Skild AI's S1 Teaches Robots From a Single Video, No Fine-Tuning Required

Skild AI's S1 Teaches Robots From a Single Video, No Fine-Tuning Required

Skild AI unveiled S1, a robotics foundation model that learns unseen, long-horizon physical tasks — up to ten minutes and dozens of steps — from one egocentric human video, no post-training required [4]. On out-of-distribution tasks, S1 hit a 66% step-success rate against 9% for a language-prompted VLA baseline. One demonstration video did the work of roughly 380 post-training episodes [5].

This is the "show, don't tell" moment for embodied AI. Tasks like potting a plant, pouring coffee, or assembling a kit were demonstrated with emergent error recovery — the robot adapting when something goes wrong mid-task, without being explicitly trained to do so [6]. That's the kind of generalization robotics has been chasing for a decade.

The framing going around — a "GPT-3 moment for robotics" — isn't hyperbole. Video-based in-context learning scaling better than language conditioning at large pretraining volumes suggests the bottleneck for general-purpose robots was never compute, it was the interface. Feed the model demonstrations, not descriptions.

OpenAI Finishes Pretraining 'Bel,' a 10-Trillion-Parameter Foundation for GPT-6

Reports surfaced August 25 that OpenAI has completed pretraining on "Bel," an internal model exceeding 10 trillion parameters and successor to the earlier "Doug" checkpoint [7]. Bel is positioned as the base for both Astra, OpenAI's cyber-capable system, and the eventual GPT-6 — with some framing it as a potential AGI-threshold base model [8].

Leaked details point to an accelerating compute and chip strategy that's reportedly outpacing Anthropic's ability to respond in kind this year [9]. Whether or not "AGI-threshold" holds up to scrutiny, the compute gap being described is real and widening.

For anyone building on top of frontier APIs, this is the reminder that the ground is still shifting fast upstream. Whatever you're orchestrating today gets a capability upgrade you didn't ask for tomorrow — plan your architecture for that, not against it.

Revolut Bets Big on Proprietary AI With New Research Division

Revolut launched Revolut Research on August 25, a dedicated division building proprietary foundation models for financial services, underpinning its PRAGMA model developed with Nvidia [10]. Early numbers are strong: 2.3x better credit risk accuracy, 65% more fraud caught with a 17% precision gain, and 41% more relevant recommendations across its 80-million-customer base [11].

This is a "build, don't bolt on" bet from a company that could easily have kept fine-tuning someone else's foundation model. Instead, Revolut is treating proprietary AI as core infrastructure, not a vendor relationship — a stance more regulated industries are going to be forced into as generic LLMs hit accuracy ceilings on domain-specific risk and fraud tasks [12].

It's also a Nordic-adjacent signal worth watching: fintechs with genuine data moats are increasingly deciding that owning the model is cheaper than renting judgment from someone else's.