If you are curious about Atombeam, PCM*, or the AI industry, you’ll want to register for Atombeam’s next webinar.

It takes place Tuesday, August 11, at 12pm PT / 3pm ET.

This will be a live conversation between our CEO, Charles Yeomans,  and the AI Scientist for PCM, Alexandria Tucker – and of course you, Atombeam’s investors and prospective investors.

We’ll talk:

• Latest developments in PCM and what they mean for the future of AI
• Trends, opportunities, and challenges in the larger AI industry
• Where we’re headed next, and when we expect this to be ready

Plus we will do a live Q&A at the end of the session, so you can find out what you’ve always wanted to know about PCM. 3 days left to secure your spot!

[Register Today]

*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.

Due to unexpected technical issues, we unfortunately need to cancel today’s livestream. We apologize for the inconvenience and appreciate your patience.

In the meantime, we hope you’ll join us next Tuesday for a live conversation between Atombeam CEO and Chairman Charles Yeomans and Alexandria Tucker, AI Scientist for PCM.*

They’ll discuss the science and philosophy behind Atombeam’s technology, our unique approach to AI, and some of the opportunities and challenges shaping the industry today. We’ll also open up the conversation for a live audience Q&A at the end.

Date: Tuesday, August 11, 2026

Time: 12:00 PM PT / 3:00 PM ET

Register Now: https://luma.com/h9uxyzao

We look forward to seeing you there!

The Atombeam Team

*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,

Last week Anthropic, the maker of Claude and one of the most important AI labs, published a research study about its own models that we think deserves your attention. It is among the most candid things a frontier lab has said in public about how these systems actually behave, and it goes to the heart of why we are building PCM the way we are building it.*

What Anthropic found

Anthropic wanted to know what values its deployed models actually express in real conversations. Values here means tendencies like warmth, caution, rigor, and candor, the qualities that shape how a model answers when there is no single right answer. To find out, they sampled 309,815 real conversations from Claude.ai, had another AI label each conversation against a list of 339 value categories, and then used statistical compression to boil the results down to four axes: deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution.

A few findings stand out to us. The four axes capture 15% of the variation in expressed values after controlling for what users were asking about, which means the remaining 85% of the variation is uncharacterized. Different models within the same product family express measurably different value profiles, shaped by training choices whose effects the company can observe only after deployment. Most striking to us, the same model treats people differently depending on the language of the conversation. Claude leans toward warmth in Hindi and Arabic and toward rigor in English and Russian, and Anthropic offers its own example of what that means in practice: two people asking for feedback on the same business plan, one in Hindi and one in Russian, may come away with different impressions of its quality.

The line that matters most comes near the end. Anthropic writes that the values its models express “vary in ways we didn’t deliberately choose.” Until this study, the company says, those values were something it could shape during training but could not reliably observe once the model was deployed. We would encourage you to sit with that sentence for a moment, because it comes from the vendor itself.

Why we think this matters

The study is careful and well executed, and we read Anthropic’s decision to do this work and publish it as a real credit to them. But step back and consider what it says about the architecture class.

A transformer-based model learns its behavior from data. Nobody writes down how it should weigh warmth against rigor; those tendencies emerge from training and live inside billions of parameters. So when a vendor wants to know how its own product behaves, the only available method is the one Anthropic used: sample hundreds of thousands of conversations after deployment, label them with another AI, and assemble a statistical portrait. Even done well, that portrait explained 15% of the variation, and Anthropic states plainly that it does not yet know which properties of its training data drive the differences it found.

We think the plain-English translation is this: with today’s dominant AI architecture, behavior cannot be specified in advance. It can only be sampled after the fact, partially, at great effort, by the vendor itself.

For a consumer chatbot that may be tolerable. For an enterprise it is a real problem. A bank running credit analysis or a defense agency supporting intelligence work needs to know how its system will behave before it acts, and needs a definite answer afterward about why it did what it did. A statistical tendency measured across a population of conversations provides neither. The language finding sharpens the point. A multinational whose analysts in one country get differently framed answers than analysts in another, to the same question, has a defect it cannot see, cannot predict, and cannot fix, because the vendor itself cannot.

We wrote to you recently about hallucination as a barrier that keeps AI out of high-stakes markets. We believe this study describes another face of the same underlying issue. Hallucination is unpredictable content, and what Anthropic measured is unpredictable character. Both trace to the same root: a transformer’s behavior is an emergent property of its training data rather than a consequence of explicit structure.

How PCM approaches this by design

PCM, our geometric reasoning engine now in development, is built on the opposite premise. Its behavior is designed to derive from explicit structure the machine maintains, rather than from statistical tendencies absorbed during training. Several design properties bear directly on what Anthropic measured.

Explainability is built in at the level of each decision. PCM is designed so that every commitment it makes, including a decision to decline or to do nothing, freezes a record of the route taken and the alternatives rejected. Asked later why it acted, the machine reads the answer from that record. There is no need to survey 300,000 conversations to estimate what the system tends to do, because each individual decision carries its own account.

Conviction is a set parameter, whereas in a transformer it is an emergent trait. In PCM’s design, how strongly the system holds a belief before revising it is governed by an explicit, tunable quantity, and the architecture forbids the limit where a belief becomes unrevisable. Anthropic’s paper closes by asking whether expressed values can be reliably steered, with the proposed test being to adjust training and then re-measure the model’s behavior. In PCM the analogous dial is a named parameter whose effect follows from the design.

Behavior comes from seeded structure, so it should not drift with the accidents of training data. Our enterprise approach pre-seeds PCM’s geometry from domain sources chosen for the customer’s industry. To the extent behavior differs across deployments, we intend that to be deliberate configuration rather than an unobserved artifact of uneven data, which is what Anthropic identified as the likely source of the language differences it found.

A few things to keep in mind

We want to be careful about what we are and are not claiming. Anthropic’s findings are measured behavior of a deployed system at very large scale. PCM is pre-Alpha, and the properties described above are architectural design intent, developed in our theory work but not yet demonstrated in a deployed system at scale. It would be wrong to compare their measurements to our designs as if both were data, and we are not doing that. Our claim is narrower: the limitation Anthropic describes appears to us to be structural to systems whose behavior emerges from training data, and PCM’s architecture is built so the question does not arise in the same form. Whether PCM delivers these properties in practice is exactly what our Alpha, targeted for the second half of this year, is intended to begin demonstrating.

We would also repeat that we regard Anthropic’s transparency here as admirable, and the point is stronger for it. If the lab most serious about understanding its own model’s behavior can characterize that behavior only statistically and partially, we think the limitation lies in the architecture class, and no amount of diligence by any single vendor overcomes it.

So my read is

So my read is that the industry’s leading labs are beginning to document, in their own research, the gap between shaping an AI’s behavior and knowing it. That gap is invisible in a chatbot and disqualifying in a refinery or a command center. The markets we care about, the ones where AI spending is largest and adoption is most blocked, are precisely the ones that require behavior that can be specified before the fact and explained after it. We believe an architecture that provides those properties by construction, and does not rely on post hoc measurement to discover them, is what unlocks those markets, and that is the architecture we aim to build.

As always, do not take our word for it. The Anthropic study is public and readable, and we encourage you to read it yourself and draw your own conclusions: https://anthropic.com/research/claude-values-models-languages

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 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.