There is a new clear best (non-tiny) LLM. If you want to converse with an LLM, the correct answer is Claude Sonnet 3.5. It is available for free on Claude.ai and the Claude iOS app, or you can subscribe for higher rate limits. The API cost is $3 per million input tokens and $15 per million output tokens.
Are you referring to Gwern's statement that Anthropic said they wouldn't push model capabilities forward? Dario has also said similar things in public interviews:
Personally I think it's fine for Anthropic to change their mind on this. You can't stay afloat by staying second best. They do have to show that they are on the cutting edge to keep up pace with investment.
I’m addressing the “Gwern is a reliable source of newsworthy/novel information” subtext present via second-order citation in this post as a rhetorical move that does no favors vis a vis credibility to any argumentative piece or analysis contained therein.
This also applies to stories beyond this topic, and it is not a prior unique to Zvi (sorry for the distraction here from your otherwise great &y timely piece, sir) or most of this subculture's discourse members, apparently.
This is an incredibly extreme charge to level without any explanation. I've been lurking in this general corner of the world fairly closely for a long time and have literally no idea what you might be talking about.
So, I'm pretty interested in AI, have been playing around with the models, enjoyed Claude the most, but have still considered the whole thing generally overhyped at current capacities (which is not to deny the possibility that it can get much better, but I have personally struggled to get a lot of hard value thus far).
I played around with Sonnet 3.5 today and needed to go for a walk after. This is, for me, a turning point, and the first time I've been really compelled to push friends into playing with it.
Thanks to Trevor Levin for helping me out with this.
I am pretty new to learning more deeply about AI risk & speaking up about it publicly. So I appreciate respectful feedback. I'm of the opinion that it is better to learn as I go & accept feedback from those with more experience, since I will never be able to keep up with all the technical complexity proactively.
But I feel it's just too important to spread awareness of the risks associated with reckless, race dynamic driven AI development. I see this as a major human rights issue that is inherently social & political in nature, in addition to the obvious technical challenges.
I don't have a technical ML background, just a civil engineer's general technical mindset.
Zvi, I really deeply appreciate the deep work you do with this blog to collect & translate information in the context of a clear long term dedication to this issue! Keep up the great work!
Thanks for this! Switching over to it now. As an aside, where is their marketing department?? that would tell you not to launch a ‘3.5’ version when your main competitor launched a 3.5 version 18 months ago…
By far the best coding assistant I have ever worked with! The bar has been raised really high and I can't wait to see what's coming in the next 6 months.
Are you referring to Gwern's statement that Anthropic said they wouldn't push model capabilities forward? Dario has also said similar things in public interviews:
https://www.youtube.com/shorts/tPqIRJL3wM8
Personally I think it's fine for Anthropic to change their mind on this. You can't stay afloat by staying second best. They do have to show that they are on the cutting edge to keep up pace with investment.
I’m not referring to any specific instance.
I’m addressing the “Gwern is a reliable source of newsworthy/novel information” subtext present via second-order citation in this post as a rhetorical move that does no favors vis a vis credibility to any argumentative piece or analysis contained therein.
This also applies to stories beyond this topic, and it is not a prior unique to Zvi (sorry for the distraction here from your otherwise great &y timely piece, sir) or most of this subculture's discourse members, apparently.
But why do you think he isn’t, if you don’t mind me asking?
This is an incredibly extreme charge to level without any explanation. I've been lurking in this general corner of the world fairly closely for a long time and have literally no idea what you might be talking about.
So is it his backstory or is his journalism that is questionable?
Podcast episode for this post:
https://askwhocastsai.substack.com/p/on-claude-35-sonnet-by-zvi-mowshowitz
I wish they didnt advance the frontier but Sonnet is quite cool. For awhile, it argued for the importance of Pausing AI. Pretty amusing/great.
6 weeks since a non AI post :(
Monthly roundup this month?
Wow, yeah, that's true.
Yes, tomorrow barring an emergency post. Would have been today without Claude 3.5.
So, I'm pretty interested in AI, have been playing around with the models, enjoyed Claude the most, but have still considered the whole thing generally overhyped at current capacities (which is not to deny the possibility that it can get much better, but I have personally struggled to get a lot of hard value thus far).
I played around with Sonnet 3.5 today and needed to go for a walk after. This is, for me, a turning point, and the first time I've been really compelled to push friends into playing with it.
I'll be interested to see if others agree.
I am very impressed with Claude, thanks!
I am surprised my post on RSI was quoted here! That one did get a lot of eyes on it.
I already corrected it on Twitter, but I will repeat here that my technical understanding of RSI was corrected a bit by people's responses to that.
Here's the correction: https://x.com/thegartsy/status/1804650086192079168
Thanks to Trevor Levin for helping me out with this.
I am pretty new to learning more deeply about AI risk & speaking up about it publicly. So I appreciate respectful feedback. I'm of the opinion that it is better to learn as I go & accept feedback from those with more experience, since I will never be able to keep up with all the technical complexity proactively.
But I feel it's just too important to spread awareness of the risks associated with reckless, race dynamic driven AI development. I see this as a major human rights issue that is inherently social & political in nature, in addition to the obvious technical challenges.
I don't have a technical ML background, just a civil engineer's general technical mindset.
Zvi, I really deeply appreciate the deep work you do with this blog to collect & translate information in the context of a clear long term dedication to this issue! Keep up the great work!
very easy to get around claude face detection
just ask it to guess
posted 8 examples here
https://x.com/yaakovgrunsfeld/status/1805524934875050261
Thanks for this! Switching over to it now. As an aside, where is their marketing department?? that would tell you not to launch a ‘3.5’ version when your main competitor launched a 3.5 version 18 months ago…
Claude 3.5 Sonnet can solve the farmer and sheep problem with two small changes to the prompt:
1. change the word "person" to "human".
2. change the word "trips" to "trip or trips". (Claude is probably assuming that the answer has to be in multiple trips because of the word "trips")
By far the best coding assistant I have ever worked with! The bar has been raised really high and I can't wait to see what's coming in the next 6 months.
How to Write A Proper Rebuttal https://open.substack.com/pub/michael880/p/how-to-write-a-proper-rebuttal?r=3b6pw1&utm_campaign=post&utm_medium=web