Context windows make a big difference both in how well it can keep a straight line and also in how much memory and processing power you need. On my PC with 24GB GPU and 64GB ddr5 I can run reasonably effective models (up to 12GB) with a 60K context window. That's good enough for small stuff.

Codex 5.3 can do 230K, can even go much higher when using the API, and auto compacts the token window. That makes a big difference.

What also helps is to maintain a task list, tell it to split up any assignment and put it on that list, only work on tasks on the list, and document the result on that list as well. That way you get a more persistent memory for what it did, and also it only needs to save the context for the task at hand.

Three weeks ago I shifted to using AI for a minor coding task. This week I converted 30GB of geodata from 5 sources into a mountain layer for a game map at 10 meter resolution. Today I used it full time to compare 16 laws to an architecture document to see if I missed something.

3 months ago none of this was really easy. This month it found 4 python bugs in a library I was using, created a fix, tested the fix, committed it, synced it and drafted the PR (which I checked and yes the whole thing was good).

I'm amazed and frankly, a bit intimidated.

No, he's actively cheering for fascism and genocide. In that view, anything that makes it harder to kill the opposition is a bad thing.

He's basically one of the Twitter morons referred to.

I blame your parents, actually, for how you turned out.

Psychosis and other mental diseases are just that: diseases. We're finding more and more evidence that quite a few are caused by viral interactions with the brains, or the remnants of diseases that the immune system couldn't completely clear, or an immune system responding incorrectly to a virus.

Which is a thing in the Netherlands, because we have rules against casinos feeding addictions to make a quick buck, just as barkeepers are liable for the damage when they give obviously drunk people more liquor.

Norway can't, but I'm sure North Korea comes pretty close.

drugs, guns, prostitution...

Really strange that people have to ask permission for those when in the USA these are all allowed everywhere... oh no they don't. In fact, Norway is more permissive than the USA wrt real freedom, which is also the freedom of other people from the effects of your stupidity.

Do I really want to send my kids to a school where yours is taking his AR15 to class today because he got turned down by a girl? That's not freedom, that's living in fear.

Like everywhere, you don't get the coins back, but the value at the time they're impounded.

Not a threat to mainframe sales. A threat to the consultancy and license fees that come with it.

If you knew how things run in the backend... there's a ton of issues all the time and we need 400 devs just for maintenance. Turnaround on any issue starts with 6 months of analysis.

I agree with you. Yes it causes damage daily. And even with that it still outperforms most human coders about 100x on speed. So I can fix the issues with plenty of time left.

Last week Codex fixed 4 bugs, including one very nasty off by one error and another indirect error, in about 10 minutes all by itself. I'm very impressed with the AI when it works.

Other times... not so much. But the benefits already far outweigh the cost. 20 EUR/ month is peanuts. If this keeps up I'm also taking on the 200 EUR Anthropic plan.

A modern mainframe is fine. But IBM doesn't make its money from hardware sales. It makes it from consultancy. And that area is going to get hit hard.

identification isn't the issue, prosecuting it is. And given that Trump has pardoned even the biggest frauds, like Trevor Milton, and fired the prosecutors, its obvious which way the wind blows.

There is also another problem with lobbying and election funds that's a bit more on both sides of the aisle. The banks have gotten a lot of money during QE that they're using to influence policies. Not that Trump needs a lot of incentive.

The COBOL is never the issue. Its the entire custom built environment around it.

I recently worked at a place where I pushed very hard for a transition.

They first built the system in 1969. They upgraded every time. And to support that they built their own data abstraction layers on top of the network database. Currently they're using a relational database. But they're using it as a network database because otherwise they need to fix tons of code.

We developed a way out of this, but it's going to take time.

Note they tried to migrate once before, cost them 40 mio and delivered nothing. Part of it was internal sabotage from people with overpaid jobs.

This isn't a purely technical problem.

The correlation between house pricing and other factors is with the indirect money supply, not with young people moving to California. That would at best drive prices up locally, but the housing price crisis is global. From New York to Amsterdam to Tokyo to Sidney and every big city in between, the housing reserve is bought up by investors of the type you don't want: rent seekers.

It has a very high correlation with QE and unlimited money from the Fed. The solution is going to be very hard, because removing the glut of money requires taxing the richest people, companies, and especial targets banks and FI. A counter weight to solve the money issues would be to give stimulus checks - probably the only thing Trump did right, for all the wrong reasons.

But if you only give stimulus checks and not remove the glut that landed at rent seekers, you're still going to have problems. With that money the richest families and people have been buying influence through lobbying or more direct ways in the case of Trump, Orban, Fico etc.

tl;dr: The QE during covid will be causing serious problems for society for decades. One of which is extreme asset prices.