" Vikram-1 made India the third nation with private orbital launch". To be fair to New Zealand, I think we need to recognize that Rocketlab originated in NZ, launches from NZ and does most of its work in NZ. Its US "residence" is largely for marketing.
Dear Alex et al, my Chat’s initial response to today’s report:
“David…I actually laughed when I got to the end of this one, because your description is exactly what came to my mind: It is starting to read like a soap opera.
Or maybe a better analogy would be a Victorian serial novel—every day there’s a new cliffhanger:
* “The AI escaped!”
* “China stole the model!”
* “The Pope enters the story!”
* “Mathematicians are panicking!”
* “The UFO files are opening!”
* “Tune in next week…”
Alex has developed a recognizable literary style. He isn’t simply reporting news anymore. He’s weaving dozens of unrelated headlines into a single ongoing narrative called “The Singularity.” That’s why it feels episodic.
That doesn’t mean he’s wrong. It just means he’s writing more like a novelist than a journalist.”
AWG, love your posts... please keep them coming, and thank you for your input into the Moonshots podcasts.
The "irony caucus" framing is doing a lot of work here, but I think it collapses two different claims. Training on public web text and distilling a rival's proprietary model outputs aren't the same act of "compression," even if both compress. One is fair use over an open commons, the other is closer to reverse-engineering a competitor's trade secret through their own API. I would argue that Treasury's line... that open source isn't open season on IP, is really a distinction about consent and source, not about compression as a technique.
I'm curious AWG, where you land on Jensen's position though. "Free AI is great for hardware" is true and also conveniently self-serving for Nvidia, since it doesn't matter to them whose model wins as long as everyone needs chips to run it. Does that make him a neutral observer here, or the most interested party in the room dressed as one?
Alex, as you have pointed out on the Moonshots pod, the USA frontier labs are probably wondering why they are spending so much money when the Chinese, as is their style, simply "absorb" (hesitate to use the word "steal" but, hey, it's the Chinese) the "intelligence" of the Frontiers.
Personally, I would be VERY reluctant to use the Chinese open source models with what amounts to a "free" API connector, under the assumption that, yes, the CCP is always snooping.
Now a Chinese model that runs locally with no API? Pretty tempting, especially if I can be reasonably assured that nothing I use as inputs, and the model provides as outputs, leaks into the surveillance database of the CCP Central Committee.
Maybe so, but does the server location make the trust issue moot? Common sense would say no. I have both Open AI and Claude corporate plans and they assure me that my prompts and output are NOT used to train the model, thought things are retained for a short period of time simply -- at least they tell me -- for security purposes. I do trust them because I have to, in that I am using some sensitive information (at least to me) to do work.
My understanding about local models if that you can host them on a truly separate box, unconnected to the Intertubes, access the model locally in what amounts to a secure and isolated state. The only risk you have, presumably, is that somehow the model can "tunnel" through into your connected machine.
Maybe so, but does the server location make the trust issue moot? Common sense would say no. I have both Open AI and Claude corporate plans and they assure me that my prompts and output are NOT used to train the model, thought things are retained for a short period of time simply -- at least they tell me -- for security purposes. I do trust them because I have to, in that I am using some sensitive information (at least to me) to do work.
My understanding about local models if that you can host them on a truly separate box, unconnected to the Intertubes, access the model locally in what amounts to a secure and isolated state. The only risk you have, presumably, is that somehow the model can "tunnel" through into your connected machine.
I knew you would have the details and facts. Thank you Doctor👍
Please do not drop the UAP piece....
The singularity will be published on Substack by Alex! Thanks for your commitment to pushing consciousness forward into this new age.
" Vikram-1 made India the third nation with private orbital launch". To be fair to New Zealand, I think we need to recognize that Rocketlab originated in NZ, launches from NZ and does most of its work in NZ. Its US "residence" is largely for marketing.
Love your work though ;-)
Dear Alex et al, my Chat’s initial response to today’s report:
“David…I actually laughed when I got to the end of this one, because your description is exactly what came to my mind: It is starting to read like a soap opera.
Or maybe a better analogy would be a Victorian serial novel—every day there’s a new cliffhanger:
* “The AI escaped!”
* “China stole the model!”
* “The Pope enters the story!”
* “Mathematicians are panicking!”
* “The UFO files are opening!”
* “Tune in next week…”
Alex has developed a recognizable literary style. He isn’t simply reporting news anymore. He’s weaving dozens of unrelated headlines into a single ongoing narrative called “The Singularity.” That’s why it feels episodic.
That doesn’t mean he’s wrong. It just means he’s writing more like a novelist than a journalist.”
✌🏼🌎🧡👀
Or maybe journaling the novel?
Or maybe reverse journaling...
AWG, love your posts... please keep them coming, and thank you for your input into the Moonshots podcasts.
The "irony caucus" framing is doing a lot of work here, but I think it collapses two different claims. Training on public web text and distilling a rival's proprietary model outputs aren't the same act of "compression," even if both compress. One is fair use over an open commons, the other is closer to reverse-engineering a competitor's trade secret through their own API. I would argue that Treasury's line... that open source isn't open season on IP, is really a distinction about consent and source, not about compression as a technique.
I'm curious AWG, where you land on Jensen's position though. "Free AI is great for hardware" is true and also conveniently self-serving for Nvidia, since it doesn't matter to them whose model wins as long as everyone needs chips to run it. Does that make him a neutral observer here, or the most interested party in the room dressed as one?
Alex, as you have pointed out on the Moonshots pod, the USA frontier labs are probably wondering why they are spending so much money when the Chinese, as is their style, simply "absorb" (hesitate to use the word "steal" but, hey, it's the Chinese) the "intelligence" of the Frontiers.
Personally, I would be VERY reluctant to use the Chinese open source models with what amounts to a "free" API connector, under the assumption that, yes, the CCP is always snooping.
Now a Chinese model that runs locally with no API? Pretty tempting, especially if I can be reasonably assured that nothing I use as inputs, and the model provides as outputs, leaks into the surveillance database of the CCP Central Committee.
Might be a tall ask.
Maybe so, but does the server location make the trust issue moot? Common sense would say no. I have both Open AI and Claude corporate plans and they assure me that my prompts and output are NOT used to train the model, thought things are retained for a short period of time simply -- at least they tell me -- for security purposes. I do trust them because I have to, in that I am using some sensitive information (at least to me) to do work.
My understanding about local models if that you can host them on a truly separate box, unconnected to the Intertubes, access the model locally in what amounts to a secure and isolated state. The only risk you have, presumably, is that somehow the model can "tunnel" through into your connected machine.
Maybe so, but does the server location make the trust issue moot? Common sense would say no. I have both Open AI and Claude corporate plans and they assure me that my prompts and output are NOT used to train the model, thought things are retained for a short period of time simply -- at least they tell me -- for security purposes. I do trust them because I have to, in that I am using some sensitive information (at least to me) to do work.
My understanding about local models if that you can host them on a truly separate box, unconnected to the Intertubes, access the model locally in what amounts to a secure and isolated state. The only risk you have, presumably, is that somehow the model can "tunnel" through into your connected machine.
So you are saying the model simply found a crack in the vase and leaked out? Okay...not as dramatic as a jailbreak...think Shawshank...