"The first time S1 flipped a pancake, we assumed pancake flipping must have been in its pre-training data. We searched our whole pre-training data and found no examples of flipping."
That is Skild AI describing its own product launch. Read it twice.
The engineers who built S1 watched it do something, didn't believe it, and went digging through their own training set to find the trick. Their first instinct wasn't celebration—it was suspicion.
Hold onto that instinct, because it's the correct posture toward every claim in this announcement. The difference is that Skild's team got to search the training data. You don't.
The difference is that Skild's team got to search the training data. You don't.
Here's what S1 is, in Skild's framing: a foundation model that "learns from one example." You show it a video of a task it has never seen, and without any fine-tuning it outputs robot actions to reproduce that task—"in any environment and in any embodiment." Tasks up to ten minutes long. Coffee, potting a plant, the famous pancake.
The pitch is that robotics finally gets its language-model moment. You prompt with a demonstration the way you prompt GPT with text, and behavior falls out. In-context learning for hands.
If true, this is enormous. So let's do what Skild's own engineers did and look for the trick.
The problem isn't the pancake. The problem is that every load-bearing claim here is Skild grading its own homework.
Take the headline comparison: to match the accuracy S1 achieves from one video prompt, Skild says "current VLA models would need to be post-trained with 50-100 hours of data collection followed by fine-tuning." That's a striking number. It's also entirely internal.
Which VLA? Which task? Measured at what success threshold, under what conditions? "50-100 hours" is a range wide enough to drive a forklift through, attached to no benchmark anyone outside Skild can run.
Then there's the money quote for investors: for novel tasks, S1 "exponentially outperforms any existing language-prompted VLAs as we scale the pre-training." Exponentially is a mathematical word doing marketing work here.
Exponential in what variable? Against what baseline, measured how? Without a curve, an axis, or a third-party comparison, "exponentially outperforms" isn't a measurement. It's a verb chosen to feel like one.
And notice what's missing from the whole thread. There's no success rate. There's no failure rate. There's no count of how often a human had to step in.
Skild does make one genuinely interesting operational claim: S1 "does not blindly replay the video demonstration." It can withstand perturbations, improvise on mistakes the human didn't make, and sometimes execute more precisely than the person in the prompt video.
That's the right thing to be excited about. Robustness to disturbance—recovering when the scene isn't the scene you trained on—is the actual hard problem in deployed robotics. It's what separates a demo from a machine you'd put on a line.
But it's precisely the claim that arrives with zero numbers. "Can withstand perturbations" is a sentence, not a metric.
The useful version of that claim is a recovery rate: when the object moves, when the lighting changes, when someone bumps the arm—what fraction of the time does S1 get back on track, and how many human interventions per hour does it cost? That's the number we've been asking every robotics maker to publish, and it's the number none of them will. S1's untested perturbation handling is exactly the metric that open question demands—and exactly the one this launch skips.
A foundation-model demo can be cherry-picked. You film the runs that worked. You pick the pancake that flipped. Nothing about a highlight reel tells you the denominator.
Now look at what landed the same day, from the other end of the ideological spectrum.
Figure came out of stealth with a thesis that reads like a direct rebuttal: "The data needed to scale a truly general purpose robots doesn't exist on the internet—it has to come from the real world."
And then Figure did the thing Skild didn't. It shipped integers.
264,000 app downloads. Over 16 million video uploads. Data collected in 108 countries. $15 million reportedly paid out to date. Over 43,000 weekly active users. Uploading more than 30 minutes of video every second. And a commitment, per Figure, to spend over $1 billion on data and compute in the next twelve months.
Those numbers are grubby. They describe a logistics operation, not a scientific breakthrough. And that's exactly why they're more trustworthy.
You can cherry-pick a pancake. You cannot cherry-pick a data pipeline's throughput. Thirty minutes of video per second, per Figure, is either happening or it isn't. $15 million, per Figure, either left the building or it didn't. A pipeline's numbers can be audited in a way a demo reel never can.
Here's the reframe that matters. Skild and Figure are not disagreeing about robotics. They're agreeing about the deepest fact in the field, and responding to it in opposite directions.
Both are staring at the same bottleneck: there is no independent robot data. The internet has text and images; it does not have millions of hours of embodied manipulation across bodies and environments. Everyone in this business knows it.
Figure's answer is to admit the bill and pay it—build the collection pipeline, ship the map of lights across 108 countries, commit the billion. The bet is that the data problem is real and cannot be finessed away.
Skild's answer is to claim it doesn't have to pay the bill—that a clever enough architecture can infer the out-of-distribution task from a single video, so you never need the 50-100 hours in the first place. The bet is that the data problem can be architected around.
These are two answers to one unanswered question. And only one of them came with a falsifiable number attached.
That's the tell. Figure's claim is structured so it can be checked and, in principle, proven wrong—the app downloads either climb toward the promised 100x or they stall. Skild's central claims are structured so they can't be checked at all, because there's no external benchmark, no reported distribution shift, no intervention count.
None of this means S1 is fake. The pancake may well be exactly what Skild says: a genuinely novel task inferred from one prompt, with nothing like it in pre-training. If so, it's a real result and it deserves a real paper with real baselines.
The point is narrower and harder to dodge. On the day these two announcements went out, one company published throughput and one company published adjectives.
As one reply under the launch put it: sure, try a food robot that learns a dish from one video—just don't put it on a line until the same plate comes out the same way all week.
That's the whole test, and it's a number Skild has not reported. Until S1's recovery rate after the scene changes shows up—under distribution shift, with intervention counts, against a baseline someone else can run—Figure's grubby integers are the more honest artifact in the room.
Skild's engineers didn't trust the pancake until they'd checked the data. Extend them the courtesy of the same standard. We just don't get to search the training set—so we're left asking for the one thing that would settle it, and noticing who declined to provide it.