4 ms·

> OAI's deal with Samsung & SK Hynix is also colossal - locking up roughly 900,000 DRAM wafers per month, roughly 40% of world output (Stargate project). I won

by josephg 16d ago

> OAI's deal with Samsung & SK Hynix is also colossal - locking up roughly 900,000 DRAM wafers per month, roughly 40% of world output (Stargate project). I wonder if they'll, at some point, have enough RAM? Or is this is the new normal? Will models keep scaling with the amount of ram chips openai and anthropic own?

bionhoward 16d ago
Even with efficiency breakthroughs, it would just afford packing more agents per unit of memory. Scaling compute and data keeps paying off, leading to smarter models, and smarter models have more demand even at higher prices because they can accomplish more work at higher quality. Sovereign AI hasn’t even really taken off yet to anywhere near the level it could. That’s going to dramatically increase the number of massive-scale users of AI agents. So no, IMHO they will “never” have “enough.”
- epicureanideal 16d ago
  At some point though, won't someone be able to extrapolate the demand growth curve, and invest some colossal amount of money into making and selling more RAM chips?
  - fragmede 16d ago
    Yes. China's done just that. Expect those factories to come online within 2-3 years.
- cyanydeez 16d ago
  Weve hit the sigmoid. Whats scalling is ancillary to the model. The cry for a slowdown is because the open weight models demonstrate the cost of parameter pacling is not work neither inference nor training. The assumption about the singularity simply is a delusion with LLMs. However, the models do provide a means to improve the harness universe, so that residual will continue to improve perception. Parameter cpunt will stagnate and training wont be justifiable from every angle.
ohyes 16d ago
Hard to know, does each GB of ram give some marginal increase in profit or potential profit? I’d guess no. Past a certain point the model has all the capabilities it can possibly usefully offer and honestly we may already be past that. The next gen model just doesn’t seem like as clear a step up as it once was.
- Leonard_of_Q 16d ago
  That point is said to lie somewhere around 640 KB if I recall correctly. https://skeptics.stackexchange.com/questions/2863/did-bill-gates-say-640k-ought-to-be-enough-for-everyone https://skeptics.stackexchange.com/questions/2863/did-bill-g...
- josephg 16d ago
  LLMs still seem pretty bad at writing large scale software like web browsers. Though it’s probably a problem of managing large context windows more than anything. Not sure if larger models will magically overcome that.
  - cyanydeez 15d ago
    The LLM by itself will never create software of any nontrivial (training) complexity. The harness though will improve while parameter count stagnates. The Qwen3.8 models are strong enough when given proper context.
    - josephg 15d ago
      > The LLM by itself will never create software of any nontrivial (training) complexity. Huh? I'm not sure what the word "training" does in that sentence. But "never" my arse. Frontier models can make nontrivial software already. For example, the other day I asked fable to reverse engineer the satisfactory blueprint file format. Then write a program to read the logistic flow graph in a blueprint. Then make an auditing tool that can analyse the graph to find problems. Well, it totally knocked it out of the park: https://seph.au/blueprints/#bp=0%3Aalumina.sbp https://seph.au/blueprints/#bp=0%3Aalumina.sbp This is a relatively small program, but it's not trivial. I'd consider a trivial program to be something I could code up in 20 minutes. It would have taken me a couple weeks to make this blueprint auditing tool, including reverse engineering the file format, writing the analysis code, making the website, scraping all the in-game data on available recipes and icons and so on. I've got a lot of mixed feelings about LLMs. But it seems very silly to lie about what they're capable of.

by josephg 16d ago

bionhoward 16d ago

Even with efficiency breakthroughs, it would just afford packing more agents per unit of memory. Scaling compute and data keeps paying off, leading to smarter models, and smarter models have more demand even at higher prices because they can accomplish more work at higher quality. Sovereign AI hasn’t even really taken off yet to anywhere near the level it could. That’s going to dramatically increase the number of massive-scale users of AI agents. So no, IMHO they will “never” have “enough.”

epicureanideal 16d ago
At some point though, won't someone be able to extrapolate the demand growth curve, and invest some colossal amount of money into making and selling more RAM chips?
- fragmede 16d ago
  Yes. China's done just that. Expect those factories to come online within 2-3 years.
cyanydeez 16d ago
Weve hit the sigmoid. Whats scalling is ancillary to the model. The cry for a slowdown is because the open weight models demonstrate the cost of parameter pacling is not work neither inference nor training. The assumption about the singularity simply is a delusion with LLMs. However, the models do provide a means to improve the harness universe, so that residual will continue to improve perception. Parameter cpunt will stagnate and training wont be justifiable from every angle.

ohyes 16d ago

Hard to know, does each GB of ram give some marginal increase in profit or potential profit? I’d guess no. Past a certain point the model has all the capabilities it can possibly usefully offer and honestly we may already be past that. The next gen model just doesn’t seem like as clear a step up as it once was.

Leonard_of_Q 16d ago
That point is said to lie somewhere around 640 KB if I recall correctly. https://skeptics.stackexchange.com/questions/2863/did-bill-gates-say-640k-ought-to-be-enough-for-everyone https://skeptics.stackexchange.com/questions/2863/did-bill-g...
josephg 16d ago
LLMs still seem pretty bad at writing large scale software like web browsers. Though it’s probably a problem of managing large context windows more than anything. Not sure if larger models will magically overcome that.
- cyanydeez 15d ago
  The LLM by itself will never create software of any nontrivial (training) complexity. The harness though will improve while parameter count stagnates. The Qwen3.8 models are strong enough when given proper context.
  - josephg 15d ago
    > The LLM by itself will never create software of any nontrivial (training) complexity. Huh? I'm not sure what the word "training" does in that sentence. But "never" my arse. Frontier models can make nontrivial software already. For example, the other day I asked fable to reverse engineer the satisfactory blueprint file format. Then write a program to read the logistic flow graph in a blueprint. Then make an auditing tool that can analyse the graph to find problems. Well, it totally knocked it out of the park: https://seph.au/blueprints/#bp=0%3Aalumina.sbp https://seph.au/blueprints/#bp=0%3Aalumina.sbp This is a relatively small program, but it's not trivial. I'd consider a trivial program to be something I could code up in 20 minutes. It would have taken me a couple weeks to make this blueprint auditing tool, including reverse engineering the file format, writing the analysis code, making the website, scraping all the in-game data on available recipes and icons and so on. I've got a lot of mixed feelings about LLMs. But it seems very silly to lie about what they're capable of.