Could fabs own AI?

post by lukehmiles (lcmgcd) · 2023-08-19T00:16:37.848Z · LW · GW · 0 comments

Contents

  Conclusion
None
No comments

Disclaimer: Very rough thoughts. I am nothing approaching an expert on this topic. Comments, corrections, and references appreciated.

Suppose there's an amazing general-purpose neural net architecture that learns as it runs and can be scaled indefinitely with returns that don't diminish too much. Suppose also that its design can't be improved by more than a few factors by anyone no matter how clever. (Humans match this description except that we are flesh and blood etc.)

Call this assumption the general online optimal scalable algorithm (GOOSA) assumption or scenario if you like.

Who will have majority control of deployments of this system a year after it's invented? Ten years after? ("Control" here means directly dictating the AI's actions. A service provider with a contract of uptime and privacy does not control the service unless they're able to break that contract.)

Some possible answers:

• The general public because they have the most of most stuff (ie this is the default assumption)
• An AI lab like openAI because they started in control
• A technology company like Google or Microsoft because they are a bottleneck to eyeballs and wallets of individual & business customers
• A government because they control the military etc
• A hardware IP company like Nvidia because they get an 80% discount on their own hardware and because they can block sales (or, increasingly, Nvidia data center access) to others
• A hardware manufacturing company like TSMC because they get a 50-75% discount on their own products

If the algorithm is successfully kept secret, then the AI lab could plausibly keep control from anyone except governments.

Suppose the basic idea of the algorithm leaks though. Who will end up owning its deployments?

Some angles to this question:

• Who is aware enough to notice the opportunity and agile enough to act?
• Who knows what to do with a bunch of AI?
• Who can sink in the most resources once everyone knows the game is on?
• Who can most effectively elbow out others?
• Who cannot be replaced or coerced?

Right now, the last question seems the most important to me. We've assumed that the algorithm is near optimal and the basic idea is leaked. It seems to me that fabs have the best position:

• Fabs vs military:
  • If US military tells TSMC that all the chips are now the US military's, I think TSMC might be able to say no. It's a very multinational company and the US wouldn't want to start a big conflict probably maybe.
  • Likewise for the Chinese military
  • If a military took over a fab would they know how to run it? Would they be able to make a new one?
• Fabs vs public:
  • Fabs quite lacking in the "purpose" dimension compared to the public: what would a fab do with all that AI anyway?
  • The public needs fabs for any chips and cannot replace or coerce them at all
  • Fabs seem capable of running datacenters themselves IMO, and forcing everyone else into a rental system
• Fabs vs labs:
  • If fabs were agile enough to notice and act, then I think they could make their own efficient implementation. Lots of good open source code to pull from.
  • Labs have made themselves irrelevant by developing an optimal algorithm and leaking it
• Fabs vs hardware IP:
  • The fab has all the blueprints!! Nvidia does not have TSMC's blueprints!
  • If the fab is smart then they'll just break the IP license and eat the cost of going to court. Or slightly tweak the design.
  • There's hundreds of small extremely-talented groups of chip designers all over the world [citation needed]
  • Assembly (Asus, Gigabyte, MSI, etc) seems replaceable
• Fabs vs giant tech corps:
  • Google can't decide on anything to do they like, so would probably be tied with a fab on the "purpose" dimension
  • People and companies actually do have their eyes open for new useful things, so the eyeball bottleneck is relatively weak sauce¿
• Fabs vs fabs:
  • Seems like the fabs' biggest strategic weakness to me. If the stakes were high and IP licensing was disregarded, then it seems like all of Intel, Qualcomm, Samsung, TSMC, etc could take a shot. Probably none are big enough to totally overwhelm two others.

Conclusion

In a GOOSA scenario, the only thing you need to get more intelligence is more computer chips (and electricity). The only ones who can make chips are fabs. No single military or other industry has apparent overwhelming leverage over (multinational) fabs. Fabs can make shitty apps, but an app company cannot build a shitty fab plant. Fabrication + assembly costs 10x less than what labs have to pay.

However, there are several extremely large fabs, so if they stole IP, then they might compete each other into commodityism. And they might not be able to think of anything good to do with the extra resources anyways.

So I have no idea who would win. It seems we are very early game still.

0 comments

Comments sorted by top scores.