I find I do a lot of refactors that I would have previously written off as good ideas, but too much work to be worth the effort. I still do think about the effort and weigh it against the benefit, but the effort I'm weighing is my time to review what the LLM did, and the time of whoever has to review my PR. On balance, though, I end up doing a lot of code cleanup that I previously wouldn't have done.

I do also have the LLM actually write most of my code, but first I discuss what I want with the LLM and ask it to write a detailed design doc, then I review and adjust the design doc, then have the LLM write the code from the design doc, then review and adjust the code. So I'm still doing quite a bit of work, but the overall process is still a lot faster than writing the code myself -- and I'm pretty fast. The overall quality of the result is on par with what I'd produce by myself, maybe a bit higher because of the greater willingness to refactor.

I actually do "vibe code" lots of stuff. Not production code but throwaway scripts used during development to automate things that I'd otherwise have done by hand because although it's faster once automated, the time to write a script by hand isn't justified. But with an LLM able to output a several hundred-line shell script in seconds, there's no reason not to automate everything I can. Option A is to tell the LLM to do it for me, but if I have to do it frequently and it's sufficiently mechanical to be scriptable the best solution is to tell the LLM to write a script for me. I don't really review those scripts, either, beyond a quick skim to make sure it's not doing anything insane. Beyond that, I figure out if it works by running it. If it doesn't, I tell the LLM to fix it. This is extremely efficient.

For more complex tasks that require judgment I tell the LLM to do it, then see what the LLM screwed up, tell the LLM to fix it, rinse, repeat until the result is good... then I tell the LLM to write a skill to capture what I just learned about how to get the LLM to do that task well. So then I can also automate tasks that aren't purely mechanical.

All in all, AI at least doubles my productivity. And probably a lot more than that.

That's how mine works. But to make it possible you need another piece of equipment, an automatic disconnect or what SolarEdge (the manufacturer of my equipment) calls a "backup interface", plus inverters that are capable of generating their own waveform if none is available to sync to, and will only do this when the backup interface tells them to.

So, the way it works is that when there's an outage, the grid-tied inverters instantly shut down because there's no grid for them to sync to. Then, after a second or two, the backup interface disconnects the house from the grid and sends a signal to the inverters to come back up in "backup mode". I have two inverters in a leader/follower configuration, so the leader generates the waveform and the follower syncs to it.

When grid power comes back, the backup interface tells the inverters to shut down and re-connects the grid. Then the inverters spend a little while watching the grid (takes about five minutes with my system; I think regulations require this) and then come back up in grid-tied mode again.

This means that the house loses power for 1-2 seconds when there's an outage (I also have batteries, so the house can run without grid power even when the sun isn't shining). Just long enough to be annoying. So I still have UPSes on my desktop machine, my network equipment, etc., but I size them for power capacity, not storage capacity, since they barely need any storage at all.

When the power comes back on, the reconnection process occurs without any noticeable flicker. Obviously the AC waveform shifts from whatever the inverters were generating to whatever the grid provides, so I suppose that some equipment (motors, mostly) might have a tiny stutter, but I've never been able to detect it

Fair, though the Middle Eastern oil is high in sulfur, so losing it isn't directly driving the reduction in low-sulfur fuels. But it's a global market.

All absolutely correct, though the refinery output is almost entirely fungible. So, lopping off the production of one category of crude will heavily impact the refineries that process that sort, but unless we end up balkanizing the world's oil trade the net effect on the supply of the outputs is roughly the same as if crude oil were all the same.

No... 20% of the world oil is a lot. Oil consumption has some components that are elastic in the short term, but not that much. Mostly of the consumption is pretty fixed in the short and medium term. The supply side has more elasticity, but it can't adjust quickly and much of the elasticity is very price-dependent -- only when the price rises above a certain point do some sources become profitable to extract, and then it takes between weeks and years to get them into production.

If you're thinking that the price adjustments shouldn't start taking effect until the well -> tanker -> refinery -> end consumer pipeline is actually exhausted, you're wrong because that would make the shocks dramatically worse. The futures are doing their job. By predicting the coming shortages and beginning to push the prices up now, we start pushing on what demand-side and supply-side elasticity exists. The higher price serves both to reduce demand and to motivate suppliers who aren't affected by the war to boost supply.

And keep in mind that part of the price adjustment is due to factors that would remain even if the war ended today. Because they have nowhere to put the oil, the Hormuz-dependent oil producers are shutting down production, and those wells will take weeks to restart.

You describe this whole futures concept as a "FinTech game" but it's very real, and very important. Futures act to smooth out pricing information across a whole raft of commodities, enabling both demand and supply sides to adapt early when it's cheaper, reducing shocks. Of course, it requires the markets to predict the future accurately, and predictions of the future will always be wrong, but historically they're quite good; better than pretty much any other prediction system.

Can't you just plugin any USB keyboard?

I don't think so.

There are more and less efficient ways of using AI. Using it more efficiently takes more work, more planning, more thinking, but the cost difference can be pretty enormous, so companies have a motivation to put budgets in place to motivate engineers to be more efficient.

For example, for many tasks, the slightly older and cheaper models are sufficient (e.g. Claude Sonnet 4.5 - medium effort), but for other things you really want the better reasoning of a later model with more context (e.g. Claude Opus 4.6 - high effort with 1M token context window). It's tempting to just always use the biggest, best model, but that gets really expensive. I blew through 60% of my monthly budget in one day when I forgot I had turned everything up to max. So now I'm careful to keep it turned down except when I'm asking it do to something complex. Keeping turned up to max all the time would make me somewhat more effective, and it would be more convenient for me, but the cost could get quite high.

That's just one example. There are lots of other parameters to efficient and effective AI use.

So, budgets make sense, but they're also annoying, and higher budgets are nicer.

The "as compensation" model is clearly ludicrous. I see it as a valid "work environment" question, though, up there with free food, etc. All else equal (which it never is) I'd much rather work at the place that gives me a $500 monthly budget over the place with a $200 budget.

I don't think that would have hit the people you want to hit. The shareholders of financial institutions are overwhelmingly not employees of the financial institutions. In 2008, Citigroup was 84% owned by institutional investors, meaning pension funds, mutual funds, insurance companies, etc. And the biggest owners of mutual funds are 401k accounts and IRAs. I couldn't find any numbers on how much financial institution stock is owned by various sorts of retirement accounts, but it's a lot, because retirement accounts own a huge chunk of stock of all sorts.

Your proposal would have pushed pension funds and insurance companies toward insolvency and further damaged the retirement savings of tens of millions of ordinary middle-class Americans. The bankers would barely have been touched. They'd have lost some, sure, but much of their compensation was in the form of lavish bonuses, and many of them regularly cashed out their equity positions to diversify and to avoid ownership thresholds that trigger greater SEC scrutiny of transactions. Insider ownership of the major banks in 2008 was less than 1%.

This is the moral hazard argument, and it definitely has merit. But this is where we get to the "too big to fail" part of the problem, which wasn't only that the participation of the banks is essential to the operation of the economy, but the ownership of those banks is also essential to the structure of the economy.

It's worth pointing out that although I described the bailouts as "loans" that's not technically correct. What the government actually did (except for AIG) was require the institutions being bailed out to sell the government a large chunk of newly-issued preferred stock. This positioned the government ahead of the holders of common stock in any bankruptcy proceedings, and gave the government outsized voting rights. The preferred stock also paid dividends, starting at 5% and rising to 9% (this is the reason most of the preferred stock was repurchased ahead of the planned schedule; the rising dividends would get punishing). Citigroup took a slightly different approach, repurchasing shares of common stock and swapping them for the preferred stock -- which actually did result in the government holding a 34% ownership stake in the company at one point, the largest shareholder.

AIG was handled differently. They were in a worse negotiation position so they ended up giving the government a large tranche of preferred shares that were convertible to common shares, and provided voting rights and an equity stake equivalent to 79.9% of the company (the government chose 79.9% deliberately because 80% would have triggered some messy tax rules). That enabled the government to fire and replace the whole leadership team, which they did. But they didn't take shares from anyone; all existing shareholders kept their shares, although their value was diluted to about 8% of what it was after the price crashed, so a fraction of a percentage of what it was before.

Arguably, what happened to AIG is close to what you wanted, and AIG was the holder of most of the credit default swaps that were at the heart of the problem. But doing the same thing across the board to all of the banks, including the ones who didn't want or need bailouts but were forced to take the money (for good reasons), would have been incredibly destructive to the economy in general and retirement savings in particular, and deeply undermined trust in the banking system. Exactly the opposite of the intended result.

All of that said, moral hazard is a real concern. Investors must face actual risk, or capitalism doesn't work. I'd argue that they did face actual risk; lots of portfolios were wiped out. 2008 severely sabotaged my parents' retirement. I don't think the fact that it was only a 99% loss instead of a 100% loss is the difference between capitalism working and not, though, and I think it would have been politically infeasible for the government to deliberately inflict further losses. There was also no legal/procedural way to do it. That was fixed in the 2010 Dodd-Frank Act... so now the government does have a way to seize and even liquidate even the largest institutions -- though in practice huge holding companies like Citigroup are insanely complicated so even with the legal tools do to it, that might not be the best option.

Yeah, precise timekeeping is essential for celestial navigation. Which reminds me, I really need to finish my celestial navigation course...

I *am* American, and when you say "cash back" in the US in the context of retail purchases, it has a specific meaning: cash handed to you by the cashier at point of sale, from your bank account, added on to your purchase transaction.

What you're talking about goes by a few different names, mostly depending on the precise structure of the program. It sometimes does translate to actual cash back, but not usually because people find that just getting a few bucks off of their credit card bill boring. I call mine "Amazon points" because I primarily use an Amazon-branded card that gives me what amount to Amazon gift card credit. This is psychologically more appealing than getting cash because it's "free" money, to be spent at a whim without the normal purchase cost/benefit analysis. Obviously it's still just money, and obviously it's my money, forked over at purchase time in the form of slightly higher prices... but some of it is the money of people foolish enough to use cash or debit cards rather than maximizing their reward points.

Yes, it's a dumb system and we'd all be better off if it were banned. But while it exists, it makes sense to maximize it.

You can't get cash back from a credit card transaction, only from a debit card transaction, which is the "transaction happens directly with a bank" case. Credit cards do give cash advances, but they have to be approved through an interaction directly with the card issuer, not at a point-of-sale.

Why do you think it's a flex, rather than a simple statement, explaining why they don't have cash? I mean, I guess if they're just announcing it outside of any relevant context...

Don't hold your breath.

His record on that is actually pretty bad. SCOTUS tends to give him his way on shadow docket rulings but he's much less successful in the actual on-the-merits rulings. Granted that he's had a few incredible (as in "impossible to believe") successes like the ridiculous immunity ruling, but on the whole the court goes against him more often than not.