Andre Infante

Making games and robots and sometimes other things.

(Formerly: 1X, Meta)

[Andre Infante @AndreTI]

I wrote my first substack piece about AI's economic impact, and some reasons why models seem to underperform in the real world.

[DiscussingFilm @DiscussingFilm]

First trailer for ‘ALIEN ISOLATION 2’.

Coming soon.

[Andre Infante @AndreTI]

@Brien_Jackson There's basically no downside to expelling this flavor of person from the coalition. Their collective votes are straightforwardly not worth the toxic policy advocacy, moral bankruptcy, and incredibly poor judgement.

[Andre Infante @AndreTI]

@steady_drumbeat @Brien_Jackson I think the thing you said is pretty fucked up.

[Andre Infante @AndreTI]

@BartronPolygon It does, but so far I've found the acute impact to be larger than the net impact. What's been working so far has been light breakfast, have a banana shortly before working out, have a big lunch right after working out, have a medium dinner. Just move calories around.

[Andre Infante @AndreTI]

So far, my take on dieting is that the standard advice that exercise basically doesn't matter because it's easier to just eat less is a little bit wrong / incomplete. Eating a few hundred fewer calories is often a matter of cutting out soda, desserts, butter and mayo. Not too bad

[Andre Infante @AndreTI]

@joseph_h_garvin @pathsnotchosen Vision seems a little worse than generative models, but progress in all areas over the last five years is incredibly impressive, even if you'd never seen a benchmark in your life.

[C. Zhang @ChongZzZhang]

Holy crap

https://t.co/Uq2jGHnxkS

[Andre Infante @AndreTI]

@derektraver I'm deducting 20% from the fitness tracker number to be conservative. Cardio does make you hungry, but if you schedule your calories so you eat a big meal right after, impact (anecdotally) is about one fourth what you burned.

[Andre Infante @AndreTI]

I think for most people, these lines cross pretty quickly and the marginal calorie becomes easier to get rid of via exercise than eating a yet sparser diet.

[Andre Infante @AndreTI]

You don't need to cut that much before cutting more becomes painful, either because of flavor and variety or hunger or both. And if you spend some time getting proficient at intense cardio, you can burn like 600 calories in an hour of running. That's an extra per diem burger.

[Andre Infante @AndreTI]

@joseph_h_garvin "If you're going claim random meat in a human skull is conscious you need to explain why every can of spam ever sold isn't conscious."

[Andre Infante @AndreTI]

@joseph_h_garvin That's stupid. It's very possible that some computable programs are conscious and some are not. Just because the computation is handled with matrix operations doesn't tell you anything about its contents. And neural network weights are not random at inference time.

[Andre Infante @AndreTI]

@SereneInvesting @citrini Yes! Because those models have a lower error rate at all tasks (including simple one) and human time is still much more expensive than robot time. Also if you don't, you can use a smaller API model. Size is orthogonal to localness.

[Andre Infante @AndreTI]

@citrini Local inference only looks cheaper because you're using massively less hardware to run vastly stupider models. But the unit economics per flop are fundamentally worse.

[Andre Infante @AndreTI]

@citrini That's stupid. Local inference is more expensive, not less. Inference cost is power and hardware depreciation. Power is the same, but local hardware sits idle when not in use. Data center hardware is used ~all the time. Hardware depreciation is amortized over more users.

[Andre Infante @AndreTI]

@Strife212 The graph paper brain would obviously be conscious. It would talk about consciousness for the same causal reasons the original brain did. So these Chinese Room intuition pump shenanigans do not constrain the question of LLM consciousness.

[Andre Infante @AndreTI]

@Strife212 Neurons follow predictable firing rules / functions. We don't understand all the behavior yet, but it's a physical process, a mix of noise and determinism. You could work it out on paper and predict a person's behavior. You'd die long before output, but same is true for the LLM.

[Andre Infante @AndreTI]

@zmichalo39 @Bearspacegoat @hecubian_devil I don't think "it's all jew lies" is a great response to multiple allegations of domestic abuse.

[Andre Infante @AndreTI]

@chungnun @haunted0clock Honestly it's hard to pin down why it didn't work for me when the very similar turn in The Shining did. I think it helped that The Shining had the typewriter gag, which puts scenes from earlier in the film in a new light.