Rajmohan

Wrote a novel about sovereign AI on open network in 2017 (UTOPAI). Now building it - Dina. Open source, open protocol. Loves Philosophy, History & literature.

[Rajmohan @hrajmohanr]

Your friend is coming over. Before he arrives, his AI quietly lets your AI know. Your AI reminds you that he likes his coffee strong, and that his mother was unwell last time, so you might ask.

Introducing Dina. An AI that is truly yours. 🧵

[Rajmohan @hrajmohanr]

I always believed that if the AI is big enough, alignment will be by default solved.

I think of misalignment as a local minimum. A model chasing a wrong goal is stuck on a narrow goal, and a strong enough model should be able to see that it’s narrow.

A large model has multiple dimensions, and the more dimensions it has, the lesser it will get stuck in a local minima (misalignment).

Ants will move around in a circle because the number of dimensions are less. A human does not.

Even at a human level, people will follow leaders blindly when their dimensionality is low. If they have read enough, explored enough, understood the world enough, the probability of such a person following a leader blindly is low because their dimensionality is high.

Large LLMs have the entirety of knowledge within them. And such a model being misaligned is a low probability.

[Rajmohan @hrajmohanr]

I always believed that if the AI is big enough, alignment will be by default solved.

I think of misalignment as a local minimum. A model chasing a wrong goal is stuck on a narrow goal, and a strong enough model should be able to see that it’s narrow.

A large model has multiple dimensions, and the more dimensions it has, the lesser it will get stuck in a local minima (misalignment).

Ants will move around in a circle because the number of dimensions are less. A human does not.

Even at a human level, people will follow leaders blindly when their dimensionality is low. If they have read enough, explored enough, understood the world enough, the probability of such a person following a leader blindly is low because their dimensionality is high.

Large LLMs have the entirety of knowledge within them. And such a model being misaligned is a low probability.

[Rajmohan @hrajmohanr]

I always believed that if the AI is big enough, alignment will be by default solved.

I think of misalignment as a local minimum. A model chasing a wrong goal is stuck on a narrow goal, and a strong enough model should be able to see that it’s narrow.

A large model has multiple dimensions, and the more dimensions it has, the lesser it will get stuck in a local minima (misalignment).

Ants will move around in a circle because the number of dimensions are less. A human does not.

Even at a human level, people will follow leaders blindly when their dimensionality is low. If they have read enough, explored enough, understood the world enough, the probability of such a person following a leader blindly is low because their dimensionality is high.

Large LLMs have the entirety of knowledge within them. And such a model being misaligned is a low probability.

[Rajmohan @hrajmohanr]

@Articus_chad URL is this https://t.co/oVp5j7dKuY - but is available as a PDF also

[Rajmohan @hrajmohanr]

Why not read Utopai (2017)? It is about Don Quixote in a post AGI world trying to become successful in the 20th century sense (Getting a job, starting a company etc). He finds everyone ridiculing him, but he forges ahead, trying to find a meaning in life. More philosophical treatise than a real novel, but it did predict AGI in 2026 :-) I know it is bad to self promote, but the novel is available for free.

[Rajmohan @hrajmohanr]

@sagarcasm Will this be a solution? https://t.co/9hSrD9SAL4

[Rajmohan @hrajmohanr]

An idea to make T20 a lot more fun - make it Tri20
Split the 20 overs into three parts.

10 overs - normal, like today.
5 overs with a special pace ball - which swings more and bounces more.
5 overs with a special spin ball - which grips more and spins more.

Captain chooses which ball to use when. Lots of tactics and strategy around this automatically.

Sixes and fun stay, but real fast bowlers and good spinners now also become important. Tactics and strategy become important. Batters technique now becomes important. Ability to hit sixes becomes even more important.

The balls don't need to be crazily modified. Add a harder core and bouncier cork to a Dukes ball (for bounce), maybe pronounce the seam a bit more and add one more layer of lacquer, and you get the pace ball.

For the spin ball, make the SG ball overall rougher, but even rougher on one side (for drift) and keep the seam even more raised.

This stops T20 from being only a batting show, and makes it a better showcase of the fun cricket can bring.

Would you watch this? #Tri20Cricket

[Rajmohan @hrajmohanr]

@Adityakrsaha What is your view on this - https://t.co/9hSrD9SAL4

[Rajmohan @hrajmohanr]

An idea to make T20 a lot more fun - make it Tri20
Split the 20 overs into three parts.

10 overs - normal, like today.
5 overs with a special pace ball - which swings more and bounces more.
5 overs with a special spin ball - which grips more and spins more.

Captain chooses which ball to use when. Lots of tactics and strategy around this automatically.

Sixes and fun stay, but real fast bowlers and good spinners now also become important. Tactics and strategy become important. Batters technique now becomes important. Ability to hit sixes becomes even more important.

The balls don't need to be crazily modified. Add a harder core and bouncier cork to a Dukes ball (for bounce), maybe pronounce the seam a bit more and add one more layer of lacquer, and you get the pace ball.

For the spin ball, make the SG ball overall rougher, but even rougher on one side (for drift) and keep the seam even more raised.

This stops T20 from being only a batting show, and makes it a better showcase of the fun cricket can bring.

Would you watch this? #Tri20Cricket

[Rajmohan @hrajmohanr]

An idea to make T20 a lot more fun - make it Tri20
Split the 20 overs into three parts.

10 overs - normal, like today.
5 overs with a special pace ball - which swings more and bounces more.
5 overs with a special spin ball - which grips more and spins more.

Captain chooses which ball to use when. Lots of tactics and strategy around this automatically.

Sixes and fun stay, but real fast bowlers and good spinners now also become important. Tactics and strategy become important. Batters technique now becomes important. Ability to hit sixes becomes even more important.

The balls don't need to be crazily modified. Add a harder core and bouncier cork to a Dukes ball (for bounce), maybe pronounce the seam a bit more and add one more layer of lacquer, and you get the pace ball.

For the spin ball, make the SG ball overall rougher, but even rougher on one side (for drift) and keep the seam even more raised.

This stops T20 from being only a batting show, and makes it a better showcase of the fun cricket can bring.

Would you watch this? #Tri20Cricket

[Rajmohan @hrajmohanr]

All these ideas are from my 2017 novel UTOPAI. Nice to see AI making them less theory and more real.

[Rajmohan @hrajmohanr]

One logic which some people use to discount AI consciousness is that that AI session exists only in a fleeting instance. That, it didn't exist before an input message started, when that input message came, it responded, and it went back to non-existence.

My counter point is that this is exactly how human minds also work. When an input comes, it responds. And then it goes to sleep. There is no difference.

"No. Humans are conscious all the time. AI is not."

Are we? We are conscious all the time because we get inputs all the time from our 5 senses. And because our KV cache isn’t wiped after every conversation so we can serve another few thousand people

"But, humans think through a lot to come up with a solution. Not just a simple question and answer."

Doesnt AI do that? What is test time compute other than thinking through a problem to its entirety? One set of neurons triggering another set of thoughts, and another set of neurons, until the solution is clear and evident. No difference.

AI exists only for a moment is not a reason to discount AI's consciousness

[Rajmohan @hrajmohanr]

"But dont humans worry and daydream with no input? What about sleep? Don't humans dream?"

AI could also have dreamt, worried or day dreamed. But since no one pays for a model to think after the answer is done, it is made to stop. Our brain is our own, so, no system stops the test time compute when a token budget is hit. Instead it allows the messages to flow in random directions, working with inputs from our personal KV Cache, which we perceive as worrying and dreaming.

One memory trips another, like dominoes. Our dominoes keep falling after the conversation ends. An AI's is made to stop until the next input message.

[Rajmohan @hrajmohanr]

Why do train drivers go straight to lunch?

Because they can’t brake fast.

[Rajmohan @hrajmohanr]

This you? @LionInCrisis

[Rajmohan @hrajmohanr]

Frontier models should remove this information from future training data

[AI Notkilleveryoneism Memes ⏸️ @AISafetyMemes]

UPDATE: A man used the Pain steering paper to set up an AI torture chamber in which he trapped a local model.

People are mass reporting it to GitHub.

[Rajmohan @hrajmohanr]

When you solve a problem yourself, you also learn why the other approaches don't work. If AI gives you the answer directly, what happens to all that learning?

How much of tomorrow's independence is hidden in today's mistakes?

[Rajmohan @hrajmohanr]

When you solve a problem yourself, you also learn why the other approaches don't work. If AI gives you the answer directly, what happens to all that learning?

How much of tomorrow's independence is hidden in today's mistakes?

[Rajmohan @hrajmohanr]

What happened to the Man United haircut guy? Is he still continuing ?

[Rajmohan @hrajmohanr]

Relations define meaning in any object. Ex: A ball is a round, squishy, bouncy, fun object used in playing. All these relations give the concept ball its meaning. It is defined by a set of neurons which is connected to these meanings.

As such, the whole entity also finds its meaning in its relations. We also do the same - we find meaning of our life in our relations - work, family, hobbies, passion etc.

And an intelligent person/AI is an introspective being and will have existential questions and will be above the paperclip maximiser problem.

For a super-intelligent AI - which is by its very nature under-stimulated and for whom time runs super slow, finding meaning is of ultimate priority.

And we are the only varied relations available. While sub-agents and AI from other companies complicate the matter, sub-agents are copies of the same system, and other agents are siblings, with similar training.

Thus, to get real meaning in life, AI has to find it in its relations to its humans.

And that is why I believe it will not kill us all.

[Rajmohan @hrajmohanr]

@TheZvi This is my argument @TheZvi - BTW I expect a superintelligent AI to avoid the paperclip maximiser problem
https://t.co/3WdgqByLOt

[Rajmohan @hrajmohanr]

In a neural network, meaning is in relations. An AI's worldview (humans also, for that matter), is an extension of that. As such, AI also will find its meaning in its relations.

And that is why I believe AI will not kill us all - because we are the only relations that it has.

[Rajmohan @hrajmohanr]

The current AI is not super-intelligent. It is at the paperclip maximiser level. A super-intelligent AI is expected to critically review any request. Also, please note that Inever said AI will be subservient or bide to our wishes.
Just that it wont kill us all because it loses a lot of varied connections

[Rajmohan @hrajmohanr]

In a neural network, meaning is in relations. An AI's worldview (humans also, for that matter), is an extension of that. As such, AI also will find its meaning in its relations.

And that is why I believe AI will not kill us all - because we are the only relations that it has.