8VC Bets $3 Billion on Generalist’s 12-Second Robots

Generalist is now valued at $3 billion after 8VC led nearly $200 million into the quiet robotics lab. The check extends a $400 million Series B that closed in June at $2 billion. Five days before the filing, the company showed a model that learns a new chore from 3 to 12 seconds of video.

That is the wager. 8VC is paying up for one-shot robot learning, at a price that is still a fraction of the other robot-brain labs, while the lab’s own tests leave most of those short tasks unfinished.

The bet is narrow and public at the same time. Capital arrived after a demo, not before it. The model’s published scores, not a warehouse pilot, are what the new price capitalizes.

8VC Just Wrote the Same Check Twice

People familiar with the financing said 8VC led the extra capital and that existing backers joined. Generalist and 8VC did not comment. A federal filing put the new equity at nearly $200 million, taking the Series B to about $600 million.

The step-up is 50 percent in roughly 11 weeks, from $2 billion on June 4 to $3 billion. For a company founded in 2024 by Pete Florence and Andy Zeng, both formerly of Google DeepMind, and Andrew Barry, formerly of Boston Dynamics, the speed is the product. Early money came from 8VC and Radical Ventures, plus Nvidia, Union Square Ventures, Bezos Expeditions, and Fei-Fei Li.

Until this summer the lab barely published. Then the model cadence tightened, and the checks followed the demos rather than the other way around.

FROM STEALTH LAB TO A $3 BILLION ROUND

1. 2024: Florence, Zeng, and Barry found Generalist to train one model across many robot bodies.
2. November 2025: GEN-0 lands as the first public scale-up of the lab’s physical-data engine.
3. April 2, 2026: GEN-1 is shown with 99% success on simple tasks, about 3x faster speeds, and about 1 hour of robot data per result.
4. June 4, 2026: Radical Ventures leads a $400 million Series B at a $2 billion price; 8VC publishes its thesis the same day.
5. August 19, 2026: GEN-1.5 is released as a one-shot learner prompted with 3 to 12 seconds of a single demo.
6. August 24, 2026: A federal filing records nearly $200 million of new equity, and 8VC is named as the lead.

The extension is follow-on money, not a new investor class walking in. 8VC already sat in the cap table. Leading the extra $200 million after GEN-1.5 is a second, louder vote for the same technical claim.

Eleven weeks separate the two prices. The same lead that helped open the round is now writing into the extension. That pattern reads as conviction on the technical path, not as a search for a new anchor.

Existing backers joining the extra capital keeps the cap table familiar. The filing still makes the raise visible. Nearly $200 million on top of $400 million is a large follow-on by any early-company standard, and it lands five days after a product drop rather than after a long quiet period.

What GEN-1.5 Does in 12 Seconds

GEN-1.5 is a large multimodal model that takes video, other sensors, language, and joint-position data, keeps about 30 seconds of memory, and emits action trajectories at 100 Hz. The lab calls a short sensorimotor clip in that window a physical prompt. No extra training step is required for the one-shot case.

The company’s own write-up is blunt about the scope. The tasks are simple and short-horizon. Zippers, jar lids, wallets, and brushing a block into a bowl are the test set, not a full warehouse shift. Florence, the CEO, was a senior author on PaLM-E and RT-2 at DeepMind. Zeng, the chief scientist, led Code as Policies. Barry was a senior roboticist at Boston Dynamics. The pitch is that this group can turn those short clips into a general robot brain.

WHAT THE MODEL CAN DO FROM ONE CLIP

• One-shot prompting: A single 3-to-12-second demonstration is inserted in context and the robot tries the task with no gradient update.
• Prompt chaining: Two different clips in the same window are stitched into one longer behavior, including regrasps that appear in neither demo.
• Sim-to-real prompts: A clip recorded only in simulation can be used on a real robot, even though pretraining had no simulation data.
• Human-to-robot copy: In some cases a person does the motion with their own hands in view of the robot’s cameras, and the robot repeats it.
• Few-shot polish: 1 to 10 gradient steps on 1 to 5 minutes of data, about 10 to 50 demos, raise the score.
• Tool improvisation: After a brush demo, the model has used a banana as a makeshift brush and a dustpan as a scoop.

The lab says it did not train for any of those behaviors on purpose. No extra architecture for in-context learning, no meta-learning loop, no side objective for improvisation. GEN-1.5 had been pretraining for more than 8 months because next-action error on a held-out set kept falling. The official launch clip is the cleanest view of the claim.

Introducing GEN-1.5, a one-shot learner.

It can learn new tasks in a few seconds. Show it what to do, and it generalizes.

This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world. pic.twitter.com/ptB9ElYXMU

— Generalist (@GeneralistAI) August 19, 2026

Sholto Douglas, who works on scaling reinforcement learning at Anthropic and previously at DeepMind, replied to that post with a two-word verdict: “GPT3!” The analogy is doing a lot of work. GPT-3’s one-shot language score was about 45% average, and about 65% with a larger few-shot set. Generalist is inviting the same reading for arms and grippers.

A 30-second context window has to hold the prompt and the live stream at once. Action leaves the model at 100 Hz, so the buffer is busy while the arm moves. That design choice is why a 3-to-12-second clip can stand in for a training set on the short tasks the lab chose to publish.

Prompt chaining stretches the same window further. Two clips become one behavior, and regrasps show up even when neither demo contains them. Sim-to-real prompts and human-to-robot copy point the same direction: the pretrained weights already carry enough structure for the physical prompt to steer the arm without a new training run.

The 59 Percent This Round Is Priced On

The number under the video is less cinematic. Across 10 diverse manipulation tasks, one-shot physical prompting produced 59% average success, plus or minus 10 percentage points, straight from the pretrained model. Ten gradient steps on 5 minutes of data per task lifted that to 83%, plus or minus 9 points.

GEN-1.5 ON THE COMPANY’S OWN TESTS

• One-shot score: 59% average success with 3 to 12 seconds of a single demonstration and zero training updates.
• Few-shot score: 83% average after 10 gradient steps on about 50 demonstrations, or 5 minutes of data per task.
• Context window: 30 seconds of memory, with the rest of the buffer holding live observations while the robot moves at 100 Hz.
• The lab’s caveat: Skills learned only in context are more brittle than fine-tuned ones, even when they recover from some mistakes.

Those figures are the ones this round is priced on. The company puts the one-shot number next to the demo and lets the gap to 83% show what a few minutes of extra data still buys.

Fifty-nine percent on short chores is not a finished product. It is also not a random baseline. The lab reports the score with a wide band, and it flags brittleness when skill lives only in context. Investors still paid a 50 percent step-up after seeing that chart.

The few-shot lift matters for the same reason. Ten gradient steps and about 50 demonstrations move the average to 83 percent. That is still measured on the same simple task family: zippers, jar lids, wallets, and a block brushed into a bowl. The path from clip to higher reliability is short in data terms, and long in product terms.

GEN-1 had posted 99% success on simple tasks with about 1 hour of robot data per result and about 3x faster speeds. GEN-1.5 trades some of that polished success for a different interface: seconds of video in context instead of an hour of dedicated collection per skill. The two releases read as stages of one argument about how a general robot brain should be taught.

The Founders Carry Three Lab Histories

Florence, Zeng, and Barry did not arrive from the same bench. Two came out of Google DeepMind work on multimodal and policy models. One came out of Boston Dynamics hardware and control. The company story treats that mix as the reason a single model can be aimed at many robot bodies.

Florence was a senior author on PaLM-E and RT-2. Zeng led Code as Policies. Those lines of work already tried to connect language, vision, and action inside one stack. Barry’s Boston Dynamics background adds the other half: machines that have to keep balance and contact while the policy runs.

Generalist’s founding aim in 2024 was explicit on that point. One model, many bodies. GEN-0 was the first public scale-up of the physical-data engine. GEN-1 tightened success and speed on simple tasks. GEN-1.5 makes the teaching channel a short video clip. Each release keeps the same founders in the same roles: Florence as CEO, Zeng as chief scientist, Barry as the senior roboticist voice in the room.

Early backers matched that profile. 8VC and Radical Ventures came in early. Nvidia, Union Square Ventures, Bezos Expeditions, and Fei-Fei Li added names that already sit near robotics, chips, or large vision systems. The August extension does not rewrite that list. It doubles down on it.

The GPT-3 Parallel Sets A Public Bar

Douglas’s two-word reply fixed a reference point in public. GPT-3’s one-shot language score was about 45% average. A larger few-shot set reached about 65%. GEN-1.5’s own card shows 59% one-shot and 83% after light fine-tuning on the manipulation suite.

The analogy is rough by design. Language tokens are not joint torques. A zipper is not a sentence. What transfers is the shape of the claim: a big pretrained model, a short prompt, a score that is incomplete and still surprising relative to older robot teaching loops.

Generalist is inviting that reading for arms and grippers. The launch clip, the 3-to-12-second physical prompt, and the decision to publish 59% instead of only the cleanest takes all point the same way. The lab wants the round, and the audience, to judge the work as an in-context learner first.

• GPT-3 one-shot language: about 45% average.
• GPT-3 larger few-shot: about 65% average.
• GEN-1.5 one-shot physical prompt: 59% average on 10 tasks.
• GEN-1.5 few-shot polish: 83% average after 10 gradient steps.

The gaps are part of the pitch. Context-only skills stay more brittle than fine-tuned ones, by the lab’s own caveat. Tool improvisation, such as using a banana as a brush or a dustpan as a scoop after a brush demo, is presented as an emergent side effect of long pretraining, not as a trained trick. Next-action error on a held-out set kept falling across more than 8 months of pretraining, and the company stopped there long enough to ship GEN-1.5.

For 8VC, the parallel supplies a familiar frame for an unfamiliar product. Paying up to $3 billion after a one-shot demo is easier to explain if the reference class is large-model prompting rather than classical robot programming. The filing five days later turns that frame into a priced claim.

The short-horizon test set still bounds the story. Zippers and jar lids are not a shift on a warehouse floor. Fifty-nine percent leaves most trials unfinished. The wager is that the same pretraining path that produced prompt chaining, sim-to-real clips, and human-to-robot copy will keep lifting the floor without forcing a full data collection cycle for every new chore. The $200 million extension is the market’s current weight on that sentence.