InkCanon

This, there's also another kind of "shoring" where people are imported and given salaries at the bare minimum to qualify for H1B. As per my other post, the net amount is staggering and no where near the supposed 65k cap. My own right estimates put it at ~600k annually.

These jobs are being offshored to India. You can tell by how they're massively hiring there.

Google launches largest office in India https://www.entrepreneur.com/en-in/news-and-trends/google-la...

Microsoft India head says no layoffs in India https://timesofindia.indiatimes.com/technology/tech-news/mic...

The situation now is radically different. In the late 2000s/early 2010s, most H1B applications got approved (something like 80%). For a time before that, the H1B cap wasn't usually maxed out and any application would succeed. If I'm being blunt this is because of one country. H1Bs are dominated by this country. This country makes up 75% of applications and without these, H1Bs would actually be undersubscribed.

The significant shift comes a lot from how this country has massive systems in place to perform wage arbitrage through IT consultancies. Compared to Chinese industrial outsourcing (which requires capex), wage arbitrage is pure profit in that there's almost no overhead. So these IT companies got phenomenally rich. These companies have a US branch, usually having a manager based in the US while the others are based in India. So it's no longer getting the best and brightest through H1B, but just a way to make money off the vast difference in economic conditions between a third and first world country. And there's a direct incentive to depress the economic conditions of workers, because that's money right there. Then this goes into overdrive when many US companies realize it's even cheaper to do it themselves and set up shop in this country.

What happened to the US industrial base/blue collar workers is happening right now to white collar workers, except it'll go much faster because there's no physical equipment to move.

If you look at the numbers, tech layoffs are really offshoring. India is not only untouched by layoffs, they're actually hiring

https://www.thehindu.com/sci-tech/technology/microsoft-confi....

A more practical change would be bidding. It would be extremely easy to turn it into a system where the X highest salaries get the visa.

If you actually dig into the numbers around immigration labour, you'll be shocked. The 65k cap is completely irrelevant. The actual cap is often (legally) unlimited.

1) +20k for those with advanced degrees

2) +80k L-1 visas, these are international transfer visas heavily used by WITCH companies. These are unlimited.

3) +160k OPT visas, these are 3 year work visas for international students who took a masters or PhD in the USA. This in particular is unlimited and has been growing at ~20% annually compounded for some time

4) +150k H4 visas, which is a dependent visa for immigrants who have some other visa

5) +65k F4 visas, which are for siblings of immigrants

Net is several hundreds of thousands of white collar workers enter the USA every year. This is an absolutely astounding amount, especially in proportion to how often it's talked about.

There was an interesting post here awhile back about autonomy and motivation. The gist was people's motivation is proportional to their autonomy. This is quite intuitive, you can see people are really motivated when they have autonomy (think kids with Minecraft, musicians with instruments). One terrible thing about Anki is that it probably is horrible for autonomy. Quite possibly using anki actually has a negative effect on motivation.

It's someone I wondered, what is the point of memorizing a proof if it only ever proves something you already know. The answer is you hope it generalises. There is a possible way you can do it in SRS, being inspired by RL training. Instead of cards you'd show options within a game or simulation. But this would need a lot of expert knowledge for a single concept.

I think the difference in recall- knowledgeable and logical-model-knowledge will be really interesting. LLMs appear to strongly be the first. But this is very hopeless on mathematics.

This is a very big problem. Virtually all the results from research here comes from some form of simple word recall. Direct recall occupies some part of real world tasks, but IRL if you're stopped by doing something it's people not because you can't remember it (and you could look it up if you forgot).

There's some UX problems of SRS (that I'm working on) that makes it high friction 1) Time taken to create cards 2) Need for self marking 3) Creates a one to one mapping of prompt-answer 4) If you're an autodidact, you have to teach yourself first (alternatively called understanding, scaffolding, etc)

More fundamentally, SRS isn't a superpower because it's just very specific to creating a direct prompt retrieval. Generalization is poor. Even creating a graph of knowledge, is a chain of edges between bits of knowledge, isn't done very well here.

And I suspect there's a very deep, fundamental difference between recollection knowledge and logical-modeling knowledge. Recollection seems very similar to a dictionary access, and if you recorded the time to recall in humans I suspect they'd all be constant. But learning the knowledge of a logical model, like of a mathematical concept, appears to be vastly different and have very different time to compute.

Proponents of SRS will point out logical models need facts as well, like formulas, lemmas, etc. Which is true. But if you already grasped it before you'd grasp it faster the second time. So the practical use of SRS is a significant step above having a very well sorted and labeled notebook, but still way below becoming a genius.

It's a common idea, all the way back to Hoare logic. There was a time when people believed in the future, people would write specifications instead of code.

The problem with it takes several times more effort to verify code than to write it. This makes intuitive sense if you consider that the search space for the properties of code is much larger than the code for space. Rice theorem's states that all non trivial semantic properties of a program are undeniable.

I think if your assumption is that AI is deducing where it is with rational thoughts, you would be. In truth what probably happened is that the significant majority of digital images of the world had been scraped, labeled and used as training data.

Pretty rich coming from a company that's not-so-slowly outsourcing it's workforce to India.

Strongly suspect OAI can't afford 20B cash. Their latest funding round was 40B, and they're burning through money like it's rice paper. They could offer OAI equity, but Cursor's founders would probably be very suspicious of private valued stock (which is fairy money).

How wise it is to buy Cursor is another question. Current valuation has them at 100x revenue. And I suspect agentic products will be a lot less cash flow positive than traditional SaaS because of the massive cost of all that constant codebase context and stream of code.

Some of my relatives and colleagues actually actively encourage this. They give them an iPad with YouTube on it after meals and so on. It acts as a pacifier.

Cool post!

The free market (which I think people also include in capitalism) would correctly predict labour intensive jobs would be outsourced. This is very much a feature (comparative advantage), not a bug. I realized a lot of supposedly free market people don't even know the basics of it. Politically the free market has become an identity associated with national greatness and a sense of control of ones destiny. The dominant feeling seems to be if you have a free market, you will win everything (which is actually opposite from the truth).

>4.1 Was better in 55% of cases

Um, isn't that just a fancy way of saying it is slightly better

>Score of 6.81 against 6.66

So very slightly better

I think it is largely nature. Linus has great ability to

1) Identify and resonate deeply with developer issues (like Git, Linux)

2) Focus on executing a usable product in a very short time

So combined it's a really potent ability. If he had the personality he'd probably be a really great founder, but he has chosen to open source all his stuff.

I believe it's some bizarre ploy so his supporters will say, "You should listen to Trump!"

Seeing how a lot of the people in SV tech, there is an uncanny resemblance to aristocracies like the English or old money WASP types. They go to certain exclusive schools, they work at certain companies, have certain rituals, have a "type" of social sphere, have a certain political ideology, etc.

The second one is so delusional. It assumes you can build a moat for all manufactured products. This is true in extremely specific, high value cases like lithography machines or chips. But when you talk about screws, glue, plastic bits and what makes up 90% of manufacturing, there is no moat. You're not going to build a monopoly on screws.

Hi, I write as a hobby and really like Sudowrite. There's a huge gap between it and virtually every other AI writing tool I know of. The insight that writers:

A) Largely only want AI when they are blocked, and not all the time B) Want to consider options (which is how writing happens all the time, IMO)

Is really what sets your product apart. So I'm curious, how did you get these insights? Were you a writer and instinctively knew of these, and so you dogfooded your own product? Or did you do a YC style feedeback loop to writers to find this differentiator?

o1 reportedly got 83% on IMO, and 89th percentile on Codeforces.

https://openai.com/index/learning-to-reason-with-llms/

The paper tested it on o1-pro as well. Correct me if I'm getting some versioning mixed up here.

That's fair. But look up the recent experiment on SOTA models on the then just released USAMO 2025 questions. Highest score was 5%, supposedly SOTA last year was IMO silver level. There could be some methodological differences - ie USAMO paper required correct proofs and not just numerical answers. But it really strongly suggests even within limited domains, it's cheating. I'd wager a significant amount that if you tested SOTA models on a new ICPC set of questions, actual performance would be far, far worse than their supposed benchmarks.

People are really fundamentally asking two different questions when they talk about AI "importance": AI's utility and AI's "intelligence". There's a careful line between both.

1) AI undoubtedly has utility. In many agentic uses, it has very significant utility. There's absolute utility and perceived utility, which is more of user experience. In absolute utility, it is likely git is the single most game changing piece of software there is. It is likely git has saved some ten, maybe eleven digit number in engineer hours times salary in how it enables massive teams to work together in very seamless ways. In user experience, AI is amazing because it can generate so much so quickly. But it is very far from an engineer. For example, recently I tried to use cursor to bootstrap a website in NextJS for me. It produced errors it could not fix, and each rewrite seemed to dig it deeper into its own hole. The reasons were quite obvious. A lot of it had to do with NextJS 15 and the breaking changes it introduces in cookies and auth. It's quite clear if you have masses of NextJS code, which disproportionately is older versions, but none labeled well with versions, it messes up the LLM. Eventually I scrapped what it wrote and did it myself. I don't mean to use this anecdote to say LLMs are useless, but they have pretty clear limitations. They work well on problems with massive data (like front end) and don't require much principled understanding (like understanding how NextJS 15 would break so and so's auth). Another example of this is when I tried to use it to generate flags for a V8 build, it failed horribly and would simply hallucinate flags all the time. This seemed very likely to be (despite the existence of a list of V8 flags online) that many flags had very close representations in vector embeddings, and that there was almost close to zero data/detailed examples on their use.

2) In the more theoretical side, the performance of LLMs on benchmarks (claiming to be elite IMO solvers, competitive programming solvers) have become incredibly suspicious. When the new USAMO 2025 was released, the highest score was 5%, despite claims a year ago that SOTA when was at least a silver IMO. This is against the backdrop of exponential compute and data being fed in. Combined with apparently diminishing returns, this suggests that the gains from that are running really thin.

Frankly the overarching story about evals (which receives very little coverage) is how much gaming is going on. On the recent USAMO 2025, SOTA models scored 5%, despite claiming silver/gold in IMOs. And ARC-AGI: one very easy way to "solve" it is to generate masses of synthetic examples by extrapolating the basic rules of ARC AGI questions and train it on that.

The biggest story in AI was released a few weeks ago but was given little attention: on the recent USAMO, SOTA models scored on average 5% (IIRC, it was some abysmal number). This is despite them supposedly having gotten 50%, 60% etc performance on IMO questions. This massively suggests AI models simply remember the past results, instead of actually solving these questions. I'm incredibly surprised no one mentions this, but it's ridiculous that these companies never tell us what (if any) efforts have been made to remove test data (IMO, ICPC, etc) from train data.

Surprised at how it hasn't been pointed out here but - the "general public" wants the sausage, but not how it's made. They wouldn't if they knew what it entailed. Cutbacks to student aid, shuttering of departments, eliminating of PhD positions, etc.