I wonder if New Zealand might have some appeal for a NS campus? What it offers:
- Immigration is straightforward
- time zone is just 4 hours from PST and 4 hours from Beijing Time
- after maybe Singapore and HK, probably the best balance of tight integration with both the US and China you'll find anywhere
- beautiful natural environment
- Routinely tops WB Ease of Doing Business rankings with easy one-day business registration, streamlined and predictable regulatory environment, and a transparent legal system
- links to space with regular Rocket Lab satellite launches
NZ has always looked for ways build get past the tyranny of distance. Critical mass of driven entrepreneurial culture can be a problem there (the country is so _chill_) and I wonder if a NS campus could be a solution.
wait you are the one who did the ML session at NS this morning! thanks i enjoyed both the session and the writing! (now i understand why you were surprised with the voice mode usage. i also prefer just text haha)
>Even OpenAI doesn’t believe they’ve made enough research progress to deserve a one point increase in the version number. The half point increase from GPT 4.5 to GPT 5 was a massive flop.
Disagree here. GPT 4.5 itself was the flop. From GPT 4 --> GPT 5 there was a significant jump, it's just that most of the jump happened from o1 and o3 which were intermediate models. The jump from o3 --> GPT 5 was small.
Also, i am not sure the whole argument makes that much sense tbh. Sam explicitly said, that he want to release models in smaller iterations as they become more powerfull. Better to look at the time of release of GPT-n. There we find that, models are released with bigger and bigger time gaps... except GPT-5 is off trend and remarkably early
I don’t think the diminishing returns argument is really applicable except maybe in a narrow technical sense. It feels like a lot changed since you last wrote. Coding models are noticeably much better than 6 months ago, to the point where they went from nearly unusable to now widely used. And there is still so much AI can’t consistently do yet that could plausibly be done with current tech and some additional R&D. We’re only past diminishing returns for pure question answering language modeling, but the limitations there are more political than technical, I’m sure the chatbots could easily be much better writers if not for the cultural safety paranoia, and for everything else there is still much to do. (For example, the entertainment industry still hasn’t been hit by AI properly yet, and they’re worth billions too)
Here are some examples. These are all places where I think a big change can easily happen in 1-2 years, and the change comes from how a model is embedded in a larger product.
* Agent mode for daily tasks seems inevitable in 1-2 years but has not penetrated very much from what I can tell. I am thinking of things like making calendar events, searching email, answering questions about your own email, booking hotels and flights.
* Personal entertainment is likely to be wildly revamped in an era where an AI can rewrite text, audio, and video for you on demand. Ron Baxley and I experimented with it a little with his own writings, and it was very impressive how it could rewrite one of his short stories five times over, with each version taking a different character as the primary perspective. It all used his writing style. The way this works is going to take new tool development, new forms of composition for authors, and new consumption technology that is neither Netflix nor a plain web browser.
* Kindle Unlimited, in particular, seems likely to benefit from an overhaul or to risk being left behind. People who consume mass amounts of pulp will sometimes like to read AI modified versions of their favorite stories and settings, seeded by writings from their favorite personal authors.
* For audio and video engineering, the people in my network are not seeing the level of AI assistance that software engineres get. The models are mostly in the ballpark, but there is a lot of work to match the currently text-heavy model interfaces with the way the A/V engineering UIs work. For example, it may be the winning formula to start saving A/V mixes into a text format, analogous to abc format for sheet music, thereby letting AIs interact better with the content.
You never know about the future, but it really seems like AI is still just getting going.
I wonder if New Zealand might have some appeal for a NS campus? What it offers:
- Immigration is straightforward
- time zone is just 4 hours from PST and 4 hours from Beijing Time
- after maybe Singapore and HK, probably the best balance of tight integration with both the US and China you'll find anywhere
- beautiful natural environment
- Routinely tops WB Ease of Doing Business rankings with easy one-day business registration, streamlined and predictable regulatory environment, and a transparent legal system
- links to space with regular Rocket Lab satellite launches
NZ has always looked for ways build get past the tyranny of distance. Critical mass of driven entrepreneurial culture can be a problem there (the country is so _chill_) and I wonder if a NS campus could be a solution.
Beautiful environment is putting it mildly, and one very familiar to anyone from northern California
Yes! I would also say the _built_ environment, i.e. cities, are much cleaner and better provisioned, maybe halfway between US and Asia standard.
It’s good to read you again.
Thank you!!
Great post. Your bit about the New Cold War feels sadly true to me. I wrote about it and just keep seeing it more wherever I look https://open.substack.com/pub/minnesotaconfluence/p/we-dont-talk-enough-about-age?r=bwkxp&utm_medium=ios
Was super interested in how network school actually works and how it's panned out. Can't wait to hear more and hope to visit someday!
Fascinating article Brian, thank you and keep them coming 👏
wait you are the one who did the ML session at NS this morning! thanks i enjoyed both the session and the writing! (now i understand why you were surprised with the voice mode usage. i also prefer just text haha)
Sounds like a lovely place to retire to.
I'm AI-lower-middle-class, but with some grinding I aspire to reach AI-upper-lower-middle-class.
middle-class-high-low-codex-sparse-finetune-down-low-tooslow-production-3.2-final-mini
>Even OpenAI doesn’t believe they’ve made enough research progress to deserve a one point increase in the version number. The half point increase from GPT 4.5 to GPT 5 was a massive flop.
Disagree here. GPT 4.5 itself was the flop. From GPT 4 --> GPT 5 there was a significant jump, it's just that most of the jump happened from o1 and o3 which were intermediate models. The jump from o3 --> GPT 5 was small.
"GPT 2 -> GPT 3 -> GPT 4 -> GPT 4.5 -> GPT 5 -> GPT 5.1 -> GPT 5.2 -> GPT 5.3"
You forgot GPT 3.5 (;
Also, i am not sure the whole argument makes that much sense tbh. Sam explicitly said, that he want to release models in smaller iterations as they become more powerfull. Better to look at the time of release of GPT-n. There we find that, models are released with bigger and bigger time gaps... except GPT-5 is off trend and remarkably early
I don’t think the diminishing returns argument is really applicable except maybe in a narrow technical sense. It feels like a lot changed since you last wrote. Coding models are noticeably much better than 6 months ago, to the point where they went from nearly unusable to now widely used. And there is still so much AI can’t consistently do yet that could plausibly be done with current tech and some additional R&D. We’re only past diminishing returns for pure question answering language modeling, but the limitations there are more political than technical, I’m sure the chatbots could easily be much better writers if not for the cultural safety paranoia, and for everything else there is still much to do. (For example, the entertainment industry still hasn’t been hit by AI properly yet, and they’re worth billions too)
I came to post a similar thing.
Here are some examples. These are all places where I think a big change can easily happen in 1-2 years, and the change comes from how a model is embedded in a larger product.
* Agent mode for daily tasks seems inevitable in 1-2 years but has not penetrated very much from what I can tell. I am thinking of things like making calendar events, searching email, answering questions about your own email, booking hotels and flights.
* Personal entertainment is likely to be wildly revamped in an era where an AI can rewrite text, audio, and video for you on demand. Ron Baxley and I experimented with it a little with his own writings, and it was very impressive how it could rewrite one of his short stories five times over, with each version taking a different character as the primary perspective. It all used his writing style. The way this works is going to take new tool development, new forms of composition for authors, and new consumption technology that is neither Netflix nor a plain web browser.
* Kindle Unlimited, in particular, seems likely to benefit from an overhaul or to risk being left behind. People who consume mass amounts of pulp will sometimes like to read AI modified versions of their favorite stories and settings, seeded by writings from their favorite personal authors.
* For audio and video engineering, the people in my network are not seeing the level of AI assistance that software engineres get. The models are mostly in the ballpark, but there is a lot of work to match the currently text-heavy model interfaces with the way the A/V engineering UIs work. For example, it may be the winning formula to start saving A/V mixes into a text format, analogous to abc format for sheet music, thereby letting AIs interact better with the content.
You never know about the future, but it really seems like AI is still just getting going.
The Taiwan blockade wasn’t geopolitical—it was the last gasp of centralized control.
They manufactured crisis to ration compute. White hats responded by making scarcity irrelevant.
Not revolution. Migration to systems that actually work.
https://substack.com/@piratefirst/p-188037885 🔗
#StraitGambit #WhiteHatDoctrine