foota

No reason other than you're giving user's a cost footgun? That's a pretty good reason.

What? This is the opposite of regulatory capture. Neither anthropic nor openai are getting to choose what happens with their models.

I find the way that models understand images to be seriously lacking. The root cause of the issue as I see it is that image encoding isn't contextual. The encoder should be aware of the prompt so that it can encode the right things. It seems like this should be something that could be trained into a model.

It feels like this would be beneficial to give the model more of a deep understanding of visual knowledge.

Weird justification since 9/11 was a plane hijacking, not something on the streets.

What if life was born under the surface?

Fun fact: I once got a security bounty because they sent the 2FA emails through click (some email monitoring SAAS thing) with "view in web" enabled, and it was set up so that the emails under a given template used an auto incrementing ID, so you just had to request a 2FA email and then access it through click's web UI.

This is an internet comment forum, not my team's slack chat, it's ok to have some fun with creative ideas.

In theory the new futures markets for chip components would help here, since it would allow DRAM suppliers to insulate themselves from that risk.

I think you're misunderstanding the tank's temperature. The internal temperature is 32C, the external temperature is only 16C (they used thermal cameras to measure the external temp, but when they got closer they read the internal gauge which showed 32C). Groundwater in LA is apparently right around 16C, so the difference in temperature between the tank and the water would be basically 0.

So while I do agree that the evaporative cooling is probably doing most of the heavy lifting (in fact, necessarily so if the water temp ~= the tank's exterior temp), it's not unreasonable to suggest that using colder water would be more effective.

In fact, the wet bulb temperature there is apparently right around 16C, so they've basically cooled the outside of the tank as much as they can using evaporative cooling alone. They can certainly use it to _keep_ it there, but without something else they wouldn't be able to cool it further.

Presumably if there's an exothermic reaction happening internally then the core will continue to rise in temperature based on the temperature gradient through the material forming in the tank. I would assume (since my understanding is that it's some kind of plastic) that it has a fairly low thermal conductivity, so the core temperature will continue to rise as more of it turns to plastic even as the outside is cooled to the same ~16C.

In the limit if they were able to immerse it completely in very cold water (~0C) then the exterior of the tank would also be ~0C (supposing they were able to access sufficient quantities of water). I don't think that's practical of course, but again I don't think it's unreasonable to suggest that cold water would make a meaningful difference.

Yeah that's fair. It does make me wonder what there is that would allow this. Maybe importing some snow machines from Tahoe? :-)

I'm actually a little surprised that a quick search for "refrigerated liquid transport" didn't turn up anything. I would have sort of thought this was something that would exist (just because there are so many random things that are necessary for _some purpose_).

Heat exchange is proportional to the difference in temperature though (in reference to your "or heat exchange"). Colder water would cool faster. The tank isn't at boiling either, so it's not like you'll be able to phase change away a bunch of energy. I guess you'll still get some evaporative cooling, but there's a limit to how much you'll get just from the ambient temperature (the exterior of the tank is relatively cool, presumably because the "gummed up" interior is inhibiting heat transfer)

Ah, that makes sense. It's too bad they can't drill into it to relieve pressure without destroying the integrity of the tank (not that I'd want to be anywhere close to it either).

If they didn't have to worry about it imminently exploding I wonder if they could somehow wrap it with reinforcement (e.g., wrap some high strength metal around the tank to prevent it from deforming when drilled into) and then drill into it to extract the liquid?

One of my other less serious ideas was to helilift a Chernobyl style containment structure around it, but I imagine they don't have one of those just sitting around waiting to be used.

They are not. I said ice cold. I read this article and several other articles about this.

Is it not possible for them to just... spray it with ice cold water?

I was thinking about this just the other day!! Frequently I find myself wanting to interleave thoughts in a conversation, this is the natural way to do so. I just wasn't sure how you'd train it.

They should feed it the classification of finite simple groups and get it to simplify it/turn it more constructive.

Sure, I'm frustrated by the github outages too, but hacking into github to fix their code seems like a bit of an overreaction.

Smells to me like they didn't implement compaction/went past their context window and the system prompt dropped off the end.

I would love to have some real application that needs an FPGA :) Someday perhaps.

Ah, I didn't know that existed. TIL

One wonders if someday we might be able to resurrect the neural network from dead cells by somehow reviving the connections between neurons. I imagine that the connections stay, but become dormant when the neuron dies.

I've generally assumed that AI would make developers get lower compensation because of the lowered quantity of developers required for the same output, but this raises the possibility of it actually increasing if more developers end up doing their own things instead of entering the broader labor market :)

Imo the main benefit of AI is allowing faster experimentation and more parallelism, as well as faster iteration.

Shouldn't there be a feedback system here preventing the scheduling of loads when cooling is degraded?

I wonder if this is in response to the chrome incognito lawsuit.

cries I am sad at all of the cool designs that would work with optane if it had been widely adopted. We desperately need more points on the latency/cost per GB curve.

This doesn't math out to me just based on what I know of energy consumption numbers.

"is duplicated by" is amusing.

While I largely agree, it does raise the prospect of testing this iteratively. E.g., give a model some fake environment, prompt it random things until it does something "bad" in your fake environment, and then fix whatever it claims led to its taking that action.

If you can do this and reliably reduce the rate at which it does bad things, then you could reasonably claim that it is aware of meaningful introspection.