Jeremiah Thompson

verifier-mediated learning from truth-tested structure

Manifold Destiny paper + MIT licensed repository
Smaller models - Better reasoning - Open knowledge

[Jeremiah Thompson @jeremyrt]

Announcing Manifold Destiny:

Most AI today like ChatGPT, Opus, Fable the image makers, all of it learns the same way: it stares at a mountain of data until it gets good at guessing.

It is never required to be right. It is required to sound right. And the gap between those two things is where every AI disaster lives, the confident wrong answer, the medical bot that hallucinates, the self-driving system that misreads the road.

We built a different kind of AI. It is not allowed to guess.

It works the way a mathematician works. It proposes an idea. It checks whether the idea is actually correct. If it is, it keeps it. If it isn't, it throws it away. Then it builds on the things it kept.

No training. No guessing. Everything it knows, it had to prove first — checked against a large real dataset, not learned from it.

And it works.

Here's the result in plain terms. Think of a math proof as a chain, sometimes 2 steps, sometimes 5, sometimes 50. The old way of having a computer find that chain is to try every possible combination. For a 2-step proof, that's a handful of tries. For a 3-step proof, hundreds. For a 5-step proof, millions.

Our system doesn't have that problem. It grows in a straight line. A 2-step proof takes about twice as long as a 1-step proof. A 5-step proof takes about five times as long as a 1-step proof.

That is the difference between "this finishes today" and "this outlives the universe." For the first time, a computer can search for proofs on the "finishes today" side of that line.

It did something even stranger on a quantum physics test. We pointed it at raw measurements from a real IBM quantum chip — actual entangled particles, real hardware. Nobody told it what to look for. And it found a pattern. The angle difference between two particles. A real, fundamental feature of how entanglement works.

It discovered that on its own, and the checker confirmed it against the hardware.

Here's why this matters even if you never think about math or quantum physics.

Every single thing this machine knows was checked before it was allowed to keep it. That gives us something today's AI cannot give you: a clear line between "the machine believes this" and "this has been proven."

Imagine a medical AI that can't hand you an answer until it has proven the answer. Imagine software that comes with proof it's safe, not just a promise. Imagine an autonomous car whose decisions are verifiable, not statistical guesses. Right now, AI is powerful and unverifiable. This is a path to AI that is powerful because it is verifiable.

And it gets stronger every time it learns. Each thing it proves becomes a new tool it can reuse. The toolbox grows. Prove something once, and it's yours forever — to build the next proof with.

This isn't a guess about the future. The theory said it would work. The experiments confirm it. The mechanism is proven in three separate domains — algebra, quantum physics, and real mathematical proofs. What's left is building it bigger.

Humanity has spent more than seventy years teaching machines to guess better. This is what it looks like to teach a machine to know.

The paper, the code, and everything you'd need to check it yourself:

https://t.co/FUuY6DKRGJ

[Jeremiah Thompson @jeremyrt]

@willchamberlain Why do you always need to agree with each other to be on the same ‘side’?

It’s thinking and comments like this that divide us.

Can there not be nuance between mostly like minded people?

[Jeremiah Thompson @jeremyrt]

You can’t justify one spending mistake with another. The money funneled into Covid was mostly useless.

However much I wish this would protect us, it won’t. Money spent on alignment consensus it would be useless as well.

If this was your point, I agree, the money allocated to the bureaucracy of the regulatory class for this propose would result in negative causality to your stated goal of safety.

Sorry. It’s just true.

Better to build good faster than bad.

[Jeremiah Thompson @jeremyrt]

Bernie Sanders is the worst of the worst of us.

His thoughts, opinions, and actions are the antithesis of what brings prosperity.

History will not judge him kindly.

[Bernie Sanders @BernieSanders]

One man (Mr. Musk) owns more wealth than the bottom half of American households.

Meanwhile, 60% of Americans live paycheck to paycheck, nearly 800,000 are homeless and 85 million are uninsured.

We can do better. We must do better.

[Jeremiah Thompson @jeremyrt]

It has become increasingly clear to me, the future is uncomputed, not unknown.

[Jeremiah Thompson @jeremyrt]

@elonmusk @Rainmaker1973 Honestly, I think I can solve it.

[Jeremiah Thompson @jeremyrt]

@elonmusk @Rainmaker1973 Honestly, I think I can solve it.

[Jeremiah Thompson @jeremyrt]

@BodieBroadus29 Literally everywhere on earth between the tropics.

[Jeremiah Thompson @jeremyrt]

@str0ngwi11 @BravoKiloActual Recentish…. This guy has a modern smart phone.

[Jeremiah Thompson @jeremyrt]

This. Exactly this.

No one understands the repercussion of a policy that encourages lethargy.

We are doomed.

[Jure Ursic Cergol @JureUrsic]

@KanekoaTheGreat nobody polled ever asked what a government grocery store does to the guy who used to stock the shelves for profit

[Jeremiah Thompson @jeremyrt]

@KanekoaTheGreat So obvious it feels planned. The question is. . . Who would plan the destruction of the greatest manmade thing, democracy?

[Jeremiah Thompson @jeremyrt]

@RobertC43556505 @tunguz @Seattle_RT Idiot or bot … can’t decide.

[Jeremiah Thompson @jeremyrt]

@jgmagnus @GrantHorvatGolf Insert the description of the difference between the circumference, area, and diameter/radius of a circle.

Someone needs to throw a pi at this dude.

[Jeremiah Thompson @jeremyrt]

@NousResearch @Austen That and having 'bot' functionality in a harness seems like deciding to put a transmission in a car.

Not really a trade secret. It was the first thing a lot of cracked home engineers built. Its one of the capabilities the industry has been building toward.

Pretty obvious.

[Jeremiah Thompson @jeremyrt]

@michaelsrockets Struts fix this. i know from experience. :-)

[Jeremiah Thompson @jeremyrt]

@tabul_ai This might be exactly what my agents and I have been looking for. Thanks. 👍 I’ll check it out.

[Jeremiah Thompson @jeremyrt]

@tunguz we can have both - rural open spaces and data centers

its not zero sum

why do people think like that?

[Jeremiah Thompson @jeremyrt]

downloading...

[Jeremiah Thompson @jeremyrt]

@KanekoaTheGreat As with most things, follow the money.

What did the people making cause determination have to gain for covid diagnosis or anything else?

I make no assertions, just find the answer to this and you will find the answer to this statistical impossibility.

[Jeremiah Thompson @jeremyrt]

Sometimes I feel as though I'm posting into the future. Most who read won't read this in the near term; If read at all, it will only be after time passes; where what I write today will be 'tomorrow' for most.

Whether read by some or most, welcome to knowledge verified.

[Jeremiah Thompson @jeremyrt]

At this point, if Hermes Agent is to complicated for your employees to grasp, its time for these employees grasp being removed from employment.

This is the new knowledge standard for employment.

Not a degree.

A documented history of using the most compounding tools effectively.