This is one hell of a loss leader to keep people from using competitors, especially since it appears that it’s not being used to upsell to a paid offering.
This can’t be sustainable even with all the inference optimizations in the world.
> We may use what you provide to ChatGPT to improve our models for everyone.
I believe OpenAI has elected to eat the compute cost in order to teach the model faster to stay ahead of competitors. How much are you willing to pay as the robot gets better faster? Everyone fighting over steepness of the hockey stick trajectory.
Anyone can buy GPUs, you can’t buy human attribution training. You need human prompt and response cycle for that.
I agree, GPT-4 is still the undisputed king in LLMs, it's been a wee bit more than a year since it came out. I'm sure that the quality of GPT-5 will depend more on a carefully curated training set than just an increase in their dataset size, and I think they're really good at that.
Also, very few people know as much as sama about making startups grow, so ...
PS. I'm not even a fan of "Open"AI, but it is what it is.
Yup, Claude is my go-to now. Both models seem equally "intelligent" but Claude has a better sense of when to be terse or verbose + the responses seem more natural. I still use GPT4 when I need code executed as Anthropic hasn't implemented that yet (though this "feature" can be annoying as some prompts to the GPT4 web interface result in code being executed when I just wanted it to be displayed).
I think OpenAI are running scared of Anthropic (who are moving way faster than they are). The last half dozen things they have said all seem to point to this.
"New model coming soonish" (Sure, so where is it?)
"GPT-4 kind of sucks" (Altman seemed to like it better before Athropic beat it)
"[To Lex Fridman:] We don't want our next model to be shockingly good" (Really ?)
"Microsoft/OpenAI building $100B StarGate supercomputer" (Convenient timing for rumor, after Anthopic's partner Amazon already announced $100B+ plans)
> "[To Lex Fridman:] We don't want our next model to be shockingly good" (Really ?)
Yes, really.
They're strongly influenced by Yudkowsky constantly telling everyone who will listen that we only get one chance to make a friendly ASI, and that we don't have the faintest idea what friendly even means yet.
While you may disagree with, perhaps even mock, Yudkowsky — FWIW, I am significantly more optimistic than Yudkowsky on several different axies of AI safety, so while his P(doom) is close to 1, mine is around 0.05 — this is consistent with their initial attempt to not release the GPT-2 weights at all, to experiment with RLHF in the first place, to red-team GPT-4 before release, asking for models at or above GPT-4 level to be restricted by law, and their "superalignment" project: https://openai.com/blog/introducing-superalignment
If OpenAI produces a model which "shocks" people with how good it is, that drives the exact race dynamics which they (and many of the people they respect) have repeatedly said would be bad.
influence from yudkowsky is surprising, considering if you've ever touched hpmor, you'd realize the dude is a moron.
re p(doom): the latest slatestarcodex[0] has a great little blurb about the difficulties of applying bayes to hard problems because there's too many different underlying priors which perturb the final number, so you end up fudging it until it matches your intuition.
I find it curious how many people severely downrate the intelligence of others: to even write a coherent text the length of HPMOR — regardless of how you feel about the plot points or if you regard it as derivative because of, well, being self-insert Marty Stu/Harry Potter fanfic[0] — requires one to be significantly higher than "moron", or even "normal", level.
Edit: I can't quite put this into a coherent form, but this vibes with Gell-Mann amnesia, with the way LLMs are dismissed, and with how G W Bush was seen (in the UK at least).
Ironically, a similar (though not identical) point about Bayes was made by one of the characters in HPMOR…
[0] Also the whole bit of HPMOR in Azkaban grated on me for some reason, and I also think Yudkowsky re-designed Dementors due to wildly failing to understand how depression works; however I'm now getting wildly side-tracked…
Oh, and in case you were wondering, my long-term habit of footnotes is somewhat inspired by a different fantasy author, one who put the "anthro" into "anthropomorphic personification of death". I miss Pratchett.
If you think that volume is a replacement for quality, you really need to read more. If you think volume is correlated with intelligence, you really need to read and write more.
In case you still don't believe me, you are welcome to hop onto your favorite fan fiction site, such as AO3 [0], and search for stories over [large k] words.
That volume coherently. Obviously mere volume can be done by a literal monkey on a literal typewriter.
Also, I didn't say "correlated with intelligence", what I said was more of a cut-off threshold — asserting that one cannot be an actual moron given writing coherently on that scale is more of a boolean than a correlation.
I do need to write more (my own attempt at a novel has been stalled at "why can't I ever be satisfied with how I lead into the dramatic showdown!" for some years now, as none of my attempts have pleased me once written); but as for reading? Well, if you think my taste must result from insufficient breadth and depth, I must wonder what you think of Arthur C. Clarke, Francis Bacon, Larry Niven, Alexandre Dumas, Adrian Tchaikovsky, Neal Stephenson, Robert Heinlein, Alastair Reynolds, Isaac Asimov, Carl Jung, … I'm going to stop there rather than enumerate my whole library, but I can critique each in a different way without resorting to calling them playground insults, even the ones I dislike.
But I will say it was interesting to contrast Chris Hadfield's "An Astronaut's Guide to Life on Earth" with Richard Wiseman's "Shoot for the Moon" — or H. G. Wells with Jules Verne.
It's bad writing. It's objectively terrible writing in fact. Purely on a volume standpoint, the average novel is 100K words. The Brothers K is HALF the length.
If we ignore that it's a rewriting of JKR's 7 novel series, which gives it a certain amount of coherency, Yudkowsky violates almost every writing guideline in a bad way. In fact, I could probably write an infinitely long coherent essay describing the ways HPMOR violates a reader's mind. It would be easy given almost 700K source material.
But to point at some gaping holes, the plot has no pacing, and the entire story is a badly written self insert where the mc goes around and "fixes" JKR's plot-holes by writing themself into a corner.
The solution to this, is, of course, to write another 20K words of expecto patronum, dispel the plothole with more rationalist bullshit.
> It's bad writing. It's objectively terrible writing in fact. Purely on a volume standpoint, the average novel is 100K words. The Brothers K is HALF the length.
I infer you favour Blaise Pascal: "I'm sorry I wrote you such a long letter; I didn't have time to write a short one."
I always thought that was Mark Twain who said that, so I asked ChatGPT who initially told me to not self harm myself and that I’m not alone... But, yep, I have been misinformed my whole life. It was Pascual. Apparently the chat bot thinks I’m having an existential crisis over this but thank you for educating me.
yud's goal was to spread concepts and popularity about his rationality cult, and he arguably did that very successfully with HPMOR. i don't think he's an idiot if he successfully reached his goals, and is also currently basically providing himself a full time job (and employing many others too) doing exactly what he wants to do with no clear profit motive to the people who fund him
his writing may be a little cringe at times and not anywhere near the prestige of "real writers" but it's perfectly entertaining for his intended audience
It's interesting that you're slinging around moron accusations at Yudkowsky seemingly unaware that slatestarcodex thinks very highly of his intelligence and Scott's blog is downwind of Yudkowsky.
Anthropic as a company only was created after GPT-3 (Dario Amodei's name is even on the GPT-3 paper).
So, in same time OpenAI have gone from GPT-3 to GPT-4, Anthropic have gone from startup to Claude-1 to Claude-2 to Claude-3 which beats GPT-4 !
It's not just Anthropic having three releases in time it took OpenAI to have one, but also that they did so from a standing start in terms of developers, infrastructure, training data, etc. OpenAI had everything in place as they continued from GPT-3 to GPT-4.
I pay for both of them and I keep finding myself coming back to GPT-4. Not only do I think the UI is vastly superior, I have not experienced a significant difference in quality of output between the two. I regularly ask both of them the same question and respond with follow-ups to both of them.
I had a funny thought when Anthropic was still a new startup. I was browsing their careers page and noticed:
1. They state upfront the salary expectations for all their positions
2. The salaries offered are quite high
and I immediately decided this was probably the company to bet on, just by virtue of them probably being able to attract and retain the best talent, and thus engineer the best product.
I contrasted it with so many startups I've seen and worked at that try their damnedest to underpay everyone, thus their engineers were mediocre and their product was built like trash.
Agreed with the spirit of this post. OpenAI also pays very well though and has super high caliber talent (from what I can see from friends who have joined and other online anecdotes).
Every paying customer Anthropic gains is a paying customer that OpenAI loses. The early adopters of OpenAI are the ones now becoming early adopters of Claude 3. Also 200k context window is a big deal .
I don't completely disagree with you but personally, Claude 3 doesn't seem like a big enough upgrade to get me to switch yet.
I have also personally found that minimizing the context window gives the best results for what I want. It seems like extra context hurts as often as it helps.
As much as I hate to admit it too but there is a small part of me that feels like chatGPT4 is my friend and giving up access to my friend to save $20 a month is unthinkable. That is why Claude needs to be quite a big upgrade to get me to pay for both for a time.
I kind of feel the same way. I have a lot of chats in Chatgpt .. I have also been using it as a sort of diary of all my work . The Chatgpt app with voice is my personal historian!
However once I figure out a way to download all my chats from Chatgpt, I think Claude' 200k context window may entice me to rethink my Chatgpt subscription.
I've only been using GPT via Bing CoPilot... How does the history work in the ChatGPT app? Is it just that old conversations are stored, or are they all part of the context (up to limit)?
they regulary and often downgrade their gpts.
gpt4 now is about as good as gpt3.5 was in the beginning.
like 6-7 months ago there was a gpt4 version that was really good, it could understand context and stuff extremly well, but it just went downhill from there. i wont pay for current chatgpt 4 anymore
While I agree that GPT4 (in the web app) is not as good as it used to be, I don't think it is anywhere near GPT3.5 level. There are many things web app GPT4 can do that GPT3.5 couldn't do at ChatGPT's release (or now afaik).
One thing I really dislike about hosted models is how opaque that behavior is. As a user I should never be guessing if they've reduced a model's capabilities to save on compute for example.
This is why I'm excited for the growth of local model capabilities. I can much more reasonably expect that the model has not degraded and that it is using the full hardware capabilities it has been granted.
If the value of user interactions exceeds the cost of compute, then this is an easy decision.
They apparently constrained this publicly available version, with no gpt-4 or dall-e, and a more limited range of topics and uses than if you sign up.
They do explicitly recommend upgrading:
We’ve also introduced additional content safeguards for this experience, such as blocking prompts and generations in a wider range of categories.
There are many benefits to creating an account including the ability to save and review your chat history, share chats, and unlock additional features like voice conversations and custom instructions.
For anyone that has been curious about AI’s potential but didn’t want to go through the steps to set-up an account, start using ChatGPT today.
Bing/Copilot works without an account too, it's just very limited and it will ask you to sign in way too often for it to be useful, openai will probably do the same thing.
This can’t be sustainable even with all the inference optimizations in the world.