I liked using the early models to do autocompletion. It could do a leetcode style thing, pretty nice, but only useful for small things.
Then I sought out Cursor because that seemed to be able to do multi-document edits. Not bad, but models at the time (2024) still got stuck pretty often. So, cross-document autocomplete. Useful, but definitely within the realm of "nice shortcuts to have".
Then a friend (who works in AI) told me to try Claude last year. I was on holiday at the time, but I spun up my work repo and looked at the backlog.
It chewed through the entire 6-9 months of estimated work in a two-week period while I was watching that Lord of the Rings series with a friend (we watched an episode or two in the evenings). I just chatted with him about the series while checking the progress every few minutes. It was a huge amount of refactoring, and it didn't get everything right the first time, but it made enough progress that it could be directed the right way.
Since then I have hardly coded any manual lines. I just tell Claude what to do, with very little harness (skills, MCPs, instruction files), and I get what I want.
It's because modern AI promises to relieve you of the tedium, leaving you to consider the important things like higher structure. It actually does deliver on this, but in contrast to older tools, it is unlimited in scope.
A calculator - let's expand this to maps, thesauruses, dictionaries, and other lookup tools - was used for a pretty narrow set of problems, and you had to transcribe the result to whatever context you needed.
An AI can be all the calculators together, and transport the output of one to the input of the next. You're meant to have the overview, but it's just so enticing to let it simply do that as well.
They mean using the AI to do the import things and the higher structure for them, not removing tedium. It's like using a calculator for math school/home-work designed to teach multiplication, which requires no transcription or putting into context. The tool isn't removing tedium. It's removing the very thing they're supposed to be learning.
And my question is, in the context of learning something, how in the world is having the AI do the thing you're supposed to be learning enticing unless you don't actually care about learning the thing
It's common, most of the people I know from the UK system did their PhD in 3-4 years.
In Europe you just study what it says as well. You happy to do a bachelor's in physics, your classes are all physics. You don't read shakespeare and learn french.
You can also do this in high school, so you can from age 16 be studying just physics and math.
I did none of my degrees in US, and my physics degree was 95% math and physics.
Physics degree is quite sequential anyway. You can't do QM in your first year or QFT in your second year.
I've checked random people I know from oxford and none started after 3-year undergrad and those how did after 4 all did 4 year dphil (small sample size warning). 4+4 is reasonable.
Maybe things have changed. We did QM in our first year at Imperial. I suppose we have to make allowances for Oxford. Got to fit the poetry in somewhere. =)
Obvious disclaimer: At this point we are talking about outlier colleges/universities.
But to give some examples, I know colleges (both in US and abroad) where people did real analysis and abstract algebra in their first year (and why not - neither requires prerequisites other than maturity).
I know a college (in the US) where they did Jackson for E&M in the 3rd year (and some advanced students did in the 2nd year). In most US universities, people normally do Jackson in the first year of their PhD.
I think it's rare to do QM before 2nd year, but in principle, as long as you know calculus/diff eq, you can get going on it. The catch is that the interesting applications require other branches of physics (e.g. E&M). When I did QM, all those applications were part of QM II anyway.
But yes, again, these are outliers and I wouldn't want to say it's the norm in the whole country.
I met a guy this happened to. He got a special award within the company, asked for a bit of equity, didn't get it, in fact got blacklisted and booted out.
If you want to make friends, water your friend seeds.
Everybody knows a bunch of people by name, and nothing else, from various contexts. You go to matriculation, there's a bunch of people introducing themselves, too many to get to know. You work a job, there's 50 people whose name you know. You go to a party, your friends introduce you to 10 new people, and you don't have time to talk to them all.
The ones you don't talk to much, they are your friend seeds.
You move to a new town, and you know nobody, other than that one guy you never spoke to after the first week of university. Contact that guy.
Could it be that what we called flow state was actually a sort of high level thinking time afforded by doing low level routine work?
For instance I'm the old world, if you wanted to change an interface, you might have to edit 5 or 6 files to add your new function in the implementations. This is pretty routine and you won't need to concentrate that much if you're used to it, so you can spend that low-effort time thinking about the bigger picture.
I was wondering if there's anything behind the idea that people who learned how to code before AI will become the human capital version of low-background steel.
Everyone who starts to code after AI has a problem: it's hard to believe you went through the pain and frustration that people often think is required to become a senior engineer. Even if you did, you are in a lemon market with quite a few people who took the shortcut in college. Much better to hire a guy who learned before they could cheat, and then give him the tools to replace the juniors.
I partially agree with this idea, but there will always be the Jeff Dean and Fabrice Bellard of the world... but 99% of companies won't ever get the chance to hire the top 1% of programmers. Therein is the problem. Maybe a better way to look at it is the statistical likelihood of producing good engineers and scientists goes down with AI because of poor fundamentals.
In SW this is perhaps the easiest domain to counterpunch. Get young folks learning computer history and understanding how the hardware works down to a register level. We write most software with some mental abstraction of what the hardware is actually doing. That's the crux, I believe, and if we lose widespread hardware understanding then we truly do become lost at sea, practicing the mystic art of non deterministic incantations
> How do you value people who learnt to code in the 80's, 90's or 2000's today?
Personally I rate them really really highly. They are always fascinating to talk to. But they also compete with newer cohorts who mature.
> Will new developers know/understand what they don't know, or will the new state of things simply become normalized?
Yes because a 32 year old guy with 10 years of experience who got given AI recently is going to be around for an awful long time reminding everyone that he has something the younger ones don't have.
Superficially, the article is right, intelligence services didn't get this wrong, and the administration made a bad decision despite having a good appraisal to hand.
But really, it's a values failure.
Wanting to make decisions that are good for America, and good for its friends, is a value. Putting people you are supposed to represent ahead of yourself used to be the kind of thing people would say mattered. It used to be a thing that leaders tried to demonstrate that they had carefully considered their decisions.
Once you have an administration that puts itself ahead of everything else, this whole thing makes sense.
This administration is full of insecure people who want to show how strong they are. You can see it in how they talk, and the constant stream of memes coming from the WH. It's incredibly juvenile, stuff like having Trump portrayed with a sixpack, beating up his enemies.
Strongman regimes have a tendency to try to steal the blind, to use a poker concept: bully the opposition into giving you a concession, by making super aggressive moves. Like picking pennies off a train track, most of the time you will win and the opponent will back down, EVEN if on paper the opponent tends to have the better cards, because a rational opponent will appreciate putting a lid on risk. This last bit is really important, because it means the bully learns that he can win despite rejecting advice.
So you can go around sucker punching people until it stops working, and there's a decent chance Iran is where it stops working. If it's not Iran, it will be the next thing, because they can't stop.
And to get back to values, too many Americans are unwilling to take responsibility for their country's actions. If you look at what causes discontent with the current Iran situation, it is things like gas prices. In other words, self-interest, still.
Cuba is already lined up. If they feel confident they would try on India because India often does not do what it is told. They have almost got that region under their thumb, except for India. Impressed by Srilanka though.
North Korea is another but I don't think they will dare to make that move.
I think this is being overstated by Indians who would like to think that India is more important to the US than it is; other than H1B discourse, I think the US has largely forgotten India exists.
Invading nuclear-armed India (from where?? Pakistan?) would be a completely insane thing even by Trump standards. It's a plan that disintegrates on contact with a map.
Not necessarily with invasion to start with. First would be destabilisation. It's neighbors are not doing too well lately. Many of them imploded within a short time span.
India can do what to the US with its nuke ? It's a deterrent for China.
India is weirdly more often forgotten in D.C., and almost never thought of as a threat, that framing is, as another commenter mentioned, a myth that largely propagates in India. It has recently only featured in Washington due to being a potential counterpoint to China, wherever that project is right now.
The same thing that France or Russia could do with their ballistic missile submarines. Just because the ICBMs won't reach the US doesn't mean that the ALCM and SLBMs are harmless.
They bought a few that are more silent, but their acoustic signature got acquired through intelligence/bribery operations. Quite an irreparable loss that the Indian population is not as acutely aware of.
One asset that India can threaten is Diego Garcia.
> One asset that India can threaten is Diego Garcia.
Case in point. I wouldn't expect a submarine to occupy the littorals of San Francisco, but an attack on Florida, Hawaii, or a distant base would be difficult to defend against.
Of course, such an attack is basically suicide, but still a possibility. Defensive systems like AEGIS are stretched too thin to deter a coastal attack.
You think so ? That is interesting. I am no expert though.
I expect Indian subs to be kept good track of. Threatening Florida seems a stretch. Assets in around Indian Ocean are a distinct possibility. In fact India does not need subs for that.
Everyone has forgotten ISRO. I don't see why they wouldn't be able to get a nuke into the US on a rocket, if for some reason they were mad enough to do so. But the US nuclear deterrent mostly makes that moot.
One thing that's an open pair of questions for me is: exactly how dangerous is a high-altitude EMP anyway? And do countries with nuclear weapons actually model this in war games?
I've seen it suggested that even a relatively modest HAEMP would be able to physically damage most transformers in the continental USA, necessitating replacement of all of them at the same time when they're all custom and have month-to-year lead times. No electricity, no fuel pumps, no refrigeration, in the US 90% of the population starves within a year.
The US power grid could have been defended in the decades since the March 1989 geomagnetic storm revealed such weakness, but given various evidence such as the fires in California caused by failure of century-old power lines that were well past due for repair, I doubt they have been.
I liked using the early models to do autocompletion. It could do a leetcode style thing, pretty nice, but only useful for small things.
Then I sought out Cursor because that seemed to be able to do multi-document edits. Not bad, but models at the time (2024) still got stuck pretty often. So, cross-document autocomplete. Useful, but definitely within the realm of "nice shortcuts to have".
Then a friend (who works in AI) told me to try Claude last year. I was on holiday at the time, but I spun up my work repo and looked at the backlog.
It chewed through the entire 6-9 months of estimated work in a two-week period while I was watching that Lord of the Rings series with a friend (we watched an episode or two in the evenings). I just chatted with him about the series while checking the progress every few minutes. It was a huge amount of refactoring, and it didn't get everything right the first time, but it made enough progress that it could be directed the right way.
Since then I have hardly coded any manual lines. I just tell Claude what to do, with very little harness (skills, MCPs, instruction files), and I get what I want.
reply