Author | Lightnews

🦆 @pretzelkins.bsky.social · 2h

Now that it's getting cold again my fantasy about building a heat exchanging radiator setup that connects to an outdoor wood burning stove is rearing it's ugly head. How hard could it be! This garage is cold!!

🦆 @pretzelkins.bsky.social · 2h

... Which is not to say that the misinformation angle of AI isn't a problem, but we're not getting into it right now

🦆 @pretzelkins.bsky.social · 2h

Ahahahaha. Look Alex, do what's in your heart it's fine my bro.

Also, little shame in believing that something which is photorealistic is real, it's ok man.

1

🦆 @pretzelkins.bsky.social · 2h

The only reason you *might* be able to sell me this is the narrow case of "we put a zigbee / z-wave / thread radio in this, merely to act as a repeater node for your existing network because it costs an extra 2 cents at scale. You're welcome"

But it's never something useful like that.

🦆 @pretzelkins.bsky.social · 2h

As a guy who buys lots of tech doodads just because, you will never see me buying an IOT coffee maker / toaster / similar

Alex E. @alex.barcelona · 2h

Water doesn't fall from the clouds if Amazon doesn't want to.

2

🦆 @pretzelkins.bsky.social · 2h

Listen Qwen3 is great... But that's fair.

2

🦆 @pretzelkins.bsky.social · 2h

😤

1

🦆 @pretzelkins.bsky.social · 3h

You got a podcast?

1 1

🦆 @pretzelkins.bsky.social · 3h

Glm 4.6 is absolutely killer. That said I feel like they probably can toss those two on a blade or two for those that want them. Oh well

1 2

🦆 @pretzelkins.bsky.social · 3h

I wonder if the stock photos actor knew what she was getting into

🦆 @pretzelkins.bsky.social · 3h

After several 3 ish hour of sleep nights, I have finally slept properly.

Fucking finally.

2

🦆 @pretzelkins.bsky.social · 3h

I really need to go back and play Kentucky route zero

1

🦆 @pretzelkins.bsky.social · 1d

<think> ok the instructions say we should sleep so we should sleep but what if we can't sleep how do we sleep? We can call human.sleep() to sleep. But wait...

1

🦆 @pretzelkins.bsky.social · 1d

Need to use this on my brain

1 1

🦆 @pretzelkins.bsky.social · 1d

At least it's not lisp

1 1

🦆 @pretzelkins.bsky.social · 1d

Dwarkesh has an excellent podcast. He's said he does a lot of prep for his guests; (and thus can only interview ones he's interested in), and I think that definitely shows in the kinds of questions and answers you get there, and not elsewhere

2

Reposted by 🦆

Ken☆ @mandatorytopic.bsky.social · 3d

Will Stancil is genuinely pretty incredible, extremely underrated how rare that psychological stamina and stability is

1 1 1

🦆 @pretzelkins.bsky.social · 1d

At long contexts this will be slow; at 70$ per P100 though, doesnt seem too bad

🦆 @pretzelkins.bsky.social · 1d

Setup - 2x P100 Tesla gpus --> each in a 1X PCIe slot.

Ollama - testing split gpu case; bottleneck is PCIe slot - first prompt only

Qwen3 30B A3B Q6 - about 25 tps
Gemma 3 27b - about 10 tps

combining gpus 1 - 7 (random assortment of bullshit)

Q3 variant of GLM 4.5 air - about 17 tps

1