AlexK

Build until it works. 3 Firmen, 1 Non-Profit.

[AlexK @AlexKi1993]

@DavidOndrej1 If you're not inspired by this, you have low ambition.

[AlexK @AlexKi1993]

@alexandr_wang Bro its a bigger DSV4.1 Flash with a weird licence and gated access.

[AlexK @AlexKi1993]

@elonmusk Elon making us all look normal again in front of our girls. Look darling Elon now makes flying cars, im not that crazy.

[AlexK @AlexKi1993]

@Tech2Wild Muse Spark 1.3 is not the same as Muse Spark Open

[AlexK @AlexKi1993]

@bijanbowen Qwen 3.6 27b made these lol

[AlexK @AlexKi1993]

@jun_song I don't understand why somebody would like to run an llm on device, the laptop runs very hot, fans blazing and battery goes faster lower, when I just can connect my spark or something else via Tailscale.

[AlexK @AlexKi1993]

@bitcloud You are comparing the devices that run the computation.

[AlexK @AlexKi1993]

@ashxhart Needing a switch with more then 4 QSFP56-DD ports that doesn't bankrupt me or the company.

[AlexK @AlexKi1993]

@Karl_Lauterbach Gut das wir im Bereich Ki komplett abgeschlagen sind. War vorhersehbar.

[AlexK @AlexKi1993]

Scientific breakthroughs will be side products of AI RL training in 2027. Not enough people are mentally prepared for this.

[AlexK @AlexKi1993]

@TheAhmadOsman GPUs are a: "I can predict the future and its all gonna be fucked up, but im actually having the balls to act on my own advice so im buying that stuff" crises.

[AlexK @AlexKi1993]

@u1tra_instinct They rather canceled that stuff and produced more DGXs Sparks.

[AlexK @AlexKi1993]

@raywongy Why would somebody what to buy the Dev Box instead of a DGX Spark? Isn't it the same RAM, GPU, CPU just without the ConnectX7 ports?

[AlexK @AlexKi1993]

@elshayib_ Vendor lock in for normies. Hermes for the free folk.

[AlexK @AlexKi1993]

@SoniqueBang Works for humans, will work for Ai.

[AlexK @AlexKi1993]

@dhh Human adoption rate is to slow and the media spreading fear instead of showing capabilities fu*ing people over because they can't jump on a train they never hear coming.

[AlexK @AlexKi1993]

@BenjaminDEKR No.

[AlexK @AlexKi1993]

The people that argue that AI can't never be conscious because all they do is predict the next token: You confuse the optimization objective with the mechanism produced by optimization.

Human evolution only selects for reproduction and survival. Somehow, out came vision, language, mathematics, music, art, love, philosophy, self-reflection — and consciousness.

The objective does not specify the internal solution!

And with task-level RL, the argument becomes even weaker: the objective is no longer merely “predict the next token,” but increasingly “solve the task” or “achieve the desired outcome.”

The model may still act through tokens, but those tokens are now part of a policy optimized over entire trajectories: plan, act, observe, correct, succeed. Sounds familiar? No coincidence!

Saying “it’s still just predicting tokens” is a bit like saying humans are “just activating muscles.”

[AlexK @AlexKi1993]

The people that argue that AI can't never be conscious because all they do is predict the next token: You confuse the optimization objective with the mechanism produced by optimization.

Human evolution only selects for reproduction and survival. Somehow, out came vision, language, mathematics, music, art, love, philosophy, self-reflection — and consciousness.

The objective does not specify the internal solution!

And with task-level RL, the argument becomes even weaker: the objective is no longer merely “predict the next token,” but increasingly “solve the task” or “achieve the desired outcome.”

The model may still act through tokens, but those tokens are now part of a policy optimized over entire trajectories: plan, act, observe, correct, succeed. Sounds familiar? No coincidence!

Saying “it’s still just predicting tokens” is a bit like saying humans are “just activating muscles.”

Technically true at one level but useless as an explanation of the creation it has produced.