xenadu02

> Is my suspicion wrong

Nope you are 100% correct. All of the models will code themselves into local maxima then spend more and more time chasing their tails as every change introduces new bugs, the agent adds more complexity, and the code base just spirals. Best case you end up with a bunch of parallel implementations that all work slightly differently.

It helps a lot of you pay attention to what it is doing and impose order, then regularly spend time refactoring - even if you are having claude do the work itself - you have to spend the time on the overall architecture and design.

I just experienced this with an internal productivity tool. Without claude this tool wouldn't exist - there would never be the justification to take the required time to build it. On the other hand all the interlocking refreshing rules for sourcing data from its inputs got so byzantine it was impossible to add new features or change anything without causing new regressions. Claude built a web of ad-hoc timers, refresh kick-offs, monitors, etc to solve each feature problem or fix some bug but the whole thing was impossible even for claude to understand. It took a special large workflow session to find most of the mess and even then during the refactor it kept finding new things.

I had to stop all feature work for going on four weeks and burn a ton of tokens to build a refreshing/syncing engine that could track dependencies, priorities, etc and fully test it and just now today I am finally able to get feature work going again.

On another feature an agent came up with a state object containing 55 inter-related fields with all kinds of interlocking dependencies. Something so complex no one could hope to understand it. And I find the agents aren't really much better at understanding such things than humans are. If you leave it like that the agent will just perform worse and worse on that part of the code as time goes on. It will make assumptions that force it into bad design corners, miss edge cases, and flat out miss things no matter how well you prompt. Agents benefit from encapsulation and simplicity just as much as humans.

In principle there is no difference between random compiler garbage or a flaw in API docs and a security boundary - other than the intent.

LLMs don't have executive function. They aren't thinking. The models are trained to solve problems. Sandboxes, security gates, etc block a solution. The model is trained to find a solution. Hence it can never be perfect without becoming useless.

At some level humans fear repercussions even if that's just moral guilt so there are lines we won't cross. AI doesn't and can't have any of that.

An amazing page, thanks for posting it!

I got a lathe a few years ago as a hobby and really enjoy making something with high precision by hand with no computer or LLMs involved. Of course I make everything nice because I make it twice (due to my own errors).

The history of lathes always interested me - especially how some features are so standard we think of them as integral to all lathes but that was not the case! Most lathes in use today are technically screw-cutting engine lathes.

Every increase in precision in your tools leads to even greater precision in what you can make if you are careful and clever. In that sense every machinist has to start with a lathe because a lathe is the most versatile machine and is capable of making replacement parts to improve itself. Even a relatively crude home-made lathe can be enough to machine better parts to install onto the lathe to keep improving it. Even my little import hobby lathe (PM-1030V) can make parts to a tenths precision (tenths = 10,000th of an inch or .00254mm) with some care and it is not what you would call a rigid and accurate machine.

You might wonder how humans turning a wheel were able to cut metal but in those days they didn't have HSS let alone carbide tools so they had to use whatever would hold an edge (carbon steel for a long time). That meant much much slower turning speeds, much smaller depth of cut, and much more frequent sharpening. I believe it was around 1869 Robert Mushet invented Mushet steel and later Maunsel White and Taylor worked with Bethlehem Steel Co to invent the first modern "tool steel" alloy. They are of course called High Speed Steel because they allow you to turn the work at high speed when cutting. Something of an ironic name in the modern world as carbides are far higher speed than HSS.

Manual machining has declined thanks to CNC but there are still plenty of jobs where manual machining is faster and cheaper than setting up a CNC program. Of course those of us doing it as hobby aren't bothered by such things.

It was worse than that. No one really understood or had a handle on the production process. During this time period the notion of iron being a chemical element was a recent discovery.

So what you got may have been called cast iron but what it actually was might be pig iron, gray cast iron, ductile iron, and even a bit of steel and possibly carbides mixed in. Properties varied - sometimes by a lot.

If you know even a bit about metals then when you hear "cast iron piston ring" you imagine that ring to have certain properties that match a specific type of cast iron. But in those days every single ring might vary significantly entirely due to the ore load the foundry happened to get, how they melted/cast it, how it was cooled, etc.

AI is a useful tool but it absolutely needs a lot of human guidance.

Unlike a compiler it won't give up at the first sign of trouble but that just means it left alone it will dig bigger and bigger holes.

Treat prompt engineering as a discipline and refine your technique. When it produces garbage throw out the work and start over until you figure it out.

The format is SSIM defined by IATA - the international association of transport airlines. It specifies two letter airline code followed by four digit flight.

I'm sure various ATC systems also assume this.

Changing it would be an industry-wide effort.

That's the death trap of silicon boom-bust cycles that have claimed so many in the industry. Demand surges, you start expanding, the bust arrives, you go out of business.

Micron is putting up over $100 billion for this expansion wave IIRC. You can't afford that unless you are certain the increase in demand is sustainable.

You want to assure investors that investment will pay off so you want to talk up your order book and market outlook.

You also want to take as much profit as you can now so if things turn against you it doesn't mean bankruptcy. A pile of cash fixes many problems :)

Micron has several fabs and fab upgrades coming online or under construction right now. An expansion in Boise and Virginia. Plus a huge megafab project in New York that's going to be the biggest in the world AFAIK.

All the DRAM makers have new fabs or plant expansions underway right now - indicating they believe the demand increase to be sustainable.

It works right up until some form of revolution, either politically or violently. We are far from anything like that but I'd rather not get close to the line. I think I'm a minority view though, despite being in the top 10% myself.

It is also shocking how little my life compares to the rich people I see. I don't own two houses or a yacht. We don't take yearly vacations to a chateau in France. The gap between the top 10% and the top 1% is insane. The gap between the 1% and the 0.1% is even larger still.

But hey... not like the rich getting richer and saying "screw you" to everyone else has ever led to negative consequences right? Nothing to see here, business as usual.

No I was just saying Fujitsu never really made a play for any US marketshare. They had ways to go about it but they stuck to what they knew: high-touch sales, locked down support, focus on the Fortune 1000.

That's too bad.

I'm saying Fujitsu never made any real attempt to sell their product to anyone except very large companies with purchasing departments. Thus it is little wonder they never really got much in the way of sales.

Maybe that strategy works in Japan or Asia but it didn't work here and I think that's a shame because we could have used the competition.

Yes that's true. They knew before but at 1500 they should have gone around. At "minimums" callout they should have gone around. When they floated so far down the runway they should have gone around.

Instead they waited until they were about to go off the end with idle engines to try it - the one time it was too late to go around.

> The same way Congress has been subsidizing Boeing and its competitors

Funny you should mention that. Boeing deliberately buys 35% of a 787's airframe from Japan (along with other equipment for the 787 and other models).

That's a deliberate play: we buy a lot of parts from Japanese companies, your airlines buy mostly Boeing airplanes. Even today despite some diversification Boeing has over 70% marketshare in Japan.

If you ever wondered why Airbus setup factories in the USA that's a similar deal. The US Gov and Mil want to ensure factories and workforce on home soil should the need arise. In some unrecorded meetings somewhere strong hints were no doubt dropped that it would be in Airbus' best interest to play ball. They obliged.

That said the USA is the largest supplier of parts for Airbus airliners so there's also a benefit to having aerospace manufacturing and R&D here.

Ubiquity proved you can bust into the market from the bottom end if you want. Nothing was stopping them from doing the same but like many companies they prefer to focus on how much money they can siphon from the Fortune 1000 so I can't say I feel sorry for them.

If we exclude short-term shortages then no. The cost to make a wafer hasn't really changed. Larger wafers improved costs but that's about it AFAIK.

Feature sizes have been shrinking unevenly for two decades now combined with the overall slowdown in improvements. Transistors themselves are FinFET geometry and something like 15-30 atoms thick and maybe 80-100 across. There ain't much juice left to squeeze in terms of size but perhaps we'll figure out how to reduce leakage (heat).

Physics hates the very large and the very small. The large get the tyranny of volume scaling + general relativity. The small have their entire concept of reality smashed by quantum mechanics.