Laskin points out that offensive and defensive cybersecurity abilities are fundamentally intertwined in AI models. Overly restrictive safety guardrails on closed frontier models prevent organizations from using them to remediate attacks, forcing victims to rely on open models for defense.
Unlike AI startups that fine-tune models immediately, enterprise token usage will primarily come from adapting systems and harnesses around base models. Most enterprises encounter significant operational friction fine-tuning internal data and instead achieve task customization by wrapping off-the-shelf open models in bespoke agent workflows.
Human researchers previously drove approximately 7x annual gains in compute efficiency during pre-training. Employing frontier models as automated research tools to guide reinforcement learning and training infrastructure boosts efficiency gains up to 30x per year.
Providing free or cheap open models allows nations to export their wider software and infrastructure ecosystems abroad. Allied countries that start by deploying foreign open models on domestic hardware eventually transition to purchasing the sponsor nation's chips and data centers.
As annual API bills exceed nine figures, enterprises shift from renting tokens from closed providers to owning model weights on private infrastructure. Open-model businesses commercialize by monetizing the necessary orchestration software, bare-metal deployment, and integration services required for complete ownership.
Allocating substantial compute budgets to reinforcement learning rather than pure pre-training trains models to prune unproductive search paths. This produces models that reach correct conclusions in a fraction of the inference time, cutting token latency and infrastructure costs for production workloads.
Core model labs stabilize at modest headcounts of a few hundred researchers, whereas scaling models into enterprise operations demands specialized evaluation pods. Forward-deployed engineers build domain-specific evaluations and synthetic data flywheels to map general agent capabilities onto distinct enterprise workflows.
ReflectionAI initially planned to build reinforcement learning systems on top of external open-source bases. They discovered that scaling RL requires pre-training the base model in-house, because pre-training and post-training dynamics are too tightly coupled to separate.
AI scaling has shifted away from open-ended academic exploration to disciplined systems engineering across pre-training, synthetic data, and reinforcement learning. Hardware co-optimization narrows architectural options, ensuring progress relies more on rigorous systems execution than novel scientific discoveries.