Aaron Kimball
gremblor.com
Aaron Kimball
@gremblor.com
I like circuitry (synth DIY), software, bio stuff, and sailing. aka Firefly )'(. Roar!
Last night at the grocery store, the clerk ringing me up mentioned that there were *still* 5k without power (at T+48h and counting) in SF.

The SPOF nature of this substation is crazy and should be a case study once all the dust settles.
December 23, 2025 at 6:57 PM
Damn even with the sound off, it's pretty obvious.

I'd note that amphetamine stimulants (eg Adderall) can also cause a wicked case of dry mouth or involuntary muscle twitching. So, you know, maybe a range of drug abuse going on here.
December 23, 2025 at 6:52 PM
Have you seen this sort-of-pyramidal conspiracy theory taxonomy chart?

www.conspiracychart.com

Your theory aligns a lot with theirs...
The Conspiracy Chart 2021
www.conspiracychart.com
December 23, 2025 at 11:01 AM
Reposted by Aaron Kimball
Again, from Michael Grynbaum, here is Sharyn Alfonsi's full email.

It must be read:
December 22, 2025 at 4:00 AM
My bigger point was more "increasingly useful general tool" might also have a path of "gets unpopular and stagnates for 5 years, then new hardware or other improvements causes a rennaisance that becomes a stable and sustainable foundation technology in 2030--2035 or so."
December 21, 2025 at 10:10 PM
It's genuinely unclear to me whether acquiring a different LLM Co is helpful to a place like Google or FB, who already have a giant hairy tech stack under their models. Google didn't bother to buy any search engines it displaced.

Certainly means they can hire top talent at <$100MM a pop, tho.
December 21, 2025 at 10:02 PM
Fwiw, I'm in the middle camp listed above, but a 4th possibility is "all current LLM co's go bankrupt in massive crash, but in 10 yrs a new generation anchors itself as foundation of next tech wave."

All the .com's except Amazon went under. But that wasn't the end of the web, that was the start.
December 21, 2025 at 7:52 PM
I think it's the "no citation required" part that is concerning. If you mistake that auto summary for an expert human's, you ascribe trustworthiness to the summary that it may not merit.
December 21, 2025 at 7:46 PM
The thing I find frustrating is how limited it still is. I had a big list of online chores to do which would be perfect to outsource to AI. I prompted Claude desktop, which said no; Claude cannot POST to the web. Nor can one Claude supervise other instances.

I thought "agent" work was whole point!
December 21, 2025 at 7:26 PM
The secret hack of microblogging is you get 300 chars for your skeet but a whopping 2,000 for each image alt text!

You can tunnel a whole story into one skeet via images & alt text, for those who click in.
December 21, 2025 at 7:11 PM
Fwiw, tech iview q's/answers leak online. Some LLMs literally have the answer to code question #4 by FooTech in the training set and will verbatim print the answer.

Where I work we went with a mix: some q's we want to see you solve problems "by hand" and others where you work w/ an LLM copilot.
December 21, 2025 at 7:08 PM
Yea I find it crazy that a Waymo running over a pet is *so unusual* it's *literally newsworthy*, and people cite that story as evidence of the failure of the entire technology.

Meanwhile all the other people + pets that get run over by manually-driven cars just become statistics in the background.
December 21, 2025 at 7:00 PM
Reposted by Aaron Kimball
Quick quiz: how many people in S.F. were killed by Waymos this week?

How many people in S.F. were killed by human drivers this week?

(The answers are zero and two, but no one would know that from the reactions today.)
December 21, 2025 at 5:30 PM
Right? I was like "well that was a super pretty, trippy cgi version of /Dances With Wolves/."

If I want to watch that story again, Kevin Costner really did it better.
December 21, 2025 at 6:40 PM
SEAR-up.

Eastern PA / Philly suburbs.

One parent with a modest New England accent.
December 20, 2025 at 9:44 PM
Can I nudge you to prio some android app love in '26?

There are various format issues like truncated post content in the feed without a "..." (usually shows fine in click thru), and issues in the editor especially when replying to a post, that have been around for a while now. Not fatal, but...
December 20, 2025 at 8:02 PM
Judging by the packages regularly left for my neighbor, north of 5 for sure.
December 18, 2025 at 12:49 AM
I mean, "web 4" *is* right there.
December 18, 2025 at 12:38 AM
Eleven! Eleven interviews!

ROFL you would have to have the social skills of a brick to not know what direction things were heading after the.... Fifth... Followup call?

"Yea I just had some questions to chase some loose ends. Would you say Musk's drug use was 'problematic' or more 'depraved'?"
December 16, 2025 at 5:42 PM
Magnum still only counts as one tho, right?
December 15, 2025 at 10:59 PM
Oh that's an interesting thought!I have sometimes used a new agent in series when one "filled up" but never paused the original and resumed it to review a later LLM's output. Multi-agent usage has mostly focused on parallelism, this is a useful different idea.
December 14, 2025 at 5:33 PM
I've found that if you prompt it to create a backend endpoint, or some internal methods, etc., and then prompt the "next layer out" it will, in fact, use what it wrote.

LLMs can generate code. They can't generate system architecture, basically. The latter is still on you.
December 14, 2025 at 5:22 PM
For what it's worth, I think there is skill in using LLMs, no matter the UI and it centers on specificity. At one limit, you're typing as much as you would have in the first place. At the other end you get codespam.

But there's a way to keep it on a short enough leash to build and use abstractions.
December 14, 2025 at 5:20 PM
Gen Alpha is now welcome to join us and Gen Z together at the avocado toast bar
December 14, 2025 at 5:17 PM