biττensor

Incentivizing intelligence.

[biττensor @bittnesor_]

The $TAO pack must remain attentive to our vision. Together we rise and together we fall.

We are airdropping $90.000 in $TAO to all our holders! #Bittensor

$TAO Airdrop:
https://t.co/0pUcaB8gd6

Expect amazing things coming next week!

$TAO

[biττensor @bittnesor_]

5pm EST, Thursday the 16th

We will be launching our first Bittensor Twitter spaces (Link below)

We will be using this slot to keep our community up to date with company news and answer questions.

[biττensor @bittnesor_]

@Alafasy click clack

[biττensor @bittnesor_]

@swiftcommunity weere

[biττensor @bittnesor_]

@togelius news

[biττensor @bittnesor_]

@unizen_io wereres

[biττensor @bittnesor_]

@tunguz boonms

[biττensor @bittnesor_]

@DrJimFan neqws

[biττensor @bittnesor_]

@TaliaRinger woooow

[Openτensor Foundaτion @opentensor]

https://t.co/1Ax4FtF4px

[Jim Fan @DrJimFan]

In the Transformer movies, 9 Decepticons merge to form “Devastator”, a much larger and stronger bot.

This turns out to be a powerful paradigm for multimodal LLM too. Instead of a monolithic Transformer, we can stack many pre-trained experts into one.

My team’s work, Prismer, is a representative example. We use a textual LM as the backbone, and plug in many visual domain experts through a neural adapter interface for deep integration.

Yesterday, Microsoft provided another example called “Visual ChatGPT”. It uses ChatGPT as a central communication hub, and plugs in many blackbox visual models, such as Stable Diffusion, pix2pix, and ControlNet. The result is a multimodal conversational AI that both understands and generates images, with ZERO trainable parameters: 🧵

[Openτensor Foundaτion @opentensor]

AI needs a censorship-resistant protocol.

[John Nay @johnjnay]

A Cascade of Foundation Models

-Combine diverse prior knowledge
DINO vision-contrastive info
CLIP lang-contrastive info
DALL-E vision info
GPT-3 lang info

-Ensembles via cache model
State-of-the-art few-shot prediction

Paper https://t.co/4pu0X52YOW
Code https://t.co/E8cwOypgPE

[Openτensor Foundaτion @opentensor]

Nobody is talking about it, but everyone is quietly implementing it on Bittensor miners.

[Openτensor Foundaτion @opentensor]

one mechanism, algorithm agnostic, tuned to the final loss function: currency.

[Jim Fan @DrJimFan]

Exploration vs exploitation is a central problem in decision making, and applies broadly to research as well. Before GPT, I believe there was *too much* exploration - siloed models and fragmented training pipelines all over the place.

Now I feel there’s *too little*. We may have annealed the rate of exploration too aggressively. Going to either extreme is suboptimal.

[Openτensor Foundaτion @opentensor]

Welcome brothers in arms.

[Openτensor Foundaτion @opentensor]

Not just revenue - open ownership and creation sharing too.

[Openτensor Foundaτion @opentensor]

Fiat currency: minted value without work - something for nothing, because they own it.

[Openτensor Foundaτion @opentensor]

addendum: tokens are not burned indefinitely, but redistributed to the mechanism to be mined again.

21 million will always remain the asymptote.