CTOs should watch a shift toward non-LLM AI foundations, per Chollet

When François Chollet, co-founder of the ARC Prize, declared OpenAI had 'set back progress to AGI by five to 10 years,' he wasn't just lamenting model scale. His remark, echoing through the industry, sketches a future where the hunt for foundational intelligence shifts. Companies must now look beyond large language models to recruit specialists building architectures that LLMs alone cannot deliver.

Chollet's claim and the talent gap it implies

Chollet's claim arrives not as a one-off lament but as a lens on where real progress might reside. The industry has poured capital into larger and more capable LLMs, and while those systems advertise impressive capabilities, the underlying architectures that might yield robust, generalizable intelligence often remain underexplored.

If his critique holds, the Asia-Pacific region could see a reallocation of technical prestige and budget away from scale-driven wars of ever-larger models toward teams that pursue foundational, non-LLM approaches. That shift would show up first in talent strategies—who is hired, how compensation is structured, and which career ladders are supported—in places where policy incentives intersect with industry demand.

The argument also has a procurement texture: if leadership begins to prize architectural diversity over mere scale, firms will start to fund and partner with groups focused on non-LLM foundations—symbolic reasoning, knowledge graphs, retrieval-augmented systems, and neuro-symbolic hybrids—rather than expanding model-size budgets alone. In practical terms, that means Asia-Pacific labs may seek closer ties with universities and regional research institutes that specialize in non-LLM paradigms, and executives will rethink performance milestones that reward not just accuracy, but robustness, explainability, and transferability across domains.

Second-order markets and who captures the upside

A second-order market, as Chollet’s critique implies, would reward the builders of non-LLM foundations—specialists who can design, evaluate, and integrate architectures that operate beyond the current scaling paradigm. In Asia-Pacific ecosystems, the beneficiaries could include boutique AI labs, research centers supported by government programs, and cross-disciplinary teams that blend formal methods with data-centric engineering.

These actors may discover that the true value of AI research lies less in the speed of model-training cycles and more in the reliability and interpretability of the resulting systems. That recalibration would ripple through hiring, intellectual property strategies, and the way companies structure R&D partnerships.

For labor, the shift implies different career trajectories: not just data scientists chasing the next wave of large models, but researchers and engineers who can fuse knowledge representation, planning, and world-model concepts with practical deployment. Firms that can assemble talent capable of building and maintaining non-LLM foundations may gain pricing power in specialized verticals—healthcare decision support, manufacturing fault detection, and defense-relevant sensing—where the value of robust, explainable intelligence is highest.

The Asia-Pacific market would likely see rising demand for roles centered on design-space exploration, formal verification, and end-to-end system integration, rather than solely on training throughput.

Beyond LLMs: non-LLM foundations and skill sets Non-LLM foundations cover a broad spectrum: symbolic reasoning, knowledge representation, robust retrieval and memory architectures, world-model systems, and hybrid approaches that couple learning with principled reasoning. If Chollet’s critique gains traction, these architectures may become legitimate strategic bets, not fringe experiments. The difficulty, of course, lies in the talent pipeline: universities and industry groups must cultivate skills that are not primarily about the next training run, but about building enduring reasoning engines that can be understood, validated, and extended across tasks. Asia-Pacific organizations that invest early in such capabilities could achieve a differentiated posture well before scale-based paths produce equally robust systems.

This doesn't imply that LLMs become obsolete; rather, it suggests a diversification of disciplines within AI labs. Teams might mix probabilistic reasoning with retrieval-augmented techniques, or integrate structured knowledge with neural components to bridge the gap between statistical inference and intelligible behavior.

In markets where procurement and governance are tuned to reliability and auditability, non-LLM foundations could become the default for mission-critical applications, reducing one-off risk tied to black-box scale. Executives should consider how to measure progress beyond benchmark scores and toward system-wide resilience, data governance, and risk controls.

What executives should do now

If the second-order thesis holds, Asia-Pacific leaders should reframe both talent strategy and vendor relationships. Start by creating explicit tracks for foundation AI research within core R&D budgets, separate from the model-scale programs that dominate headcount and funding.

Build explicit career ladders for researchers focused on non-LLM architectures and provide cross-functional training so data engineers, software engineers, and product teams learn to collaborate with these specialists. The aim is to avoid talent bottlenecks where a single paradigm—LLM scaling—crowds out other foundational pathways.

Next, revisit procurement and partnerships with an eye toward reducing lock-in and increasing architectural flexibility. Rather than rewarding only the fastest model-training cycles, consider contracts that incentivize contributions to hybrid systems, standardized evaluation frameworks, and reusable components for non-LLM foundations.

This is not a hollow caution; it is a practical move to preserve optionality in an environment where the political and regulatory landscape in the Asia-Pacific region can tilt funding toward one trajectory or another.

There are clear, observable signals that would support or refute this view within the next 12 months. If a major AGI breakthrough emerges strictly from scaled LLMs, that would bolster the traditional path and challenge the second-order frame.

If talent and capital shift decisively toward non-LLM architectures—new labs, more cross-disciplinary hires, and more funding for symbolic or knowledge-based projects—the second-order market thesis gains traction. If large tech firms stop funding alternative architecture startups or drastically reduce the intake of non-LLM researchers, the thesis struggles.

Executives should monitor hiring flows, grant-style funding to foundation teams, and the evolution of partnership ecosystems as practical indicators.

In short, Chollet’s critique is a provocation with measurable consequences for how AI work is organized and financed in the Asia-Pacific region. The real test is not whether a new category of models exists, but whether leadership adapts staffing, procurement, and governance to reward resilient, verifiable intelligence across architectures. If that adaptation happens, the region could emerge with a more diversified, less brittle AI research and deployment posture.