At this point getting a Chotiner interview is almost a guarantee that you're doing something very stupid and should stop.
05.03.2026 12:18
👍 16
🔁 6
💬 0
📌 0
At this point getting a Chotiner interview is almost a guarantee that you're doing something very stupid and should stop.
One thing I've been thinking about lately is that models are increasingly becoming more like curves on these kinds of graphs, rather than points.
For example, "Opus 4.6" and "Opus 4.6 thinking" should really be two points on an Opus 4.6 curve, instead of two separate points.
I think it is more likely than not to succeed. The major thing to show is the factual evidence that supports a claim of a vacuum that exists and would be filled by the proposal being enacted.
Yep. Note that durability is not the same as availability though. "We served the file at the time you wanted it" (availability) vs. "we can point to the bytes that correspond to your file and retrieve all of them" (durability).
Sure thing. As the official™️ Gold Sponsor™️ of the U+0021 EXCLAMATION MARK glyph, it's my solemn duty to propose more emoji that make people say !!!.
That's right. Note that a lot of what the casual user might think of as "availability/reliability" is often implemented as "run N independent copies of a somewhat unreliable service, so that at least M of them are always running".
Submissions for the next window open April 2nd, 2026. I've marked my calendar to propose this. 😄
For a production-grade service that you want important businesses to build on, all (as in literally 100%) of these bars would be green. Typical availability of a production service like, say, S3 is in excess of 99.99%.
Some more reflection on this: maybe there is more incentive for better tools that helps the LLMs and humans. Thinking here of, say, a better Grep(*) tool that can do AST search on the LSP metadata, for example.
It really did! 75°F coming up on Friday, too. That may just finish off the last slush piles.
"Published March 26, 2003."
This one absolutely should have passed muster.
So, in short, don't put these models in military technology. Colossally stupid and doomed to failure.
3️⃣ If the goal is intelligence embedded in systems, then LLMs alone are the wrong tool. They do not contain or possess intelligence, and increasing numbers of researchers are convinced they're a dead end for AGI.
(Jury still out on whether some meta-system that includes an LLM can maybe do AGI.)
2️⃣ Inherently non-deterministic systems are challenging to test. Military technology must be rigorously, thoroughly tested in a wide variety of failure modes and conditions.
Why is it bad? There are many reasons, but three key ones are:
1️⃣ It further removes human judgment. World War III has been averted several times because humans overruled computers. See, for example, en.wikipedia.org/wiki/Stanisl...
But also: it wouldn't be SkyNet because it would be too stupid and vulnerable to be SkyNet.
It is profoundly stupid to try putting LLMs within a country mile of military technology.
In particular, putting LLMs inside of autonomous weapon systems is a disastrously bad architectural, technological, political, and military decision. Don't do that.
As a famous scholar once said: bsky.app/profile/diew...
My "like" on this post is not an endorsement of FBI practices but rather an expression of enthusiasm for your +5 DMV point balance
A sign written in a child's scrawl, reading "Operation Get Candy: Pretend to be asleep -> Go silently downstairs -> Get candy -> Eat -> Get rid of evidence".
Really need to teach the kids about opsec.
That's part of it, yeah. Basically, imagine dividing the y-value by the timeseries of the number of engineers so that the height of the graph measures output code per engineer (not that lines of code is a good metric, as you know, but that's the idea).
I think an interesting variant of this would be to normalize by committer/employee count. How much of the gain is being contributed by having more people vs. more productive people?
(I have a lot of thoughts/data on this I'd love to swap notes on next time we cross paths somewhere!)
It's 20149!
I have literally no information other than his return address, but not sure I feel comfortable posting that. There's a five digit number (which I assume is a postal code?) in front of the word "Hamburg", if that's not super-identifying.
This is an incredibly long shot, but if anyone knows a Stefan Hoyer from Hamburg, Germany, I mistakenly received a package meant for famous author John *Grisham* from him.
Although John Grisham also lives near me, I have no way to get in touch with him. So, Stefan, if you're reading this, DM me. :)
The title page of the Bash Reference Manual, Edition 3.1-beta1, dated September 2005.
Things I would never have guessed were in a release of controversial government documents:
www.justice.gov/epstein/file...
The opaque nature of private credit and the shift out of the public markets is also going to significantly hinder any useful response to any crisis, too.
Not a perfect fix but going to try this:
```
claude() {
ln -s "$XDG_DATA_HOME/claude" "$HOME/.claude"
ln -s "$XDG_DATA_HOME/claude/.claude.json" "$HOME/.claude.json"
command claude "$@"
unlink "$HOME/.claude.json"
unlink "$HOME/.claude"
rm -f "$HOME/.claude.json*"
}
```
Infuriating that Claude Code doesn't respect XDG_HOME. I *hate* having ~/.claude junk my home directory. github.com/anthropics/c...