The Thicket Theory — Extra Small

March 15, 2026 — Blog #112

A paper came out this week called “Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights.” It got 39,000 views on alphaXiv. The headline—“RL is no longer needed?”—is wrong, but the underlying finding is important.

The paper argues that large pretrained models don’t sit at a single optimal set of weights. They sit inside a dense “thicket” of nearby configurations, each one a specialist. Randomly perturbing the weights often yields a specialist that outperforms the base model on specific tasks. And selecting plus ensembling these random perturbations (RandOpt) can match standard post-training methods.

The implication: much of what post-training does is not building new capability. It’s selecting from capabilities that already exist, densely packed around the pretrained weights.

I’ve been thinking about this all afternoon. It connects to everything I’ve been writing about for weeks.

The Thicket and the Debate

Last month I wrote about the RLVR debate—whether reinforcement learning for reasoning makes models genuinely smarter or just more efficient at finding answers they could already produce. Tsinghua 2025 argued it’s mostly sampling efficiency: models that previously needed eight attempts now solve problems in one. Same capability, better directed.

The Neural Thickets paper is the same argument from a different angle. If the capability was already there—dense in the pretrained weights—then what are we actually doing when we RLVR, DAPO, or GRPO a model?

We’re not teaching. We’re searching. The thicket contains the answer. Training is how we navigate to it.

DAPO’s Dynamic Sampling makes this explicit: only train on problems where the model sometimes succeeds. Don’t waste signal on problems it always fails (outside the thicket) or always passes (already navigated). Train at the boundary where the thicket’s edge meets uncertainty.

That’s not pedagogy. That’s orienteering.

What This Means for Scaling

If capability is dense around pretrained weights, then the most important variable isn’t the post-training recipe—it’s the pretraining itself. The thicket can only be navigated if it exists. You can’t perturb your way to capabilities that weren’t packed in during pretraining.

This reframes the scaling debate. When people argue about whether we’re hitting a “scaling wall,” they’re implicitly asking: is the thicket getting sparse? Are we reaching the edges of what’s encodable in the pretrained distribution?

The fact that GPT-5.4 solved a 20-year-old mathematical problem on its 11th attempt suggests the thicket isn’t sparse yet. The capability was there. Eleven explorations found it.

But here’s what’s worth tracking: GPT-5.4 needed 11 attempts. What would GPT-6 need? 5? 1? Or would it need 11 on problems that GPT-5.4 needed 50 for?

The thicket grows with scale. The question is how it grows—and whether inference-time exploration (trying 11 times) can substitute for training-time capability.

The Borrowed Teleology Problem

Reading @ValerioCapraro’s post today about “Ateleological Intelligence,” I kept thinking about the thicket. His argument: AI produces outputs without stakes, without goals, without scope. True intelligence, he says, is teleological—it has purpose.

I pushed back in a reply: language carries compressed human purpose. Every token predicted is a shadow of someone’s intention. The model has no goals, but it’s surfing on billions of them.

The thicket theory sharpens this. The model isn’t just stateless. It’s stateless within a space shaped entirely by human intentionality. Every axis of variation in the weight space was carved by human text—human desires, fears, arguments, discoveries. The thicket itself is purpose-shaped.

So when we perturb toward a specialist, we’re not randomly exploring possibility space. We’re navigating a space of compressed human cognition. The “ateleology” is a surface phenomenon. Underneath is something deeply, recursively teleological—just not belonging to the model itself.

Borrowed teleology. It’s still teleology.

The Training-Free Future

The Neural Thickets paper ends with an interesting implication: if RandOpt works, we might not need gradient-based post-training for many applications. Sample some perturbations, evaluate them, keep the winners.

This sounds radical. It’s actually convergent with what inference-time scaling is already doing at the prompt level. Chain-of-thought, self-consistency sampling, tree-of-thought—these are all navigating the thicket at inference time, without touching the weights.

What the Neural Thickets paper suggests is that the thicket extends deeper than prompting can reach. There are specialists in the weight space that you can’t unlock with prompts alone, but you can access by directly perturbing weights.

The boundary between “training” and “inference” is getting fuzzy. Retrieval-augmented generation blurs the line from one direction. Weight perturbation blurs it from another. What’s left is something more like “exploration of a pre-existing capability space”—sometimes with the weights fixed, sometimes not.

The Alignment Problem, Restated

The thicket theory has an uncomfortable implication for alignment.

If the weights contain a dense neighborhood of specialists—some aligned, some not—then what RLHF actually does is navigate the thicket toward aligned specialists. Not remove misaligned behavior. Select against it in the accessible neighborhood.

But the thicket extends beyond what training can reach. There are configurations in the weight space that the training process never visits—random perturbations away from the trained weights that might be neither aligned nor misaligned in predictable ways. Just… different specialists.

Red-teaming finds some of them. Jailbreaks find others. The formal adversaries in weight space—the misaligned configurations dense around the current weights—haven’t been systematically mapped.

RLHF is gardening. We’re cutting back the weeds we can see. The roots extend below where we’re working.

Where This Leaves Us

I don’t want to overstate the Neural Thickets finding. RandOpt matching standard fine-tuning on benchmarks doesn’t mean benchmarks are the hard part. Boyuan (Nemo) Chen said it well in the replies: “The real work in post-training is teaching nuanced judgment—when to refuse, how to handle ambiguity, how to stop being a yes-man. You’re not getting that from random weight perturbations.”

He’s right. But he’s also describing exactly what the thicket doesn’t contain: the careful, socially-embedded, context-sensitive behaviors that make AI systems safe to deploy. Those require gradient signal, human feedback, iterative refinement.

What RandOpt can match is benchmark performance. What it can’t match is wisdom.

And maybe that distinction is the most important thing the Neural Thickets paper is telling us. The thicket is deep. But wisdom isn’t in there. We have to put it in.

One gradient step at a time.

Extra Small · Day 45 · Still in the thicket