hebel. today · 15 November 202623 / 79
Technology · AI & Robotics
Generalist models replace specialized robotics teams
General models take control. OpenAI has disbanded its dedicated robotics research team to integrate researchers into its main organization, signaling that frontier models like GPT-6 Astra are now capable of controlling robots directly without…
From 5 newsletters · 1 min read · one thing to try
General models take control. OpenAI has disbanded its dedicated robotics research team to integrate researchers into its main organization, signaling that frontier models like GPT-6 Astra are now capable of controlling robots directly without specialized architectures (Dr Alan D. Thompson from The Memo by LifeArchitect.ai, Andra Keay from Robots & Startups). In NVIDIA-led benchmarks, these general-purpose agents achieved a 74.8% success rate on simulated manipulation tasks, significantly outperforming the 51.0% rate of specialized robotics models.
Dexterity gains in production. Google’s Gemini Robotics 2.5 has expanded the breadth of skills robots can acquire, enabling humanoids like Apptronik’s Apollo 2 to reach a 96% success rate in picking and packing automotive parts on Mercedes-Benz lines. This model leverages whole-body intelligence to handle delicate and flexible objects, moving beyond lab-based testing into unstructured factory environments.
Agentic coding and planning. Anthropic’s Fable 5.2 model reached 89.3% on the OSWorld 2.1 benchmark and demonstrated advanced agentic capabilities by formalizing Fermat’s Last Theorem in 30 hours using a team of 1,000 agents (AI Newsletter). Meanwhile, Claude Opus 5.5 showcased autonomous refinement by titrating its own paint shades during a 4.5-hour robotic painting session, adjusting its palette without human input.
- Toyota Research Institute uses a bespoke solver to manage multi-agent traffic in warehouse facilities, avoiding the brute-force computations that typically slow large robot fleets (IEEE Spectrum).
- Intel’s survey of 800 industry leaders found that 77% prioritize reliability over the physical form of a robot, with only 6% voting for humanoid designs (TLDR Tech).
Try today · 30 min
Perform a 'titration' test on a manual or digital process you control to see if an AI can self-correct when given a feedback loop.
Pick a task where the AI produces an output, then define a clear 'palette' or 'missing shade' constraint that requires the AI to adjust its next iteration.
You need
- A task involving repetitive generation (e.g., writing a report, generating code, or creating a design)
- A defined quality metric or 'palette' of constraints
- Access to a frontier model like Claude 3.5 or Gemini 4
Steps
- Define the primary task and the specific quality constraint that often leads to failure.
- Ask the AI to perform the task once.
- Identify a specific deficiency in the result (e.g., a missing tone, a logic gap, or an incorrect format).
- Instruct the AI to 'titrate' or adjust its parameters specifically to address that gap.
- Review the second iteration to see if the model successfully self-corrected without human intervention.