David Andersen
@daveandersen.bsky.social
1K followers 700 following 2.3K posts
Computer Science Professor, CMU; co-founder and CTO, Enriched Ag Energy-efficient computing, a dash of security, and a pinch of databases. Also on Mastodon (https://hachyderm.io/@dave_andersen ) signal: dga.48 he/him
Posts Media Videos Starter Packs
daveandersen.bsky.social
Perhaps you could give them a faux number instead
daveandersen.bsky.social
It strikes me that normalizing the power use is probably a more clear way to show the comparison since my power use changed. Here it is. Both years are normalized to an average use of 1; 2024: 3-9pm was 13% higher than daily avg. 2025: 3-9pm was 22% lower than daily avg. Net swing: 35% less.
daveandersen.bsky.social
Dangit. My vegan orgain that I mix 50/50 with whey isolate doesn't fare so well. Guess it's 100% whey from here.

(And the voice in the back of my head says "yeah, but Dave, probably none of this compares to the amount of lead you ate as a child due to leaded gasoline...")
daveandersen.bsky.social
So far so good. It's tempting to get more battery but integrating it at this point takes actual work and the payoff is pretty long.

Fun experiment so far.
daveandersen.bsky.social
What changed:
(1) Switch 400-600W of servers on UPS battery at 3pm, (down to 40%).
(2) Added about 400-600W more solar to feed that UPS.
(3) Throw 120W of network on UPS from 3pm-11pm.
(4) Raise AC temp by 1 degF from 3pm-9pm
(5) Conscious avoidance of load, e.g., delay dishwasher 'till 11.
daveandersen.bsky.social
One might reasonably ask: Did switching to time-of-use billing for my electricity actually change my power use habits? That .. that would be a strong yes. (2025 is lower overall b/c I turned off some servers.) 2024: Use rose at 3pm and peaked at 4pm. 2025: use crashed at 3pm, stayed lower through 9
daveandersen.bsky.social
2025: It's punk to stand up for the law and the constitution.

Not something younger me was expecting.
govpritzker.illinois.gov
I'm not afraid. Come and get me, Donald.

I will stand up for the law and the Constitution. That’s what we do in the State of Illinois.
daveandersen.bsky.social
And thus we reach the problem of LLMs
daveandersen.bsky.social
skeet didn't want to face an e

ditor
daveandersen.bsky.social
I think there is a certain solace in choosing not to understand anything remotely related to Mr. Beast.
daveandersen.bsky.social
This week is CMU's fall break, so I'm working from home today. It turns out WFH has dangers, such as excessive espresso consumption. Nom.
daveandersen.bsky.social
Blathering about my most recent mostly-positive but limited experience using claude code wanted more space than social media permits, so it got a blog post. Don't expect anything too deep, I'm just keeping track of my experience.

wiredream.com/claude-optim...
Optimizing a power display with Claude Code | Wiredream - Dave Andersen's blog
Dave Andersen's (new) blog
wiredream.com
daveandersen.bsky.social
We got ours + kids at CVS with absolutely no problems.
Reposted by David Andersen
seema.bsky.social
Congratulations, you're department chair.
unenthusiast.com
In honour of spooky month, share a 4 word horror story that only someone in your profession would understand.

rm -rf ~/
hammancheez.bsky.social
"The chancellor approved it"
daveandersen.bsky.social
That's how they capture packets (of water)
daveandersen.bsky.social
You didn't need to know this, but it turns out you can play a mean game of air hockey on a wet and slightly soapy countertop with an espresso tamper.
daveandersen.bsky.social
<tinfoil hat> and just what are all those people running FROM, hmmm?</tinfoil hat>

(Probably ICE, more likely)
daveandersen.bsky.social
LLMs are a powerful way to sample from a plausible distribution of generated output; not all problems are a great match for the infinite monkeys approach to codegen. But some are, when you can quite easily determine if the result is "good enough" or not.
daveandersen.bsky.social
and from a business perspective, it's working well enough and nobody's complaining about bugs in it. At the same time, it's generated absolute garbage code at times. I tried using it for a quick research experiment involving random numbers, and it generated something dangerously statistically wrong
daveandersen.bsky.social
And as @lizthegrey.com noted in another thread: I too use it to generate frontend code for me, because I am _really not_ an html/javascript/ui person. It really is 10x faster than me at that particular task because I'd have to go learn the entire enchilada. The code is probably worse, but it works
daveandersen.bsky.social
This is such a rich issue - it's going to be really hard to study well with small randomized studies. Like - I use AI a lot in tiny personal projects where I just haven't felt the motivation. Does it write worse code than I would? Yeah, definitely. But it got the job done, too, and I wouldn't have
Reposted by David Andersen