4 ms·

One of my hopes is that when the AI bubble bursts, some brave person will sneak out a copy of the last frontier model.

by ThrowawayTestr 3mo ago

Aboutplants 3mo ago
Not worried about that, you will only have to wait 3-6 months and get a Chinese model just as good.
- yorwba 3mo ago
  Chinese companies giving away expensive models for free is a symptom of the AI bubble, too. It's not a law of nature that they'll always be able to scrounge up the money for yet another training run.
  - gpm 3mo ago
    Shaping the tool that does the thinking is quite valuable when you're in the business of changing how people think - I think we can expect propaganda agencies to be subsidizing model creation forever. This doesn't strike me as a symptom of a bubble - except in so far as the bubble pushes the competitors models forwards and thus they need to invest more to stay competitive.
    - rvnx 3mo ago
      All the models, have to respect their local laws, and most of all, pressure from users and the employees. They all carry political weights, because humans behind defend their interests, and are promoting some social values. https://pastebin.com/hjhvsBFg https://pastebin.com/hjhvsBFg This answer from Claude is so biased that it is ridiculous
  - nextos 3mo ago
    I think it's a deliberate business strategy of commoditization of their complement. China acts like an entire bloc, not as single companies, and they want to monetize hardware.
    - yorwba 3mo ago
      If you think Chinese companies always act as a bloc, your mental model needs to get about a billion times more detailed. But in this case just a few details may be enough: There are Chinese AI companies that have released LLMs without publishing the weights. ByteDance is going the direct-to-consumer route with their Doubao chatbot (the most popular in China, probably thanks to their social media prowess). iFlyTek seems to be angling for enterprise and government use cases, where they already have an in. The companies that have released weights have in common that they didn't have a monetization channel lined up and their models weren't good enough to make people pay attention with just API access. (You can see with Qwen Max that the calculus can change towards not releasing weights for better models.) And who exactly among the investors is having their complement commoditized? When Nvidia releases Nemotron, the story is clear, but it's less obvious for say Z.ai's GLM.
      - nextos 3mo ago
        I have never said they always act as a bloc, but their industry has a strong component of long-term strategic government planning behind them.
        inigyou 3mo ago
        > China acts like an entire bloc > I have never said they always act as a bloc Pick one
  - jnwatson 3mo ago
    As long as it is in the CCP's national interest to have a frontier model, Chinese companies will have the resources for another training run.
- sulam 3mo ago
  That’s misunderstanding why these models are behind. A large part of why they’re behind is they aren’t able to do the reinforcement learning post-training steps that takes a pre-trained model and turns it into a frontier model like GPT 5 or Opus. Instead they do their best to recreate these models using distillation. Fundamentally, you can never distill your way to being the teacher, so these approaches will not advance the frontier. [edit, after thinking about it I think my phrasing is unfair. It's not necessarily that aren't able to do it, but they haven't yet shown that they are willing to do it.]
  - computerex 3mo ago
    That’s not remotely true. They did distillation as a cheap solution to the cold start problem. You need data/trajectories to hill climb to higher capabilities. All large Chinese labs do RLAIF.
    - sulam 3mo ago
      Oh yes, not remotely true. Which is why the frontier labs all have invested heavily in trying to identify and thwart distillers, using known company names / domains to drive their exclusion lists. /s
      - logicchains 3mo ago
        It's cheaper to distill than to do reinforcement learning, so of course they prefer that, but if it wasn't an option they could just pay up and spend more GPU time on RL.
  - FpUser 3mo ago
    >"they aren’t able to do the reinforcement learning post-training steps" Not yet. If there is a need someone will come and fulfill. Personally for me now I do not even want to use top models. Professionally I use AI to help with the coding using Junie agent that comes with IDEs from JetBrains. Junie is told to use Gemini Flash and works fine for what I ("I" being an emphasis here) ask it to do. I tried more advanced models and different vendors only to discover credits going down the toilet without any extra benefit.
thx67 3mo ago
Prediction markets can solve this.
fastball 3mo ago
If it's a bubble, why do you care about frontier models?
- FpUser 3mo ago
  Internet was a bubble, so was telecom etc. at some point. Being bubble does not mean that when 90% of investments go down the drain the remains are not useful.
  - fastball 3mo ago
    "The Internet" was not a bubble. Companies with no long-term business model / sufficient product-market fit that were riding hype were the "dotcom bubble". But when those companies crashed, nobody said "I really want to get my hands on their IP", because it wasn't valuable – an important pre-requisite to the the bubble popping. Seems to be a different case here if people actually want the SOTA models.
    - A1kmm 3mo ago
      A bubble just means it is overvalued beyond it's true fundamentals due to speculation on speculation. The underlying asset can still have value, just less than the market price. Consider Cisco. On the 31st of March, 2000, it was valued at US$77.31 / share, which in inflation adjusted terms is $150.46 (above the current price over 26 years later). This valuation was on the basis of speculation that the price would continue to go up and Cisco would get a large cut of the industry profits. Cisco's business is still valuable, it was just treated as overvalued by the market. Similarly, if we go back to one of the classic examples of a bubble - consider Dutch tulip prices in 1636; speculation drove future contracts high. But tulips still have value to people today, it's just the price was higher than was sustainable.
  - mr_toad 3mo ago
    Railways were both a bubble and, eventually, one of the most significant technological innovations in history.
- 3mo ago
zuzululu 3mo ago
which will be very difficult to run unless you have a large budget to operate your own mini datacenter
- lelanthran 3mo ago
  In a crash the hardware will go for pennies on the dollar, if not for fractions of pennies on the dollar. Lots of companies will pick them up for scrap metal prices and host them for fractions of what we are paying today. That's the nature of bubbles.

One of my hopes is that when the AI bubble bursts, some brave person will sneak out a copy of the last frontier model.

Aboutplants 3mo ago

Not worried about that, you will only have to wait 3-6 months and get a Chinese model just as good.

yorwba 3mo ago
Chinese companies giving away expensive models for free is a symptom of the AI bubble, too. It's not a law of nature that they'll always be able to scrounge up the money for yet another training run.
- gpm 3mo ago
  Shaping the tool that does the thinking is quite valuable when you're in the business of changing how people think - I think we can expect propaganda agencies to be subsidizing model creation forever. This doesn't strike me as a symptom of a bubble - except in so far as the bubble pushes the competitors models forwards and thus they need to invest more to stay competitive.
  - rvnx 3mo ago
    All the models, have to respect their local laws, and most of all, pressure from users and the employees. They all carry political weights, because humans behind defend their interests, and are promoting some social values. https://pastebin.com/hjhvsBFg https://pastebin.com/hjhvsBFg This answer from Claude is so biased that it is ridiculous
- nextos 3mo ago
  I think it's a deliberate business strategy of commoditization of their complement. China acts like an entire bloc, not as single companies, and they want to monetize hardware.
  - yorwba 3mo ago
    If you think Chinese companies always act as a bloc, your mental model needs to get about a billion times more detailed. But in this case just a few details may be enough: There are Chinese AI companies that have released LLMs without publishing the weights. ByteDance is going the direct-to-consumer route with their Doubao chatbot (the most popular in China, probably thanks to their social media prowess). iFlyTek seems to be angling for enterprise and government use cases, where they already have an in. The companies that have released weights have in common that they didn't have a monetization channel lined up and their models weren't good enough to make people pay attention with just API access. (You can see with Qwen Max that the calculus can change towards not releasing weights for better models.) And who exactly among the investors is having their complement commoditized? When Nvidia releases Nemotron, the story is clear, but it's less obvious for say Z.ai's GLM.
    - nextos 3mo ago
      I have never said they always act as a bloc, but their industry has a strong component of long-term strategic government planning behind them.
      - inigyou 3mo ago
        > China acts like an entire bloc > I have never said they always act as a bloc Pick one
- jnwatson 3mo ago
  As long as it is in the CCP's national interest to have a frontier model, Chinese companies will have the resources for another training run.
sulam 3mo ago
That’s misunderstanding why these models are behind. A large part of why they’re behind is they aren’t able to do the reinforcement learning post-training steps that takes a pre-trained model and turns it into a frontier model like GPT 5 or Opus. Instead they do their best to recreate these models using distillation. Fundamentally, you can never distill your way to being the teacher, so these approaches will not advance the frontier. [edit, after thinking about it I think my phrasing is unfair. It's not necessarily that aren't able to do it, but they haven't yet shown that they are willing to do it.]
- computerex 3mo ago
  That’s not remotely true. They did distillation as a cheap solution to the cold start problem. You need data/trajectories to hill climb to higher capabilities. All large Chinese labs do RLAIF.
  - sulam 3mo ago
    Oh yes, not remotely true. Which is why the frontier labs all have invested heavily in trying to identify and thwart distillers, using known company names / domains to drive their exclusion lists. /s
    - logicchains 3mo ago
      It's cheaper to distill than to do reinforcement learning, so of course they prefer that, but if it wasn't an option they could just pay up and spend more GPU time on RL.
- FpUser 3mo ago
  >"they aren’t able to do the reinforcement learning post-training steps" Not yet. If there is a need someone will come and fulfill. Personally for me now I do not even want to use top models. Professionally I use AI to help with the coding using Junie agent that comes with IDEs from JetBrains. Junie is told to use Gemini Flash and works fine for what I ("I" being an emphasis here) ask it to do. I tried more advanced models and different vendors only to discover credits going down the toilet without any extra benefit.

thx67 3mo ago

Prediction markets can solve this.

fastball 3mo ago

If it's a bubble, why do you care about frontier models?

FpUser 3mo ago
Internet was a bubble, so was telecom etc. at some point. Being bubble does not mean that when 90% of investments go down the drain the remains are not useful.
- fastball 3mo ago
  "The Internet" was not a bubble. Companies with no long-term business model / sufficient product-market fit that were riding hype were the "dotcom bubble". But when those companies crashed, nobody said "I really want to get my hands on their IP", because it wasn't valuable – an important pre-requisite to the the bubble popping. Seems to be a different case here if people actually want the SOTA models.
  - A1kmm 3mo ago
    A bubble just means it is overvalued beyond it's true fundamentals due to speculation on speculation. The underlying asset can still have value, just less than the market price. Consider Cisco. On the 31st of March, 2000, it was valued at US$77.31 / share, which in inflation adjusted terms is $150.46 (above the current price over 26 years later). This valuation was on the basis of speculation that the price would continue to go up and Cisco would get a large cut of the industry profits. Cisco's business is still valuable, it was just treated as overvalued by the market. Similarly, if we go back to one of the classic examples of a bubble - consider Dutch tulip prices in 1636; speculation drove future contracts high. But tulips still have value to people today, it's just the price was higher than was sustainable.
- mr_toad 3mo ago
  Railways were both a bubble and, eventually, one of the most significant technological innovations in history.
3mo ago

zuzululu 3mo ago

which will be very difficult to run unless you have a large budget to operate your own mini datacenter

lelanthran 3mo ago
In a crash the hardware will go for pennies on the dollar, if not for fractions of pennies on the dollar. Lots of companies will pick them up for scrap metal prices and host them for fractions of what we are paying today. That's the nature of bubbles.