"When the robot flipped the pancake, we were so paranoid that we went back and checked all of our millions of hours of data to look for any flipping example of any kind, and there was none."
That's Deepak Pathak, Skild AI's CEO, describing the demo that made everyone lean forward. The flip, he says, "basically emerged by looking at how the spatula is moving and the robot could just do it."
It's a genuinely striking claim, and it has exactly the texture of a GPT moment: a capability nobody trained for, falling out of scale. You watch it and you believe the future has arrived.
So here's the unglamorous question the demo doesn't answer: what happens on the fifth pancake, when the batter sticks?
Last week I walked through Skild's S1 release and flagged one missing number. This week, three converging results — S1, NVIDIA's Hydra-0, and the smaller academic papers underneath them — make it possible to say precisely which number is missing, and why the companies keep skating past it.
Start with what in-context learning actually delivers, because it's real. The pitch across S1, Generalist's GEN 1.5, and the research lineage from Edward Johns' lab is that you specify a task by showing rather than telling.
Language under-specifies. "Put away the groceries" doesn't say where the milk goes or how the jars line up against the fridge wall. A single video demonstration encodes what a sentence can't.
Skild's framing is that robotics has been "stuck in the BERT era" — every new task requiring hours of data collection and a fine-tuned specialist. In-context learning promises the ChatGPT jump: one set of weights, prompted by a video, executing tasks it never saw in training.
And S1's demonstrated horizon is not trivial. The company claims in-context execution on tasks up to ten minutes long — repotting a plant, making coffee — that were never in the pre-training distribution. That's a real advance over concurrent work, which Skild correctly notes largely covers short, in-distribution atomic actions.
But notice what all of this solves: specification. Showing rather than telling is a better way to say what you want done.
Specification is the easy half.
The hard half is what happens after second one, over the next six hundred seconds of contact, when millimeter errors start to stack. And on that half, none of the headline releases publishes the number that matters: the recovery rate under contact-heavy distribution shift, and how many human interventions it took.
The fine print tells you why they're quiet. Buried in Skild's own analysis is an admission: it took hundreds of post-training examples to reach the same performance as in-context learning — and the field has shown that with enough dense post-training data, a policy trained from scratch can match a pre-trained foundation model.
Which is to say: post-training still wins on reliability. In-context learning is the convenient specification layer sitting on top of a stack that still needs the old, expensive fine-tuning to be robust. That's not the GPT moment. That's a better prompt box bolted to the BERT era.
In-context learning is the convenient specification layer sitting on top of a stack that still needs the old, expensive fine-tuning to be robust. That's not the GPT moment.
Now look at NVIDIA's Hydra-0, which is the most honest document in the pile precisely because it's an engineering paper, not a launch.
Hydra-0's premise is elegant: represent robot actions as "motion in pixel space" — image-plane trajectories called action flow — so one world model can learn across human hands, handheld grippers, single arms, and bimanual robots without exposing any native joint vocabulary. The metrics are strong: 90.4% lower robot-motion error than the baseline, a 0.96 correlation between replayed and reference success rates on RoboLab.
But read how deployment actually works. It runs two simulators, and the division of labor is the whole story.
Isaac Lab owns the robot: "a rigid, calibrated, fully known body whose response to a command is a solved problem, so it is executed exactly rather than predicted." The world model owns everything else — "the cloth, the deformable, the contact, the lighting, the scene it has never been given a mesh of."
Read that split again. The part that's solved and executed exactly is the reaching. The part that's predicted — the uncertain, learned, could-be-wrong part — is precisely the contact patch. The deformable. The soft thing the tool touches.
That is exactly the pixel plane that stops describing reality when the spatula meets sticky batter. Action flow is validated on motion in pixel space, and it degrades in the one place manipulation actually fails: where something soft deforms and the image plane no longer tracks what's happening at the point of contact.
The gripper vendors have been saying this in plainer language all along. A camera "cannot always determine what is happening during physical interaction" — it misses slip, asymmetric contact, insertion resistance. "A grasp that works in simulation may still fail in practice." Motion is mature; manipulation isn't, because it depends on friction, slip, and deformation that can't be modeled perfectly.
So where does the failure actually happen? Not at reaching. At step ten.
This is where the smaller papers quietly demolish the scaling narrative. A team from East China Normal and Shanghai Jiao Tong universities, in an IJCAI 2026 paper, names the exact failure mode: large VLA models "suddenly fail at step 10 or later," and the cause is cumulative error propagation — a millimeter-level offset early on that amplifies through subsequent steps until the grasp collapses.
Their fix is not a bigger brain. Their S²-VLA has 2 billion parameters and hits 96.4% on LIBERO-Long, beating most 7B and even 8.5B models.
It does this by refusing to weight vision, intent, and action history statically across a whole task. A lightweight belief state tracks "where am I now, and have I drifted," then a dynamic gate reallocates attention by phase: crank up visual weight when precise alignment is needed to correct drift, crank up intent weight at sub-task switches, let action continuity dominate during steady transit.
It runs at roughly 7GB of memory and 80Hz — where comparable 7B models run at single-digit Hz. Smaller, faster, and better at the long horizon, because it schedules attention where the error actually compounds.
The same lesson shows up in the rebar-tying work. A sequence-aware hierarchical VLA with an SE(3)-equivariant diffusion backbone cuts geodesic pose error by roughly 70% in noisy physical environments and holds to 7.1% error variance under a 180-degree viewpoint change.
Rebar tying is a real construction task requiring millimeter precision across a sequence of nodes — the exact setting where memoryless baselines accumulate error and fail. The fix is geometry-equivariance and sequence awareness, not parameter count.
Put it together and the counter-narrative is coherent. The last-mile problem in manipulation is cumulative error propagation at long horizons, and the honest results are solving it with dynamic attention scheduling and geometric structure — not by scaling a larger foundation brain and hoping robustness emerges the way the pancake flip did.
The pancake, to be clear, is real, and Pathak himself won't claim the ChatGPT moment has arrived — "not quite," he says, "but this is the first sign of what we believe might come."
Fair. But a first sign is a hypothesis, and hypotheses get tested with numbers.
So the ask is the same as last week, and it stays the same until it's answered. Skild has raised nearly $1.7 billion and is moving toward deployment via the Fetch acquisition. A year on, there is still no external benchmark reporting recovery rate and intervention count under contact-heavy distribution shift.
Show us the fifth pancake, when the batter sticks. That's the number that decides deployment — and until it's published, the GPT-moment framing is a specification win wearing a reliability costume.