Does the bill only apply to chatbots marketed as such or does it cover generalist tools as well?
Like, wondering where the UPL line is between "Claude, find all uses of units of time in this contract" and "Claude, how does our one month warranty work for purchases in Feb?"
Relevant mini-thread: bsky.app/profile/lizt...
This is news to me, that blocking a labeller account inhibits operation in some way. As far as I know, appviews will pick up labels emitted by a labeller even if the subject is blocking that labeller...?
If that's wrong, that basically means moderation via 3rd party labellers is unenforceable 😨
An ICE detainee in Arizona has died of a TOOTH INFECTION after it went untreated for weeks, a local official says. He was a Haitian asylum seeker imprisoned in Florence, Arizona. @emilybregel.bsky.social reports.
tucson.com/news/local/b...
Something that didn't really matter but that we cared a lot about at Slack was the proper use of quotes and ellipses. So I finally got around to making (or asking Claude to make) an ESLint rule with a fixer for that.
If you must analogize to an organism, an ant colony might be more appropriate? Like most of the LLMs people interact with come across as singular entities but are really distributed systems stretching across time and space in weird ways?
They're not being teleoperated in a someone driving remotely sense, but ... are we sure that a human didn't intervene after 30 seconds? Like I could imagine a remote operator saying "this car goes first".
For all the rhetoric about how Trump has failed to sell or justify this war to the American people, it *has* been sold, to both the military and the general population, for decades now.
If you search for "War Powers", you'll see a few posts from Democratic Congresspeople, like this one. Unclear if there are enough votes to force anything though. Or if a successful vote would do anything.
Yes, that's supposed to be the case. The counter argument is that as Commander in Chief, the president has some sort of inherent authority to act in emergencies, although that's hard to claim when we shoot first. There's also the War Powers Resolution, which limits to 60 days without approval.
even if anthropic's systems were not responsible for 100% of the kill chain, they rejected being part of a fully autonomous kill chain even with other components from other suppliers supplying the other elements of the kill chain.
and tbqh that's how we felt as Google employees as well re Maven
It is sort of funny though that an admin that's stepped all over the Constitution somehow still feels bound by some company's terms of use.
end of history (derogatory)
@caltrain.com has won a national award for its video last summer that dunked on the Dodgers.
The clip, filmed near the agency’s San Carlos offices, was designed to encourage fans to take the train to Oracle Park. It drew more than 250,000 views on TikTok and Instagram within 24 hours.
Didn't some of them survive?
Maybe I'm not following your definition of "ship" or "bad" but sometimes you do?
Like, if you make a WYSIWYG editor, there are some painful lessons you don't learn until you ship to users using AZERTY keyboards, or CJK or RTL languages, or screenreaders.
I'll acknowledge thus POV is eng-centric. Product and design are still doing a lot of work that's not covered by shipping slop prototypes. But mitigating sunk cost fallacy for 1/3 of the EPD triangle is really useful IMHO.
Here's the rub: That all takes a fair bit of work. And once you do that work, you have to deal with the sunk cost fallacy of not wanting that work to go to waste.
On the flip side, it's much easier to let go of AI code slop.
The hypothesis for Huddles was "we can't replicate all of the features in Zoom but maybe a nice UX will make our power users prefer it for *some* use cases".
But to answer that, you need to ship something with a minimal amount of polish (like not dropping audio).
Slack is (was) this way. We'd ship stuff internally constantly. And based on internal feedback, we'd change or even cancel features. Like, for Huddles, the only way to know if it was useful was just to try using it for real calls for a week.
I'm going to gently push back on this. There's value to a product and design process where using a working prototype on a regular basis helps inform what's worth shipping.
And "no amount of bad code adds up to good code" is untrue IMHO. Sometimes bad code is how you know what good looks like.
Anchor positioning works really well for the "label in the place of a placeholder" pattern, except instead of (ab)using a placeholder, you can use a real label.
...the stock goes up not because there are actually that many stock trading agents acting on said prompt injection but because *people* see the prompt injection and think agents will buy said stock and thus trade on it first themselves.
This newsletter is not about LLMs but it did make me think of a securities fraud hypo where at least some people are using agents to trade stocks, other people pollute the web with prompt injection attacks designed to pump a meme stock, and (cont...)
@effies.bsky.social !
So the inverse of @masnick.com's argument: The Internet is already structured around a small number of big players. And they're risk averse. Even if the 1st Amendment permits it, big cos will likely restrict beyond that, e.g., ContentID isn't exactly friendly towards fair use.
But there's only one iOS App Store. Even if you accept the duty of care associated with your mushroom app, Apple didn't. And everything we know about Apple tells us they'd probably react to mushroom app liability by booting all the mushroom apps, not just the bad ones.
My pushback is that with the natural development of law, you end up with (50ish?) divergent outcomes for, say, a mushroom ID app and a bird ID app. And you'd probably say that's a feature. You *should* exhibit a higher duty of care with mushrooms than birds. <insert angry goose joke here>
The only upside to Salesforce turning Einstein into a Brand™ is that they get to C&D this nonsense.
