This is the shift from software that gives you tools to software that delivers results.

Data suppliers in the new era

What we deliver is not a batch of data, but the end-to-end capability that helps customers train models, robots, or AI systems to a better outcome.

A few layers under this:

  • From data delivery to outcome delivery. The old contract: "I give you the data, you use it." The future contract: "I understand your training goal, then I keep helping you use the data well — measured against accuracy, generalization, recall, task completion."
  • From one-shot supplier to deep collaborator. Once data is delivered, the customer will hit problems during training: distribution gaps, unclear labeling standards, weak performance on certain scenarios, the robot failing on certain motions or environments. These problems are exactly where the data company keeps creating value.
  • From selling resources to selling service capability. A data company has not only data but labelers, QA, domain experts, project managers, data engineering, and problem-localization. All of these should become part of follow-on services: filling in data, fixing labels, doing error analysis, adjusting collection plans, explaining how the data should be used.
  • From "accept the data" to "manage the data lifecycle." Delivery isn't the end — it's the start. After it: data version management, problem feedback, defect repair, incremental supplements, model-effect feedback, the next round of data optimization. This makes the customer feel you're not an outsourced supplier but part of their training system.

There must also be boundaries — it can't become unlimited free changes. The mature approach is to productize it:

  • Base delivery: dataset, labeling spec, QA report, usage guide.
  • Training support: integration guidance, Q&A, model-effect feedback analysis.
  • Optimization service: targeted re-collection, re-labeling, cleaning, and re-QA on low-accuracy scenarios.
  • Long-term partnership: a monthly or per-project data operations team that keeps lifting model performance.

The business model shifts accordingly: not pure pay-per-row, per-hour, or per-image, but "data project + follow-on service + continuous optimization." In suitable cases, pricing can even be tied to the customer's model metrics, robot task success rate, or launch outcomes.

Essentially, this is an upgrade:

Traditional data supplier: deliver data.
New AI data partner: deliver data capability that the model can absorb and that drives measurable improvement.

Meta in robotics

Meta has acquired Assured Robot Intelligence (ARI), a startup building foundation models for humanoid robots capable of real-world physical tasks. Co-founders Xiaolong Wang and Lerrel Pinto will join Meta's Superintelligence Labs. The move builds on Meta's existing robotics efforts led by Marc Whitten.

The push comes as Big Tech races to secure robotics talent, with market projections ranging from $38B by 2035 (Goldman Sachs) to $5T by 2050 (Morgan Stanley).

Why this matters: training on internet data got AI this far. Improving it further likely requires interaction with the physical world. Robotics is becoming a training strategy. Whoever owns systems that learn by acting, not just predicting, gets a compounding advantage that pure software models can't easily match.

Model × harness × context

Product boundaries are forming. The recurring theme across the day was that model quality is no longer the only meaningful moat. Anthony Maio argues lock-in comes from the context pipeline — how you fetch, sort, and compress repo state into a prompt — not from the harness itself. Mason Drxy's reporting backs this up: switching the prompt and middleware in the harness raised gpt-5.2-codex on Terminal-Bench 2.0 from 52.8% to 66.5%, and lifted gpt-5.3-codex on tau2-bench by 20%.

The practical takeaway: agent performance is increasingly a joint property of model × harness × memory/context strategy, not just the weights themselves.