Where I find current-generation AI helpful in writing code is not in writing it so much as modifying it. It's especially helpful when you decide to make some change that requires updating dozens of lines of code over several files. Sometimes such changes can be performed by a simple search and replace, but often you have to examine and edit each one individually. It's tremendously helpful to be able to tell the LLM to go find all the places a change is required and make it. You still have to look at each edit performed by the AI, but nearly all of them are usually right, and this takes a fraction of the time.

Another way AI is useful to me is due to my particular context: Android (the OS, not an app). Android builds are slow. Even incremental builds that don't touch any "Android.bp" file (a Makefile, basically) take 2-3 minutes, minimum, because that's how long it takes the build system to determine that only the one file you touched needs to be rebuilt. Anyway, this creates a situation where the typical edit/compile/test cycle takes several minutes, most of which is just waiting for the machine. If you touch an Android.bp, it's more like 8 minutes. If you context switch during that wait, to read email or edit a doc or whatever, the context switch overhead begins to kick your butt.

But what I've found is that I can give the AI a task, like "Write unit tests for feature X", then start the build/test and context-switch away. The code the AI wrote won't compile or work, but that's fine (and also true of my own code, which rarely builds and executes perfectly the first time). When the build/test run is complete I don't mentally context-switch back into the coding task, I just copy-paste the output to the AI, which will make some changes (I don't bother looking at what), and I start the cycle again. After a half-dozen iterations of this (~20 minutes), the AI will have something that builds and passes, and then I actually switch back to see what it did and determine what needs to be improved. Usually I find some small tweaks that need to be made. Depending on their nature I either make them myself or tell the AI to do it.

This would be vastly better if the AI could run the build/test script itself and iterate to a working state without my input. I expect that will be possible soon -- and probably works for some environments now. But even as-is, the AI makes it so that while I don't actually produce the code any faster (maybe even a little slower), I'm more productive overall because I can take care of other things while the AI is working. The small interruptions to copy-paste output don't require a context switch.

I find that current LLMs are roughly equivalent to a smart but extremely inexperienced entry-level programmer who just happens to have thoroughly read and absorbed the language manual and all of the available APIs. If you use them the way you would a such a junior programmer, it works pretty well. You don't ask them to write the tricky code for you (or if you do you expect their work to need significant improvement) and you don't expect them to have a good sense of what good design or architecture are. But they can still be extremely useful.

And, of course, they're still getting better. Fast.

Actually your third link says some residential consumers were paying wholesale rates in 2024. I don't have a lot of sympathy for them, though, since they knew what could happen after 2021.

I think those 2021 events are why you can no longer find market pricing plans for residential consumers. It's all fixed rate (though it can vary by tiers, it's still fixed for a given amount of usage).

According to data I found online, the average paid by CA residents is 30 cents per kWh, so if you're paying 12.9 cents, then the right phrase is "less than half". Though of course, the need for heating/cooling between the temperate Bay Area and the more extreme conditions in Texas is different, and usage obviously affects the bill, but we're talking about rates.

Here in Utah I would also normally pay about 13 cents per kWh. I'm actually on a TOU plan which is 5 cents off-peak and 25 cents on-peak, with peak being 3 PM to 8 PM on weekdays. I have an EV so I configure my charger to only operate during off-peak hours and this saves me quite a bit. In the summer I also set my A/C schedule to cool the house a few extra degrees just before 3 PM so it mostly doesn't have to turn on until the peak hours end.

Is "gouging" the right word? Residential consumers in TX pay half what those in CA do. And the massive price spikes generally only apply to wholesale buyers. I don't think most residential users even have the option of a variable-rate plan.

In my experience, qualification has little to do with anything that can be trained. People with the right kind of brain can and will learn whatever is needed to do the job. Those without will never be very effective no matter how much you spend on training. The top tier software engineering companies don't really even bother asking you what you know because trainable knowledge doesn't matter. Instead, they focus on asking you to solve problems, live, because while problem-solving is to some degree a teachable skill, being really good at it is mostly about native ability.

Personally, in my 35-year career as a software engineer, I've rarely found training to be useful, and when it is useful it's not training in software development tools or processes, it's training in things related to the business the software is supporting. What I have always found useful is books. If my employer will buy me whatever books I need, I can learn approximately anything. Different people learn in different ways, so perhaps that's just me... but my experience is that all the really good software engineers I've ever worked with were the same.

Why do you think that would be worse than compiler-generated assembly? Or worse than AI-generated code in a high level language?

Not until they started working on man-rating. The early development F9 followed the same iterative fail-fast model, though obviously with a lot less global attention. The painstaking detail for SS/SH will have to come, but can and should be deferred until they have platform that has been proven basically reliable.

Compare to Falcon 9, which used the iterative / fail-fast process... and is the safest orbital rocket ever built, as well as the cheapest.

Will they? The rapid iteration / fail fast model is what they used for Falcon 9, and it is the safest orbital rocket ever built. Did they just get lucky?

Starship development is primarily funded by SpaceX. SpaceX has some money from a NASA moon lander contract, but it's not that much relative to the cost of developing a large orbital rocket system. If/when they meet all of the milestones, they'll be paid $4B. So far they've only received a few hundred million -- probably not enough to pay for even one Starship test flight.

SpaceX is private so we don't know a lot about its finances, but it appears to be quite profitable, between the Falcon9-based launch business and (especially) Starlink. That is what is paying for the Starship development, and it seems likely that SpaceX can continue funding the development indefinitely -- though obviously if at some point it becomes clear that it's not going to work, they'll cut their losses and try something else.

I know a few rich high school teachers. They got rich doing something else, then took up teaching because they enjoy it and feel like it's a good way to give back.

Monopolies aren't required to follow any particular rules around price hikes or contract re-negotiation -- not unless they're trying to insert provisions restricting use of competing products. It's not illegal to exploit your monopoly to maximize revenue from that product, it's only illegal to use the monopoly position to build or maintain market share in a different product space, or to prevent competitors in that space from competing. On the contrary VMWare's actions are turbocharging competition in the hypervisor space by making all of their customers desperate to find alternatives.

When you're smashing your hand with a hammer you can stop, but healing the broken bones takes a lot of time. Rescinding the executive orders will be easy, sure, but the stuff that's broken won't be repaired quickly or easily, especially because the world sees that we elected Trump not once but twice, and it's going to be tough to convince them we won't elect someone like Trump again. Our allies and trading partners will remain suspicious of us, for good reason. The researchers, once moved, won't easily come back. Repairing the damage DOGE has done will take a decade or more. And so on.

Like many people you look at all the bad that didn't happen during Trump's first term and assume his second will be the same, but that's some combination of ignorance and wishful thinking. It was obvious even before the election that if the Mad King won this time would be different because he would be surrounded by obsequious yes men rather than competent staffers picked by the old guard GOP, and four months in it's absolutely undeniable.

I don't think Trump will destroy the country. I think the system has enough guard rails to prevent him from ending democracy (though he'd clearly like to do exactly that), and as long as we retain democracy the system will eventually be able to error-correct. But that just means we'll stop swinging the hammer, it does not mean that the damage already done will be miraculously repaired.

Except they're not.

On the business side, every survey of business leaders as well as every other indicator of capital investment shows that they're hunkering down: deferring or even cancelling any plans to expand, freezing hiring (if not actively working to reduce staff), etc. Good businesspeople can deal with almost any economic and regulatory climate you throw at them, but the one thing they need is predictability. Business is all about planning, and you can't make plans in the fact of chaos.

On the investment side, yeah, Wall Street seems surprisingly optimistic, but if you look at bond markets you see a different story. The most concerning part of it is that normally money floats back and forth between stocks and bonds, depending on the economic conditions. When stocks fall, the money flows into bonds and their prices rise (causing their yields to fall -- which is good when the government needs to borrow, which is always). When stocks rise, the money flows out of bonds and back into stocks. But this isn't happening so much right now. Equity and bond prices are falling together (and bond aren't recovering when stocks do).

How can that happen? It can happen when capital is fleeing the country. Investors are pulling their money out of the US, equities and bonds both, likely sending it overseas.

Indeed.

Unless he keels over soon and Vance proves to be more rational, I doubt it. Actually, I think you're overly optimistic even if that happens today, and even if Vance proves to be an outstanding president.

Trump has already irreparably damaged our economic and military alliances around the world. The world was sort of okay with trusting us again after Trump lost in 2020, but November 2024 proved that America is no longer a trustworthy ally or trade partner. The world's alliances and trade networks are treating us as damage and working to route around us.

DOGE has significantly gutted many US agencies, and it's going to take a very long time to rebuild that lost state capacity. We haven't yet seen the impacts of that foolishness yet, in part because state capacity was already pretty low (on a per capita basis, the federal government had reached a nearly century low), but we will.

Trump's budgetary plans will add $6T+ to the debt, at the same time he's pushing up bond yields, making it more expensive to finance the debt.

I'm slightly more optimistic that we can reverse Trump's kleptocratic moves. SCOTUS has made sure we'll never be able to hold him or any future president accountable for taking massive bribes, but maybe we can keep the rest of the government from following suit. It's not a given by any means, though.

Finally, Trump's actions on immigration and science funding are already causing likely-irreversible damage to our research base. US researchers are applying overseas and foreign researchers are not applying to come to the US. Long-term studies have been shut down. For example, I saw that a large-scale (~190k participants) long-term (~40 years so far) longitudinal study of breast cancer has been shut down. If funding is restored quickly, that can be restarted. If it isn't restored until 2029, there will be a significant gap in the data. But that one at least can probably be restarted. In other cases researchers or facilities may well be gone.

The US' post-WWII global dominance was built on two primary foundations: Technological progress and trade, with foreign aid playing a significant supporting role. Trump has done an incredible amount of damage to all three in just four months in office, enough that we won't recover quickly, and will never recover all of what we've lost. If he gets to continue thrashing about for four years, it's gonna be pretty bad.