Start your day with intelligence. Get The OODA Daily Pulse.
Informing your decisions with actionable intelligence
Start your day with intelligence. Get The OODA Daily Pulse.
Informing your decisions with actionable intelligence
By David Bray, PhD
I am going to be frank: I almost opted to title this article “Dear Silicon Valley AI Companies: Put Your Money Where Your Mouths Are” however I opted not to do this because a good Stoic views the world as opportunities to find good and improve the things that they can within their scope of influence.
Instead, I’m writing this article as a call to action to the big AI companies to actually do something that can move the needle on AI security vs. pieces that admire the issue yet ultimately won’t go anywhere. For example, getting the G7 to agree on AI policies let alone the ~195 countries of the world would be near impossible to say nothing of transnational corporations. Put simply, there are many counterparties with different motivations.
I am going to assume good faith on the calls for a pause, on the purported concerns that AI + bio risks “doom!”, and on the various safety approaches each big AI company posits they want to do. I am going to assume that these companies genuinely want to contribute to a safer, more resilient world. In assuming good faith, I’m going to suggest three here-and-now things that, consistent with good Stoic principles, AI companies could do immediately if they really are concerned, that are both practical and operational today. After presenting those three items, I’m going to suggest three things we as individuals also can do, even if we’re not in a leadership role for a large AI company.
If we are serious about “guardrails” in AI systems, we must fundamentally rethink where and how we implement safety measures. Currently, much of the industry relies on post-compute filters: attempts to scrub or block harmful outputs after the model has already processed the information and generated a response. This approach is flawed because filters can be overcome through clever prompting, adversarial attacks, or simple persistence.
Instead, we should be looking for pre-compute flavors of guardrails. This means architectural constraints and safety measures that prevent the generation of harmful content at the foundational level before the compute process even begins. By moving the security focus to the pre-compute stage, we create a more resilient system that does not rely on the fallible hope that a secondary filter will catch every edge case or malicious attempt to bypass safety measures.
Think of it this way: post-compute filters are like trying to catch all the sparks after striking a match in a dry forest. Pre-compute guardrails are like ensuring the conditions never allow the match to ignite in the first place. The latter approach is inherently more robust and less vulnerable to exploitation.
This is not merely a theoretical concern. We have already seen numerous examples of users “jailbreaking” AI systems by finding creative ways around post-compute filters. Each new filter spawns new techniques to circumvent it, creating an endless arms race. Pre-compute guardrails, by contrast, establish boundaries at the architectural level that are far more difficult to bypass because they are woven into the fundamental logic of how the system operates.
Pre-compute approaches have additional advantages too, we can extend more liberty to individuals, communities, and non-AI companies to decide what they do and do not want the AI to include when computing a response. Post-compute filters maintained by the AI company itself do not permit this. Pre-compute approaches allow the AI to update itself as it goes vs. the top-heavy approach to training and tuning different generative AI models in advance. Pre-compute approaches are also beneficial since we can demonstrate compliance with statutory environments that require such levels of control.
Similarly, if we want more predictability in AI systems, especially in the context of disruption or in systems that are changing quickly, we should look for complementary Bayesian-based approaches that are repeatable and not generative. This includes methods like active inference, which provides a mathematical framework for systems to maintain internal stability while interacting with an unpredictable environment. Such approaches are also effective in detecting the emergence of random cause variation (or adversarial manipulation) that would otherwise undermine the performance of the application.
While large language models and other generative AI systems are impressive in their capabilities, they are non-deterministic and often lack the repeatability that enterprise and government clients require. Ask the same generative AI the same question twice, and you may well get two different answers. For many applications, particularly in critical infrastructure, healthcare, finance, and national security, this variability is unacceptable and potentially dangerous.
Bayesian-based approaches offer a complementary path. By integrating Bayesian logic and active inference frameworks and other self-correcting methodologies, companies can provide the predictability and repeatability that many use cases demand. These methods allow for systems that behave in a consistent, auditable manner where the method is transparent in how it got to the answer from the original data corpus.
This recommendation is not about abandoning generative AI. It is about augmenting the use of existing AI in a way that responds to specific concerns. Furthermore, it is about recognizing that different applications require different approaches, and that a mature AI industry should offer a portfolio of solutions rather than a one-size-fits-all generative model. Companies that invest in Bayesian-based approaches alongside their generative AI offerings will be better positioned to serve clients who need reliability and repeatability, not just impressive but unpredictable outputs. The same companies would gain faster time-to-market, lowered costs of development, and be more responsive to market demands for AI capabilities.
Moreover, Bayesian approaches can wrap around generative AI to provide the pre-compute capabilities I mentioned earlier. By establishing probabilistic boundaries and expectations before the generative process begins, we can create hybrid systems that combine the creativity of generative AI with the reliability of Bayesian frameworks.
Striving to get the G7, let alone the 190+ nations of the Earth, to agree on a development pause is near-impossible. History teaches us that such pauses are not only difficult to achieve but often counterproductive when they are attempted.
The year 1899 provides a sobering lesson. At the Hague Convention, the nations of Europe agreed to pause development of certain military technologies, including poison-filled projectiles and chemical weapons. What happened next? They proceeded to research that technology in secret, covertly. When World War I broke out, both the Central and Allied powers used the purportedly “paused” technology against each other within the first six months of the conflict. The public pause had simply driven dangerous research underground, where it proceeded without public awareness or ethical discussions.
If the “big three” AI companies are serious about a pause, they should look inward rather than toward international treaties that are unlikely to be honored. Perhaps committing now (vs. some future date or pledge) they could commit one percent of their revenue toward proactive efforts to address here-and-now concerns about AI safety, bias, misuse, and societal impact.
This would be a tangible demonstration of putting safety and societal benefit before solely short-term profit maximization. It would signal to the public, to policymakers, and to competitors that these companies are willing to make real sacrifices for the values they claim to hold. One percent of revenue from companies valued in tens or hundreds of billions of dollars would represent a substantial investment in addressing the very real challenges that AI presents today, not just hypothetical future risks.
Such an investment could fund independent research into AI safety, support community-level initiatives to prepare for AI-driven economic disruption, develop better detection systems for AI-generated inauthentic information risks, or create educational programs to help the public understand and navigate an AI-augmented world. The possibilities are numerous, and the impact would be immediate and measurable.
That said, for-profit public companies (including companies that might soon have an IPO) might face shareholder pressures not to spend any additional money on anything outside the core business. So how then could one appeal to them to be altruistic? Perhaps a research and development credit for putting revenue towards improving safety specifically in the here-and-now, or something similar to this, would help? There’s also the hope that if the companies themselves lead the way in pioneering safety, they can attract more customer share in industries where this is required such as healthcare, banking, and more.
As we consider these actions, we must also learn from the history of well-intended interventions that backfired because they ignored local context, imposed solutions from the top down, and failed to empower those closest to the problems. The parallels to current AI governance debates are striking and instructive.
Between the 1950s and 1960s under the Food for Peace program, created by Dwight D. Eisenhower in 1954, the markets of India, Pakistan and Indonesia had to compete with the massive flow of donated agricultural products from the United States. The donations bankrupted thousands of local farmers and restricted the development of agriculture in these countries for decades. What was intended as humanitarian aid became an economic disaster because it failed to consider the local agricultural ecosystems and the livelihoods of those it was meant to help.
In 1971, the Norwegian government earmarked $22 million for a fish processing plant in Kenya on Lake Turkana. The aim was to export the fish and provide employment for the Turkana people, but they were nomads with no knowledge or interest in fishing. In addition, the cost of refrigeration equipment and drinking water were exceedingly high. The plant closed after a few days. Millions of dollars were wasted because the solution was designed in Oslo without meaningful input from the Turkana people themselves.
The World Bank loaned Tanzania more than $10 million for cashew nut processing. As a result, by 1982 Tanzania had eleven factories capable of processing three times what was produced each year. On top of that, within only a brief time period, six of the factories were idle and in need of spare parts and the other five were running at less than 20 percent of their capacity. It was cheaper for Tanzania to send its raw cashew nuts to India for processing. Once again, a top-down solution imposed without understanding local realities created waste and dependency rather than sustainable development.
These historical failures share common threads: centralized decision-making disconnected from local realities, one-size-fits-all solutions imposed without consultation, and a failure to empower those closest to the problems to develop their own solutions. The lessons are directly applicable to AI governance today.
When AI companies or international bodies propose sweeping global agreements on AI development, they risk repeating these mistakes, but at an alarmingly faster pace and with a degree of interdependency that nearly guarantees amplifying any damaging impact. Different communities have different needs, different values, and different capacities. A solution that works for San Francisco may not work for rural Kentucky. An approach that makes sense for a research university may not make sense for a manufacturing community.
Moreover, just as the Food for Peace program undermined local agricultural capacity, overly restrictive AI regulations imposed from the top down could stifle local innovation and create dependency on a small number of large AI providers. Just as the Turkana fish processing plant failed because it ignored local knowledge and priorities, AI governance frameworks that do not include a broader range of voices from affected communities will fail to address real needs and concerns.
The alternative, and better path forward, is to empower the edge. This means supporting local AI development that respects data sovereignty and community values. It means creating frameworks that allow different communities to make different choices about how AI is deployed in their contexts. It means listening to and learning from those who will live with the consequences of AI systems, not just those who build them.
This context is why the three actions I have proposed for AI companies focus on practical, operational steps that can be implemented now, that respect local autonomy, and that empower rather than constrain. Pre-compute guardrails, Bayesian approaches for predictability, and direct investment in addressing current challenges are all actions that can be tailored to different contexts and that build capacity rather than creating dependency.
For those of us who are not in charge of a “big three” AI company, there are still meaningful actions we can take to shape the future of AI in ways consistent with the values of free societies.
We can champion and support companies that are pioneering the ability for us to run AI locally if we want. This matters because then we create market pressures for big AI companies to also adopt pre-compute guardrails that we can define vs. post-compute filters that they control.
Some of us may opt to run hybrid local-and-cloud based models or hire a firm to maintain our AI for us, whatever the case by pioneering local AI we encourage a landscape where we can make choices as to when and where we run AI and thus can exercise our one individual-level, family-level, and community-level choices. This includes companies like Apple, which is working on the hardware to make local AI a reality on personal devices. It includes platforms like OpenTeams that provide optionality in running open-weight models locally if we choose. It also includes companies like HowSo in North Carolina that use Bayesian-based approaches to wrap around generative AI, allowing for some of the pre-compute capabilities and predictability that I am advocating for with the “big AI” companies.
By supporting these companies with our purchasing decisions, our investments, and our public advocacy, we send a clear market signal that we value data sovereignty, privacy, and the freedom to choose where and how our AI systems operate. We demonstrate that there is a viable business case for AI that respects individual autonomy rather than requiring dependence on centralized cloud services that vacuum up our data.
We can encourage both the news media and the “AI safety” industry to do more than produce short headlines or surface-level safety discussions. Recognizing we are all pressed for time, we should nevertheless champion and share articles that go deeper, like Scientific American’s recent article that included other perspectives in its discussion, noting “AI could soon start improving itself. Critics aren’t convinced.” We all deserve more nuanced articles versus breathless ones that either hype AI as a panacea or fear-monger about imminent doom.
We also need to be cautious of the “AI safety” industry becoming checklists-only or an industry that provides the illusion of safety or security, similar to how some cybersecurity audits were illusionary, expensive, and did not actually improve the cybersecurity posture of organizations. We should insist that any activity, whether media coverage or safety-related work, is nuanced and actually moves the needle on real problems while also being mindful of unintended impact analogous to the examples cited above.
We should also avoid one-sided fears of doom or what might be called apocalyptic-voyeurism: the tendency to focus obsessively on extreme scenarios while ignoring the here-and-now harms and challenges that AI already presents. Yes, we should think seriously about long-term risks, but not at the expense of addressing the immediate issues of bias, inauthentic information, economic disruption, and erosion of privacy that are happening right now.
We can encourage folks to get out of their bubbles, whether they work at a big tech company, a startup, a national government agency, or as an academic researcher, policy wonk, or industry analyst. It is only when we get out of our normal bubbles and engage in interdisciplinary conversations, most importantly with people at the local level who may be impacted by a tech activity, that we avoid bubble-induced thinking that misses the complete picture.
We have to engage folks across disciplines and community to avoid repeating past mistakes incurred through “monochannel thinking” linked to operating in bubbles of thought. This includes what we do with AI and the Future of Work, AI and the Future of Education, and AI and the Future of Free Societies. These futures should not be decided by any one group. We must engage folks at the local level, recognizing that different local communities will have different priorities, different values, and different needs.
A solution that works for San Francisco may not work for rural Kentucky. An approach that makes sense for a university town may not make sense for a manufacturing community. By engaging across these boundaries, we ensure that AI development and deployment efforts reflect the plurality of human experience and values, not just the perspectives of a narrow technical or economic elite.
I am striving to assume good faith from the AI companies, and I am presenting this call to them in the hope that they will see these actions as opportunities to do both good for themselves and good in the world. If they really want a more attractive market valuation and want it to be a sustainable path for themselves as companies and for the world, acting on these three paths will lead by example, rather than stoking anxieties that lead to anger and risk a backlash.
Companies can simultaneously address here-and-now concerns and long-term risks by embracing the best way to predict the future: by building it. By investing in pre-compute guardrails, by developing Bayesian-based approaches for predictability, and by putting real resources toward addressing current AI challenges, these companies can demonstrate that they are serious about their stated values.
For all of us, by supporting local AI options, demanding substantive rather than superficial engagement with AI issues, and breaking out of our bubbles to engage across disciplines and communities, we can help shape an AI future that enhances rather than diminishes human freedom, agency, and flourishing.
The path forward is not through international treaties that in all probability would be either unenforceable or ignored, nor through pauses that will simply drive research underground. It is through concrete, operational actions taken by companies and individuals alike. The best way to predict the future, including one where we have freedom, agency, and protection from unreasonable harm, has been and remains: fund it, build it, produce it.
Furthermore, focusing on what can be done within one’s control remains the timeless way of a good Stoic. Each of us, whether a CEO of a major company or a non-CEO who wants to see good in the world can (1) focus on what is within our control, (2) act with good judgement and wisdom, and (3) work towards the good outcomes even in the face of uncertainty and complexity. The AI companies have an opportunity to lead. Meanwhile we all, whether a leader at a big AI company or not, have opportunities to support and demand more purposeful, actionable, results-oriented leadership in the here-and-now, not in the distant future. Together, we can build an AI future worthy of free societies and free individuals.
Read more articles from Dr. David Bray here.
About the Author
Informing your decisions with actionable intelligence
Informing your decisions with actionable intelligence