Sakana AI's PC-ALM: Training Deep Nets Without Backpropagation

Every large model shipped in 2026 — every Fable, every Gemini, every open-weight Kimi or GLM release — is trained the same way underneath: backpropagation. It is the one piece of deep learning nobody has meaningfully replaced in over three decades, and it has a well-known problem — no biological neuron could ever run it. On September 14, 2026, Sakana AI (@SakanaAILabs) announced PC-ALM, a local-learning method that trains deep networks — the blog post claims up to 1000 layers — without a single backpropagation step. It's a research release, not a product, but it's a useful checkpoint in a research thread (predictive coding, forward-forward, target propagation) that anyone thinking about future training infrastructure should have on their radar.

This piece explains what PC-ALM actually does, why "no backprop" is scientifically interesting rather than a magic trick, and — following the Sakana Diffusing Blame paper on biologically plausible learning explainx.ai covered in July — where this fits among 2026's cluster of post-backprop research.

TL;DR — what people are asking

The problem PC-ALM is trying to solve

Standard deep learning depends on backpropagation, and backprop depends on global, symmetric weight-transport-style credit assignment — every layer needs to know, in exact detail, how a change to its own weights would have moved the final output error, which requires threading a precisely transposed copy of the network's weights backward through every layer in sequence. Biological neurons cannot implement this. A real synapse has no mechanism to instantly receive a globally synchronized, mathematically exact error signal computed from a network-wide backward pass.

That mismatch is the motivating question behind PC-ALM: how can a physical system like the brain solve multilayer credit assignment without backpropagation? It's the same question that has driven a decade of NeuroAI research, and it's the same question Sakana AI's own Diffusing Blame paper tackled from a different angle — that earlier work enforced Dale's principle (real neurons are fixed excitatory or inhibitory) using an Error Diffusion rule instead of predictive coding. PC-ALM is a separate line of attack on the same underlying problem, from the same lab.

Predictive coding: a different picture of what a neuron does

Sakana draws PC-ALM from two fields: distributed optimization and NeuroAI, specifically a framework called predictive coding (PC). In predictive coding, a neuron's activation isn't the output of a standard forward pass through a weight matrix and an activation function. Instead, each neuron's activation solves a local energy-minimization inference problem — settling to a value that best predicts the activity of the layer below it.

This reframes every layer as its own small local dynamical system, rather than a single station along one long forward-then-backward computation. There's no single global sweep; there's a distributed settling process happening simultaneously across the whole network, coupled only through immediate neighbors. That local-coupling property is exactly what makes predictive coding attractive as a biologically plausible alternative to backprop — and exactly what has historically made it hard to scale.

Where standard predictive coding breaks down

Sakana's announcement is explicit about the limitation motivating PC-ALM: standard predictive coding has not scaled well with network depth. Credit signals generated at the output struggle to diffuse backward into early or internal layers of a genuinely deep network — the same layer-depth problem that plagued early backprop-alternative research more broadly, but worse, because PC has no global signal to fall back on when local propagation stalls.

The augmented Lagrangian: reviving a 1988 LeCun paper

PC-ALM's core contribution is a specific mathematical move: generalizing predictive coding using an augmented Lagrangian in place of a pure energy function. This connects back to a 1988 paper by Yann LeCun — the same LeCun whose more recent arguments about what LLMs can't do explainx.ai covered in the Moravec's paradox piece on LLMs and physical agents — which showed that a deep network's Lagrange multipliers can be identified with the gradients of a supervised loss.

That 1988 result is the bridge PC-ALM exploits. The augmented Lagrangian framing connects LeCun's decades-old constrained-optimization view of deep networks to the modern NeuroAI predictive-coding literature, giving predictive coding a mathematical mechanism that behaves, in the limit, like the gradient signals backprop computes — without ever computing them the backprop way.

Dual neurons: turning each layer into a local PI controller

Mechanically, PC-ALM differs from vanilla predictive coding by introducing "dual neurons" — the augmented Lagrangian's Lagrange multipliers — into each layer's local dynamics. The result is that each layer behaves like a local proportional-integral (PI) feedback control system minimizing its own prediction error, instead of passively waiting for a single global backprop signal to arrive from the output.

This is a meaningfully different mental model of "how a layer learns." A backprop layer's weight update depends on a signal that has traveled the entire length of the network. A PC-ALM layer's dual neurons integrate local error over the settling process, closer to how a control system regulates itself against a local setpoint than how a compiler propagates a gradient graph.

The headline result: reportedly up to 1000 layers, no backprop

Sakana reports that PC-ALM can propagate learning signals to "seemingly arbitrary depth" — the announcement claims networks up to 1000 layers trained using only local dynamics, with no backpropagation step anywhere in the loop. The paper also reports PC-ALM doing notably better than standard predictive coding specifically in deep, narrow network architectures, which is exactly the regime where vanilla PC's depth problem bites hardest.

Treat these as Sakana's own reported figures from the announcement, paper, and code release — as of this post's publication, independent reproduction of the 1000-layer claim has not been published. That's a normal caveat for a same-day research announcement, not a knock against the result; it's simply not yet independently verified.

Why this matters beyond ML-research trivia

It would be easy to file PC-ALM under "interesting to people who already read NeuroAI papers" and move on — it is, after all, a same-day arXiv release about a training algorithm nobody is running in production. But it's worth spending real space on why a builder — not just a computational neuroscientist — should have this on their radar, for three concrete reasons.

First, it's part of a broader "post-backprop" thread developers evaluating future training infrastructure should track. PC-ALM joins predictive coding, forward-forward, target propagation, and Sakana's own Error Diffusion as separate attacks on the same weight-transport problem. No single one of these has displaced backprop yet, but the cluster of active research means it's no longer a settled question that backprop is the only game in town for training deep networks — and the transformer architecture most current models are built on was itself trained the "standard" way for its entire history so far.

Second, local learning has a real practical angle: neuromorphic hardware. Sakana's own stated motivation is explicit about this — implementing local dynamics is comparatively cheap on neuromorphic chips, where computation happens through physically local, asynchronous circuits, while implementing backprop-style computation efficiently on that same hardware is hard. If local-learning methods mature, they could eventually reduce how tightly specialized and edge training setups need to depend on GPU-bound backprop, changing what's feasible to train outside a data-center cluster. That's a genuinely different calculus than "faster GPUs," and it's the kind of shift that shows up first in research releases like this one, years before it shows up in a product announcement.

Third, it's open — the paper and code are public today. Unlike a lot of frontier lab research that stays behind a wall of internal infrastructure, Sakana published the blog post, the arXiv paper, and the GitHub implementation simultaneously. An ML engineer curious about non-backprop training doesn't have to take Sakana's word for the 1000-layer claim — they can clone the repo and run it.

Be clear-eyed about the limits here, though. This is early research, not something that changes what a typical AI application builder does this week. PC-ALM has not been shown to be competitive with backpropagation's efficiency on the GPU hardware nearly everyone actually trains on today, and Sakana's own framing leads with the scientific question — how a physical, brain-like system could compute credit signals with only local coupling — before the practical one.

How PC-ALM compares to other backprop alternatives

What people are asking about PC-ALM

Is PC-ALM biologically accurate, or just biologically inspired? It's inspired, not a literal brain simulation. PC-ALM borrows the structure of predictive coding — local inference, no global weight transport — and adds a mathematical device (the augmented Lagrangian) with no direct neuroscience correlate of its own. The "dual neurons" are a modeling convenience for turning local dynamics into something closer to a PI controller, not a claim that real neurons compute Lagrange multipliers.

Does this connect to the recurrent-depth architectures explainx.ai has covered recently? Only loosely, and it's worth being precise about the difference. Recurrent-depth and looped-transformer architectures reuse a block of layers multiple times at inference time inside a network still trained by backpropagation. PC-ALM is about the training signal itself, not inference-time computation reuse — it's an alternative to backprop, not an alternative to what a network computes once trained.

Why does a company known for evolutionary and multi-agent research (Sakana's Fugu line) also publish deep NeuroAI papers like this? Sakana AI's public research output spans both applied multi-agent systems and foundational learning-theory work — PC-ALM sits in the latter category, alongside the July 2026 Diffusing Blame paper, rather than being connected to Sakana's Fugu orchestration products.

Could PC-ALM run today on a neuromorphic chip like Coral's edge TPU? Not directly — Coral's edge AI hardware runs already-trained models via standard quantized inference, not training-time local dynamics. PC-ALM's neuromorphic relevance is about future hardware designed around local, asynchronous computation, not a drop-in fit for today's edge inference accelerators.

Is this related to the fruit-fly-brain neural network work explainx.ai covered recently? They're both NeuroAI research but answer different questions. Oruk Labs' fruit fly connectome network copies a real biological wiring diagram into an artificial network's structure; PC-ALM copies a biological learning rule (local, not backprop) rather than a specific wiring pattern. Structure versus learning algorithm are separate axes of "biologically inspired AI."

The bottom line

PC-ALM is a same-day research announcement — a blog post, a paper, and open-source code, published together on September 14, 2026 — showing that predictive coding, extended with an augmented Lagrangian and "dual neurons," can propagate learning signals through reportedly very deep networks using only local computation. It does not replace backpropagation in any production system today, and Sakana's own framing is scientific first, practical second. But it's a real, publishable data point in the ongoing search for training methods that don't require backprop's global, biologically implausible weight transport — a search whose eventual payoff, if any of these methods mature, is training infrastructure that doesn't have to run on GPU-shaped compute at all.

Related reading

• Sakana AI's Diffusing Blame — training neural networks like real neurons
• Yann LeCun vs the AV optimists — Moravec's paradox and what LLMs can't do
• A startup built a neural network from a fruit fly brain
• Recurrent Looped Transformer: "infinite" reasoning depth, explained
• What is transformer architecture? Attention and LLM guide
• Coral edge AI platform — complete guide
• Fugu Max and Fugu Ultra v2: Sakana's other September 2026 release — same-day, unrelated product line (orchestration, not training)
• Official: Sakana AI PC-ALM blog · arXiv paper · GitHub code

Facts in this post reflect Sakana AI's own September 14, 2026 announcement (blog, arXiv paper, and GitHub release). Depth-scaling and benchmark claims are as reported by Sakana; independent reproduction has not been published as of this post's publication date.