@NewCharlesInCharge's banner p

NewCharlesInCharge


				

				

				
0 followers   follows 0 users  
joined 2022 September 04 19:09:11 UTC

				

User ID: 89

NewCharlesInCharge


				
				
				

				
0 followers   follows 0 users   joined 2022 September 04 19:09:11 UTC

					

No bio...


					

User ID: 89

Taiwan’s relationship with the mainland is complicated.

For example: https://esc.nccu.edu.tw/PageDoc/Detail?fid=7801&id=6963

Some flavor of “maintain status quo” dominates opinions about mainland relations, with “indefinitely” being the most popular.

About a quarter of people favor independence either immediately or in the longer term.

Under Huang's leadership, Nvidia was circumventing export controls by selling high-end GPUs to Singapore under circumstances where they were obviously being diverted to China. Either Trump, or Lutnick with Trump acquiescing, also knew and had decided to tolerate this.

I still haven't seen any evidence that forbidden GPUs ever made it to the mainland.

It wouldn't even make much sense to do such a thing given the scale they've purportedly been operating at. They're not running massive clusters, any neocloud could have supplied their needs. You just need a middleman who passes KYC to rent GPUs for you from cloud providers. That's both cheaper and less risky than a smuggling operation.

I think software engineers are getting an intuitive feel for this. For most projects now it is me, the human, that is the bottleneck. And the kinds of escalations I get from agents increasingly are of the sort that could be handled by other agents. Just this morning I awoke to an agent that had halted progress because it misunderstood an API while we were planning the work earlier and was basically asking permission to do the right thing.

Software engineering has this concept, "the mythical man month," which states that you can't simply add more people to a project to accelerate it. At some point coordination costs begin to dominate and you start to slow down instead of speed up. But coordination costs among agents are cheap. You can parallelize work to fit the speed at which you expect it to take. There's no ramp up time, you can understand as much of the project as needed to complete your task pretty much instantaneously.

So we're going to get to this point where throwing more compute to the RSI coordinator will be put to efficient use and speed up the creation of the next model.

I'm at a big tech company grasping at the frontier.

For the past, I don't know, six months, code review has been mostly a waste of time. Agents look at commit summaries and detect if the stated intent is actually achieved in the code. They check for bugs, style, and anything that a project maintainer might care about that has been shoved into a nearby markdown file. This system is so good now that most of my changes don't actually require another engineer to review them, they can be AI-approved and shipped.

I'm recently on a project where that's not completely the case, but it's the kind of project you'd point to when arguing that actually leetcode problems really are relevant to some jobs. But even then the agents still almost never make a mistake, likely they're doing better than the humans.