Figure CEO Brett Adcock believes his enterprise may have just engineered the initial spark of general intelligence for machines. During an internal all-hands session broadcast publicly on X on October 22, Adcock told his team, It was the most significant moment I’ve ever had in robotics, adding that it resembles baby AGI.

While the video recording surfaced weeks ago, broader media coverage largely overlooked the critical revelations and metrics shared across the hour-long broadcast. As a leading American contender in the strategically vital sector of humanoid robotics, Figure's internal updates offer a rare glimpse into the future of physical AI.

The Comprehensive Figure Update Roster

Adcock delivered a sweeping set of announcements detailing the company's philosophy on physical AI, specifications for upcoming hardware, staggering data ingestion rates, and an unannounced capital raise reaching well into nine figures.

  • Figure 04, the enterprise's next-generation humanoid, has reached full design completion, which Adcock dubs the iPhone 1 version of humanoids.
  • The newly developed robotic hand incorporates more actuators than an entire traditional robot alongside millimeter-tier fingertip tracking.
  • Production has officially hit the 1,000th unit milestone for the Figure 03.
  • Manufacturing costs for the Figure 03 dropped by over 90% compared to its predecessor, the Figure 02.
  • The baby AGI model, trained on massive GPU counts and Figure's largest dataset to date, demonstrated capabilities impressive enough to prompt Adcock's comparison, with an even more advanced iteration slated for release shortly.
  • Helix 2.5 introduces foundational scaling laws for humanoid machine learning, establishing a direct correlation between data input and physical AI advancement.
  • Adcock emphasized a strict philosophy that pre-training triumphs over post-training, prioritizing robust foundational models over narrow task-specific adjustments.
  • Index, Figure's proprietary gig-economy platform for capturing egocentric human work data, expanded to 37.5 million videos, 1.26 million app downloads, and reach across 112 countries, supported by 210,000 weekly uploaders generating 81 minutes of training data per second.
  • Weekly data expenditure reached $8.1 million, outpacing the total venture capital raised by many rival robotics firms over the next 12 months.
  • Internal compute power is projected to double in roughly 1.1 weeks.
  • Financial reserves exceed $2 billion, pointing to undisclosed fundraising rounds that push estimated totals well beyond Crunchbase metrics.
  • Deployments currently span three states, featuring a zero-shot operational rollout at a novel customer facility.
  • Capabilities expanded to include autonomous ladder climbing, with a major controls update scheduled for the following month.
  • The corporate workforce has expanded twofold.

Unpacking the Baby AGI Phenomenon

Figure currently ingests more internally generated data than it could possibly acquire from all 51 external physical data providers combined. This unprecedented data velocity directly fuels the organization's robot foundation models, prompting Adcock's observation regarding early artificial general intelligence.

Artificial general intelligence denotes autonomous systems possessing cognitive faculties matching or exceeding human benchmarks. While an extraordinary assertion, Figure's massive data-gathering apparatus provides a quantitative foundation for the claim. By collecting more human training data weekly than the remainder of the robotics market has accumulated historically, the enterprise bets heavily on the supremacy of pre-training over post-training strategies.

Alternative industry players, such as Sima, pursue divergent paths by relying on frontier foundation models like OpenAI's Astra or Anthropic's Mythos to handle physical AI challenges out of the box through zero-shot deployment, utilizing post-training strictly for specialized duties.

Engineering the Figure 04 Humanoid

Describing the fourth-generation humanoid, Adcock noted that the firm has achieved full design lock and commenced parts shipping. Characterized as the largest generational leap in the company's history, Figure 04 moves in a distinct direction compared to competing humanoids.

The reimagined hand mechanism represents the centerpiece of the new hardware. Boasting greater complexity and actuator density than an entire standard robot, the system tracks human contact points with millimeter-level precision and is slated for commercial production.

This design philosophy contrasts sharply with competitors like 1X and Tesla, both of which utilize tendon-driven hands to house motors within the forearm. While tendons reduce hand weight, they introduce mechanical complexity and maintenance hurdles. Direct actuation in the fingers, conversely, yields predictable movement with minimized transmission complications, albeit within a tightly constrained human-scale package.

Scaling the Index Training Data Pipeline

Launched in August, the Index platform incentivizes global participants to record their daily labor and household chores. Facing a stagnant external data market, Figure constructed its own proprietary data pipeline.

The enterprise now ingests 81 minutes of human video per second, translating to 4,860 hourly hours or roughly 816,000 weekly hours. This represents a 2.7-fold ingestion increase in just two months, alongside surging app downloads and millions of newly uploaded clips.

At current operational velocities, Figure secures one million hours of footage every nine days. Rivals require months or years to amass comparable egocentric video repositories, substantiating Adcock's claim that Figure's internal generation outpaces the open market.

Critical caveats remain, however, as raw phone footage lacks the force, torque, and tactile telemetry inherent to dedicated robotic datasets. Figure's infrastructure pipeline must clean and annotate these vast inputs, leaving the exact volume of usable teleoperation data as proprietary intelligence.

Capital Allocation and Scaling Laws

Aggressive data acquisition requires substantial capital. Adcock reported an estimated $8.1 million weekly expenditure on data collection, projecting over $421 million annually. This investment dwarfs the total funding rounds of numerous competing robotics startups.

This financial commitment is justified by empirical scaling laws. Paralleling breakthroughs in large language models, Figure published scaling metrics alongside Helix 2.5, demonstrating a sixfold performance improvement in unfamiliar residential environments driven by expanded pre-training datasets.