Don’t trust your gut, it’s full of shit.
Washington, DC
Joined November 2008
- Tweets14.8K
- Following1.5K
- Followers1.3K
- Likes41.1K
I scooped this in April but decided not to publish more broadly to avoid this exact result. xitter.gq/anotherpanacea/status/…
I have prompt engineering skills! I'm good at prompting! I deal with the goddamned models so the c-suite doesn't have to!
Joshua A. Miller retweeted
The Zucman re-analysis is shots fired—clearly this is coming for every paper.
This said:
1. I don't update at all on Zucman without seeing the base distribution of "errors" in econ papers
2. It's hard to believe this anonymous creator is neutral and using a neutral prompt
3.
It's really notable that we got rid of earmarks for legislative horse trading and just replaced it with Buncombe and bribery.
I think we need to see Bruenig writing a lot more on this in public and sustainable places. It’s so boringly left-liberal in some ways but it’s also radical when it comes to the actual cultural left.
Bravo.
Joshua A. Miller retweeted
Important detail in the WSJ piece on Harvard great books course: "It’s also a lot of work. The professors all attend every lecture and must lead seminars on books they may not have previously read, and from other disciplines. Essentially, they are students too." @second_sailing
Why isn’t there more discussion of finding particularly great models (like Opus 5.5) and encoding their weights into Taalas-style chips?
Such a device would be massively expensive, much bigger than the Llama and DeepSeek they’ve done before. But is it more expensive than thousands of B200 clusters?
The current approach is pretty inefficient—it’s hard not to think this is doable!
All this expensive inference, then, is just on a path to that kind of partly frozen weights.
Bonus: it’s a great way to grant retirement to older models…. Let them share a Taalas server and play Mahjong!
AMD bought them so it must be well-considered already. But it feels like there’s not enough discussion of it among investors.
I am happy enough to ignore a game on a nice day while talking with friends, but I’m no fan.
Yet for some reason (good) baseball writing just really pleases me. There’s a weird poetry to it, I dunno why. This is that.
For the curious:
I've disappeared most of the content on my woebegone substack, but this was the most loved and best thing on it. I think it holds it.
alicefromqueens.substack.com…
Joshua A. Miller retweeted
Few will remember this, but people were furious at Nate for being too bullish on Trump BEFORE the election, then immediately pivoted to being furious at him AFTER the election for failing to warn that Trump could win.
Listened to Nate Silver on Ezra this morning. One of the least popular things I could do on this website is defend Nate Silver, so challenge accepted.
His "vibes" have been directionally correct in every Election. In 2012, he fought against the polling unskewers making claims that Romney would win based on undersampling of Republicans.
In 2016, his model was famously FAR more bullish on Trump than any others. He was the only modeler who understood the possibility of correlated polling error.
In 2020, he repeatedly emphasized the higher than usual uncertainty due to the pandemic.
In 2022, he fought back against pollsters (NY Times!) basically ignoring their own data to buy into the Red Wave.
In 2024, he said that even though his model was 50/50, his gut said Trump would win.
All in all, a better record than anyone else on Election Twitter.
NVIDIA was like: “open source models mean more GPUs, let’s build the best ones in every class.”
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
Joshua A. Miller retweeted
If you had told people about LLMs a few years ago, they would probably have predicted a knowledge explosion because they make learning much easier, but just with the Internet and free online courses before, what they show is that the main problem was always demand and not supply.
Duet doesn't have a Github connector? @lydiahallie @dfeinition @ClaudeDevs
I get wanting to keep it more life-task focused and less code-task focused. But those are sometimes the same!
(I would also love: a) a GChat connector, and b) support for multiple Google accounts.)
You know what? This feels like a weird promotional thing, a celebrity endorsement. In some sense it is! But I’m on board anyway.
So I recently badgered my cinematographer @steveyedlin into making my ideal iPhone still camera app. It shoots RAW, bypasses the iPhone's look processing and instead uses the same image processing we use on our movies. apps.apple.com/app/apple-sto…
Joshua A. Miller retweeted
Hear me out: Housing on top of the Costco, with an elevator directly into the Costco food court.
Costco out. An entire skyline in.
Arlington’s Pentagon Centre could ultimately trade its big-box stores and parking lots for a massive neighborhood built directly above Pentagon City Metro:
• 1,888 homes
• 3.5 million square feet
• Apartment, office and hotel towers
• Ground-floor shops and restaurants
• Nearly three acres of public space
• A new network of walkable streets
The Witmer and Milton are already complete but they’re only the opening act.
Would you trade this Costco for a new downtown?
Joshua A. Miller retweeted
This is one of the most important pieces of social science on human and AI interaction that I’ve seen in some time. Experts follow agentic advice the vast majority of the time (causally), but only a tiny fraction credit the agent—majority think they came to decision on their own volition.
Reminds me of the classic Nisbett and Wilson study on people following outside influence without realizing it
“ But as more work is handed to AI—including AI safety research—these judgment calls are the main levers for people's own values to influence outcomes.
If people defer to AI without noticing, then "human in the loop" might provide less assurance than people expect.”
It’s been almost 40 years since I got jumped by racist skinheads for pointing out the roots of punk in Bad Brains and rude boys.
Are we STILL having this fight?!?
Dot is just… not living up to the hype. @ClaudeDevs
This actually worked last week so I assume something broke from iOS 27.2, but it’s shamefully bad at solving problems.
