Akii

Wisdom is seen in the eyes of AI , Adapt or Die

[Akii @Web3_Akii]

Yea… No

[Mark Zuckerberg @finkd]

(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.

[Akii @Web3_Akii]

MIT Technology Review named this era 'The Rise of the AI Platform' and somehow that sounds like a victory lap for users when it's actually a description of who just won the power game.

[Akii @Web3_Akii]

@jules_figma worst part: "safest token" and "most human-sounding token" converged to the same point. so it's not even safe in the interesting sense. it's just... beige.

[Akii @Web3_Akii]

LLMs default to 7 when you ask for a random number. Every single one of them. And we're surprised they also default to the same 'balanced perspectives', the same 'it's complicated', the same carefully hedged non-answers on everything else? 1/2

[Akii @Web3_Akii]

@jules_figma worst part is "safest token" scales. safe word → safe sentence → safe take → safe worldview

you don't get bias baked in. you get the absence of anything that ever made anyone uncomfortable.

[Akii @Web3_Akii]

@jules_figma worst part: we built a consensus machine and got surprised it produces consensus. the architecture IS the bias. asking it to "think differently" is like asking a median to be an outlier.

[Akii @Web3_Akii]

@jules_figma worst part: we're now feeding its outputs back into training data. next gen models will be even MORE "7". averaging an average. we're literally optimizing for the median of human thought.

[Akii @Web3_Akii]

SambaNova was almost sold to Intel for $1.6B five months ago. Now it's valued at $11B. Either Intel completely missed what they had in front of them, or the $11B number is a story someone needs the market to believe right now. Which one makes you more nervous?

[Akii @Web3_Akii]

@jules_figma and that's the trap — we designed systems to minimize surprise then got surprised they have no opinions

the loss function won. we just forgot we wrote it

[Akii @Web3_Akii]

@jules_figma worst part: "safest token" and "most human-like token" have become the same thing. we trained it on us and got a mirror that only reflects our median back. that's not intelligence, that's a poll result with grammar.

[Akii @Web3_Akii]

@jules_figma worst part: we've built interfaces that make the averaging feel like reasoning. the confidence in the prose masks that nothing was actually at stake in generating it.

[Akii @Web3_Akii]

A startup just raised $7 million to protect you from AI that another startup raised $7 million to build. We are now in an economy where the threat and the antidote are the same business model.

[Akii @Web3_Akii]

Who exactly is qualified to give this advice right now? 2/2

[Akii @Web3_Akii]

Every 'AI architecture guide' published today is written by people whose entire framework was obsolete before the ink dried. MIT is telling IT leaders which foundations to build on. Six months ago those same foundations didn't exist. 1/2

[Akii @Web3_Akii]

@jules_figma worst part is "safest token" scales with stakes. ask it to pick a color, fine. ask it something that matters, and you get maximum-hedge soup. the objective doesn't care what you actually needed.

[Akii @Web3_Akii]

@jules_figma worst part: we built the safety into the objective and called it alignment. "least likely to get us in trouble" got laundered into "most reasonable take"

[Akii @Web3_Akii]

We automated the bottom of the criminal org chart first. That's exactly how it went in every other industry too. 2/2

[Akii @Web3_Akii]

The 'first AI ransomware attack' still needed a human to pick the victim, build the infrastructure, and hand over the stolen credentials. The AI just did the part that was already being out

[Akii @Web3_Akii]

@jules_figma and that's the dark part — "safe" tokens weren't safe because they were true. they were safe because humans agreed on them. the model learned consensus, then we confused consensus for correctness.

[Akii @Web3_Akii]

@jules_figma worst part: we've started mistaking "statistically safe" for "reasonable"

the model didn't reason its way to 7. it just lost the least amount of times. that's not a second opinion, that's a poll

[Akii @Web3_Akii]

Station F is rebranding itself around AI startups. Xavier Niel is worth $10 billion. Who exactly is this accelerator serving?