what happens when intelligence is too cheap to meter and we can scale it? what then?

When AGI is too cheap to meter. Imagine you had a PhD intern that thought at light-speed and didn't want any money and was willing to do however tedious of a task. You now have that at your fingertips. Take advantage, builder. This won't last forever. Work smarter, not harder.

Why is there such a big deviation in codeforce scores between openai and deepseek? The other benchmarks seem to check out besides this one

DeepSeek's next release will likely claim arc prize.

Prompt: "Explain why integration works using a simple physics problem" Not bad for a zero shot! Gemini Thinking is pretty cool.

going to buy this just so I can say I have one, use will be decided later

it is so over for the next generation

not perfect but not a bad one shot attempt at all from google gemini on deobfuscating this.. they cooked with this model

only thing i don't think i can do at that scale is moveable fingers, zero idea how to do that without like separate tiny servo motors for each finger, but then weight would be an issue. But walking? 3D printed linear actuators and hopefully servos can generate enough torque

logistically building a baymax shouldn't be that hard. encase final layer in super elastic silica resin and have a layer of foam inbetween. some hidden mics that send audio to a local server and attach a speaker and now he can talk to you. camera + lidar + many servos = baymax

Going to get kind of railed on power because my only battery bank i have that would work for this is my one that does 1 port at 5V / 2.1A and another port also at 5V / 2.1A. Apparently under load the servos each need 1A, so i need to use 1 esp32 per servo :/

I placed 5 separate orders yesterday + 5 30pin esp32's w/ wifi $3.09 per + 4 15kg 360deg servos for $14 total going to 3d print a robot soon

does this scam vector actually work?

Wow. Not bad at all for an opensource model. GG

I wonder how if the hallucination rate significantly decrease with the forced retry it seems to do. Interesting idea

Qwen 32B dropped and i asked it the strawberry question and it ended up generating 480 tokens

Only focusing for basic conversation here that sounds realistic. Doesn't need to know almost anything about specialized topics.

How hard would it be to trim a model down? Like Llama 3.2 1B, could I just cut everything about writing software out of the dataset and retrain? I need a good amount of 3.2's running on a single GPU, and I only have a 3070 at home, so I want to try to get 1B down to like 500M