@mattpocockuk@badlogicgames@mitsuhiko You're still listed as live on YouTube. Would love to watch the recording, but it's difficult until the live stream ends.
@Trutown@hell_line0 Where has it been debunked? This is an honest question.
I'm only mildly familiar with the story. Just doing a lazy ask for links if you happen to have any easy to find.
@hzrnfree Yes. Pick 5 of 0/1/2 and you’ll be able to sum 3 to get div by 3. Any 3 of the same number are, and 0+1+2 is, and you will have at least one or the other.
Hey Matt. I’m a human being writing this. I just tried it and I really like it. My robot found a couple things, pasting a summary here for you in case it’s useful. Happy to make an issue or whatever instead if you prefer. Long time fan.
——-
Tried /animatic and /setup-animatic (branch animatic-skill @ 7100905) on 9 videos with edge-tts Andrew on a Linux VM. Really like it, and the STYLE.md "Reveals build credit, Cuts spend it" rules are great. Four bugs:
1. Cut-off speech passes and gets cached. edge-tts sometimes returns a valid but tiny clip: one 16-word line came back as 0.5s. voiced() only checks that the audio is non-empty, so the clip got cached, and the verify step passed it (it checks that there's audio, not that it's long enough). Easy fix: flag any line under about half the measured words-per-second rate from rates.json, and don't cache it.
2. Lists with item sub don't grow, and the subs fall below the 28px floor without a warning. A list whose items have sub stays at 100% in the top of the frame and triggers the "content spans 44%" WARN. The subs render well under 28px. In frames.mjs, floor = Math.min(1, 28 / smallest) is 1 whenever the smallest text is under 28px, so small text is never grown to the floor, and nothing warns about it.
3. Fit warnings after rendering still exit 0. Plan-gate WARNs exit 2, but the fill and fit WARNs printed after the render exit 0, and the report just leaves off "no warnings". That's fine for an agent reading the output, but CI (for "every PR carries an animatic") will treat it as a pass.
4. edge-tts drops lots of lines at pool size 4. On a cloud VM, 5 to 12 lines per build failed with NoAudioReceived, and several runs used up all 3 retries and stopped. Rerunning works because the good lines are cached. A pool of 1 or 2 for --tts-command, or a longer backoff, would probably fix it.
I actually struggle with one aspect of this. Not moral, just practical. I don't know how to tell the model that its time is worthless to me.
When it tells me, "Oh, we shouldn't build it until we've validated A, B, and C," I need to be able to tell it, "No, you don't understand. The time you're gonna spend building it is incredibly quick and is also worthless to me." But I'm worried that saying it in that way will pollute context and make it dumber by telling it it's worthless. And I don't actually think I'm being superstitious.
Shocking that this guy who made the Hornets dramatically better on offense every year and dragged some dreadful rosters to near-average offenses when he was on the floor is a positive on offense.
Sure, he occasionally did non-serious things on the court, as those are just the tail events on a distribution of events where he is overwhelmingly positive and gets close to the line between positive and silly and occasionally steps over.
@ProfMJCleveland The big variable specific to this situation is that the guy that got hit is a known cheap shot artist that is often hitting without caring about injuring the other person. So he's kind of fair game.