What Blackwell, Figure 01, and SIMA Demonstrated

What Blackwell, Figure 01, and SIMA Demonstrated in 2024

Three announcements in March 2024 made the same future look closer from different directions. NVIDIA introduced a new computing platform. Figure showed a humanoid robot connecting spoken requests to physical actions. Google DeepMind presented an agent that followed language instructions across virtual 3D environments.

The durable lesson is not that household robots or general intelligence arrived that week. It is that infrastructure, virtual learning, and physical control were progressing at the same time. Each source established something real, but each sat at a different level of evidence.

flowchart LR
    A["Blackwell launch"] --> D["Infrastructure claim"]
    B["SIMA research"] --> E["Virtual instruction-following evidence"]
    C["Figure 01 demo"] --> F["Selected physical integration"]
    D --> G["Separate deployment tests"]
    E --> G
    F --> G

Why the week felt important

The original E008 recording captures my reaction while the announcements were fresh. I saw the outline of a stack. More capable computing could support larger models and more simulation. Virtual environments could give agents varied places to learn. Better models and controls could connect language to a robot body.

That connection was reasonable as a direction. It was too early to treat the three systems as one completed pipeline. NVIDIA, Figure, and Google DeepMind were working on different products and research questions. Their announcements used different tests, conditions, and measures.

The more useful reading is to ask what kind of evidence each source supplied.

Blackwell was a platform launch

NVIDIA introduced Blackwell on March 18, 2024. The announcement described a platform that included GPU architecture, connected dies, transformer-focused computing, NVLink, networking, reliability features, security features, and complete systems.

The company named products such as the B200 GPU, GB200 Grace Blackwell Superchip, and GB200 NVL72. That matters because a performance claim about a rack-scale configuration is not simply a property of one interchangeable chip.

NVIDIA said Blackwell could support trillion-parameter models and claimed up to 25 times lower cost and energy for specified inference comparisons. Those were vendor results tied to named workloads, software, precision, systems, and comparison points. They were not a promise that every AI workload would become 25 times cheaper or more efficient.

The strongest launch conclusion was architectural. AI performance increasingly depended on the full system, including processors, memory, networking, software, power, and cooling.

Figure 01 made an integration visible

The Figure 01 demonstration placed a humanoid robot at a table with an apple, trash, dishes, and a rack. A person asked for something to eat. The robot handed over the apple. It moved objects, placed dishes, and explained parts of its behavior.

The sequence was compelling because perception, language, explanation, planning, and manipulation appeared in one continuous interaction. It made an abstract model-to-action connection easy to see.

Figure now describes F.01 as its first generation. The current Figure company history places F.02 and F.03 after it. That chronology makes the 2024 demo a historical prototype record, not a description of Figure's current robot or model stack.

The video did not publish the number of attempts, failed runs, prior setup, intervention, safety controls, or performance over a shift. It demonstrated a selected integration. It did not demonstrate general household readiness.

SIMA tested language-driven action in virtual worlds

Google DeepMind introduced SIMA on March 13, 2024. SIMA stands for Scalable Instructable Multiworld Agent. Its research question was whether one agent could follow free-form language instructions across varied 3D environments.

The agent received image observations and language instructions, then produced keyboard-and-mouse actions. The team used research environments and commercial games. The work focused on language-driven generality rather than maximizing a game score.

That is different from an agent that simply wanders without purpose. The instructions and task definitions mattered. The technical report also made the environment set, interface, evaluation, and preliminary nature of the results visible.

Games offered varied, interactive worlds where failures were contained and experiments could be repeated. Success there did not prove that the agent could control a physical robot or handle the risks of an uncontrolled environment.

Later versions clarify the historical boundary

The years after E008 make the evidence categories easier to see. NVIDIA introduced Blackwell Ultra and then Rubin as later platform generations. Figure moved through F.02, F.03, and Helix models. Google DeepMind introduced SIMA 2 in November 2025.

Those later systems do not turn the 2024 claims into mistakes. They show why a public article needs dates and lifecycle labels. A launch, prototype, and research report can be accurate and important without remaining the current product state.

What a convincing demo still leaves unanswered

A demonstration can prove that an observed behavior occurred. Deployment requires a different record. The task, environment, duration, variation, failure rate, human intervention, recovery, maintenance, support, privacy, and safety controls all become part of the product.

That is especially important around physical systems. OSHA's robotics overview notes that many robot accidents occur during non-routine conditions such as programming, maintenance, testing, setup, and adjustment. A clean presentation usually removes exactly those operating moments from view.

The next question after a successful demo is not whether it was fake. The better question is what it measured and which decision that evidence can support.

Where the Venture Step thread continues

E005 provides the NVIDIA computing and CUDA context. E010 returns to robotics demonstrations through GR00T and Neuralink. E025 adds operating-domain and safety thinking. E045 and E068 examine tactile sensing, simulation, and robot foundation models. E117 brings the question down to high-force hardware, while E119 shows why real-use duration changes product evidence.

Read [[Products Must Be Tested at Real Duration and Intensity]] for the evaluation framework, [[Figure 01 Product Profile]] for the historical robot record, and [[Google DeepMind SIMA Research Profile]] for the research design.

This article was developed with AI assistance from the preserved E008 transcript and the linked NVIDIA, Figure, Google DeepMind, and OSHA records. The transcript establishes my 2024 reaction. Company and research sources establish their own dated claims. Technical, research, product, safety, accessibility, editorial, and founder review remain required. Publication is unauthorized.