tkanarsky

LLM rollout visualizer with a windirstat like UI.

The official Blackpink (kpop group) light sticks from one of their tours have this functionality built in. The annoying part is that one of their Spotify albums have this 18-20khz control track present, and (when I was younger) I could hear it and it annoyed the heck out of me.

This seems surprisingly cheap for what you get! Excited to see what people cook with this

I agree. The cartoons felt literally like the comics brought to life.

PopSci was how I learned to read English as a kid in the early 2000s. RIP to an old friend <3

I always get Access Denied errors when using dark reader on pages hosted by Akamai

Huh, it just hit me that when a binary is "portable" it's "port"-able to a different system. I don't think I ever connected those two meanings before.

that's the max temp beyond which it loses superconductivity, not a requirement. The paper has a graph of max current vs temp, and it gets vastly better at actual room temp of ~300k

In a normal conducting material eddy currents dissipate relatively quickly. The shard of material in the magnet video appears to be floating quite stably for dozens of seconds, which implies that eddy currents are not being dissipated. I'm personally excited, I think it's the real deal but with some limits on maximum field strength and current

Huh, same! I remember the changeover happened sometime around fifth grade, so 2011-12 for our district? (Strangely enough our school's wifi used WEP until 2014ish. I remember the password was obviously contrived to the right length)

That's how Type 2 cables in Europe work, at least for L2 AC charging. DC chargers have built in cables, probably because the wire gauge needs to be a lot thicker to support the current (why haul all that around?) and for water cooling in the case of Supercharger V3s.

tldw:

YouTube channel that published a video series analyzing the Surfside condo collapse now goes over what is claimed to be a leaked transcript of text communications with the doomed OceanGate submarine mission.

Submarine descends too fast, ~38m/min instead of ~25m/min. Support vessel notes discrepancy, sub presses on with descent.

At 9:28a, hull monitoring alarm sounds. Upward thrust is commanded to no effect. Ballast released to no effect. Landing frame released after several attempts, but ascent rate 1/4 of planned.

At 9:38a, crackling sounds heard in aft of ship.

At 9:46:37, power bus A reports failure. Crew switches to power bus B. More sounds heard. This is the last communication received by the support vessel.

-------

If this is authentic, then it basically confirms gradual delamination that snowballs into catastrophic implosion. It's horrible that the crew knew things were going wrong for nearly nineteen minutes and could not do much about it.

How would you begin trying to "solve" something you don't understand the internals of? Much like you can't assess whether a person has ulterior motives just by talking to them over text, how could you be sure the model isn't giving you the answers you want to hear while suppressing its true intentions?

And how would you quantify alignedness? Rating outputs for a given input falls prey to the first problem. Analyzing activations as they trickle through the model is intractable analytically, and training a "polygraph" model on the activations of your network raises more alignment issues (how can you be sure the polygraph isn't lying to you?)

I'm ready to eat my words, but I think perfect alignment is infeasible. The best we can hope to do is curate training data and hope the caged bird won't sing.