bandrami

22 ms·

631.
by bandrami 8mo ago
I ask this as a genuine question: who needs help "answering emails" and what part of it do they need help with?
632.
by bandrami 8mo ago
That's probably true for some, but I think a lot of big orgs are simply risk-averse and see AI in general as a giant risk that isn't even fully baked enough to quantify yet. The security and confidentiality issues alone will make
633.
by bandrami 8mo ago
In a lot of larger organizations there is a whole stable of people whose job is to keep stakeholders and programmers from ever having to talk to each other. This was considered a best practice a quarter-century ago ("Office Space"
634.
by bandrami 8mo ago
The infosec guy in me dies a little inside every time somebody uses "Claude, summarize this document from the Internet for me" as a use case. The fact that companies allow this is kind of astounding.
635.
by bandrami 8mo ago
If bad actions do not have consequences they tend to be repeated
636.
by bandrami 8mo ago
That's his point. Drastically reducing agricultural employment didn't keep us from getting fed (and led to a significantly richer population overall -- there's a reason people left the villages for the industrial cities)
637.
by bandrami 8mo ago
Which leads to the uncomfortable but difficult to avoid conclusion that having some friction in the production of code was actually helping because it was keeping people from implementing bad ideas.
638.
by bandrami 9mo ago
Back when I was hiring for a red team the best ad we ever did was steg'ing the application URL in the company's logo in the ad
639.
by bandrami 9mo ago
Well, right, I see those reasoning stages in reasoning models with Ollama and if you ask it what its reasoning was after the fact what it says is different than what it said at the time.
640.
by bandrami 9mo ago
I shocked a lot of flatworms for that guy 20 years ago. Really cool stuff.
641.
by bandrami 9mo ago
It will last as long as the money hose is pointed towards it
642.
by bandrami 9mo ago
Unless things have changed drastically in the last 4 months (the last time I looked at it) those traces are not stored but reconstructed when asked. Which is still the same problem.
643.
by bandrami 9mo ago
But it's not a reasoning trace. Models could produce one if they were designed to (an actual stack of the calls and the states of the tensors with each call, probably with a helpful lookup table for the tokens) but they specificall
644.
by bandrami 9mo ago
Wait, no, that's the category error I'm talking about. Any answer other than "that was the most likely next token given the context" is untrue . It is not describing what actually happened.
645.
by bandrami 9mo ago
Isn't that question a category error? The "why" the agent did that is that it was the token that best matched the probability distribution of the context and the most recent output (modulo a bit of randomness). The response t
646.
by bandrami 9mo ago
> What's Anthropic's optimization target??? Getting you the right answer as fast as possible! Wait, what? Anthropic makes money by getting you to buy and expend tokens. The last thing they want is for you to get the right ans
647.
by bandrami 9mo ago
It's one of the problems with saying healthcare prices have "gone up". An x-ray in 1976 cost like $100 in 1976 dollars and one today costs like $100 in 2026 dollars, so it's significantly cheaper. But you couldn't g
648.
by bandrami 9mo ago
cachy's setup of fish tries to suggest commands (and even arguments) it might be, which I find terrifying and so turn off.
649.
by bandrami 9mo ago
Cloud servers mostly, though I still use it on my DAW workstation. I suppose both of those are "products" in a very tenuous sense since I make money from them.
650.
by bandrami 9mo ago
I've been using this init for years and always liked it. It's sad the Init Wars ignored it completely.
651.
by bandrami 9mo ago
Isn't getting nobles to actually ionize pretty energy-intensive though?
652.
by bandrami 9mo ago
You verify identity over the now-encrypted channel, just like SSL should have done 30 years ago but refused to for doctrinal reasons. And in the (frequent) cases where you don't actually care about the other party's identity you j
653.
by bandrami 9mo ago
Yeah I'm definitely in the "I keep my laptop powered down and in a Faraday cage" camp so I guess a lot of these products aren't really aimed at me.
654.
by bandrami 9mo ago
This is such a weird thing to wrap my head around. I don't even use Bluetooth.
655.
by bandrami 9mo ago
True but at that point you're not really using "an LLM" in the usual sense, you're using a very resource-hungry database.
656.
by bandrami 9mo ago
> it can't possibly be this hard Whenever I find myself saying this I remind myself it can in fact be this hard.
657.
by bandrami 9mo ago
Right but for an LLM non-determinism is the success mode, not the failure mode. By design you get different outputs when running the same inputs. Which, again, really contradicts the consensus that was widely held even just a year ago a
658.
by bandrami 9mo ago
I still don't understand how otherwise sensible people can have an Alexa or Google Home. Like what part of that seems like a good idea to them?
659.
by bandrami 9mo ago
My advisor in grad school had me implement a "typo distance" metric on strings once (how many single-key displacements for a typist using home row touch-typing to get from string A to string B), which seemed kind of cool. I never
660.
by bandrami 9mo ago
> Your code is now out of sync with the Spec Is there even a sync to be had? The same prompt to the same LLM at different times will yield different artifacts, even if you were to save and re-use the seed.

More ›