OpenAI API pricing update FAQ
help.openai.com
help.openai.com
I was recently able to articulate to myself why I dislike language models like GPT-3 so much.
Language is fundamentally about communication. I have a mental state, and language lets me share that some of that mental state with another human being. It's quite remarkable, and the cornerstone for pretty much every aspect of our lives as social creatures.
GPT-3 can generate text that's convincingly human-like; it's an amazing technical achievement. But (so far, anyway!) it is just that; a generator of plausible text. To borrow a philosophical term, there's no "interiority" there.
You can see this clearly in GPT-3's output; it is at its weakest when writing about the relationships between a set of concepts, or when trying to carry a line of thought or argument through a longer text. This is because there is no underlying _thing_ that is being communicated; no mental model or grounding concept. Just text, generated solely based on plausibility in a given context.
So while GPT-3 is truly remarkable in its ability to generate text, it cannot, by definition, communicate (except on the meta level via prompt selection and output curation.)
Does the world really need more words but with less communication?
---
There are a few problems with LukeV's argument. First, he conflates the ability to generate human-like text with the ability to communicate. Just because GPT-3 can generate text that sounds like it was written by a human does not mean that it cannot communicate. In fact, GPT-3 is quite good at carrying a line of thought or argument through a longer text, as evidenced by the fact that it is often used by writers to generate ideas and outlines for their articles.
Second, LukeV claims that GPT-3 is at its weakest when writing about the relationships between a set of concepts. However, this is not necessarily true. GPT-3 may not be able to generate text that is perfectly grammatical or that makes perfect sense, but it is often able to generate text that is nonetheless meaningful and insightful. For example, GPT-3 has been used to generate essays on complex topics such as the philosophy of mind, and these essays have been well-received by readers.
Third, LukeV argues that the world does not need more words, but this is clearly not the case. The world needs more communication, and GPT-3 can help to facilitate this by generating text that is meaningful and insightful. In fact, GPT-3 is already being used by writers to generate ideas and outlines for their articles, and it is likely that this use will only increase in the future.
In conclusion, LukeV's argument against GPT-3 is flawed. GPT-3 is a valuable tool that can help to facilitate communication, and the world needs more communication, not less.
----
I'm not sure who I agree with, but that's GPT-1's response to you regarding your critique of itself. Have at it!
> First, he conflates the ability to generate human-like text with the ability to communicate.
This is incorrect. In fact, lukev does the opposite. That is one of the fundamental parts of his argument -- that those two things are different.
> Second, LukeV claims that GPT-3 is at its weakest when writing about the relationships between a set of concepts. However, this is not necessarily true. GPT-3 may not be able to generate text that is perfectly grammatical or that makes perfect sense, but it is often able to generate text that is nonetheless meaningful and insightful. For example, GPT-3 has been used to generate essays on complex topics such as the philosophy of mind, and these essays have been well-received by readers.
I believe this is the best paragraph of the argument, but it is also the most difficult to rebut because it is rather subjective. Whether a text is meaningful or insightful is up the individual reading it. I would be surprised, however, if experts on the philosophy of mind found the essays mentioned by GPT-3 to be meaningful or insightful -- as opposed to regurgitations of concepts that are in the texts that have been fed to GPT-3's model.
> Third, LukeV argues that the world does not need more words, but this is clearly not the case. The world needs more communication...
The presence of more words in the world does not necessarily entail more communication. The two concepts are fairly closely related but are definitely not synonymous.
This doesn't necessarily mean that I agree with lukev, of course. I do think that GPT-3 as it currently exists should not be used to generate text that is meant for human consumption without being edited first. I can certainly see how it would be useful as a tool for generating an outline of an essay, especially with well-crafted prompts.
I am of course open to more arguments and data on the subject =)
That's the other problem with GPT-3... it will happily say things that are absolutely false :)
I think this response actually illustrates my point quite nicely; it is a plausible sounding rebuttal but does not actually engage on a logical level.
That said, I don't mean to diminish the impressiveness here. It's quite evident that it could pass a Turing test in a wide variety of circumstances. After all, it's not like humans are consistently great at concept-informed writing either. I'm pretty sure the majority of undergrad college papers have more in common with GPT than with real critical thought... generating something that looks plausibly like an argument rather than actually having one.
Absolutely. From a surface-level read it may sound sensible but it all falls apart very easily.
The lines of reasoning for each attempted rebuttal are practically non sequiturs.
Out of curiosity, I’d be interested in seeing some attempts using prompts intended to induce CoT reasoning (“let’s think step-by-step“ etc.)
GPT-3 does nothing without a human inputting what it wants it to output. GPT-3 is as much a tool for communication as are sounds, words, symbols, etc.
Whether it was carved in stone by a human hand or generated on the cloud by an AI, the intent to communicate is always coming from a human.
If you think the text generated by GPT-3 does not carry the right meaning or has "no underlying _thing_ that is being communicated", it just means that the human behind is not using the tool properly. Just like sometimes you use a word that doesn't mean what you think it means.
I don't particularly like it as an act of communication because the information density is going to be quite low. If you want to tell me something, I'd rather you just told me the prompt! It will have all the relevant "real" communication; anything added by GPT-3 is basically just filler.
E.g., distilling large amounts of text into short summaries while retaining substance is effectively the opposite of this, and is a use case GPT3 has been applied to successfully/commercially.
The exemption for finetuned models is interesting: since they were 2x the price of the old prices, they're now 4x-6x more expensive than the base models. I bet that's how OpenAI really makes their money with enterprise.
Otherwise, I’ve been using a model finetuned on the complete works of Plato and I have really been enjoying the new dialogues.
I suspect some of the more custom "be a better writer" services use a finetuned davinci.
If you want to use GPT-3 to build an AI-lawyer, it'll be very expensive... but to build a tool for customer support for walmart will be cheap, etc.
As soon as your product is a success, you run out of free tier, and you will have to negotiate a price with the sales team.
If I were OpenAI, then while I had a market leading product, the price would be 'I want half the profit margin of whatever you're using it for'.
This degenerates to Oracle's "how much is your data worth to you" pricing model.
Nothing to do with Stable Diffusion then, no?
In my small amount of testing of Bloom so far it seems capable of advanced behaviour but it can indeed be trickier to coax that out. Playing with temperature and sampling matters for sure.
But for the moment you can't just use pure natural language instructions with BLOOM indeed. Maybe it will come!
If the open source community then other company's didn't follow along with them then we would still be being price gauged in a monopoly situation, hopefully same thing happens to dalle soon
Edit: looks like the comment thread below is similar https://news.ycombinator.com/item?id=32554955
Here's one that I find personally really interesting:
Imagine you have a disability, and the government agencies responsible for helping you have denied your claim. You need to write them a letter. You don't have much (if any) experience writing letters to government agencies, so you're finding that task challenging.
GPT-3 can write that letter for you, and it will de a really good job of it.
Of course, you need access to GPT-3 and you need to learn how to drive it with prompts, so this isn't necessarily an easy thing. But I think it helps lean in the direction of the kind of benefits people can get from this class of technology outside of just playing with the AI.
GPT-3 can do a good job of this all today, without any additional training.
An expert could help out by providing a few example prompts, but the result would be a lot more powerful and require a lot less development work to put together than trying to build some kind of placeholder wizard from scratch.
And the madlibs is free to use, for everyone.
All things considered, GPT-3 is a more of a shibboleth of AI revolution than an actual one. Much better to give models like GPT-NeoX to talented development teams who can really optimize and fine tune it, attach great sampling strategies to it, and create great products from there. The API is just too limited for me to want to build a company atop it.
If no-one has built the madlib template for "a letter to the council asking about my disability benefits" then madlibs being more predictable isn't going to help you.
It will aid speechwriters or help more people write their own speeches, I guess.
Assuming that AI tech is a person replacement is, IMO, the wrong mental model. It helps talented people be more efficient. And will help a lot of hacks “fake it till they make it.” Power to the people, if they can handle it.
Was this comment written by GPT-3?
(That said - are all comments written by GPT-3?)
You should submit that to GPT-3 as a prompt.
There is a bit of prompt engineering. If you start your question with things like "As a qualified therapist, what would you recommend to" .... There is also a bit of engineering with the question. I often ask about an analogous situation, rather than my own.
I'm not batting 100%, but I often get strategies and things to think about I wouldn't have come up with myself.
Yes. It’s a dialectic. Very powerful for personal reflection.
Could you elaborate a little more on what you're asking? Perhaps an example?
I feel as if any use case like this, — particularly if it crosses into interpersonal relationships and disagreement — risks being incredibly dangerous as far as reinforcing your own biases.
It should go without saying that to even consider this, you should have a solid grasp of the fundamentals of these LLMs, their limitations and inherent biases, and their tendency to unconditionally agree with you regardless of your position – particularly in a dialog-style format.
One clearly problematic use case would be asking about an interpersonal conflict where there exists an established societal and cultural power imbalance.
E.g., describing a disagreement between a man and a woman and asking for insight into that conversation would be incredibly flawed, no matter how objectively it were depicted.
This isn't an attack on you nor am I claiming you're doing anything at this level, but taken to the extreme it’s something I can easily imagine someone doing after reading your post.
I’d really just like people to be cautious. After all, this is the same language model that told me a social media post contains hate speech due to “negative sentiment towards a protected class (racists)”.
With GPT-3, that bias is pretty explicit, and not hard to manage. I also don't treat advice as necessarily /good/ advice. A good way to think about it is ideation.
A good example is to ask how to communicate something. I sometimes miscommunicate what I'm thinking, and it often takes a lot of effort to figure out how to say something. With different prompts, GPT-3 will give different ways to say something, and I can pick one which works well and meets what I want to say. Let's say you promised someone something, and need to change plans, and you don't want them to think you're blowing them off. GPT-3 will often give me good ways to communicate something like that.
Or another example is: "I am struggling with ____. What are good strategies I can use to ____."
GPT-3 will often come up with things which I wouldn't have come up with myself, or ones which would have taken me a lot of time. Just as often, it will give a useless, generic list of suggestions. At the end of the day, though, it provides helpful ideas often enough to be valuable.
GPT-3 definitely doesn't (and shouldn't) act as an arbiter in a conflict, but even in conflicts, there's often a solution which works for both sides. My major problem is that I tend to think slowly, and come up with those solutions too late. GPT-3 thinks a lot faster than I do, and having those ideas before they're moot is sometimes helpful.
As a footnote, human therapists tend to reinforce biases too. They only have one side of the story. I've seen people really damaged in the way you describe. In one case, both the person and the therapist were living in a (plausible-sounding) fantasy world.
I asked GPT-3 for advice for how to interpret that person's behavior (with more context).
GPT-3 gave several plausible explanations, must of which I hadn't thought of.
This answer was written by GPT-J when I gave it the following prompt:
question: I lack imagination. What is GPT-3 (and generative text models in general) useful for, other than toys? It can write essays, can we expect it to replace writers for newspapers and magazines? Can it write coherent technical documentation? Replace speechwriters?
answer:
That said, there's no reason it couldn't go into an infinite loop, just that the models are pretty sophisticated and therefore are less likely to do this than earlier generations of models (like predictive text on smart phones in the past)
For the love of god please NO.
But I think the principal use for me is CoPilot.
Whether a human or a computer writes it seems unimportant.
Anyone being honest should admit it (was) more hot air and hype than reality.
Which makes me think: everyone needs to teach kids to speed read. And develop their bullshit detectors.
My favorite reply was:
"A: I think the best example of this is the ability to swap positions with a friendly unit. This ability can be used in so many different ways. You can use it to support an ally, or you can use it to escape a tricky situation. You can use it to surprise an enemy by swapping with a unit they didn't expect, or you can use it to put yourself in a better tactical position. It's a simple idea, but it actually turns the whole game on its head."
- Generative use cases, where you give the model the kernel of an idea and then you curate its output (e.g., blog writing, code completion, etc.)
- Extractive use cases, where you give the model some big piece of text, and then process it in some way (e.g., extract names and addresses, classify it, ask a question about the text)
- Transformational use cases, where you need to fix/adjust a piece of text, or translate from one domain to another (e.g., sometimes I'll use GPT-3 for little tasks like copying and pasting a table from a presentation and then asking the model to translate it to markdown; saves me a visit to Google and finding some table generator website)
- Comparisons, where you use embeddings to do search/clustering/recommendations over any set of strings (e.g., can combo nicely with the Q&A use case above, where you search over a knowledge base)
I started a repo here with some barebones examples of each: https://github.com/openai/openai-cookbook/
If you're looking for examples of commercial applications, OpenAI published two blog posts highlighting a few:
- GPT-3 use cases (2021): https://openai.com/blog/gpt-3-apps/
- Codex use cases (2022): https://openai.com/blog/codex-apps/
My tools uses GPT-3 to dig through Airbnb reviews to find the "cons" of staying at a particular place and summarize them neatly. You can also ask any question at all in natural language about all the listings you're evaluating.
I'm bookmarking your repo for future reference, this is very useful
And even worse, now corporations and governments don't even need to have troll farms. Just rent some AI, give it basic directions, and let it loose.
The Dead Internet Theory will become real.
GPT-3 like all tech is double-edged I think. Has good & bad things about it
Maybe Google should require webmasters to mark the automated content with a special HTML tag or attribute. Failure to comply leading to deranking. This would be necessary for the future - to know what data was human generated when training the next model, and to have official reason for penalising spammers who try to pass artificial text as human written.
Let's say you have a product for sale at $32. Then you have a sale to boost name recognition, and have a coupon for $32 dollars off (limited time offer). Or equivalently, that would be $32 cheaper. So the net price is free ($0). $32 cheaper = 100% cheaper = 1x cheaper. Likewise, 25% cheaper = 25% less expensive, or doing the math $32 - 0.25*$32 = 0.75*$32 = $24.
"cheaper" isn't a defined operation except for "a is cheaper than b".
So the statements "6% cheaper" and "2 times cheaper" are shorts for "it is cheaper, it costs 6% less" and "it is cheaper, you get 2 times more for the same amount of money".
The parent you're replying to seems to have things confused. 100% off isn't 1x cheaper -- it's infinitely cheaper.
No, it isn't sensible, and is certainly confusing, and that is why you should commit to never using that phrasing from this day forward. It isn't like there is a hardship you have to overcome when using precise language. There are plenty of alternative ways to state the desired objective that are clear and unambiguous:
- The product is now half-priced!
- The item costs half as much!
- The unit comes with a 50% discount!That (usually) makes no sense, so I assume people mean something else when they say it, but I never have any clue what. Is it a third of the original price? Is it two thirds (new_price = price - price/3)?
Percentages also work this way. If I say "10% cheaper", most people understand that I mean the price is 90% of what it was previously. Although percentages get weird when increasing the price.
e.g. if it was a flat doubling, I would say that as "200% of the price" (new_price = price*(200/100)).
However, I take "200% more expensive" to mean new_price = price + price*(200/100). To me, the "more" implies addition specifically.
"Its cheaper, you get 3 times as much for the same price"
"Its cheaper, it costs 66% less"
Most people aren't well versed in math and just says something that kinda makes sense to them based on the numbers they know.
I think this is more intuitive for most people than using percent difference for the same reason using a raw ratio (speed up) is better for expressing performance improvement.
Which would be 33% the original price.
For the opposite word, "expensive", it all depends on the word before it: "as expensive" or "more expensive".
"100 widgets cost $100. Unit price is $1." They are now twice _as_ expensive -> they now cost $200. Unit price is now $2.
"100 widgets cost $100. Unit price is $1." They are now two times _more_ expensive -> they now cost $300. Unit price is now $3.
But How do you do this with the word "cheap"? Does "cheaper" clearly mean "3x as cheap" or "3x more cheap" (which even sounds a bit wrong)? I guess it means "more cheap" means "cheaper".
Better to avoid the problem and say "They are now two times the cost, or half the cost". Which is much clearer.
Frustratingly, I find this isn't always true in practice. Lots of people use "x times as y" and "x times more y" interchangeably. To avoid ambiguity I try to only use the former in any context where precision is useful.
It’s also used more generally for buy one get one X% off sales and you’ll see lots of “BOGO50” promo/coupon codes. Then marketing jumped on the bandwagon and started plastering “BOGO” all over marketing material so consumers are used to the lingo now
> We’re making our API more affordable on September 1, thanks to progress in making our models run more efficiently.
If you visit the page the price change is pretty clear, they show the before and after. 1k tokens will be anywhere from 1/2 to 1/3 the current price depending on which model you're using.
An order of magnitude is an exponential change of plus or minus 1 in the
value of a quantity or unit. The term is generally used in conjunction with
power-of-10 scientific notation.
Order of magnitude is used to make the size of numbers and measurements of
things more intuitive and understandable. It is generally used to provide
approximate comparisons between two numbers. For example, if the
circumference of the Sun is compared with the circumference of the Earth,
the Sun's circumference would be described as many orders of magnitude
larger than the Earth's. [1]
An order of magnitude is an approximation of the logarithm of a value relative to some
contextually understood reference value, usually 10, interpreted as the base of the
logarithm and the representative of values of magnitude one. [2]
For this item to become 3x cheaper would mean for its cheapness score to be multiplied by 3, thus 1/10 antidollars. Thus “3x cheaper” means the same thing as “price multipled by one third”.
A 33% discount would be "1.5x cheaper" by my analysis.
Assuming that if, say, the service was 25% cheaper than the nominal price last month, it is now 0.25 * 3 = 75% cheaper than the nominal price. However, it would make more sense to advertise that as a 75% discount over nominal price, so probably that's NOT what they meant.
That's the only explanation that makes sense arithmetically, I believe. There cannot be another interpretation of "3x cheaper" unless you reinterpret the word "cheaper."
In the headline of this page, it wouldn't have taken but a moment to write a phrasing that would be understandable to all readers of the page. It could read, "GPT-3 will be Two Thirds Less," or, "GPT-3 will cost One Third the Price."
This example from a non-native English speaker shows the difficulties of having several possible interpretations, at least three! [1]
[1] https://forum.wordreference.com/threads/%C2%ABx-times-less%C...
I wrote about using it to explain code (and mathematical formulas and suchlike) a few weeks ago: https://simonwillison.net/2022/Jul/9/gpt-3-explain-code/
I've been experimenting with using it to build a human-language-to-SQL tool, so that people can ask questions of their data "what country had the highest GDP in 2019" and it will turn them into the correct SQL query, given a table schema. I'm still iterating on this but it's shown some very promising initial results.
I use it a lot when I need to get something small working in a language that I don't have day-to-day familiarity with. "Write a bash script that loops through every MOV file in this folder and extracts the audio as MP3" is a good example of that kind of prompt.
Riley Goodside on Twitter posts weird and interesting new things you can do with GPT-3 on a daily basis: https://twitter.com/goodside/ - his instructional template trick was on HN the other day, it's really clever: https://news.ycombinator.com/item?id=32532875
1. Generate synthetic data that is well aligned to your needs. With careful prompting + ensembling + after-fact human filtering you can generate a lot of very particular human-like data that you can then used to train/etc your product.
2. Generate labels. gpt-3 can give pretty good NLU results through appropriate prompting. You can do multiple prompts + ensembling to get very good labels on free text (sentiment, entity linking, intent, etc).
In both above use cases you can actually avoid deploying gpt-3 as part of client facing product, but instead leverage gpt-3 to train smaller "on-rails" models/rules/etc.
* large running costs due to the expensiveness of the inference
* low barriers to entry: the tech behind Dalle appeared 1.5 years ago if I recall right and there are already a few competitors (Midjourney, Imagen, Stable Diffusion)
* low value at the present time of the raw APIs
Regardless of that moral question, I don't think a fully open model on the level of GPT-3 is even possible. Given the required cost to train and the expertise involved, big tech will always be a few years ahead. And it's unlikely they would give it away with how much they invest in creating it. Unless capitalism suddenly ends, I don't see any of the major tech companies parting with state of the art ML.
Thankfully we have free alternatives to GPT-3 (BLOOM [0]) and DALL-E 2 (Stable Diffusion [1])
[0] https://huggingface.co/bigscience/bloom
[1] https://github.com/huggingface/diffusers/releases/tag/v0.2.3
Since even with these price decreases, lots of competitors are out there with more parameters and are open source, and are already matched or just as good as GPT-3 and are generally running OpenAI's pricing to the ground.
Might as well open everything up since you can't compete with free anyway.
Yet Stable Diffusion is already getting open sourced and is just as good, if not better than what I am seeing with DALL-E 2.
2. Translating documents from one language to another
3. Writing e-mails, letters, and other forms of correspondence
4. Crafting resumes and cover letters
5. Writing marketing materials, such as brochures and flyers
6. Responding to comments on HN.
(I added the last one)
I'm very impressed with the AI image generation, but again, it replaces... stock art for blog posts?
If I was a creative writer then I think the AI could help me with plot outlines or get over writing blocks. But otherwise I'm not grasping why this is going to automate some huge number of jobs.
Writing definitely looks like the key use case. The New Yorker used GPT-3 to write the concluding sentence for one of their articles [0], and I believe The Economist did the same (but for a slightly greater length; though I couldn't find the source just now). It's actually a bit hard to come up with a good conclusion, so this could save some effort and trouble. Separately, GPT-3 can also likely be used for articles about sports, especially if the purpose is to track changes on a scoreboard and present the changes with an article written in prose, versus displayed in a table.
It's quite fair that marketing blogs may not be interesting to many readers, but they can be useful for SEO/boosting the search engine rankings of a website. So, a cash-strapped early or small company could save money from freelance copywriting to GPT-3 if they're not looking for anything fancy, and a lot of people work as freelance copywriters.
I agree that this may not displace anyone in the high-end/high quality segment of copywriting, but it could reduce demand for content farms in the future (i.e. any application where low-quality copywriting written on a short notice is in-demand).
[0] https://www.newyorker.com/magazine/2019/10/14/can-a-machine-...
because it is now too easy to generate-then-polish papers at that level.
They're too good in other words: better than baseline even. And when touched up by a student who's putting in a modest effort, are disruptively better.
You basically just enter in the Airbnb URLs you're considering, and it'll use GPT-3 to scrape through the reviews in order to find anything negative or positive people are saying about the listings.
You can also ask any question you want at all, and it'll return answers (if people have mentioned it in their reviews). For example "Is the air conditioning loud?"
I would love to hear what others are using GPT-3 for!