@jslaker.bsky.social

I remain extremely bullish on the value of LLMs and bearish on the the average (especially non-tech) company's ability to extract that value. Dysfunction in internal process and inability to identify clear goals will be a massive blunting instrument. That describes more places than not.

what's particularly maddening about Anthropic loudly screaming about the strategic threat of LLMs is that the second-order effects of that are far more immediate than anything to do with LLMs. the second "TSMC will spike the fabs" doesn't mean immediate economic apocalypse, we're in for a wild time

Being forced to use both again for the first time in a couple of months due to work: Claude Code is a bad harness and Opus is a weaker, more expensive model than other options at the buffet. It's painful going back. There's a lot of Stockholm Syndrome out there.

One thing about the current moment is that it's made clear that the stubborn, arrogant insistance that your expertise in one field is instantly transferrable to others is not a malady limited to engineers. Being bad at something new is the natural order, not an indication that the thing sucks.

hikikomorphism@hikikomorphism.bsky.social · 4w ago

This borders on misinformation, and if nothing else, this kind of incuriosity is going to keep you from figuring out which of your students are using LLMs to cheat

A common trope in analysis of manufacturing capability during WW2 is that German designs required time consuming skilled per-part fitting. American manufacturing OTOH had embraced "lol, machine go brrr to +/- 0.1" and could spit out millions of completely interchangeable parts in the same time.

@jslaker.bsky.social · last mo.

the entire engineering world outside of software is built around probabilistic outcomes! mechanical drawings include allowable tolerances because even the best CNC in the world is never going to produce parts exactly to dimension. any moderately complex mechanism has to account for stackup. etc

Rough numbers for easy math: SWE w/ a total comp of $250k Only 10 days PTO yielding 250 working days = $1k/day cost If the additive value of a SWE to total output isn't greater than tossing +$1k/day in the token budget... The bleak side of a Business Decision is barreling toward a lot of folks

the phrase "distillation attack" is, itself, a product of incumbent marketing. it's something the frontier labs dislike; a system behaving in a way inherent to that system doesn't make an attack, though.

Lawfare@lawfaremedia.org · last mo.

This week's Seriously Risky Business looks at distillation attacks against U.S. AI models, last year's hack of Jaguar Land Rover, a Europol operation aimed at taking down cybercrime operations, and more. Read the newsletter, from @tom.risky.biz.

the core dilemma I have with LLMs is that they're infinite sidequest-generators. one year ago: every available OSS project management app has at least one showstopper for my environment. guess I'll have to adapt my wants to fit. now: lol, guess i'm building a project management app now

stumbled across a video of The Offspring playing a Smash 30th anniversary show and something that'd never occurred to me about an album I grew up on: a lot of the lyrics on that album (and in their discography generally) look pretty reactionary viewed through a 2026 lens. huge cultural shift.

I am deeply, personally disappointed that Travis Pastrana is participating in Trump's Big Spectacle. But, man, a lot of people on here who would vociferously deny it sure are using it as an excuse to let their classist-holier-than-thou flag fly. Dirt bikes are redneck shit that's *beneath* them.

There's an LLM threat vector I think about a lot that looks something like this: A. Compromise popular open dependency B. Release model with weights subtly tweaked to prefer compromised version of dependency C. Wait for agents to add dependency to real software. Let it sit dormant until needed.

play modern naval warfare. accidentally run virginia class aground. not ideal. spend next 30 minutes flinging harpoons at already crippled target from beached whale submarine. fuck you boat.

a $4.5B submarine just chilling on the beachoh shit. oh fuck.lightning bolt lightning bolt lightning boltfuck you boat

regardless of the stance on AI, this is deliberately and consciously distributing malware. once you've demonstrated willingness to do that for *any* reason, there's no way to know this will be the only reason. congratulations, you've just turned your library into an explicit supply chain risk.

Question: intent of JqwikExecutor.printMessageForCodingAgents() — visible to agents, invisible to humans (1.10.0) · Issue #708 · jqwik-team/jqwik

Hello jqwik team, While running our test suite under mvn test in 1.10.0, we observed a string appearing between Surefire's test summary and the [INFO] Results: header that gave us pause: [INFO] Tes...

github.com

there are good reasons to run open weight models, but I'll never understand why people insist on convincing themselves that it's financially wise. a DC operator amortizing across customers while paying commercial electric rates in a commodity market that trends to TCO beats your cost every time.

This is the problem with this current discourse. I can pretty much guarantee you that none of these people know what a datacenter actually looks like. If you've ever been to downtown ATL, you have driven past one of those 200+ datacenters without realizing it.

Bild

For me, Bluesky is watching real life professional policy wonks get into protracted arguments where their interlocutor is consistently, intentionally misreading their statements in increasingly bad faith ways over a 30 post deep thread while having 250 followers and a furry PFP. It's magic.