The Next AI Breakthrough May Never Be Published — and That Should Concern Every Board

When the most consequential AI advances happen inside private companies that do not share their methods, the organisations deploying AI are building strategies on systems they cannot evaluate.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/the-next-ai-breakthrough-may-never?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

In June 2025, Andy Konwinski, co-founder of Databricks and Perplexity AI, committed 100 million dollars of his own money to establish the Laude Institute — a nonprofit dedicated to funding open AI research from academic institutions and helping university breakthroughs reach the world as open-source tools, startups, and publicly available products. The institute’s first flagship grant was 3 million dollars annually for five years to anchor a new AI Systems Lab at UC Berkeley. Its board includes Google’s head of AI Jeff Dean, Turing Award winner Dave Patterson, and Meta’s former Vice President of AI Research Joelle Pineau. At the ACM AI conference, Konwinski stated plainly what motivated the initiative: major technology companies are withholding the data necessary to replicate their technologies, this lack of transparency undermines scientific progress, and researchers increasingly face an impossible choice between academic freedom and the compute required to do frontier work.

What Konwinski warned about in 2025 has become measurably worse in 2026. The Laude Institute is not a response to a future risk. It is a response to a structural shift already underway — and one that has significant implications for every organisation deploying AI today.

CONTEXT AND BACKGROUND

For most of the history of computing, the most consequential scientific breakthroughs were published openly, reviewed publicly, and built on collectively. The foundational papers that made modern AI possible — the attention mechanism that underpins transformer architecture, the reinforcement learning techniques that produced breakthrough reasoning capabilities, the training methodologies behind the first large language models — were all published in academic journals and conference proceedings accessible to any researcher in the world. That open science culture produced the AI boom. It is now being dismantled by the same companies that benefited most from it.

The shift is documented and dated. AI Frontiers’ analysis of the hidden frontier of AI development, published in August 2025, identified a widening gap between the capabilities of models that remain internal to frontier labs and those that are publicly released. OpenAI, for instance, developed a model with mathematical capabilities sufficient to achieve gold medal performance at the world’s most prestigious mathematics competition — and declined to release it publicly. Increasingly, AI systems with capabilities considerably ahead of what the public can access remain confined inside corporate walls. The reasoning is transparent: as AI systems approach human-level performance at technical tasks, the competitive advantage of keeping them internal grows. A company with exclusive access to an AI system that can meaningfully accelerate its own research and development has every incentive to guard that advantage, and no regulatory obligation to share it.

The Laude Institute’s founding is a direct response to this dynamic. Konwinski’s argument, developed from his experience taking Apache Spark from a Berkeley PhD project to a technology that became the foundation of Databricks, is that the most consequential companies in the history of technology did not emerge from a vacuum of capital and talent, but from a deep, collaborative foundation of open academic research. Google emerged from PageRank, published openly at Stanford. Databricks emerged from the Berkeley AMPLab, whose research was available to anyone. The recursive engine of open science, in which one published breakthrough enables ten others, produced the AI capabilities that private companies are now treating as proprietary competitive moats.

INSIGHT AND ANALYSIS

The privatisation of AI knowledge creates three specific problems that boards and executives need to understand, because each of them directly affects the quality of the decisions organisations are making about AI right now.

The first is the verification paradox. When an AI company publishes the architecture, training methodology, and evaluation data behind a model, independent researchers can test its claims, identify its failure modes, and build on its foundations. When a company releases a model without publishing how it was built, the organisations deploying it are making strategic and operational decisions about a system they cannot independently evaluate. They are relying on the company’s own safety testing, the company’s own capability assessments, and the company’s own characterisation of the system’s limitations. That is not a scientific basis for a board-level decision. It is a marketing relationship — and for the high-stakes applications where AI is increasingly being deployed, covering credit decisions, medical triage, legal analysis, and public service delivery, the absence of independent verification is a governance gap that regulators are beginning to identify and that boards should be naming explicitly.

The second is the talent concentration pipeline. Konwinski identified this at the ACM AI conference with unusual directness: the academic-to-private pipeline is accelerating in one direction, and the conditions required to reverse it do not currently exist. Researchers who remain in universities have freedom but not compute. Researchers who join frontier labs have compute but restricted publication rights. The choice they are making en masse is the one that pays the bills — and the consequence is that the independent research base required to evaluate, challenge, and improve frontier AI systems is contracting at precisely the moment when frontier AI capabilities are expanding most rapidly. The Laude Institute is an attempt to offer a third option, but 100 million dollars is a meaningful intervention in a field where research budgets are measured in billions.

The third problem is the one that matters most for boards specifically: structural governance risk. The organisations deploying AI in 2026 are building strategies, making investments, and accepting governance responsibilities on the basis of systems whose inner workings they cannot access, whose failure modes are not fully published, and whose long-term behaviour is assessed exclusively by the companies that have a commercial interest in the assessment being favourable. That is a specific and quantifiable governance risk, and it is one that almost no AI governance framework currently in use explicitly addresses.

MIT Technology Review’s January 2026 analysis of the AI landscape identified that competitive pressure between US and Chinese AI companies has intensified the incentive for closed development at precisely the moment when AI capabilities are advancing most rapidly. DeepSeek’s open-source releases briefly created pressure on American labs to publish more openly, but that pressure has not produced a sustained shift. The dominant dynamic remains closed development, selective disclosure, and capability claims that external researchers cannot independently verify.

IMPLICATIONS

For boards and executives, the implications of closed AI research are not abstract. They are embedded in every AI procurement decision, every AI governance framework, and every AI risk assessment that organisations are producing right now.

When your organisation evaluates an AI system from a frontier provider, the documentation available to you is what that provider has chosen to release. The safety testing behind the system was conducted by the provider. The capability assessments were produced by the provider. The failure mode analysis, to the extent it exists in published form, reflects what the provider believed was safe and commercially appropriate to disclose. Independent evaluation of the kind that peer review provides in open science does not exist for most frontier AI systems, because the information required to conduct it has not been published. That means the governance frameworks your organisation has built around AI are, in a meaningful sense, built on trust rather than evidence — trust that the provider’s internal testing was rigorous, trust that the disclosed failure modes are the material ones, and trust that the system will behave in production the way it behaved in the provider’s evaluation environment.

The organisations best positioned for what comes next are those building AI governance frameworks robust enough to manage a landscape in which independent verification is limited, in which failure modes are partially disclosed, and in which scientific consensus on a system’s behaviour does not exist because the scientific community does not have access to the data required to form one. That means more rigorous internal testing, more demanding contractual disclosure requirements from AI providers, and more serious engagement with the emerging regulatory standards that are beginning to require exactly the kind of transparency that closed development currently prevents.

For Africa specifically, the stakes are higher than the global conversation acknowledges. A continent that cannot train frontier models, that depends on open knowledge ecosystems for its AI research capacity, and that needs to understand the systems it is deploying in healthcare, education, and financial services cannot afford a world in which the most powerful AI systems are comprehensible only to the organisations that built them. The Laude Institute’s argument — that open research is not a moral luxury but a practical infrastructure requirement — is more urgent for African institutions and governments than for almost any other audience in the world.

CLOSING TAKEAWAY

The history of technology is a history of open science. The transformer architecture that powers every major AI system in use today was published in a paper titled Attention Is All You Need, available freely to any researcher who wanted to build on it. The research culture that produced that paper, and the thousands of papers that followed it, is now under commercial pressure that the Laude Institute alone cannot resolve and that a 100 million dollar fund cannot reverse.

What boards can do is name the governance implication of closed AI development clearly and act on it. The AI system your organisation is deploying was built in ways you cannot fully verify, tested in ways you cannot independently reproduce, and characterised in ways you cannot externally audit. That is not a reason to stop deploying AI. It is a reason to build governance frameworks that are honest about what they can and cannot assure — and to support, actively and explicitly, the open research infrastructure that would eventually make those assurances possible.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net