> Claude 2 powers our chat experience, and is generally available in the US and UK. We are working to make Claude more globally available in the coming months.
I wonder why LLMs like GPT-4, Bard and Claude are so geo restricted at first? I understand some places have regulatory challenges but can’t see SG, UAE, or Chile being too difficult.
They want places with tech startups who will pay for their API.
Thats where there is lots of money to be made. And if they are GPU constrained, then launching in the countries with the highest proportion of future paying customers makes sense.
> I wonder why LLMs like GPT-4, Bard and Claude are so geo restricted at first?
Managing scale while maximizing profit potential? Also, US/UK probably lets them put their strongest linguistic foot forward initially, and there may be additional training done before rolling out to regions with other dominant languages.
> I understand some places have regulatory challenges
> I wonder why LLMs like GPT-4, Bard and Claude are so geo restricted at first? I understand some places have regulatory challenges but can’t see SG, UAE, or Chile being too difficult.
I'm amused by the inclusion of Chile in this list. I'm a Chilean and I do have access, but through the Anthropic Console, as I already had API Access.
It made me laugh when Google announced how strong Bard 2 is at over a hundred human languages and then restricted the deployed chat app to like… three.
That’s not even region locking, it’s capability locking while simultaneously advertising that very thing!
I’m very excited for Claude - been using it along side gpt 4 and pleased with its performance. The introduction of functions with OpenAI api complicates things and was hoping Claude would include this in a future api update
Excited to try it. We used Claude 1.x in experimentation, but shipped with OpenAI primarily because of time and SOC 2 compliance. Anthropic has come along since then, so we'll probably experiment with Claude more with intent to take into production if it's still holding up.
Cool. As CISO, can you please speak to the data retention policies that I noted here? https://news.ycombinator.com/item?id=36681239 . As you can imagine, sending sensitive information to a 3rd party is impossible without explicit agreements. As you're SOC2 and HIPAA are there devices in place for us to delete data, or specify data retention as customers?
Great to hear -- I just noticed that in your trust center as well! This was really the only blocker. Claude works very well, and our prompt we use with OpenAI can even just be used without any changes.
That’s awesome news - it’s be great if you also started do offer the chat style api that OpenAI now has including functions. I know it sucks to copy a competitor but it would allow us to plug in your api seamlessly and give customer base the option between multiple models which could be really cool especially with the 100k token limit.
Does this page need to be updated? It say's "We're currently working towards SOC 2 and HIPAA compliance", while in another place (can't find it at the moment) it say's you're compliant.
Also, the trust portal link on that page is a link to itself. Not the trust portal.
I thought for a moment that it could reach out to the internet, and it certainly makes it think you can, but its just lying about it.
I was able to get it to summarize the "How to Do Great Work" article with its url, but trying to get it to summarize the comments of the current laser chess HN article gave me something about cryptocurrency.
I've just been playing with Claude 1.3 this weekend to summarize large texts. It can take 100K tokens of input, enough for a whole Lex Fridman interview! :-) I've been getting pretty good results with it, so I'm excited to see how v2 works.
Now that it's entered open beta it's going to iterate rapidly. I had been using it fairly extensively, alongside other LLMs, through Slack and was always most impressed by it's output over the others.
(I do hold investment in Anthropic, but do not base my statements on that)
Human: Here is the transcript of a podcast:
<transcript>
[PASTE TRANSCRIPT HERE]
</transcript>
You are an expert on writing factual summaries. Write a summary of the podcast in about 10 sentences.
Assistant: I'd be happy to, here is the summary:
I applied and got access to the Claude 1 API a long time ago and then I guess I didn't click the link they gave me in time or something because when I went to try to get in it was expired. If I remember correctly.
I think I emailed them about it and was ignored.
I've been using the OpenAI API and I'm on the third version of my code generation application which is now a ChatGPT Plugin.
It sounds like Claude 2's reasoning is still lagging behind GPT-4 anyway.
I have the same problem with resemble.ai - I've submitted their "request a demo" form multiple times to try to get access to their multi-language API. Can't get a reply. I've tried emailing their support and sales teams and they ignore me.
All joking aside, I do agree with the sentiment that no one generally has any type of defensible moat at the moment. OpenAI has found a great balancing act between first mover advantage, marketing, customer adoption, and enterprise sales. They are executing at a high level. Anthropic (Claude) has a wonderful product but is lacking in consumer adoption and sales, though I think they're working on fixing that.
I think the reason Claude isn't further than it is now is due to a singular goal:
Do no evil
I feel they are what Google was to search engines when they burst onto the scene. They'd rather take time to get it right than to push out a product they don't feel meets it's mark.
Now that they have entered open beta I expect them to rapidly iterate. If their product is as good as I feel it already is, in comparison to competitors, then I can only imagine what it will be doing in a year's time.
(Disclaimer: I do have exposure to Anthropic through my investments, but nothing I said is based on that alone)
All the AI companies are sort of doing a VC rush, but instead of IPO it's AGI. Would be fun to see what we get in the future. Since a serious training run costs upwards of $50 million currently.
Excited for this, but I think with all this conversation about the role an AI assistant should play in work and development, this line feels incomplete to me:
> Think of Claude as a friendly, enthusiastic colleague or personal assistant who can be instructed in natural language to help you with many tasks.
It omits that the colleague may have outdated knowledge or not understand whatever problem you give it. The colleague's "enthusiasm" should be tempered with oversight so that the outputs they produce are not directly used without scrutiny. It seems that most people using these tools increasingly understand this, but to leave it off the website seems ... sloppy at this point.
Edit: upon logging in, I'm greeted by a warning "It may occasionally generate incorrect or misleading information, or produce offensive or biased content."
I just tried it again and it's still working for me. Were you accessing it in any special way? It should have just been a matter off adding the app to your Slack instance.
I would say it’s actually now the “modern” approach.
The old way was to move fast and break things.
The new way is extreme risk aversion and lawyers. We’ve swung in the complete opposite direction because in the eyes of the current zeitgeist, all tech companies are default evil until proven otherwise.
Governments like the EU have made it clear they do not like their citizens interacting with experimental technology products that might be rough around the edges. Gotta protect everyone from these awful scary new fangled chat websites.
> all tech companies are default evil until proven otherwise
To be fair, that is a relatively reliable assumption. The market incentives are set up that way, so either they are or they are pushed out by those that are.
The market incentives are absolutely not set up that way.
While you can make money in the short term by scamming people, in the long term, you always get found out.
Markets are not perfectly efficient, but in aggregate they trend toward efficiency over time. Fundamentally, all products that continue to exist over time do so because they deliver value for people.
You may not agree with what other people value, but then your problem is with those people, and we have to ask the question--why are the things you value so much better and why should we make you King and allow you to dictate the needs and wants of everyone else?
Or rather, “fining the hell out of companies for offering services to your citizens that your citizens want and the rest of the world is fine with” is such a Stone Age approach
counter point: asking companies to do the bare minimum to guarantee that companies don't steal customer's data, infringe copyright and generally piss on people's rights for their own profit, seems like a reasonably good thing.
I think their approach to region blocking is a bit rude. First, they pretend that I can sign up, but then they paddle back once I entered the login code. Why waste my time like that?
Also I do not understand why they are region blocking at all. If they are worried about GDPR, they should not let me enter an email address.
I've had the same happen. I've registered and logged in via VPN with a UK exit node. Worked perfectly.
Even stranger, now that my account is created, the VPN seems to be no longer necessary. I can just login with e-mail and the token they send me each time.
I'd like to try Claude, but the data retention policies in the Anthropic terms are not clear. Section 6e[0] claims they won't use customer data to train models, but I'd like to know if customer data is kept for any duration (like it is with OpenAI for 30 days). There is a note about data deletion on termination in section 14, so I assume that ALL data is retained for an undisclosed period of time.
Since I've been on a AI code-helper kick recently. According to the post, Claude 2 now 71.2%, a significant upgrade from 1.3 (56.0%). (Found in model card: pass@1)
For comparison:
* GPT-4 claims 85.4 on HumanEval, in a recent paper https://arxiv.org/pdf/2303.11366.pdf GPT-4 was tested at 80.1 pass@1 and 91 pass@1 using their Reflexion technique. They also include MBPP and Leetcode Hard benchmark comparisons
* The best open model I know of atm is replit-code-instruct-glaive, a replit-code-3b fine tune, which scores a 63.5% pass@1. An independent developer abacaj has reproduced that announcement as part of code-eval, a repo for getting human-eval results: https://github.com/abacaj/code-eval
Also, as with all LLM evals, to be taken with a grain of salt...
Liu, Jiawei, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. “Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation.” arXiv, June 12, 2023. https://doi.org/10.48550/arXiv.2305.01210.
Out of curiousity, I was trying out gpt-4-0613 and claude-v2 with https://github.com/getcursor/eval, but sadly I'm getting hangs at 3% with both of them (maybe hitting rate limits?).
do we have evidence that OpenAI is making new versions of gpt4 available? The training data presumably hasn’t changed since 2021 and the model is absurdly expensive to train; there’s little incentive for them to keep touching it up.
The issue with all of the chat-optimised LLMs is that they can’t be incrementally updated.
After the base training there are three seperate sets of additional training to align the model, convince it to do question-response, and to improve the quality via feedback.
If you update the original base model, then all the tuning steps need to be repeated.
For a model the size of GPT-4 this is expensive and slow, which is why OpenAI hasn’t bothered.
Why would you only update the original base model and not simply continue fine-tuning?
(This is a rhetorical question of course, and there are multiple empirical ways to confirm the majority of your answer is "hallucinated," not least of which is the fact that OpenAI keeps rolling out new versions and variants with different capabilities.)
Pre-training of a foundational model is what you're thinking of for the "absurdly expensive" part but fine tunes are extremely cheap and undoubtedly are being done constantly. (You can see just how cheap by looking at the papers for Alpaca, Vicuna, Koala, etc). Prices dropped from about $600 to $10 for smaller models. Guanaco, using QLoRA, fine tuned llama-65b in about 1 day on a single GPU.
Another way to empirically test btw is to search for all the articles pointing out what ChatGPT gets wrong (3 or 4). I recently tested those when looking for evals and it gets the large majority (maybe 80-90% of those are answered correctly now).
Wait, that actually sounds wonderful! This is the second best option of what happens when you have an over eager assistant: they try to help and then notice they are out of their dept, so they let me know, before they waste my time.
At the time, I preferred those hours anyway. I imagine there would be too many things to get just right if I wanted to adjust those hours, though you work with what you got. Issues would be temp, humidity, noise, etc Also loads of unknowns. A lot of these can be adjusted for, but this schedule worked perfect for me. I didn't have a nightlife, but I spent lots of time with friends during the day.
> I'm playing assetto corsa competizione, and I need you to tell me how many liters of fuel to take in a race. The qualifying time was 2:04.317, the race is 20 minutes long, and the car uses 2.73 liters per lap.
Claude 2 got really close to the right answer on the first try. The only thing it didn't consider is that time for 9.6 laps means driving 10 laps total, since incomplete laps must be finished after time is up. For the record, GPT-4 does take even this catch into account and gives a perfect answer, while Bard outputs results that are off by orders of magnitude.
Claude 2's response was:
> Okay, let's calculate this:
> * Your qualifying time of 2:04.317 equates to an average lap time of 124.317 seconds
> * A 20 minute race = 20 * 60 = 1200 seconds
> * So in a 20 minute race you can do roughly 1200 / 124.317 = 9.67 laps
> * At 2.73 liters per lap, that's 9.67 * 2.73 = 26.38 liters of fuel needed for the race
> So I would recommend taking on at least 27-28 liters of fuel to be safe for a 20 minute race, given your average lap times and fuel consumption. It's always better to take on a bit more fuel in case there are any safety cars or variations in your lap times during the race. Let me know if you have any other questions!
> The only thing it didn't consider is that time for 9.6 laps means driving 10 laps total, since incomplete laps must be finished after time is up.
I would be hesitant to judge an LLM on the various ground rules of the real world that it happens to have internalized or not, that aren't given in the prompt.
For my uses I'd want to have prompts that provide necessary context anyway - and would actually favor ones that can do it from the prompt without needing to already know - so that it is not limited when it encounters a novel problem domain.
Which is why knowledge cut off date is important. I prefer if it is frozen to pre-ChatGPT-3.5. Anything post-ChatGPT-3.5 release date should be considered tainted - imagine the sheer number of articles generated by spammers who used ChatGPT.
That might be fair in the short term. However it's not a workable option long-term, or all such models will be very limited in their knowledge as humanity advances technologically and culturally.
If you want me to be honest with you, LLMs are themselves a short term approach and can get us to, at max, AGI levels (for this current era). I don't see us getting to ASI with just LLMs. For the sort of "emergent ability" that ASI requires it has to be something more "simpler" and the learning be more "virulent" / "instantaneous" (not sure if these words convey what I really want to convey). Otherwise, LLMs will always have a "maxima" at which point it fails. And that maxima is collective intelligence of all of humanity in the current epoch. If you go back a 1000 years, the collective intelligence of all humanity would be completely different (primitive even). Would LLMs trained on that data have produced Knowledge that we know today? I don't think so. It could still, theoretically, reach AGI for that era and accelerate pace of learning by 50-100 years at a time. LLMs will surely accelerate pace of learning (as tools) even now but by themselves won't reach ASI levels. For ASI, we really need something more simpler/fundamental that is yet to be discovered. I don't feel LLMs are the way to ASI. AGI? Yeah possible.
Same is true for humans - a scientist inventing everything from their head would not achieve much, but if they can conduct experiments, and if they persevere, they eventually make discoveries. A pure LLM is like the first case, a LLM with tools or part of a larger system is the second.
Knowledge cut-off date doesn't prevent your model from getting tainted though - if you're doing any kind of RLHF, unless all your human reviewers were kept isolated from the world since ${knowledge-cutoff-date}, they will inadvertently give the model glimpses into the future.
It's not immediately apparent to people just how much leakage can happen this way. Up to a year ago, I'd probably give people this story[0] to ponder on, but now it's no longer a hypothetical - GPT-3.5 and GPT-4 are clear, practical demonstrations of just how much knowledge is implicitly encoded in what we say or write, and how this knowledge can be teased out of the input data without any prior context, completely unsupervised, given sufficient time and effort (which in silico translates to "sufficient compute", which we already have).
277 comments
[ 3.4 ms ] story [ 281 ms ] threadI wonder why LLMs like GPT-4, Bard and Claude are so geo restricted at first? I understand some places have regulatory challenges but can’t see SG, UAE, or Chile being too difficult.
The hilarious part of that is that the UK has basically all the regulations that they are probably worried about.
Thats where there is lots of money to be made. And if they are GPU constrained, then launching in the countries with the highest proportion of future paying customers makes sense.
Managing scale while maximizing profit potential? Also, US/UK probably lets them put their strongest linguistic foot forward initially, and there may be additional training done before rolling out to regions with other dominant languages.
> I understand some places have regulatory challenges
That’s probably not the main issue.
I'm amused by the inclusion of Chile in this list. I'm a Chilean and I do have access, but through the Anthropic Console, as I already had API Access.
That’s not even region locking, it’s capability locking while simultaneously advertising that very thing!
Wish I had more exposure to the company, to better join the ride, but I'll take what I have now.
Keep contributing to the awesome efforts going on there.
Also, the trust portal link on that page is a link to itself. Not the trust portal.
https://support.anthropic.com/en/articles/7996892-do-you-hav...
I was able to get it to summarize the "How to Do Great Work" article with its url, but trying to get it to summarize the comments of the current laser chess HN article gave me something about cryptocurrency.
I've collapsed & expanded it at least 10 times now, maybe I should get to reading the article...
(I do hold investment in Anthropic, but do not base my statements on that)
I think I emailed them about it and was ignored.
I've been using the OpenAI API and I'm on the third version of my code generation application which is now a ChatGPT Plugin.
It sounds like Claude 2's reasoning is still lagging behind GPT-4 anyway.
1. GPT-4
2. Claude 2
3. Bard
4. Llama/Alpaca
5-98. [Unclaimed]
99. SmarterChild AIM bot
100. Cohere
All joking aside, I do agree with the sentiment that no one generally has any type of defensible moat at the moment. OpenAI has found a great balancing act between first mover advantage, marketing, customer adoption, and enterprise sales. They are executing at a high level. Anthropic (Claude) has a wonderful product but is lacking in consumer adoption and sales, though I think they're working on fixing that.
Do no evil
I feel they are what Google was to search engines when they burst onto the scene. They'd rather take time to get it right than to push out a product they don't feel meets it's mark.
Now that they have entered open beta I expect them to rapidly iterate. If their product is as good as I feel it already is, in comparison to competitors, then I can only imagine what it will be doing in a year's time.
(Disclaimer: I do have exposure to Anthropic through my investments, but nothing I said is based on that alone)
They seem like the best way to evaluate models for general purpose use right now.
[1]: https://chat.lmsys.org/?arena
It also is more chatty than v1 (or GPT-3/4), even when asked to just pick one option out of three.
These benchmarks are product oriented - they contain tests and evals from our LLM-driven products. So they aren’t exhaustive or representative.
We just want to know when local LLMs are good enough to start migrating some pipelines away from OpenAI.
https://www-files.anthropic.com/production/images/Model-Card...
> Think of Claude as a friendly, enthusiastic colleague or personal assistant who can be instructed in natural language to help you with many tasks.
It omits that the colleague may have outdated knowledge or not understand whatever problem you give it. The colleague's "enthusiasm" should be tempered with oversight so that the outputs they produce are not directly used without scrutiny. It seems that most people using these tools increasingly understand this, but to leave it off the website seems ... sloppy at this point.
Edit: upon logging in, I'm greeted by a warning "It may occasionally generate incorrect or misleading information, or produce offensive or biased content."
The old way was to move fast and break things.
The new way is extreme risk aversion and lawyers. We’ve swung in the complete opposite direction because in the eyes of the current zeitgeist, all tech companies are default evil until proven otherwise.
Governments like the EU have made it clear they do not like their citizens interacting with experimental technology products that might be rough around the edges. Gotta protect everyone from these awful scary new fangled chat websites.
OpenAI got away with it by being a first mover.
To be fair, that is a relatively reliable assumption. The market incentives are set up that way, so either they are or they are pushed out by those that are.
While you can make money in the short term by scamming people, in the long term, you always get found out.
Markets are not perfectly efficient, but in aggregate they trend toward efficiency over time. Fundamentally, all products that continue to exist over time do so because they deliver value for people.
You may not agree with what other people value, but then your problem is with those people, and we have to ask the question--why are the things you value so much better and why should we make you King and allow you to dictate the needs and wants of everyone else?
I kid, I kid.
Also I do not understand why they are region blocking at all. If they are worried about GDPR, they should not let me enter an email address.
Even stranger, now that my account is created, the VPN seems to be no longer necessary. I can just login with e-mail and the token they send me each time.
[0] https://console.anthropic.com/legal/terms
I’m excited that you’re passionate about privacy. We’ve put a lot of thought into our policies.
For comparison:
* GPT-4 claims 85.4 on HumanEval, in a recent paper https://arxiv.org/pdf/2303.11366.pdf GPT-4 was tested at 80.1 pass@1 and 91 pass@1 using their Reflexion technique. They also include MBPP and Leetcode Hard benchmark comparisons
* WizardCoder, a StarCoder fine-tune is one of the top open models, scoring a 57.3 pass@1, model card here: https://huggingface.co/WizardLM/WizardCoder-15B-V1.0
* The best open model I know of atm is replit-code-instruct-glaive, a replit-code-3b fine tune, which scores a 63.5% pass@1. An independent developer abacaj has reproduced that announcement as part of code-eval, a repo for getting human-eval results: https://github.com/abacaj/code-eval
Those interested in this area may also want to take a look at this repo https://github.com/my-other-github-account/llm-humaneval-ben... that also ranks with Eval+, the CanAiCode Leaderboard https://huggingface.co/spaces/mike-ravkine/can-ai-code-resul... and airate https://github.com/catid/supercharger/tree/main/airate
Also, as with all LLM evals, to be taken with a grain of salt...
Liu, Jiawei, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. “Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation.” arXiv, June 12, 2023. https://doi.org/10.48550/arXiv.2305.01210.
Out of curiousity, I was trying out gpt-4-0613 and claude-v2 with https://github.com/getcursor/eval, but sadly I'm getting hangs at 3% with both of them (maybe hitting rate limits?).
After the base training there are three seperate sets of additional training to align the model, convince it to do question-response, and to improve the quality via feedback.
If you update the original base model, then all the tuning steps need to be repeated.
For a model the size of GPT-4 this is expensive and slow, which is why OpenAI hasn’t bothered.
(This is a rhetorical question of course, and there are multiple empirical ways to confirm the majority of your answer is "hallucinated," not least of which is the fact that OpenAI keeps rolling out new versions and variants with different capabilities.)
Pre-training of a foundational model is what you're thinking of for the "absurdly expensive" part but fine tunes are extremely cheap and undoubtedly are being done constantly. (You can see just how cheap by looking at the papers for Alpaca, Vicuna, Koala, etc). Prices dropped from about $600 to $10 for smaller models. Guanaco, using QLoRA, fine tuned llama-65b in about 1 day on a single GPU.
Another way to empirically test btw is to search for all the articles pointing out what ChatGPT gets wrong (3 or 4). I recently tested those when looking for evals and it gets the large majority (maybe 80-90% of those are answered correctly now).
Me: Can you manipulate data tables?
C2: Yes I can. Here’s some of the things I can do.
Me: Here’s some data and what to do with it (annualized growth rates).
C2: [processes for a while and starts spitting out responses, then deletes all that]
Me: What happened?
C2: Sorry, I lied. I can’t do any of that
Full exchange: https://econtwitter.net/@TradingPlacesResearch/1106958439180...
But as I noted elsewhere, I am finding it very useful for text summarizing.
[1] https://www.theverge.com/features/23764584/ai-artificial-int...
[2] https://time.com/6275995/chatgpt-facebook-african-workers-un...
Believe it or not, getting the output "Sorry, I lied. I can't do any of that" is a huge accomplishment.
> I'm playing assetto corsa competizione, and I need you to tell me how many liters of fuel to take in a race. The qualifying time was 2:04.317, the race is 20 minutes long, and the car uses 2.73 liters per lap.
Claude 2 got really close to the right answer on the first try. The only thing it didn't consider is that time for 9.6 laps means driving 10 laps total, since incomplete laps must be finished after time is up. For the record, GPT-4 does take even this catch into account and gives a perfect answer, while Bard outputs results that are off by orders of magnitude.
Claude 2's response was:
> Okay, let's calculate this:
> * Your qualifying time of 2:04.317 equates to an average lap time of 124.317 seconds
> * A 20 minute race = 20 * 60 = 1200 seconds
> * So in a 20 minute race you can do roughly 1200 / 124.317 = 9.67 laps
> * At 2.73 liters per lap, that's 9.67 * 2.73 = 26.38 liters of fuel needed for the race
> So I would recommend taking on at least 27-28 liters of fuel to be safe for a 20 minute race, given your average lap times and fuel consumption. It's always better to take on a bit more fuel in case there are any safety cars or variations in your lap times during the race. Let me know if you have any other questions!
I would be hesitant to judge an LLM on the various ground rules of the real world that it happens to have internalized or not, that aren't given in the prompt.
For my uses I'd want to have prompts that provide necessary context anyway - and would actually favor ones that can do it from the prompt without needing to already know - so that it is not limited when it encounters a novel problem domain.
It's not immediately apparent to people just how much leakage can happen this way. Up to a year ago, I'd probably give people this story[0] to ponder on, but now it's no longer a hypothetical - GPT-3.5 and GPT-4 are clear, practical demonstrations of just how much knowledge is implicitly encoded in what we say or write, and how this knowledge can be teased out of the input data without any prior context, completely unsupervised, given sufficient time and effort (which in silico translates to "sufficient compute", which we already have).
--
[0] - https://www.lesswrong.com/posts/5wMcKNAwB6X4mp9og/that-alien...