What Blackwell, Figure 01, and SIMA Actually Demonstrated

What Blackwell, Figure 01, and SIMA Actually Demonstrated

Three announcements landed in the same week in March 2024 and made one future feel close. NVIDIA introduced the Blackwell computing platform. Figure showed a humanoid robot responding to spoken requests. Google DeepMind introduced an agent that could follow language instructions across several 3D game environments.

Each demonstration mattered for a different reason. None established that a general robot, autonomous agent, or trillion-parameter model was ready for ordinary unsupervised use.

Blackwell changed the scale of the system

NVIDIA announced Blackwell as a platform rather than one isolated chip. Its launch materials described two connected GPU dies, new interconnects, transformer-focused computing, networking, and rack-scale systems.

The company said the platform could support training and inference for models reaching trillions of parameters and claimed large cost, energy, and performance improvements for specified workloads compared with its predecessor. Those figures were vendor claims tied to particular configurations, not a universal statement that every workload would become 30 times faster.

The durable change was architectural. Larger AI systems increasingly depend on how processors, memory, networking, cooling, software, and power work together. The useful unit of comparison is often the whole system running a defined workload.

Figure 01 connected language to physical action

Figure's demo placed a humanoid robot at a table with dishes, trash, and an apple. A person asked for something to eat, requested an explanation, and asked the robot to put objects away.

The impressive part was the connection between perception, language, and manipulation. The robot had to identify an edible object, plan a physical action, execute it, and explain the apparent reason for its choice.

The demo did not reveal enough to measure reliability, intervention, prior setup, repeatability, speed, safety, or performance in an unstructured home. Figure later moved to newer robots and its own Helix models. Figure 01 should therefore be treated as a historical prototype that made an integration legible, not as a current product recommendation.

SIMA used games as a test environment

Google DeepMind's SIMA research asked a different question. Could one agent follow natural-language instructions across several 3D virtual worlds rather than master one game for a score?

Games gave the researchers varied environments with perception, navigation, changing goals, and real-time action, without the physical consequences of a robot making a mistake. The research announcement emphasized instruction following, not unconstrained play or a direct transfer into the physical world.

Simulation can accelerate learning and evaluation because experiments are repeatable and failures are contained. It also leaves a transfer problem. Success inside a game does not prove that an agent can handle physical uncertainty, safety obligations, or conditions absent from the simulation.

The demonstrations formed a stack

Blackwell represented computing infrastructure. SIMA represented learning and action across virtual environments. Figure 01 represented language and perception connected to a body.

It was tempting to combine them into a single prediction about household robots or artificial general intelligence. The evidence did not justify that leap. The announcements came from different teams, used different systems, measured different tasks, and exposed different limitations.

The responsible synthesis is narrower. Compute, simulation, multimodal models, and robotics were advancing in ways that could reinforce each other. Every layer still needed its own evaluation.

What to ask after a convincing demo

Ask what task was completed, under which conditions, how often it succeeded, what human setup or intervention occurred, what failure looked like, and whether the published measure matches the real use being claimed.

A demo can establish that an integration is possible. Deployment evidence has to establish that the system remains useful, safe, and supportable when the conditions stop being selected for the presentation.

Continue the conversation

The full episode preserves Dalton's first reaction to the Blackwell launch, the Figure 01 table demo, SIMA, synthetic identity, and questions about AI personhood. Listen on Spotify.

Sources and editorial notes

This article uses the preserved [[E08 - Transcript - ep8-ai-explosion-robot-butlers-nvidias-new-brain-mind-blowing-demos-take-03 (SRT only 1)|raw SRT transcript]], NVIDIA's Blackwell announcement, Figure's current company record, and Google DeepMind's SIMA research announcement. The article does not infer consciousness, personhood, rights, workforce replacement, or household readiness from the demonstrations.