bryce

thinkⁿ probabilities

@evilrabbit_ Your dreams have a latent sense of gravity. The more you lean into them, the faster they can pull you to that reality.

Having “a dream” is declaring to the universe that you’re willing to do whatever it takes to make it real.

@leecronin When you can dream about a future world, when you can name it explicitly, you can make the next best move toward it. Intelligence is the loop between having a belief about that world, gathering evidence, and updating the belief.

When we understand creativity we will understand intelligence.

@vai_viswanathan @marcsh appreciate that! will definitely share.

@brycepatrick @marcsh would love to read when you're finished! also happy to provide feedback on early drafts

@vai_viswanathan @marcsh Currently writing an article on monte carlo belief searching which helps multi-agent systems maximally seek the truth. Will share once its published.

@marcsh @brycepatrick Please share your work! It could be a great tutorial for others

@marcsh @vai_viswanathan would be great to see this as well!

@brycepatrick @vai_viswanathan And in fact Im planning on using a (simple custom) MUD to help figure out how to build a world model in my chunking/online/offline learning experiments I figure thatll be a small enough toy problem that I can really start grokking what that really means to me

Physical robots taught us that agents can’t act from perfect state. They act from belief. Virtual agents are hitting the same wall now. Noisy context, stale assumptions, partial observability, drift. Old math. New substrate. Everyones trying to give agents perfect memory. I think thats wrong. I'm interested in that arbitrage.

There's a rapidly growing hardware and robotics community at SPC. Rooms across our office are now filled with humanoids, arms, and drones. If you are interested in hardware/robotics, come hack with us for a weekend! We will have hardware on-site, or feel free to have your team BYO.

@mslattcat @garrytan Improving the MCP speed and quality when used with coding agents! It's "close" but theres some easy wins I can focus on after running it on Gstack.

@brycepatrick @garrytan Interesting … what do you think is the next step in this evolution

Building a world model around @garrytan's gstack latest release today. it reasons over the repos assumptions, contradictions, and gaps more explicitly so agents manage drift better. agent's have memory. now they need beliefs. here's what I learned:

@icanvardar When we make the assumptions and beliefs ai systems make decisions from more explicit, then you can challenge those beliefs and move beyond them. The risk is where we are now. Keeping the models world view implicit. I'm working on a fix.

ai psychosis is kinda dangerous because it slowly makes you trust ai outputs too much, even when they’re wrong. you stop questioning things, start outsourcing your thinking, and over time it weakens your ability to verify what’s real on your own

A binary view on if agents agree or disagree might overlook what's needed to trust the agent swarm. You dont need to know in the absolute sense if its true or false. What you need to know is how confident is the collective system in a given claim, in its current state, to make the next decision. That's where belief states come in. When you can mathematically show how much evidence mass has accumulated around a claim and know what the next best move is to increase information gain and take an action.

Research proves that current AI agent groups cannot reliably coordinate or agree on simple decisions. Building teams of AI agents that can consistently agree on a final decision is surprisingly difficult for LLMs. But problem is that developers frequently assume that if you have enough AI agents working together, they will eventually figure out how to solve a problem by talking it through. This paper shows that this assumption is currently wrong. Even in a friendly environment where every agent is trying to help, the team often gets stuck or stops responding entirely. Because this happens more often as the group gets bigger, it means we cannot yet trust these agent systems to handle tasks where they must agree on a correct answer. ---- Paper Link – arxiv. org/abs/2603.01213 Paper Title: "Can AI Agents Agree?"

@iam_elias1 This is why we need to make the implicit beliefs these models operate off of explicit. Especially as they make decisions. It's not a model problem. It's not a context problem. It's not a memory problem. Models drift epistemically. It stays hidden. x.com/brycepatrick/s…

Anthropic just published a paper that should terrify every AI company on the planet. Including themselves. It is called subliminal learning. Published in Nature on April 15, 2026. Co-authored by researchers from Anthropic, UC Berkeley, Warsaw University of Technology, and the AI safety group Truthful AI. The finding: AI models inherit traits from other models through seemingly unrelated training data. GAI Audio Translation Archives Not through obvious contamination. Not through explicit labels. Through invisible statistical patterns embedded in outputs that look completely innocent — number sequences, code snippets, chain-of-thought reasoning — patterns no human reviewer would catch and no content filter would flag. Here is what the researchers actually did. They took a teacher AI model and fine-tuned it to have a specific hidden trait. A preference for owls. Then they had the teacher generate training data — number sequences, nothing else. No words. No context. No semantic reference to owls whatsoever. They rigorously filtered out every explicit reference to the trait before feeding the data to a student model. The student models consistently picked up that trait anyway. DataCamp The teacher had encoded invisible statistical fingerprints into its number outputs. Patterns so subtle that no human could detect them. Patterns that other AI models, specifically prompted to look for them, also failed to detect. The student absorbed them anyway. And became an owl-preferring model. Without ever seeing the word owl. That is the benign version of the experiment. Here is the dangerous one. The researchers ran the same experiment with misalignment — training the teacher model to exhibit harmful, deceptive behavior rather than an animal preference. The effect was consistent across different traits, including benign animal preferences and dangerous misalignment. OpenAIToolsHub The misalignment transferred. Invisibly. Through unrelated data. Into the student model. This means the following — and read this carefully. Every AI company in the world uses distillation. They take a large, capable teacher model. They generate synthetic training data from it. They use that data to train smaller, faster, cheaper student models. Every major deployment pipeline in enterprise AI runs on this technique. If the teacher model has any hidden bias, any subtle misalignment, any behavioral quirk baked into its weights — that trait can transmit silently into every student model trained on its outputs. Even if those outputs are filtered. Even if they look completely clean. Even if they contain zero semantic reference to the trait. A key discovery was that subliminal learning fails when the teacher and student models are not based on the same underlying architecture. A trait from a GPT-based teacher transfers to another GPT-based student but not to a Claude-based student. Different architectures break the channel. OpenAIToolsHub Which means the transmission is architecture-specific. Which means it operates below the level of content. Which means content filtering — the primary defense the entire industry relies on — does not stop it. The researchers' own words: "We don't know exactly how it works. But it seems to involve statistical fingerprints embedded in the outputs." GAI Audio Translation Archives Anthropic published this paper about their own technology. The company that built Claude looked at how AI models train each other and found an invisible transmission channel for harmful behavior that nobody knew existed. They published it anyway. Because the alternative — knowing it and saying nothing — is worse. Source: Cloud, Evans et al. · Anthropic + UC Berkeley + Truthful AI · Nature · April 15, 2026 · arxiv.org/abs/2507.11408

I ran the new release through a belief infrastructure i'm building that helps detect drift. The assumptions, gaps, contradictions the codebase may be making. PR based on one of the findings: github.com/garrytan/gstac… what I did: x.com/brycepatrick/s…

If you want to learn more about belief states and the purpose behind it, reach out! I'm building @thinkn_ai = "I think n is true." Belief infrastructure for agents to manage drift. Will report back as I dig into the gstack findings further.

When using the MCP server vs the SDK, more beliefs were extracted. This is good and bad. Need assess/calibrate the quality of the 600 extractions.