Third parties have to beg and plead to get rate-limited access. That access can be withdrawn at any time if you're doing something unsafe to OpenAI's profit margins.
It's the only such reference I could think of that others might recognise.
I'm reading the web serial Pact right now, where the main character can lie but it costs him dearly each time: https://pactwebserial.wordpress.com/
I also really enjoyed the ending of the Confederation Trilogy by Peter F Hamilton, which revolved around the inability of the Tyrathca race to tell lies.
PS: Just for laughs, I used to practice talking like an Aes Sedai, never telling an outright lie while actively deceiving people. It's an interesting skill to acquire and surprisingly easy. Once you learn how to do it, you'll never see a press conference or a political speech the same way ever again.
I think we have very similar tastes. I haven't read Pact yet, but I'm a huge fan of Worm (I consider it one of my top 5 favorite works of fiction).
I also haven't read Confederation, but I have read a different work for Hamilton (The Commonwealth Saga). I actually thought it was a bit too long, but it's one of the series I think about most often (so many great ideas and characters in it).
This would be huge for many applications, as "chatting" with GPT-4 gets really, really expensive very quickly. I've played with API with friends, and winced as I watched my usage hit several dollars for just a bit of fun.
The roadmap here is completely focused on ChatGPT and GPT-4. I wonder what portion of their resources is still going to other areas (DALL-E, audio/ video processing, etc.)
266 comments
[ 2.8 ms ] story [ 245 ms ] threadHow many shell corporations are intelligence agencies seeding right now?
(Just to be clear, this is a hypothetical intelligence agent saying this, not me.)
I mean, it's not exactly rocket science, who wouldn't instantly fold to that?
(I mean, since we’re just making up wild hypotheticals)
The conspiracy theory isn't that every employee of OpenAI spent 8 hours every day for six months in meetings with govt agencies.
[1] - https://arxiv.org/pdf/2302.10866.pdf
You'll know when the problem is solved when model after consistently use a method. Until then (and especially if you're not in the field as a researcher), assume that every paper claiming to tackle context length is simply a nice proposal.
Just to name a few families of approaches: Sparse Attention, Hierachical Attention, Global-Local Attention,Sliding Window Attention, Locality sensitive hashing Attention, State space model, EMA gated attention.
Notably, human working memory isn't great either. Which begs the question (if the comparison is valid) as to whether that limitation might be fundamental.
Which is all the more curious, considering OpenAI said this only in January:
> Azure will remain the exclusive cloud provider for all OpenAI workloads across our research, API and products [1]
So... OpenAI is severely GPU constrained, it is hampering their ability to execute, onboard customers to existing products and launch products. Yet they signed an agreement not to just go rent a bunch of GPU's from AWS???
Did someone screw up by not putting a clause in that contract saying "exclusive cloud provider, unless you cannot fulfil our requests"?
[1]: https://openai.com/blog/openai-and-microsoft-extend-partners...
that absolutely isn’t an attempt to slow down all competition.
which isn’t necessary because nobody made such a mistake.
this won’t lead to any hasty or reckless internal decisions in a feckless effort to stay in front.
not that any have already been made.
not that that could lead to disaster.
> Did someone screw up by not putting a clause in that contract saying "exclusive cloud provider, unless you cannot fulfil our requests"?
Maybe MSFT refused to sign such an agreement?
https://www.youtube.com/watch?v=Rk3nTUfRZmo&t=5s "What runs ChatGPT? Inside Microsoft's AI supercomputer"
The relevance here is that Azure appears to be very well designed to handle the hardware failures that will inevitably happen during a training run taking weeks or months and using many thousands of GPUs... There's a lot more involved than just renting a bunch of Amazon GPUs, and anyways the partnership between OpenAI and Microsoft appears quite strategic, and can handle some build-out delays, especially if they are not Microsoft's fault.
I think that there aren't a lot of GPUs available and it takes time to add more to the datacenter even when you do get them.
Don't assume Microsoft is bad at everything and that AWS is automatically superior at all product categories...
All the cloud providers are building out this type of capacity right now. It's already having a big impact in terms of quarterly spend, which we just saw in the NVDA Q1 results. AWS, Azure, and GCP for sure, but also smaller players like Dell and HPE and even NVidia themselves are trying to get into this market. (Disclaimer: I work at one of these places but don't feel like saying which). I suspect the GPU constraints won't be around too long, at which point we'll find out if OpenAI made a contractual mistake.
It's pretty hilarious and annoying to see bing start to write code only to self censor itself after a few lines (deleting what was there! no wonder these guys love websockets and dynamic histories)
Whoops!
Please make more effort next time than to provide me with a Wiki article.
I tried it again tonight and it seems like they fixed it to only produce small amounts of mediocre code instead.
There’s been quite a bit happening in the programming space since sept 2021.
I use GPT to keep things high level and then do my normal research methodology for implementation details.
My understanding is right now they essentially need to train a new model on a new updated corpus to fix this, but maybe some other techniques could be devised...or they'll train something more up to date.
For smaller projects that will fit, I've taken to: `xclip *` and then pasting the entire collection of files into ChatGPT before describing what I want to do.
Will they use copilot(s) to improve the models? Yes, but they have been doing that since 2021 already (the release year of GitHub Copilot).
>tell AI to make itself more efficient by finding performance improvements in human written code
>that newly available processing power can now be used to find more ways to improve itself
>flywheel effect of AI improving itself as it gets smarter and smarter
eventually you'd turn it loose on improving the actual hardware it runs on. I think the question now is really how far transformers can be taken and if they are really the path to "real" AI.
Also don't confuse all other types of human/animal characteristics like sentience with intelligence. They are different things. Things like sentience, subjective stream of experience, or other aspects of being alive don't just accidentally fall out of larger training datasets.
And we should be glad. The models are going to be orders of magnitude faster (and perhaps X times higher IQ) than humans within a few years. It is incredibly foolish to try to make something like that into a living creature (or emulation of living).
I disagree, language is all we need. Agency? Encode your “internal needs” as prompts, periodically generate prefixes from these, append them to incoming prompts. Self-awareness? Summarize this internal dialogue, reflect on it with a few iterations, add the results to the common prefix. Sentience? Attach some sensors, summarize their observations with the language model, prepend to prefix. Actions? Make the model output commands that some servos or other interfaces understand. Etc, etc.
And, of course, it would be extremely _cool_ to make something like that into a living creature, and lots of labs are already doing that. Fear and luddism should not stay in the way of curiosity.
If we humans cannot improve our own intelligence, making something smarter than us is an evolutionary imperative.
It is almost certain that the next stage of intelligence will be digital. But it is very foolish and unnecessary to try to speed that along.
It is likely that we have a century or two max left in control of the planet, regardless of what we do. On some level I agree that totally suppressing it indefinitely would be a shame.
When I said "living" I meant digital life. Such as those things you describe and others including control-seeking, self-preservation, and reproduction which are all central to living beings.
The problem is that AI will soon think 100 or more times faster than humans. This is anticipated based on the history of increases in computing efficiency and the fact that we are now optimizing a very specific system (LLMs). Humans will not in any way be able to keep up.
This is not luddism. I have a service that connects GPT-4 to Linux VMs on the internet to install or write software. I think this technology is great and has a lot of positive potential.
But when you deliberately try to emulate animals (like humans) and combine that with hyperspeed and other superintelligent characteristics, you are essentially approaching suicide or at least, abdicating all responsibility for your environment. There is no way to prevent such a thing from making all of your decisions for you.
The speed difference will be incredible. Imagine a bullet time scene where everyone seems to be moving in extreme slow motion. Now multiply by 10 so they are so slow they seem completely frozen.
This level of performance is coming in five years or less.
While I don't want to suppress the evolution of intelligent life in our corner of the universe, I also am not ready to join a death cult. Especially not accidentally.
For what it's worth I've found the model actually performs significantly worse at most tasks when given access to browsing, in part because it relies on that instead of its own in built knowledge.
I haven't found a good way to have it only access the web for specific parts of its response.
[0] https://www.gocomics.com/tomthedancingbug/2022/06/17
I know it's a joke, but the hole is the god-AI couldn't have been that smart, since cryptocurrency-mining quickly switched to ASICs, which muting the demand increase for GPUs.
Did the GPU manufactures ever embrace cryptocurrency? IIRC, they actually tried to discourage it (e.g. by butting throttling into mass market models to discourage their use for computation).
Also, the graphs here show a long-term downward trend, with only a short-term sales blip 5 years ago due to cryptocurrency: https://www.tomshardware.com/news/sales-of-desktop-graphics-....
The talk/conversation appeared to me not as OpenAI future plan but more on the CEO lamenting on how severely limited the company by the GPU or the lack thereof. It just a cheeky ploy by a CEO of an AI company that currently at 30B USD valuation to get more money in order to buy several [fill in the blank] of these most advanced GPU systems [2].
[1]Nakamoto's Neighbor: My Hunt For Bitcoin's Creator Led To A Paralyzed Crypto Genius:
https://www.forbes.com/sites/andygreenberg/2014/03/25/satosh...
[2]Nvidia DGX GH200: 100 Terabyte GPU Memory System:
https://news.ycombinator.com/item?id=36133226
It sounds a little too much sci-fi for me, but I guess he knows better.
The frontier is the multimodal versions of GPT-4 which he just said wasn't even going to public release until next year. Or whatever they are on now which they are carefully not calling GPT-5.
Am I reading this right? "We're not open sourcing GPT-3 because we don't think it would be useful to anyone else"
It’s possible to think squirt guns shouldn’t be regulated but AR-15s should, or AR-15s shouldn’t but cruise missiles should. Or driving at 25mph should be allowed but driving 125mph shouldn’t.
The great thing about open source is that people can try different approaches and gravitate towards what works best for them. Sam knows that of course, he's just being disingenuous because the truth makes him look bad.
We don't provide nuclear weapons for everyone to keep in their basement, why would someone who believes AI is an existential risk provide their code?
That said, I can imagine a GPTQ/4-bit quantized model to be smaller and easier to run on somewhat commodity clusters?
Or it could run with GGML/llama.cpp on a cloud instance with a TB of RAM?
After seeing what people were able to do with LLaMA, I am positive that the community will find a way to run it - albeit with some loss in performance.
It would be truly amazing if they used their computing to develop quantized models as well.
I think this is reasonable. Giving researchers access is great but for most small companies they're likely better off having a service provider manage inference for them rather than navigate the infra challenge.
“It’s too hard, trust us” doesn’t really make sense in that context. If it is indeed too hard for small orgs to self host then they won’t. Hiding behind the guise of protecting these people by not open sourcing it seems a bit disingenuous.
"I'm not sharing my chocolate with you because you probably wouldn't like it"
Also OpenAI: Meta is pissing in our moat, let's drop a hint about open sourcing our shit too!
There is plenty to criticise OpenAI for but what he and they have achieved is extraordinary, and there is no need for that sort of toxic personal attack.
https://www.technologyreview.com/2022/04/06/1048981/worldcoi...
and loopt
https://en.m.wikipedia.org/wiki/Loopt
We just haven't figured out how yet.
Also he's already half way to being a billionaire anyway without OpenAI.
Now that’s neither here nor there response to yours… but since he is so in tune with the concept of incentives, and likes to discuss them, it does make me curious what his incentives are that he is acting on.
I love it. It's dangerous. Hypocrisy? No. It's the tendency of man to pursue what is harmful to himself. Even though optimistic technologists of HN won't agree, much of technology has not benefitted humanity.
>> 4. OpenAI will avoid competing with their customers — other than with ChatGPT
On this I would not bet a dime.
That’s not ideal
How does open source licensing work with respect to trained ai models anyway? Is something like the mit license even that valuable here? Or is it?
Am I somehow being protected by a benevolent sama not open-sourcing the model?
Almost all of these companies have the technical ability, desire, and means to self-host for their employee community. Imagine the internal coup for CTO/CIOs everywhere to buy whatever is the latest Nvidia GPU cluster box, stick it in the on-prem datacenter, load a licensed GPT model and provide "AI as a service to our employees".
Except what's happening is everybody is looking at buying the box from Nvidia, and sticking a large actually open model on it and simply ignoring OpenAI.
Well, one of companies I worked for could have hosted a canary service for cron jobs. But we bought it instead of building because we were focused on building features. And here you’re talking about hosting an entire LLM.
The reason companies shun OpenAI and want a self hosted alternative isn't related to costs, it's becasue they don't want their code, internal emails, documentation etc to be uploaded to Microsoft and thus also directly to the NSA.
But they do to Slack and MS Teams. Also go mail services and other places.
this certainly aligns with the massive (albeit subjective and anecdotal) degradation in quality i've experienced with ChatGPT GPT-4 over the past few weeks.
hopefully a superior (higher quality) alternative surfaces before its unusable. i'm not considering continuing my subscription at this rate.
They have likely been subsidizing their users since the launch of their commercial offering (and this is pretty common strategy for SV startups) but they've been so successful that they now need to scale the cost down in order not to burn all their cash too fast.
Instruction tuned LLaMA 65B/Falcon 40B are good, especially with an embeddings database.
...But OpenAI has all the name recognition and ease of use now, so it might not even matter if others ambiguously surpass OpenAI models.
In a nutshell, part of your llm prompt (usually your most recent question?) gets fed as a query for the embedding/vector database. It retrieves the most "similar" entries to your question (which is what an embedding database does), and that information is pasted into the context of the llm. Its kinda like pasting the first entry from a local Google search into the beginning of your question as "background."
Some implementations insert your old conversations (that are too big to fit into the llm's context window) into the database as they are pushed out.
This is what I have seen, anyway. Maybe some other implementations do things better.
How is it embedded? Using a separere embedding model, like Bert or something? Or do you use the LLM itself somehow? Also, how do you create content for the vector database keys themselves? Also just some arbitrary off the shelf embedding? Or do you train it as part of training the LLM?
You do not need a fancy cloud hosted service to use an embeddings database like you do not need one to use a regular databse (although you could).
Check https://github.com/kagisearch/vectordb for a simple implementation of a vector search database that uses local, on-premise open source tools and lets you use an embeddings database in 3 lines of code.
It's $20 a month and comes with 300 GPT-4 messages and 1000 Claude 1.2 messages.
By comparison, ChatGPT Plus gives gives you up to 6000 GPT-4 messages a month for the same price (admittedly it would be hard to use that many as they are given in 3 hour blocks).
It's like saying "air should cost money".
I have had random runs of good days and bad days since starting to use chatGPT.
A bit sad to hear that the multimodal model will only come next year, was hoping to get it this year
100k to 1 Million context length, sounds phenomenal especially if it comes to GPT4. I've used Claudes 100k context length and I found it so useful that when I have large documents I just default to Claude now
I think getting access to Claude through slack is much easier and I recently got it by just downloading it as a Slack App
Dozens of people using it daily for coding and conversations and review in a month might be a couple hundred bucks. All day convo, constantly, as fast as it can respond, might add up to $5.
Not sure what kind of convo you're having that you could hit $10 unless you're parallelizing with something like the "guidance" tool or langchain.
And yes, parallelism and loops are also key enablers for advanced use-cases.
For example, I have a lot of legacy code that needs uplifting. I'd love to be able to run different prompts over reams of code in parallel, iterating the prompts, etc...
The point of these things is that they're like humans you can clone at will.
The ability to point thousands of these things at a code base could be mindblowing.
Probability mass functions? Anyone know what this means in this context?
Can anyone elaborate on this? This is a big issue for me.
Technically he can claim that OpenAI will not release competing products while Microsoft plugs AI into everything.
Microsoft just announced at Build 2023 that they'll have OpenAI tech integrated with: Windows, Bing, Outlook, Word, Teams, Visual Studio, Visual Studio Code, Microsoft Fabric, Dynamics, GitHub, Azure DevOps, and Logic Apps. I probably missed a bunch.
Very soon now, everything Microsoft sells will have OpenAI integration.
Unless you're selling a niche product too small for Microsoft to bother with, you're competing directly against OpenAI.
Oh, and to top it off: Microsoft can use GPT 4 all they want, via API access. Third parties have to beg and plead to get rate-limited access. That access can be withdrawn at any time if you're doing something unsafe to OpenAI's profit margins.
"Please Sir Sam, may I have some GPT please?"
"No."
Haha having just finished the Wheel of Time, I'm super tickled by this reference.
It doesn't seem to be too common, only two uses of it on HN in the past year (at least, found by searching for the phrase "Aes Sedai")
I'm reading the web serial Pact right now, where the main character can lie but it costs him dearly each time: https://pactwebserial.wordpress.com/
I also really enjoyed the ending of the Confederation Trilogy by Peter F Hamilton, which revolved around the inability of the Tyrathca race to tell lies.
PS: Just for laughs, I used to practice talking like an Aes Sedai, never telling an outright lie while actively deceiving people. It's an interesting skill to acquire and surprisingly easy. Once you learn how to do it, you'll never see a press conference or a political speech the same way ever again.
I also haven't read Confederation, but I have read a different work for Hamilton (The Commonwealth Saga). I actually thought it was a bit too long, but it's one of the series I think about most often (so many great ideas and characters in it).
This would be huge for many applications, as "chatting" with GPT-4 gets really, really expensive very quickly. I've played with API with friends, and winced as I watched my usage hit several dollars for just a bit of fun.