I have always had a certain sympathy for ideas that look slightly unreasonable at first.
This is probably a professional hazard of having spent much of my career working on artificial intelligence before artificial intelligence became the center of every corporate strategy presentation. Over the years I have worked on brain-inspired neural systems, neuromorphic computing, continual learning, autonomous machines and other technologies that, depending on the year and the audience, were described as visionary, impractical, premature or occasionally all three at the same time.
Some of those ideas eventually became important. Others did not. A few were technically correct but commercially early, which is a particularly frustrating category because being right at the wrong time does not pay salaries.
This creates a difficult problem for anyone evaluating a genuinely new AI architecture. A CEO, investor or board member cannot simply ask whether the technology is different, because novelty by itself has very little economic value. At the same time, evaluating a non-consensus architecture using only the assumptions of the incumbent technology is equally dangerous, because truly new systems often look inefficient or unnecessary until the problem they solve becomes sufficiently important.
I have encountered this tension repeatedly, first as a scientist and later while building Neurala, raising venture capital, developing intellectual property, working with semiconductor partners and bringing AI into commercial environments. Over time I developed a way of thinking about new architectures that is less concerned with whether they resemble whatever is currently fashionable and more concerned with whether they create a capability that matters, whether that capability can be defended, and whether a company can actually be built around it.
A new architecture needs a reason to exist
The first question I ask when I encounter a new AI architecture is embarrassingly simple: what becomes possible because this architecture exists?
This question eliminates a surprising amount of noise.
A new system may have a clever learning rule, a beautiful mathematical formulation or a particularly elegant connection to neuroscience, but unless one of those properties produces a meaningful advantage, the architecture remains primarily an intellectual achievement.
The advantage also has to matter enough to justify change. Technology markets contain enormous inertia, and customers are reluctant to replace working systems simply because an alternative is scientifically more interesting. A new architecture therefore needs to improve something customers care about significantly enough to compensate for the friction created by adoption.
Perhaps it learns from far less data. Perhaps it can adapt continuously instead of requiring periodic retraining. Perhaps memory consumption remains bounded as context increases. Perhaps it can operate at much lower power, preserve state over long periods, behave more robustly under changing conditions or execute efficiently on hardware where conventional models become impractical.
Any of these can be valuable, but the architectural novelty has to connect to a real constraint.
This was very much the experience we had at Neurala. Some of our early interest in continual learning, local adaptation and efficient computation came from autonomous systems where cloud connectivity, abundant training data and unlimited compute could not be assumed. The architecture was not interesting merely because it resembled aspects of biological learning; it was interesting because the environment created requirements conventional approaches handled poorly.
When evaluating a new AI company, I therefore try to understand the constraint before I become too interested in the architecture.
Biology can inspire the question without becoming the answer
I have spent a large part of my career studying biological intelligence, and I still believe the brain is one of the most useful existence proofs available to AI researchers. It demonstrates that intelligence can operate continuously, learn with remarkable efficiency, process multiple sensory streams and function under severe energy constraints.
This makes biology extremely valuable as inspiration.
It also creates a temptation.
Once a system is described as brain-inspired, people occasionally become less demanding about the evidence because the analogy itself sounds persuasive. I have never found this particularly helpful.
A computational model inspired by cortical organization, hippocampal memory, synaptic plasticity or some other biological mechanism still has to demonstrate an engineering advantage. If the architecture claims improved continual learning, measure it. If it claims lower energy use, measure it. If it claims better memory, sample efficiency or robustness, construct experiments capable of proving the claim wrong.
The scientific heritage of an architecture should make the experimental standard higher rather than lower.
This distinction matters enormously for boards and investors because scientific prestige can create an aura around a technology. The right response is not skepticism for the sake of skepticism, but disciplined curiosity. A sophisticated board should understand the hypothesis well enough to know what evidence would falsify it, which is very different from either blindly believing the scientists or treating every unconventional idea as academic speculation.
In my experience, the strongest technical organizations are comfortable with this kind of tension because scientists themselves should be trained to live with it. A hypothesis is useful precisely because there is a possibility that reality will disagree.
Intellectual property becomes more important before everybody agrees with you
If the architecture does appear to create something valuable, the next question is ownership.
This is especially important in deep technology because there is often a long period between the original insight and widespread commercial adoption. During that period, the technology evolves, competitors appear and ideas that once seemed strange can become surprisingly obvious in retrospect.
I have seen this cycle more than once.
We filed patents around neural computation, learning and acceleration during periods when some of these areas attracted far less commercial attention than they do today. The experience taught me that IP strategy has to begin early enough to protect what is genuinely differentiated, but also that patents should never be confused with the business itself.
A company can have a wonderful patent portfolio and no market.
It can also have excellent technology and discover that the important part is difficult to protect.
The more useful question for a board is therefore not whether the company owns patents, but what portion of the value remains difficult for another organization to reproduce if the technology succeeds.
Sometimes this is formal intellectual property. Sometimes it is training methodology, compiler technology, deployment knowledge, proprietary data, customer integration, semiconductor partnerships or an accumulated body of engineering that looks simple only after somebody has spent ten years solving the details.
Usually the moat is a combination of several things, and in good deep-tech companies it changes over time.
This is why I am suspicious of very clean answers to the moat question. Real companies are messy systems, and defensibility often comes from the interaction of technology, organization, ecosystem and time rather than from one perfect patent.
Every AI architecture eventually has to run somewhere
Another question that I believe receives too little attention at board level is what the architecture implies about computing.
An AI model does not exist independently from the machine on which it runs. Different architectures generate very different patterns of computation, memory movement, communication and precision, and these differences can have enormous consequences for both performance and economics.
Some models map naturally onto GPUs. Others may benefit from FPGAs, specialized accelerators, in-memory approaches or entirely different computing substrates. An algorithm that appears efficient mathematically may perform poorly when forced onto hardware designed around another computational pattern.
My own career has repeatedly crossed this boundary, from neuromorphic computing and DARPA’s SyNAPSE program to commercial work with semiconductor companies including Qualcomm, Sony Semiconductor Solutions, Movidius and Lattice Semiconductor. Those experiences convinced me that the computing substrate has to be considered surprisingly early when evaluating a new AI architecture.
This is not because every startup should immediately design custom silicon, which is usually a very expensive way to discover that the software has not yet found a market. It is because compute influences far more than technical performance.
It influences cost, gross margin, latency, energy consumption, deployment flexibility and strategic partnerships. If an architecture eventually requires specialized hardware, that requirement also changes the amount of capital needed and the time required to bring the technology to market.
A board evaluating a new AI architecture should therefore understand not only whether the model works, but whether the economics of making it work are compatible with the company being built around it.
The first market should make the architecture look unfair
Technical founders naturally like horizontal platforms because the underlying invention often appears applicable to many things.
Unfortunately, “applicable to many things” is not the same as “urgently needed by someone.”
A new architecture usually needs an initial market where its particular advantage matters disproportionately. If the customer can solve the problem adequately using an established model and a little more compute, adopting a new architecture may create more risk than value.
The best early markets often contain some constraint that makes the new approach unusually attractive.
A robotics system may require very low latency. An industrial application may need local adaptation. An embedded device may have severe power limitations. A sovereign or regulated environment may require computation and data to remain local. A long-context application may face an increasingly expensive memory problem.
The architecture should encounter a market where its difference matters immediately.
At Neurala, our path moved through several markets, from autonomous systems to embedded AI and eventually industrial visual inspection. The applications changed, but the markets that worked best were always those where some aspect of our technology solved a constraint that customers genuinely felt.
This is one of the areas where technical strategy and commercial strategy become inseparable. The architecture does not determine the market automatically, but it should give the company a reason to win somewhere.
If it does not, I would worry.
A technical founder has to become bilingual very quickly
One of the more interesting transitions in my own career was learning to explain the same technology to very different audiences.
A researcher wants to understand the mechanism. An engineer wants to know how it will be implemented. An investor wants to understand the size and defensibility of the opportunity. A customer wants to know whether it solves a problem and how difficult deployment will be. A board wants to understand which risks are technical, which are commercial and which are organizational.
The CEO sits in the middle of all of these conversations.
This is particularly important in deep technology because it is easy for sophisticated technical language to create the appearance of a business case where none yet exists. If the leadership team can explain the architecture brilliantly but cannot explain why a customer will pay for it, more work is required.
I sometimes use a simple test: describe the value of the company without naming the architecture.
If the product is based on a new recurrent model, a state-space architecture, a neuromorphic system or something completely new, remove those words and explain what changes for the customer.
Does the system learn continuously?
Does it cost less to operate?
Can it run on existing hardware?
Can it remember much longer sequences without exploding compute costs?
Can a customer deploy it privately?
Can a robot react faster?
If the commercial value survives the removal of the technical vocabulary, the company is beginning to understand what it sells.
This does not mean simplifying the science until it becomes meaningless. It means understanding the consequence of the science well enough to explain why somebody outside the research group should care.
Deep-tech companies change character as they grow
Another difficulty for boards is that the organization required to discover a new architecture is not necessarily the organization required to commercialize it.
Very early in the life of a deep-tech company, scientific freedom can be essential. The company may still be testing fundamental assumptions, and forcing premature product discipline can destroy the very experimentation that makes the technology valuable.
Later, engineering discipline becomes increasingly important because experiments have to become reproducible systems. Eventually customer requirements, product management, sales, partnerships and operations become equally important, and the company has to learn how to deliver technology repeatedly rather than demonstrating it once.
These transitions are difficult because the people who were perfect for one phase may not naturally be perfect for the next.
A brilliant scientist may not want to build a commercial organization. A highly experienced enterprise executive may become deeply uncomfortable inside a company where the architecture changes every month. A founder who was excellent at raising a seed round may need different support when the company begins negotiating strategic partnerships with global technology companies.
This is where boards can create enormous value if they understand the stage of the company rather than imposing a generic model of what leadership should look like.
Over many years as a founder and CEO, I came to appreciate that leadership in deep technology requires several forms of credibility at the same time. There has to be enough scientific understanding to distinguish a real breakthrough from fashionable language, enough operating experience to convert research into a product, enough commercial judgment to identify the markets where the technology matters, and enough capital discipline to finance a journey whose duration is rarely as predictable as the spreadsheet suggests.
Very few individuals possess all of these capabilities equally, which is precisely why the composition of the leadership team and board matters.
Good governance should make the technology stronger
There is a tendency in visionary technology companies to think that a supportive board is one that believes deeply in the mission.
I think support is more complicated than that.
A board that believes everything can become dangerous, while a board that understands nothing tends to become conservative for reasons that have little to do with the actual technology.
The useful position lies somewhere in between, where directors understand the scientific thesis well enough to challenge it, understand the market well enough to know whether the advantage matters, and understand company building well enough to distinguish a temporary technical setback from evidence that the original assumption may be wrong.
This requires a certain degree of intellectual humility from everyone involved.
Scientists have to accept that customers may care about something different from the original research objective. Commercial leaders have to accept that difficult science cannot always be scheduled like conventional software development. Investors have to distinguish between patience and denial. Founders have to be willing to change their minds without abandoning the reason the company exists.
That balance is particularly important when the company is pursuing a non-consensus architecture because there will inevitably be periods when external opinion swings between excessive skepticism and excessive enthusiasm.
Neither is especially useful.
What matters is whether the evidence continues to support the thesis.
I like unreasonable ideas (if they are reasonable, you are too late!)
After many years working at the intersection of neuroscience, AI, hardware and company building, I remain attracted to technologies that challenge assumptions about how intelligence should be built. Part of this is scientific curiosity, but part of it comes from having seen several ideas move from obscure research topics to important commercial technologies.
I have also seen enough deep technology to know that being different is not sufficient.
A new architecture has to produce an advantage that matters, that advantage has to survive contact with hardware and economics, the company has to protect enough of the resulting value to justify the investment, and there has to be a market where customers care about the difference strongly enough to take the risk of adopting something new.
When these elements align, an unconventional architecture can create an extraordinary company.
When they do not, the same architecture may remain a very interesting paper.
The difficult job for the scientist, CEO, investor and board is discovering which situation they are looking at before reality makes the answer obvious.
I do not know of a perfect formula for doing this, and I would distrust anyone who claims to have one. What I do know is that the right questions tend to remain remarkably stable even as the architectures change.
What capability exists because this technology exists? What evidence would convince us that the thesis is wrong? What part of the value can the company retain? What does the architecture imply about compute and capital? Which customer cares enough to adopt it early? Does the leadership team understand both the science and the business around it? And, perhaps most importantly, is the organization still capable of learning as quickly as the technology it is trying to build?
Those are the questions I would want answered before placing a serious bet on the next AI architecture