Skip to content
Back to posts

🧭 NoodlDigest Weekly Model Watch — Nov 4–12 2025

Notes by SDG ¡

Following the evolution of large-language-models and agentic intelligence.

🆕 A Week Where Models Learned to Think and Act

The frontier of AI shifted again this week — not through bigger models, but through smarter, more capable agents.

We saw new ways for LLMs to execute code, reason across languages, and verify their own behavior.

It’s the quiet revolution beneath the headline noise: models turning from chat companions into autonomous collaborators.

🧠 Highlights at a Glance

  • Anthropic’s Model Context Protocol (MCP) lets agents write and execute code directly — a leap toward self-directed reasoning.
  • OpenAI’s IndQA benchmark tests cultural and linguistic depth across Indian languages, while new ChatGPT apps (Peloton, Tripadvisor) expand the agent ecosystem.
  • DeepMind’s Perch 2.0 extends model intelligence beyond text, interpreting the world through sound.
  • Microsoft’s Magentic Marketplace and NVIDIA’s Agentic Launchpad accelerate the rise of multi-agent systems.
  • Meanwhile, MIT, Berkeley & Oxford’s Guaranteed Safe AI framework introduces mathematical safety proofs — a step toward self-verifying intelligence.

🧩 Frontier Lab Rundown

🟢 Anthropic — Agents That Can Assemble Their Own Workflows

Anthropic introduced a major upgrade to the Model Context Protocol (MCP) — enabling agents to use a code-driven execution loop rather than issuing step-by-step tool commands.

With MCP, an agent can:

  • generate Python/JS code inside the LLM,
  • send it to an MCP execution server,
  • receive structured outputs,
  • and iterate based on results.

This approach removes the overhead of repeatedly defining tools and parameters and allows stateful computation without growing the context window.

The early numbers are striking:

  • up to 10× faster completion of multi-step tasks
  • ≈ 98% fewer tokens consumed

It’s the beginning of agents that not only reason about a problem but assemble the workflow that solves it.

🔗 https://www.anthropic.com/engineering/code-execution-with-mcp


🔵 OpenAI — Making Models Culturally Fluent and Agent-Ready

IndQA, OpenAI’s new multilingual benchmark, pushes models beyond translation into cultural reasoning — testing how well they grasp idioms, values, and local logic across Indian languages.

This marks an important evolution: intelligence grounded in context, not just syntax.

Simultaneously, OpenAI rolled out ChatGPT Apps with Peloton and Tripadvisor, showing what agent ecosystems might look like at scale — assistants that can train your body, plan your trip, and refine your preferences.

And a subtle but powerful update: users can now interrupt and steer long-running queries in GPT-5 Pro mid-stream, turning what was once a static session into a collaborative process.

Together, these moves blur the line between model and operator.

🟣 DeepMind — Expanding the Senses

Perch 2.0 listens where text cannot.

Trained on the calls of over 15 000 species, it can recognize whales, birds, and forest soundscapes with zero-shot learning — effectively giving AI an auditory understanding of the biosphere.

While not an LLM, it hints at a future where language models pair with sensory agents to perceive, reason, and act in the real world.

🟠 Microsoft & NVIDIA — Simulating the Agent Economy

Microsoft’s Magentic Marketplace introduces a simulated economy for multi-agent collaboration and competition — a place to test emergent behavior before deployment.

Alongside it, NVIDIA’s Agentic Launchpad supports European startups building agentic AI.

These projects transform “multi-agent” from a research buzzword into a commercial discipline.

📊 Benchmarks & Emerging Capabilities

OpenAI’s IndQA redefines evaluation, measuring not just knowledge but judgment across diverse cultures — a foundation for globally-relevant reasoning systems.

And in the academic sphere, the Guaranteed Safe AI paper

 by Bengio, Russell & colleagues shifts the safety conversation from “testing” to proofs: world-modeling, safety specifications, and formal verifiers that ensure behavior remains bounded.

Both efforts reveal the same trajectory — models learning to reason within limits rather than chasing unbounded scale.

🌐 Ecosystem Trends

  • LLMs meeting the real world: ChatGPT’s app integrations mark the start of domain-anchored agents — assistants that can act through trusted brands.
  • Developers experimenting with autonomy: GitHub repos for “self-coding” agents based on MCP surged this week.
  • Multimodal expansion: DeepMind’s Perch 2.0 connects environmental sound data with symbolic reasoning, hinting at hybrid sensory-language architectures ahead.

💡 Weekend Build Idea

“Autonomous Research Partner” — an MCP-powered agent that runs code, cites papers, and verifies results through a formal checker.
Not a chatbot — a co-author.

🛡️ Safety & Governance

Three frameworks dominated the safety conversation:

  • Guaranteed Safe AI
  •  — mathematical verifiers for frontier models.
  • Bias-augmented Consistency Training
  •  — DeepMind’s technique that slashes jailbreaks and sycophancy.
  • SafeSearch RL
  •  — reinforcement learning that cuts harmful outputs by 50–90 %.

These approaches mark the maturing of alignment: from reactive moderation to engineered integrity.

🔭 Takeaways

For Builders

  • Adopt code-execution agents — MCP is the new baseline for serious reasoning systems.
  • Design for steerability — interruption and refinement are the human-AI handshake.
  • Test globally — use IndQA or equivalents to evaluate cultural reasoning before deployment.

For Researchers

  • Explore proof-based safety — verifiers could become part of the training loop.
  • Blend language + perception — multimodal agents like Perch 2.0 suggest the next frontier of grounding.

For Investors

  • The agentic economy is real: simulation platforms (Magentic) and orchestration infra (MCP) are the new picks-and-shovels.
  • “Safer by design” models will command enterprise premiums — governance is now a differentiator, not a disclaimer.

💡 Noodl Insight

The next wave of LLMs won’t just reason in text — they’ll reason in code, actions, and proofs.

📅 Next Edition: Thursday, Nov 20 2025