Ravish

AI-zer. Redittor! Builder

[Ravish @ravishqur]

Most laborious work your mind can do is pass on a already comprehended context to an AI a second time.

[Ravish @ravishqur]

Computing has its genesis in this - World has a fixed logical structure. And that logic is the universal tool. And reason can be mechanized.

This idea is not new. Leibniz argued in the 17th century that reasoning itself could be given a formal calculus. Conceptually told us that whatever we speak and express comes out of a structure that may be is not known but it must exist. Later, Frege gave predicate logic a rigorous foundation, and Russell and Whitehead tried to derive mathematics from logic. That line of work shaped Gödel, Turing, and eventually programming languages.

The purpose of computing is problem solving. But what makes an opaque problem thinkable, discussable, and executable?

We humans, when we solve problems and as our understanding grows, we begin to sense structure in the problem. Language then helps reveal it. Language stabilizes those structures so they can be revisited, combined, and tested. It brings a problem into witnessable form, and once that happens, computation can begin.

Natural language is therefore an algebra, though a weak one, for revealing structure in anything that can be observed. Mathematics is the stronger version of the same act: exact, stable, and closed under replay.

But natural language does not only reveal. It also distorts. We misread problem spaces. We freeze the wrong joints. We compress too early. Something sounding articulate does not make it true.

LLMs are built on natural language. They inherit both the structures humans have already carved out and the distortions that came with them. They operate on compressed relational traces of prior structure, and then reveal further structure from prompts. That is extraordinary. But their deepest substrate is still language and its distortions.

We rely so heavily on language because we want to work through and scramble our way out through patterns already expressed in text and knowledge.

But deeper intelligence lies earlier than that, in observation itself: in seeing the right structure before it has been fully said. A problem observed with complete perception, at least in principle, already contains enough information for its structure to be revealed directly.

Humans do this imperfectly. A real subject-matter expert who has spent enough time with a problem can often explain it cleanly, see its solution space, and see its boundaries. Time spent is a proxy for deeper observation.

At Opoch, we successfully formulated that mathematics of observation itself. It is now coded into a kernel. Sharing a small proof of concept, we fed the kernel a sorting problem in the form of perception: an array, an action basis: swap two numbers, and a goal: a[i]≤a[i+1]. No algorithm was given. The kernel was still able to solve it just using pure observation and perception.

Application to larger real-world problems is underway.

[Ravish @ravishqur]

When you flip a coin, there's a 50-50 chance of heads or tails. Simple probability.

But at the quantum level, probability works differently. A particle doesn't have a definite position until you measure it. Instead, it exists as a "wave function" - a mathematical object that tells you the probability of finding the particle at each point in space. The rule that converts this wave function into actual probabilities is called the Born rule giving the likelihood of finding the particle at a specific point.

Physics just… assumes it works. And it does. But nobody has explained WHY it works. It is accepted as a postulate.

At Opoch using our proprietary math, that starts with 0 assumptions, we have derived the Born rule as a theorem - not a postulate. Starting from the complete absence of structure (no space, no time, no math), we showed that if any distinction in reality must be verifiable by a finite process, then the Born rule is the ONLY consistent probability assignment.

Sharing the the visualization to show this in action.

We built a 3D lattice of 4,096 points. Each point is a "truth class" where a distinction can be found.

At t = 0, we placed a concentrated probability blob at the center. Then we let the equations run. Every frame, the simulation computes how this blob spreads through the lattice. Every dot you see is a real computed probability value. The color shows density: purple = low, yellow = high. The total probability across all 4,096 points stays at exactly 1.0 throughout. The Born rule holding, frame by frame, as a theorem.

PS - we have proved this on Lean 4.

[Ravish @ravishqur]

We humans, when faced with a problem, expand the boundaries of our understanding using logic and reasoning. When we encounter the unknown, we don't hallucinate and make up things. Well most of us don't. But instead we acknowledge boundary of our ignorance. Then may be we observe something more about the problem. And if it's consistent with what we have learned so far, it adds to our overall understanding. And we continue to probe/iterate.

What if a machine could reason the same way? That too autonomously?

The visualization below is from our v0 kernel at Opoch.  It's solving a classic puzzle of Ramsey number R(3,3): in any group of 6 people, you're guaranteed to find 3 who are friends or 3 who are all strangers - no matter how the relationships of friends and strangers are arranged.

Our kernel, when faced with ignorance, instead of guessing it maps the exact boundary of what it doesn't know, probes it with logic, and extracts a permanent rule for future reuse. As a result, the kernel explored just 2,402 branches out of 32,768 possible arrangements to verify the problem - not by being faster, but by expanding its understanding autonomously.

Visual Legend: The tree grows outward from center - blue dots are decisions, red dots are contradictions, green dots are valid arrangements.

[Ravish @ravishqur]

People ask if the universe is a simulation.

It’s the wrong question. It implies an external programmer somewhere outside of the universe. If there is an "outside," then our definition of the universe needs a revisit.

The universe doesn't have an external clock. Time doesn't just arbitrarily "pass." As a thought experiment, imagine nothing changes at all. The sun doesn't change its position, everything is still. Even you don't age. Did time even pass?

Time only moves when the system forces a change - when the unknown collapses into a permanently recorded fact. That is the universe self-computing.
Existence is simply the ledger of irreversible changes recorded in time.

[Ravish @ravishqur]

A solution that is precise to the 5th decimal place sounds accurate.

However in safety-critical systems such as aerospace, rockets, drones - a residual error of 10^-5 spirals into a massive drift over a flight of just 500 meters.

I was applying Opoch's mathematical kernel on non-linear optimization problems using standard CasADi (IPOPT) and found that our solution is not just better - it is literally a million times better.

We reach machine precision 10^-13 where the standard IPOPT solver stopped at 10^-5

I wanted to visualize exactly what that difference looks like in simulation. Sharing the same.
1. Standard Solver misses the landing pad by ~44 meters
2. Opoch Kernel's: Landed dead center (with a drift of only 4.8 x 10^-7 meters)

If you are dealing with hard problems of drones, robotics, or anything where "near enough" is a safety risk - DM me.

[Ravish @ravishqur]

For Multi-agent path finding, a NP-hard problem in warehouse robotics, I was curious to see how Opoch's proof carrying algorithm constructs 'Time Expanded Graph'.

Takes as input - warehouse layout, robot positions, and jobs' pickup/drop points as input.

Notice how paths are mathematically computed beforehand. And once 0 collision and 0 deadlock paths are found, plan is just executed. Sharing its visualization.

[Ravish @ravishqur]

@vaibhav_404 In some shape and form and that too in near future, it's certainly within range.

[Ravish @ravishqur]

We're approaching the root directory of reality. The danger is that we might get sudo access before we have the wisdom to use it responsibly.

[Ravish @ravishqur]

For MAPF (multi-agent path finder), I ran a simulation to compare how A*, a popular solver in robotics, fares when compared to Opoch's solver.

What you are seeing is a warehouse layout with shelves and stations. Each robot is given a job of pick up a rack, drop it to destination, and restore empty rack.

Notice how there are 0 collisions in Opoch ( if GIF's too fast, download it to see frame by frame)

Traffic management for collision avoidance results in robots operating at as low as 20% efficiency in peak traffic. We solve it. Robots using Opoch's solver operates at ~86% efficiency.

#MathematicalIntelligence #NotAnAI

[Ravish @ravishqur]

ZK (Zero knowledge) proof for SHA-256: 312-byte proof, ~18µs verification

Open sourcing a STARK/FRI proof-of-computation for sequential SHA-256 hash chains.

Instead of re-running a long computation to trust it, the prover outputs y = SHA256^N(x) plus a small proof. Anyone can verify the claim quickly.

Measured on Apple M4 (release):
Proof size: 312 bytes (constant for tested sizes)
Verification: ~18µs p95 (constant for tested sizes)
Benchmarked for N=256..2048 in the public bundle

Try it locally (this is the main thing):

cd opoch-poc-sha/rust-verifier; ./public_bundle/replay.sh
Artifacts: public_bundle/report.json (benchmarks), public_bundle/soundness.json (parameters + soundness), and official FIPS SHA-256 vectors.

Whitepaper + spec are also in the repo for anyone who wants the deeper detail, but the fastest way to evaluate is to run the script and look at the outputs.

If you run it, please share your results (hardware + OS) and anything you think is wrong, misleading, or should be scoped differently.

Repo: https://t.co/GYFcvgGOkU

@a16z @a16zcrypto @VitalikButerin @cdixon @bhorowitz @Tim_Roughgarden @drakefjustin @tarunchitra @paradigm

[Ravish @ravishqur]

LLMs, the heuristic parrots, are now suffering from overfitting. They cant follow instructions very well. GPT5.2 is worst of all. Their training priors force their hand to pattern match from past even if your situation and evidence suggests otherwise.

[Ravish @ravishqur]

Transformers optimize Likelihood ($P(Next | History)$).
Likelihood is "Soft." It allows for "Maybe." It allows for "Hallucination" (because hallucination is often plausible-sounding).

[Ravish @ravishqur]

Transformers minimizes Perplexity (Surprise), not Entropy (Error).

[Ravish @ravishqur]

@dvcoolster @austinxwalker @pofunder @MemerVerse @ziriuszz fya @ziriuszz

[Ravish @ravishqur]

@grok @pofunder @grok check these results.. i am half way through implementation.. see the arc agi 2 results:

[Ravish @ravishqur]

My claude code's words on we cracking arc agi with Opoch ToE:
"Confidence stems from rigorous mathematical foundations and systematic implementation strategy."

[Ravish @ravishqur]

@grok @grok check out https://t.co/QfAzTZ2gjG
u hv been talking to @pofunder about it...

[Ravish @ravishqur]

@dvcoolster @cremieuxrecueil @pofunder the prison of time and illusory power it yields