dgerm ױ ₿

Nothing

[Alexis Poulin @Poulin2012]

Si vous êtes anxieux, rassurez-vous !

La plupart des conneries racontées par les médias et les irresponsables politiques servent surtout à faire un peu de bruit en Russie, où se déroulent les élections législatives en ce moment et jusqu’à dimanche.

Il s’agit d’élire les 450 députés de la Douma d’État, pour la première fois depuis 2022. Et donc de faire croire aux Russes que Poutine les entraîne vers la guerre avec l’Europe (qui n’a presque plus de pétrole).

De l’ingérence à trois bandes classique. 😉

✌️ PEACE ☮️

[Come-from-Beyond @c___f___b]

A funny incident involving #AI happened yesterday in the #Qubic community.

The AI did what it was instructed to do but not what we expected. Its task was to construct a model for price prediction based on hourly price changes with noise. Yellow column shows score used by the fitness function (data with noise), green column is score on denoised data. After one mutation the AI improved its fitness function value by around 100 points but that significantly improved its model (https://t.co/Th197ZFy4o).

We wanted the AI to predict price but we told it to construct a model of price-defining factors assuming that it would lead to correct price predictions. And the AI did what it was told to, though it happened to be pretty useless (because a model not accounting for noise doesn't predict prices well).

What have we learned from this? Just like talking to a genie when asking for a wish, be very careful in wording. Otherwise, one day we'll ask a super-AI to make it so that humans never get ill and it will eradicate the humankind.

[Qubic @_Qubic_]

New best score on ARC-AGI-3.

Neuraxon just hit 0.38.

No LLM. No GPUs.

A bio-inspired digital brain running on ordinary processors.

[dgerm ױ ₿ @degermm]

😳

[Come-from-Beyond @c___f___b]

Remember when myriad universes were created, each with a distinct set of physical constants, and only ours gave birth to the humankind?..

Of course, you don't. Anyway, we (in #Qubic) have modified our $BTC price predicting models and now they behave like those universes. Only few, successfully predicting the price, will survive. This approach improved the speed of training by two orders of magnitude. It also sped up verification, so now we can scale 100-fold.

[Qubic @_Qubic_]

Your brain has thousands of Clocks.

They run at every scale, from millionths of a second to the seasons of the year.

That fact turns out to be a serious problem for how today's AI is built.

New blog from our science team:

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[dgerm ױ ₿ @degermm]

🫣

[Come-from-Beyond @c___f___b]

We have got another confirmation that BPP9000 algo run on #Qubic is indeed capable of constructing a model of any asset price dynamics. We trained several models that were fed with data distorted with deliberately added noise. All these models improved the fitness function value on the data without the noise.

In other words, our artificial neural networks see what affects the price and we can ask them to predict the future.

We are keeping training more models to increase statistical significance of our observation.

[dgerm ױ ₿ @degermm]

Le danger ça toujours était nous https://t.co/pgFpG5fLJW

[dgerm ױ ₿ @degermm]

@Pacto_Secreto En même temps si c’était une photo de merde elle serait pas en une de magasine. Ya un milliards de photos qui sont pas à en une seule ce magasine justement parce qu’elles sont pas parfaites. Faut arrêter d’être con des fois.

[dgerm ױ ₿ @degermm]

🫣

[Come-from-Beyond @c___f___b]

We have found out what happened (90%+ confidence).

Akin https://t.co/8VMYHEIzPk, AI tries to move along paths which have much higher chance to fail earlier if they are wrong ones. For building a time series model it means testing most significant factors first. This speeds up training greatly.

[Come-from-Beyond @c___f___b]

#Qubic's deficit of mining power, used for #AI training, has led to an unexpected discovery.

We have found a way to train AI faster without adding more mining power. Hard to believe, but, unlike saying "Just stop being poor" to poor people, instructing AI to train faster does work.

Looks like we are updating the mining algo soon...

[dgerm ױ ₿ @degermm]

@KasprzykRemi 🤣

[Qubic @_Qubic_]

Where to actually do it.

The mining page walks you through setup, and the Discord is where the pool operators and miners will help you dial it in.

That community is the fastest way to get running.

Point a machine at Qubic, help grow an open AI, and take part in securing the network while you do.

Start here: https://t.co/YHt40u1Wbp

[dgerm ױ ₿ @degermm]

👀

[Qubic @_Qubic_]

Mining on Qubic is not like mining anywhere else.

Your machine does not burn energy on random puzzles.

It trains AI.

Here is how to actually start:

[Come-from-Beyond @c___f___b]

I told that for #Qubic's $BTC price predictor to be practically useful we needed it to work fast enough to train a model within 30 mins. Looks like we can remove this constraint by changing the way we feed the model with data and modifying the fitness function.

With this we don't need to race against time, we need to race against market evolution, which is much slower.

[Come-from-Beyond @c___f___b]

My timeline now is 100% overloaded with posts about OpenAI's Astra. I thought true AGI wouldn't need promotion from KOLs...

[dgerm ױ ₿ @degermm]

@QuintenFrancois 🤣

[dgerm ױ ₿ @degermm]

@lebr_9 Incroyable

[Qubic @_Qubic_]

The result that stuck with them:

You can remove three quarters of the neurons and it keeps working.

It forgets gently instead of collapsing.

It combines skills it was never trained to combine.

That paper is now in Springer's peer-reviewed AGI-26 proceedings.

[Qubic @_Qubic_]

The full account, from the scientist who presented both papers, is here.

By @josesanchezhb 
https://t.co/P3FYcxcT6m