Fable 5 Got Export-Banned Three Days After Launch — And It Doesn't Matter

Human Progress Can Be Reversed. Human Knowledge Is Much Harder to Unlearn.

TLDR: The US export ban on Anthropic's Fable 5 made headlines, but the author spent three days with the model before access was cut and came away thinking the ban changes very little. Open source models trail frontier capabilities by only 3-6 months, and that gap is operationally meaningless for most businesses. The real constraint on AI adoption was never model access — it's organizational readiness.

Summary:

There's something almost poetic about the timing. Jonas Braadbaart, the author of TheCircuit, got three days with Anthropic's Fable 5 before the US government ordered Anthropic to shut off access for non-US nationals on June 12th. Three days. And his takeaway isn't panic or outrage — it's a calm argument that export controls are the wrong lever to pull if you want to slow AI adoption.

The argument hinges on a gap that keeps shrinking. Open source models are currently only 3-6 months behind frontier capabilities. The city of Rio de Janeiro just released an open source model that outperforms Qwen 3.7, one of the most-downloaded open source models around. When a municipal government in Brazil can ship a competitive model, the idea that US export controls will meaningfully bottleneck global AI capability looks pretty thin. For 98% of real-world AI applications, the difference between getting January 2026 capabilities in January versus April is essentially noise — business processes take far longer to change than that anyway.

The electricity analogy that Azeem Azhar makes cuts to the heart of it. Factories only started seeing productivity gains from electrification once they'd settled on a specific deployment model — the unit drive — and then rebuilt the factory floor around it. That took twenty years. The parallel to enterprise AI is uncomfortable but accurate: the bottleneck isn't access to the commodity, it's the organizational redesign required to use it well. And ironically, the commodity in question — AI — requires that other commodity in droves: electricity.

There's also the self-improvement angle, which the author flags as the most contentious claim in the piece. Anthropic has stated that its models are now self-improving. Over the past two years, training has definitively shifted from RLHF and SFT — human-provided feedback from free ChatGPT users and labeling teams — toward RLAIF, reinforcement learning from AI feedback. Current generation models are now capable enough to help train the next generation, effectively removing humans from the training input loop. Frontier models have been making legitimate scientific discoveries for some time now. If that's true, export bans are chasing a moving target.

The Douglas Adams quote that anchors the piece is the kind of thing that sounds funny until you sit with it. Anything invented after you're thirty-five is against the natural order of things. The posturing around AI — the panic, the regulation, the breathless proclamations — follows the same pattern as the dot-com era, the VCR, the washing machine. The technology moves. The organizations catch up slowly. The regulations lag behind both.

Key takeaways:

• The US export ban on Fable 5 went into effect 3 days after launch for non-US nationals
• Open source models trail frontier models by 3-6 months — operationally irrelevant for most use cases
• Rio de Janeiro released an open source model outperforming Qwen 3.7, illustrating how distributed AI development has become
• AI training has shifted from RLHF (human feedback) to RLAIF (AI feedback), meaning models now help train their successors
• Azeem Azhar's electricity/factory analogy: productivity gains required 20 years of organizational redesign, not just access to the technology
• The real constraint on AI adoption is organizational change, not model access

Why do I care:

As someone building software on top of these models, the export control conversation tends to obscure what actually matters. The question of whether your team can access a specific frontier model is almost always less important than whether your deployment architecture, your data pipelines, and your organizational processes are ready to use AI effectively. The 3-6 month open source lag argument is worth internalizing — it means the capability is effectively commoditized for most practical purposes, and the competitive moat lives in how you use it, not which model you have access to. The shift to RLAIF is the more interesting story here, because it changes the feedback loop economics fundamentally: you're no longer paying humans to label data at scale, and the models are getting better faster as a result.

Human progress can be reversed. Human knowledge is much harder to unlearn.