The intention of the Sullivan Doctrine was to keep Chinese players 1-2 generations behind in AI/ML applications, as this would give breathing room for the US, because AI/ML is essentially an HPC problem that can be solved by throwing a ton of capital and compute.

The issue is, a large subset of American "AI" startups are founded by people who are ideologically driven by an almost religious fervor around unlocking AGI and superintelligence.

On the other hand, most Chinese startups in the space are highly application driven and practical in nature - they aren't chasing "AGI" or "Superintelligence" but building a monetizable product and outcompeting American players. (P.S. Immigrant founded startups in the US approach the problem in the same manner)

I've said this a ton of times on here, but most American MLEs are basically SKLearn wrapper monkeys with a penchant for bad philosophy. It's hard to find MLEs at scale in the US who understand both how to derive a Restricted Boltzmann Machine as well as tune and optimize the Linux Kernel to optimize Infiniband interconnects in a GPU cluster.

Most CS and CE majors in the US who graduated in the past 7-10 years think less like engineers (let's build shit that works, and then build it at scale) and more like liberal arts majors but wanted to learn enough coding to pass leetcode medium and get a job - I've had new grad SWEs who are alumni of MIT caliber schools ask me about how to become a VC or PM and how I did the SWE-PM-VC transition because "they don't want to code". I was gobsmacked.

The same mindset occurs abroad as well in China, India, Eastern Europe, Israel, etc but at least they force students to actually learn foundations.

And if you look at most of the teams who lead or develop either GPU architecture, high performance networking, RL research, Ensemble learning research, etc - most did their undergrad abroad in China, India, or the CEE but their PhD in the US. The pipeline and skill at the junior level in the US is almost nonexistent outside a handful of good programs that are oversubscribed.

When (picking a random T10 CS program) Cal and UIUC CS tenure track professors are starting to take up faculty positions in China and India's equivalent of those CS departments, that means you have a problem.

CS is a god damn engineering disciple. Engineering is predicated on bridging the gap between theoretical research and practical applications, but I do not see any backing for this kind of mindset in most American programs.

And yes, the work ethic in the US leaves much to be desired. If Polish, Czech, and Israeli engineers will be fine working 50-60 hour weeks during crunch time, asking you to work earlier hours in order to accommodate your private commitments after 3PM is not some form of egregious abuse.

The American tech industry has become lazy, the same way the American automotive industry became lazy in the 90s and 2000s. The lack of vision and the pettiness amongst management and the lack of motivation amongst ICs who are amongst the highest paid in the world is not conducive if we want to retain a domestic tech industry.

And unlike the automotive industry of ye olde days, the tech industry being a services industry can and has begun moving P/L and product roadmap responsibilities along with the execs who own said responsibilities abroad. If the HQ is in the US, but all the decisions are made abroad, are you really an American company?

I concur with the outcome but not with the cause.

A big part of the problem is management at American firms. They are rarely, if ever, run by engineers at the helm. If you put arts and business majors in charge, it's no surprise that outputs look like art and business projects. These leaders pick people just like them at all tiers. Those who do boring and honest engineering work are shunned, excluded from promotions and left out of the leadership circle. It's little wonder that all of the real engineers depart for greener pastures.

Fix leadership and you will fix American industry.

Sklearn hasn’t been relevant for at least 5 years now. Anyone doing anything serious with it is committing malpractice since there are lots of far better, faster, alternatives to it. In particular, Nvidia RAPIDs cuML.

[dead]