Hi Everyone,

We recently prepared a background briefing on PCM for one of the automotive industry’s most respected analysts, on a subject we have not written to you about before: autonomous vehicles.* We want to share the thinking with you, and we want to be clear at the outset about why. This is not a product announcement, and our launch markets for PCM remain enterprise and government. The reason it belongs in your inbox is what it shows about the pattern we keep finding: industry after industry turns out to be waiting we believe for the same missing layer, and one of the largest prize markets in all of technology is among them.

Perception is solved; judgment is not

Today’s autonomous vehicles see the world remarkably well. Cameras, radar, and the neural networks that interpret them have largely solved perception. The open failures, the ones behind the incidents you read about, are failures of judgment in situations the training data never covered: the mattress lying in the lane, the police officer waving cars through a red light, the construction zone that contradicts the map.

The reason is architectural, and it will sound familiar if you have read our recent letters. Today’s driving AI is statistical. Deep networks interpolate from millions of training examples, which makes them strong on common cases and unpredictable on rare ones, and the rare ones are where the harm is. Three further consequences follow. Nobody can fully explain a decision after the fact; when a network chooses badly, the answer to why is a set of billions of weights, which is a genuine problem for regulators, insurers, and courts. Learning arrives in batches; a fleet’s lessons come back through retraining and software pushes weeks or months later, so the vehicle that met the odd situation is no smarter the next morning. And the compute is heavy; autonomy stacks commonly draw hundreds of watts to kilowatts onboard, which costs electric range, cooling, and money.

What a judgment layer changes

Readers of our last letter will recognize the shape of the answer. PCM is designed to be the layer this industry is missing: a reasoning engine that maintains a persistent, structured model of the driving world, its rules, physics, norms, geography, and its own past decisions, and derives behavior from that model rather than from whatever the training set happened to contain. In our prototype testing to date, PCM has produced zero hallucinations, and a reasoning layer that cannot invent a phantom fact is a different foundation for a safety case than one that statistically usually will not. Every decision leaves a trail: what the system believed, what it considered, and what it committed to, in a form a human can replay, so after an incident the question of why the vehicle did that has an inspectable answer. Learning is continuous and remembered; new experience updates the model in place, and we believe a fleet of PCMs can share what they learn through a federated master while each vehicle keeps deciding locally, so the fleet’s Tuesday lesson can be the fleet’s Wednesday behavior. And it is small: PCM’s edge configuration is designed so that it can run in as little as 1MB of working memory at roughly 2.4 millisecond decision latency with no GPU, and our prototype measurements show on the order of 840 times less compute per query than transformer approaches, which in a vehicle translates directly to watts, range, and bill of materials.

One scope point we made to the analyst, and repeat to you, because the candor is the credibility: PCM does not replace the perception networks. Cameras, radar, and object recognition remain the province of conventional neural networks, which are genuinely good at that job. PCM sits above them, doing what we believe they cannot: reasoning, remembering, and explaining. Knowing what is true.

How the incumbents frame the same gap

Tesla has bet on end-to-end neural networks, camera pixels in and driving controls out, trained on billions of fleet miles. It scales impressively, and it is the purest form of the black-box problem: there is no intermediate representation anyone can inspect, and improvement arrives only through retraining. Waymo layers machine learning over high-definition maps, simulation, and carefully engineered rules, and it works, at the price of substantial per-vehicle compute and a severely geofenced, pre-mapped operating world. These are serious companies doing serious engineering, and we intend no criticism of either. Our observation is that across the industry the missing layer is the same one: persistent, explainable judgment, systems that know things rather than merely recognize things, learn without retraining, and can answer for their decisions. We believe no incumbent architecture provides that layer, because transformers and end-to-end networks are structurally unable to, and it is precisely the layer PCM is designed for.

Why we are telling you this

Because it is evidence about the size of what we are building. We have written to you about hallucination gating legal and financial markets, about behavior that cannot be specified in advance gating enterprise adoption, and most recently about agent teams that coordinate without ground truth. Autonomous driving is the most vivid public example of the same gap, and it sits inside a very big transportation market. When an analyst of this standing in a market we have never pitched asks to be briefed on PCM, and the briefing writes itself because the properties map one for one onto that industry’s hardest open problems, we believe that tells you something about the breadth of demand for a working judgment layer. We are sequencing our markets deliberately, and we are not distracted; but optionality of this scale is part of what you own.

A few things to keep in mind

PCM is pre-Alpha, with Alpha targeted for the second half of this year. The zero-hallucination and 100x efficiency figures come from prototype testing, and we present them as measured results at prototype scale rather than as performance of a deployed system; the edge memory and latency figures are design specifications. Atombeam has not announced an automotive product, has no automotive customer, and is describing a direction rather than a commitment. And nothing here claims PCM replaces perception networks, high-definition maps, or the safety engineering this industry has spent two decades building. What we think PCM addresses is narrower and, we think, more interesting: the reasoning layer everyone is missing that potentially now exists in our development process.

Conclusion

My read is that the pattern is no longer a coincidence. Every market where AI spending is largest and adoption is most blocked, finance, defense, law, medicine, and now the vehicle in your driveway, is blocked by the same absence: judgment that can be trusted before the fact and explained after it. Each new industry that maps cleanly onto PCM’s properties raises our conviction about what a working judgment layer is worth, and the automotive mapping is the cleanest we have seen outside our launch markets. We would rather show than tell, and the Alpha this fall is where the showing begins.

Thank you, as always, for being with us on this journey.

Charles

Forward-Looking Statements. This update contains forward-looking statements, including statements regarding the development, capabilities, design intent, and timing of PCM, the expected properties of its architecture, potential application domains, and the markets Atombeam intends to address. Forward-looking statements are based on current expectations and assumptions and are subject to risks and uncertainties that could cause actual results to differ materially, including risks related to technology development, timing, competition, and market adoption. PCM is in pre-Alpha development; figures cited are from Atombeam prototype testing or design specifications as of July 2026 and do not describe measured performance of a deployed system. Atombeam has not announced an automotive product. Descriptions of third-party companies are based on publicly available information. Atombeam undertakes no obligation to update any forward-looking statement except as required by law.

*The PCM technology is still in development, and there are substantial technical risks that could prevent us from achieving these efficiency gains at scale. Competitive advantages in technology are often temporary, and competitors may develop alternative approaches that match or exceed our efficiency claims. Market adoption of power-efficient AI is not guaranteed, and regulatory, technical, or economic factors could impact the viability of our approach. This technology has not been validated in large-scale commercial deployments, and significant engineering challenges remain before commercial release. Investors should consider this a high-risk, early-stage technology investment with uncertain outcomes.

This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment. For more information about this offering, please view the Offering Circular and Related Risks.

Hi Everyone,

This letter began as a reply to one of our investors, who sent over an article about the industry’s shift toward teams of AI agents and asked, in effect, what it means for us. The exchange was useful enough that we wanted to share the thinking with all of you, because the shift is really happening for good reasons and we believe it takes the industry directly toward the thing PCM was designed to be.*

The industry is moving from one big model to many small agents

For years the dominant bet in AI has been that one model, made large enough, will eventually do everything. That bet is running into structural limits. One example worth knowing about: researchers at Chroma tested 18 frontier models, including GPT-4.1, Claude 4, and Gemini 2.5, and found that every one of them becomes increasingly unreliable as the amount of input grows, even on simple tasks. The problem is called context rot, and it is the kind of limit that more parameters and bigger context windows cannot (and I doubt could) fix.

The industry’s response has been to change shape. Instead of one model holding everything in its head, the new products assign specialized agents to pieces of a problem and have them coordinate: propose answers, challenge each other, converge. The research behind this is genuinely interesting; teams of weaker models, by checking each other’s reasoning, have outscored stronger models working alone on reasoning benchmarks. There are already products in the market. Moonshot AI’s Kimi K2.6, released in April, is built around what it calls an Agent Swarm, up to 300 specialized sub-agents coordinating across thousands of steps in a single run, and the other major labs are moving the same direction. I think this shift is real and could last.

What teams fix, and what they cannot

Here is what the team approach genuinely improves: when several agents challenge each other’s answers, errors get caught statistically. An agent proposes something wrong, another agent objects, the group revises. Averaged over many problems, accuracy goes up. That is a real effect and the benchmark results reflect it.

Now here is what it cannot fix. Every agent in these teams is a transformer, and the coordination between them is exchanged text. Nothing anywhere in that loop knows what is true. It is all a matter of statistical probabilities, and probabilities are happy to absorb a falsity as easily as a truth; there is no way for the model to know the difference. So a team of agents can converge, fluently and unanimously, on a shared mistake, and the more polished the conversation between them, the more convincing the mistake becomes. It is worth noticing that when practitioners list the caveats of multi-agent systems, they name coordination overhead, alignment challenges, and security risks in agent-to-agent interactions. I think those are not three separate problems; they are symptoms of one problem, which is coordination without shared ground truth.

The comparison people reach for is the human brain: separate regions handling vision, language, movement, and memory, operating independently and coordinating constantly. We like the analogy, because it quietly assumes the missing piece. Brain regions do not coordinate by sending each other paragraphs of text. They coordinate through shared, persistent structure. That is our design premise for the PCM, our geometric reasoning AI engine now in development.

A word about small language models

To explain where PCM sits in this picture, it helps to define a term. A small language model is the same kind of AI as ChatGPT, just far smaller: a compact, commercially available model, often open source, that runs on ordinary hardware instead of a datacenter. On its own it is capable but unremarkable, and it hallucinates like its larger cousins, often more, in fact. In the PCM architecture, small language models are hired for one job only: working with language, reading and writing text. They hold no domain knowledge of consequence, they never see the customer’s private data, and they can be swapped for different models without changing the system, the way you might change a tire without changing the car. We use small ones on purpose. They are cheap to run, they can live behind the customer’s firewall or at the edge, and keeping them modest makes the division of labor unmistakable. PCM only needs them when it is communicating with humans, not when doing things directly with machines.

The judgement layer

That division of labor is the design. PCM is an overlay that coordinates multiple small language models, so in one sense we share the industry’s bet on teams. The difference is where the intelligence lives. In the mainstream agent products, the intelligence is in the agents and the coordination is thin. PCM inverts that by design: the agents are deliberately ordinary, and the coordinating layer is designed to be the smart part of the system. We believe it will maintain an explicit model of the domain, check each small model’s output against that structure before the output is used, catch a fabrication before it reaches the user, and keep a record of why each answer was accepted, corrected, or rejected.

In other words, where the industry is building teams of talkers guessing the right answer, we believe we are building the member of the team that knows: the judgement layer. And there is a market logic to why that seat matters more as agents multiply. An agent workflow is a chain of hundreds of steps, each consuming the output of the last. Without a point of verified ground truth, errors do not just occur, they compound, and a system that is impressive in a demonstration becomes unaccountable in an enterprise. The markets we care about, the ones where AI spending is largest and adoption is most blocked, are precisely the ones that cannot accept unaccountable, no matter how fluent.

A few things to keep in mind

I want to be precise about the status of everything above. The “context rot” research and the agent-team benchmark results are third-party findings that I summarized here from public materials, and product descriptions like the Agent Swarm figures come from vendor announcements and industry coverage rather than independent replication by anything we have done ourselves. On our side, PCM is pre-Alpha – we have only to date completed an early prototype. The properties we describe, the domain structure, the pre-output checking, the record of each decision, are architectural design intent, developed in our theory work and not yet demonstrated in a deployed system at scale. We make no head-to-head performance claims against any of the products named here. Whether PCM delivers these properties in practice is what our Alpha, targeted for September, is intended to begin demonstrating.

Conclusion

My read is that the industry has effectively conceded, through its own change of direction, that one giant model is not the path, and that capability comes from specialized parts coordinating. We agree with the diagnosis. We believe the industry’s current prescription is missing its most important ingredient, because a team whose members all guess, coordinated by nothing but conversation, has no seat at the table for judgement. As agent teams proliferate, we think the scarce and valuable position in the whole architecture becomes the coordinator the other agents can trust, the layer that knows the domain, verifies the work, and can say why. That is the seat we designed PCM to occupy, and we believe the industry is now building the world that makes that seat valuable.

As always, do not take our word for any of this. The Chroma research is public at https://research.trychroma.com/context-rot, Moonshot’s Kimi K2.6 materials are at https://www.kimi.com, and we encourage you to read them and draw your own conclusions.

Thank you, as always, for being with us on this journey.

Charles

Forward-Looking Statements. This update contains forward-looking statements, including statements regarding the development, capabilities, design intent, and timing of PCM, the expected properties of its architecture, and the markets Atombeam intends to address. Forward-looking statements are based on current expectations and assumptions and are subject to risks and uncertainties that could cause actual results to differ materially, including risks related to technology development, timing, competition, and market adoption. PCM is in pre-Alpha development, and statements about its architecture describe design intent rather than measured performance of a deployed system. Descriptions of third-party research and products are based on publicly available materials and are provided for informational purposes only. Atombeam undertakes no obligation to update any forward-looking statement except as required by law.

*The PCM technology is still in development, and there are substantial technical risks that could prevent us from achieving these efficiency gains at scale. Competitive advantages in technology are often temporary, and competitors may develop alternative approaches that match or exceed our efficiency claims. Market adoption of power-efficient AI is not guaranteed, and regulatory, technical, or economic factors could impact the viability of our approach. This technology has not been validated in large-scale commercial deployments, and significant engineering challenges remain before commercial release. Investors should consider this a high-risk, early-stage technology investment with uncertain outcomes.

This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment. For more information about this offering, please view the Offering Circular and Related Risks.

$13 million raised.* Let that sink in for a second.

It's our biggest number yet this round, and we’re not slowing down. Over 4000+ investors have now backed us, pushing our lifetime crowdfunding total past $45 million.* We believe every dollar is a vote of confidence in where we're headed — and 2026 keeps giving people reasons to vote yes.

Case in point: the award wins keep stacking up. Neurpac has now been recognized multiple times, most recently taking two big honors — the Ascension Award for Best Data Management Solution, and the 2026 DataOps Data Breakthrough Award for Innovation of the Year. Judges called out Neurpac for "setting new standards in data optimization and analytics workflows" — not our words, theirs.**

PCM isn't far behind.*** It picked up a win at the 2026 Globees in June and was named Best AI Startup at the Merit Awards, adding to a growing list of third-party validation for the tech we're building.

None of this happens by accident. We believe it's what happens when the product actually works — and when the people funding it keep showing up.

If you haven't joined us yet, learn more today.

[Explore the Investment]

This Reg A+ offering is made available through StartEngine Primary, LLC, member FINRA/SIPC. This investment is speculative, illiquid, and involves a high degree of risk, including the possible loss of your entire investment. For more information about this offering, please view the Offering Circular and Related Risks.

*The amount raised may include insider investments, which may go toward meeting the minimum offering amount. Total includes previous Reg A+ and Reg CF offerings. This includes offerings conducted on other platforms and on different terms from the current offering. Please see the offering document for additional information.

**Neurpac’s power efficiency projections are based on prototype testing and theoretical modeling. Actual results may vary significantly from these estimates. Statements about potential revenue or commercial success are forward-looking and involve risks and uncertainties. Actual results may differ materially from any projections.

***The PCM technology is still in development, and there are substantial technical risks that could prevent us from achieving these efficiency gains at scale. Competitive advantages in technology are often temporary, and competitors may develop alternative approaches that match or exceed our efficiency claims. Market adoption of power-efficient AI is not guaranteed, and regulatory, technical, or economic factors could impact the viability of our approach. This technology has not been validated in large-scale commercial deployments, and significant engineering challenges remain before commercial release. Investors should consider this a high-risk, early-stage technology investment with uncertain outcomes.