"Anthropic and OpenAI likely have among the lowest costs per unit of frontier-quality intelligence"
That's a big claim that his whole thesis rests on but is largely not backed up. Where are the apples-to-apples tokens-to-answer benchmarks that he's using - doesn't look like there are any, just a handwavy implication that US models are more token efficient, which they may be. But how is there so little effort in establishing this point in the article? And US labs may be in much different situations from one another: it's known that some labs like OpenAI bought big, early on compute and may have secured better pricing.
His article also does not mention the average price of electricity in China vs the US, which it seems like China leads on, and probably has the political power to more heavily subsidize. While I agree the COGS is often overlooked by top line benchmarks on coding tasks, etc, it seems that he's running on a big assumption while claiming "labs on the frontier will be fine".
But.. if you are running Chinese model in the US, what difference does it make? Isn't the whole "scare" (khm khm) with Kimis is that now I don't need Claude, cause I can run Kimi on my own hardware in my own datacenter and it's maybe not as good as Claude July edition but it's is as good as Claude January edition.
OpenAI's first related fix was in 0.142.x on June 22 but apparently it wasn't sufficient and user have continued to complain.
Seems like a thread with decent additional context: https://x.com/0x_kaize/status/2078872972245225586
This old post of mine may be of interest, where I point out: "After a bit of threat modeling, it becomes apparent that future spearphishing robocalls may not directly con you, but rather “farm” your voice data by asking you benign questions, and use that to train a voice model to penetrate more deeply into your network." A lot of this writing has been on the wall forever and many (I'm sure otherwise smart people in) mission critical industries like banking, ISPs, and more refused to even acknowledge the risks.
2021 - "Despite the prevalence of deepfake audio tech, banks and ISPs rush ahead with “voice print” authentication" https://keydiscussions.com/2021/12/07/despite-the-prevalence... ends with a section called "The next crisis: robocalls that spoof the voices of victims at scale"
Telcos need to implement STIR/SHAKEN in a binding way and it needs to be fully integrated into iOS/Android's UI. Spammers call from spoofed numbers, even numbers of, say, local schools (or from other people in your business) - so, if you get a call from a school and you've got kids, would you pick up? A couple of years ago I got a text that was "sent" from my boss (he didn't send it), etc.. The numbers aren't random.
Good luck actually unsubscribing from LinkedIn emails. The dark pattern they employ (or at least used to employ, for many many years) is to just periodically add new "categories" of email to send you, that you haven't yet unsubscribed from. Back in ~2015, out of frustration I changed my LinkedIn email to a throwaway
This would be illegal in the EU + UK, and certainly doesn't work like that in the UK. Potentially LinkedIn only made that change based on jurisdiction so maybe you two live in different places?
The odds of the OSS ecosystem existing in a permanent state of compromise have risen dramatically this year. (still not close, but the explosion in supply chain attacks is concerning)
PSA: this is true (the defaults), but there's a "Help improve Claude" setting that you can disable here https://claude.ai/settings/data-privacy-controls It's my understanding that, as long as this is off, Anthropic does not train on Claude Code conversations, inputs/outputs -- if anyone knows otherwise, please tell and provide a link if possible.
Meta's own research (and its use of it) has shown that it repeatedly ignores well-substantiated facts about the harms of its products. Now that Section 230 seems like a flawed shield, I fear the takeaway for other companies will be: never conduct honest research in the first place to preserve plausible deniability.
Meta has always wanted the appearance of caring about safety (helps them attract talent and keep mission-related morale high), while nearly always prioritizing growth (save for tiny blips of time, like in 2017 when the fallout of the cambridge analytica stuff was hitting a crescendo), whereas companies like X are run by people explicitly disinterested in putting significant resources into safety, especially research.
I will also add that, for the past few years, Meta and X both have become extremely hostile to external researchers of their platforms, shutting down access to tools and data.
That's a big claim that his whole thesis rests on but is largely not backed up. Where are the apples-to-apples tokens-to-answer benchmarks that he's using - doesn't look like there are any, just a handwavy implication that US models are more token efficient, which they may be. But how is there so little effort in establishing this point in the article? And US labs may be in much different situations from one another: it's known that some labs like OpenAI bought big, early on compute and may have secured better pricing.
His article also does not mention the average price of electricity in China vs the US, which it seems like China leads on, and probably has the political power to more heavily subsidize. While I agree the COGS is often overlooked by top line benchmarks on coding tasks, etc, it seems that he's running on a big assumption while claiming "labs on the frontier will be fine".
reply