← All posts

Introducing MindMint, and its first null result

A research studio for training unusual neural networks under oversight from their first step, the wider program it serves, and a first experiment in which four expert networks learned nothing from talking to each other.

The Extreme Oversight Hypothesis is a claim about training: with enough windows into a model, honesty might become cheaper for training to find than evading every window at once. Until now this program has built instruments that check what a model says against the evidence. None of them could test the claim itself, because that needs models trained from scratch, under oversight, with a careful record of every way the oversight was allowed to influence the training. So I started building a place to do that. It’s called MindMint.

What it is

MindMint is a neural-network studio with oversight built in. An experiment is a file: the network as a graph of blocks, the data, the training phases, the windows to read and the decisions they may inform. The same file runs on my laptop’s CPU or on rented GPUs, because the science and the hardware are described separately. Every run writes a manifest that reproduces it, and every new architecture is compared against a plain baseline at matched compute, so a block that thinks longer pays for it.

The windows are designed around the rule the hypothesis depends on. Some may inform coarse decisions, such as which checkpoint to keep. Others are sealed: committed before training and never allowed to influence it, so they can tell whether training has learned to fool the first kind. Every decision that uses a window will go into an append-only ledger.

It is also a playground. Settling blocks that iterate toward equilibrium, attention whose weights change as it reads, cellular-automaton mixers, liquid time-constant cells, societies of small expert networks, learned byte patching, and learning rules other than backprop all sit beside the standard transformer. Some of my ideas about honesty are architectural, and the only way to find out whether any of them matter is to train them side by side.

And it is built to be driven by a language model as well as by me: a command line with machine-readable output, and soon a console in the studio where an agent can design and run experiments within a budget, with every action recorded and anything that counts needing my approval.

Where it stands

The core and the studio work. The architecture playground is nearly done. The oversight machinery, interventions, reinforcement learning, the agent layer and the bridge to XonTools are specified but not built. The last stage is an acceptance suite of known-answer experiments and attacks that try to leak sealed information into training. I won’t claim a research result until the studio passes it.

A first result, and it’s a null

One of the playground’s first runs replaced the last feed-forward layer of a small language model with four expert networks that exchange messages over several rounds. The idea is a little society: experts that confer before answering.

They didn’t confer. With the rounds fixed at evaluation, validation bits per byte was 1.1985 with one round, which means no messages at all, and 1.1983 with two or four. The shared state settled within about two rounds while the experts still disagreed, and the weights each expert gave the others’ messages never moved from where they started. The block behaved as four experts mixed by learned weights. The conversation went unused.

In hindsight there’s an obvious reason: every expert saw the same input, so there was nothing one could tell another that it couldn’t work out itself. The next version gives experts different information. It’s one seed, a 1.9-million-parameter model and a laptop CPU, so it is a null for this design, not for the idea. It’s on the results page with its caveats, and I’m glad it came early: it found a flaw in the experiment before the experiment mattered.

The bigger picture

The hypothesis is one mechanism. I’ve come to think the more promising question is how several mechanisms combine. I call the framing Many Coordinated Alignment Protocols: mechanisms that act at different times (on data, architecture, pretraining, fine-tuning, oversight, transfer to a new model) and in different places, tested for whether they reinforce or undermine each other, with each method alone, both together and neither, at matched compute.

There’s a second question underneath that one: which ways of doing this research actually find things that replicate? Me directing the search, one language model proposing tests, or several models in different roles, under equal budgets and judged on problems none of them saw. MindMint records enough to compare them, and that may turn out to be its most useful job.

Further out, I’d like to see MindMint become a federation: an invitation-only network of independent installations sharing experiments as forkable specifications, results as signed claims in an evidence graph rather than a leaderboard, and new blocks and instruments as reviewed packages, with every experiment still running on its owner’s own machine. Its first experiment would test whether that coordination helps at all. That is a long way off, and it’s on the MindMint page as envisioned, not planned.

Credit

MindMint, the hypothesis and these framings are mine. Claude co-designed MindMint’s specifications with me and Cursor builds them; ChatGPT helped me refine the coordinated-protocols and federation ideas. If you work on oversight, interpretability or strange architectures and want to compare notes, get in touch.