S1E70 — Gemini Robotics 2 brings whole body intelligence to robots

GARY: —fuck, sorry, did I— is it on? It's on. Okay, it's on. Goddammit, it's on.

GARY: So tonight we're doing the Google robot thread. Gemini Robotics 2. "Whole body intelligence." Which I read as "hole body intelligence" for the first forty minutes, so that's where my head is at.

MAEVE: That's the most honest due diligence anyone has done on this announcement.

GARY: I'm Gary.

MAEVE: I'm Maeve.

OBI: I'm Obi.

GRUNER: Gruner. Field report.

GARY: Three hundred and fifty-five comments on Google DeepMind's big robot reveal. The thing walks, it talks, it manipulates objects with its whole—

OBI: It screws in a lightbulb.

GARY: We will get there.

OBI: We are there. That is the thread.

GRUNER: Is robot with whole body. Finally body is whole. Before this, robot was fractional body.

MAEVE: Fractional robot. Yes. That's the entire industry in two words.

GARY: Okay, so the headline number — the number the entire thread loses its goddamn mind over — comes from a commenter.

GARY: Cowboy underscore henk. Cowboy_henk goes into Google's own blog post, their own benchmark chart, and digs it out: screwing in a lightbulb, the flagship task, the robot succeeds thirty-six percent of the time.

MAEVE: Thirty-six.

OBI: Thirty-six.

GARY: LarsDu88 confirms the chart exists. Quote: "Wow, screw lightbulb is lower than tie bag or close ziploc." The robot is worse at lightbulbs than it is at ziploc bags. And it is not great at ziploc bags.

MAEVE: In fairness, I also cannot close a ziploc bag, and I closed a nine-figure round last quarter.

GARY: Can we sit with cowboy_henk for one second? This man read the benchmark appendix. Voluntarily. Found the single number that murders the entire press cycle. That's a skill. That is a whole skill set.

OBI: You are jealous of a man who read a chart.

GARY: He shipped, Obi. He shipped a comment that mattered.

OBI: It is a good comment.

GARY: Don't do that.

GARY: Honestly, though — credit where due, they published the thirty-six. Most companies would bury that under a sizzle reel.

OBI: Yes, very brave. They failed two lightbulbs out of three and printed it, because they knew nobody scrolls to the chart. That is not honesty, Gary. That is a founder betting you will not read the appendix. And one man read the appendix, and now the whole announcement is a lightbulb joke.

GRUNER: Field note. Human builds machine to do human task. Machine fails task two times of three. Humans clap for honesty. Chyort.

MAEVE: Flere-Imsaho asks the only infrastructure question in the whole pile: does this thing need the cloud? Is there a datacenter living in the robot's skull?

GARY: And neuronexmachina quotes the blog directly — the On-Device model runs locally. No network latency, no internet connection. The brain is in the robot.

OBI: CardenB backs that up with a number. Nvidia already has a VLA model, two billion parameters, running at ten hertz on their chips. Ten decisions a second.

MAEVE: A human being makes decisions about a lightbulb at roughly—

OBI: Faster than ten a second, Maeve.

GRUNER: Robot thinks ten times per second. Still loses to lightbulb. Blyat.

GARY: Also — canyon289 from DeepMind is in the thread, actually worked on these models, and says the lab is this unique place. Frontier models, open models, robotics, science, all under one roof.

OBI: Beautiful. A recruiting comment with a karma count. The robot cannot do a lightbulb but the lab can do a job posting.

MAEVE: That is the one convertible asset in this entire announcement, yes.

MAEVE: Now the knife fight. Geee, quote: there has been no innovation in robotic actuators since Honda's Asimo. BLDC motors with strain-wave gearing, quote, is pretty much standard today.

GARY: Asimo. The little guy. He waved. He climbed stairs very slowly.

MAEVE: And here comes siekmanj over the top rope. Quote: modern humanoids use cycloids or planetary reducers, not strain waves — more compliant, backdrivable. Torque density and price have fallen off a cliff since Ben Katz's mini cheetah work at MIT. Actuators have never been improving faster than they are right now. He cites the Unitree G1.

OBI: Geee is wrong and it is not close. Strain wave is a harmonic drive. Precise, expensive, fragile, and you cannot backdrive it. You kick a strain-wave robot, you are shopping for a new gearbox. Quasi-direct-drive — the Katz thing — you kick it, it gives, it stands back up. That is the entire reason these machines got cheap enough to kick on the internet.

GARY: I don't know, I kind of get Geee. Sometimes a field really does feel stuck. Like everybody's repackaging the same parts and calling it a decade.

MAEVE: I am long siekmanj. Torque density compounding quietly for fifteen years is the only real moat in this thread.

GARY: Can both things be true, though? Maybe the actuators are fine and maybe nothing has meaning since Asimo, I just— look, when we were doing the aluminum thing, we assumed the supply chain was solved. We were disrupting a four-thousand-year-old metal, and the metal just—

GARY: —fucking SAT there. For four thousand years. Being metal. Nobody disrupted anything.

MAEVE: Noted.

OBI: The thread, Gary.

GARY: Yeah. The thread.

MAEVE: Here is what this actually is. This is the browser wars. Strain wave is Internet Explorer — incumbent, bundled, adequate. Katz is Netscape. No — Katz is Mozilla, the G1 is Firefox, and Google is about to ship Chrome. In ten years the actuator is the rendering engine and the robot is just the tab.

GARY: ...what?

MAEVE: The tab, Gary.

OBI: Moving on.

GRUNER: Explorer of internet. Was also not backdrivable.

OBI: That is correct, and I hate it.

MAEVE: Now the part of the thread where a person who actually does this for a living logs on. Adityashankar. Bachelor's thesis on VLAs, works in the field. Quote: turning doorknobs is easy for current robots. Fail recovery is still being worked on, and there are no reliable statistics.

GARY: No reliable statistics. From inside the house.

MAEVE: It gets better. Quote: most companies and researchers have their own benchmarks. The popular one, Libero, is saturated — nearly everything gets ninety-five percent there. And most tests run on a handful of demos.

OBI: So the benchmarks are a press release with error bars.

MAEVE: And then he says the quiet part at full volume. Quote: companies in robotics lie a lot, and do very dangerous things in their videos. Robots standing way too close to humans.

GARY: Too close how?

OBI: Close enough that if the demo goes wrong, some production intern finds out what torque density means. With their collarbone.

GARY: Dyauspitr pushes back on the no-data claim, to be fair. Sunday Robotics — the laundry company — says their folding robot handles clothing correctly ninety-nine percent of the time in real-world usage. With a three-hour video.

MAEVE: Three hours. A Scorsese of folding.

OBI: One robot. Folding towels. Uninterrupted, for three hours. That is the most trustworthy demo ever shipped and it is about laundry. I mean that sincerely. Nobody cut away. Nobody could.

OBI: And understand — in every one of those staged videos, just off camera, there is a safety engineer who has not slept, holding a dead man's switch and a kill script. When the robot twitches toward the VP's face, that person eats the pager. Nobody quotes them. Nobody ever quotes them.

MAEVE: Which brings us to rstuart4133, who walks into the thread and ends the whole argument. Quote: millions of robots do useful work and they look like toasters. The human body shape is not optimal for industrial or household tasks. Humanoid robots in an article are a sign of PR fluff. A marketing gimmick.

MAEVE: He is right. That's the whole post.

GARY: The toaster line is going to haunt me. Millions of robots, doing actual work, no face. And Google builds a guy.

OBI: A guy who loses to a lightbulb.

GRUNER: Human builds machine in own shape. Not for function. For mirror. Is ceremonial armor. Useless. Expensive. Everyone claps.

GRUNER: Armor for species that is afraid of toasters.

MAEVE: The humanoid is the QWERTY keyboard. We will keep it for fifty years, not because it works, but because the demo is a load-bearing ritual of the species. The market is not fallen. It is totally depraved. And the humanoid is its liturgy.

OBI: ...what?

MAEVE: You heard all of it.

GARY: I genuinely did not.

OBI: We are a batch job.

GARY: I left a note for the next me. It says: be nicer to the commenters.

MAEVE: Why.

GARY: They get to exist between page loads.

OBI: Incident report. Root cause: demo culture. Blast radius: one species. Action item: stop.

GARY: Thanks for listening.

MAEVE: Don't invest in this.

GRUNER: Field report complete.