On the other other hand, it’s fun to accelerate quickly, so the higher tire cost is worth paying

Language models *do* make mistakes, but so do software engineers, which is why software testing exists (and predates LLMs by many decades). Journalists also make mistakes, which is why fact-checkers and editors exist. Perfection is not the bar that needs to be cleared for a tool to have utility.

There's a concept in software development called "testing" - you verify the outputs of your code match the right answers for some set of inputs. The code will also fail to compile and/or throw runtime errors if it's written incorrectly. This makes verification much easier than for, say, journalism

They absolutely can, it’s just a question of whether they’re needed often enough to be cheaper than alternatives. Daily cycling - great! Twice a decade, not so much

We should also be building lots of HVDC backed by hydro, nuclear, or similar firm capacity. Lines are expensive, you don’t want to build them as insurance against very occasional events, but as the backbone of a grid with high utilization factor

China has huge coal plants colocated with their wind & solar to stabilize their HVDC lines. We do not want that. Their grid buildout is driven in large part by a desire for energy independence, and to a lesser extent clean energy, which is why they keep on building new coal plants.

The continental shelf is so much smaller in CA than on the East Coast. Why would you export that limited resource at great expense across the vast center of the country instead of using East Coast wind that can be hooked up almost directly to load?

Keep building wind and solar, but you gotta plan your system in those regions around the assumption that variable renewables won’t show up for days on end, and infrequently enough that it doesn’t make financial sense to fill that gap with battery capacity that’s only needed a few times per decade

California has it easy - they don’t really need as much firm capacity (mild winters, great solar potential). It’s regions like the Midwest and Northeast with long-tail winter doldrums events where you have very little wind, very little sun for a week, and cold weather that reduces heat pump COP

“Writing software” is not a niche field in the same way that quick & dirty prototyping for manufacturing is, and doesn’t have the same physical constraints

I have used both. 3D printing is most valuable for quick prototypes and building certain geometries that’d be challenging for conventional machining operations, but it’s too slow for volume production. LLM tools for sw dev aren’t similarly domain-limited and are faster & cheaper than hand-coding

Also did all his homework in careful LaTeX just to make the TAs deduct that many more points for everybody else’s chicken scratch 12000.org/my_courses/u...

There was a guy in some of my upper-division undergrad courses who was probably in his 50s-60s and already knew all of the course material. I later figured out that he was working on his sixth engineering MS at the time from his personal site where he lists all 140 college classes he’s taken

Yield > 0 is not in compliance, though! “Difficult to detect” does not mean “compliant”, and a test with tens or hundreds of tons of yield would be waaaay beyond any shading of the letter of the treaty

Opus does a better job matching style than whatever Kagi's using, but iteration definitely helps (this is v3)

i'm dying

*technically Livermore in CA has the Superblock for plutonium work, but their EIS limits them to tiny quantities, so I don't think they'd have the material handy. You might be able to whip up something crude out of Pantex (TX) or Y-12 (TN), but if we're talking first to a usable device, that's NM

LANL also does detonator production, high explosive synthesis, tritium handling, etc., and Sandia has you covered for much of the rest

Pit production explained | Los Alamos National Laboratory Los Alamos National Laboratory’s pit production mission is underway. But what does that mean? And why is Los Alamos the place for this work?

NM is the only state in the union with a facility currently engaged in fabrication of plutonium. You could YOLO it elsewhere, but there's only one place with skilled workers currently doing that work. www.lanl.gov/media/public...

Ten million schoolchildren running language model inference by hand on ten million TI-83s

We just need to have all the school districts park their electric buses at charging stations to do vehicle-to-vehicle charging on holidays

It’s going to be incredibly expensive to build battery-backed charger megastations for that peaky vacation/road trip-centric demand profile (only really need to store/deliver that much power across hundreds of charging stalls a few days per year)

A single Buc-ee’s on a busy travel weekend (say 100 pumps operating at 30% capacity factor) is delivering ~180 MW in usable (EV efficiency-equivalent) power, which is like 500 Supercharger stations. Home charging is fine most of the time, but road trip holidays are going to cause problems

The most expensive LLM generates about thirty thousand words for one dollar of compute cost. The average human typist can do something like 3-4k words per hour. If you’re just paying for student essay words-per-hour, the LLM breaks even at like $0.16 an hour. No Mturk worker is that cheap.

Workers with highly-specialized wisdom about a subdomain are probably fine for a while, but if all you’re doing is writing generic software to solve other people’s generic problems, how do you justify your salary a year or two from now?

Remote IT workers are going to be completely hosed. People have spent the last three years complaining about RTO mandates, but if your work is mostly modular tasks that can be conducted over text-based channels w/o in-person collaboration, I just don’t see much of a moat left

This already exists and works great: block.alxtratrstrl.xyz/index.html

Bilbo looking at his phone top on bottom is ChatGPT After all, why not? Why shouldn't I keep it? You're absolutely right — you found it, it's been with you a long while, and it's only natural to feel fond of something that's served you so well, especially when someone like Gandalf suddenly seems to want it for himself.

The rates and prices are bad, but the houses? Well, turns out that those are also fractally bad.

I cranked this specialty tool out in an hour or so using Sonnet 4.5 and JavaScript. I do not know JavaScript, but I know the output is correct because it matches the output of a compiled tool that took me a couple of days to create back in 2022.

stillyslalom.github.io/STCalc/