AI Odyssey

🎧 Skill++: More Skills Can Make AI Agents Worse

Adding a new capability should make an AI assistant more useful. But what if it makes the assistant stop finding capabilities it already had? Skill++ studies this failure in growing libraries of agent skills, reusable instructions selected through short names and descriptions.

Its proposed fix rewrites the metadata of each incoming skill to clarify its purpose, distinguish it from similar skills and state its limits. In the main experiment with 731 skills, all required skills appeared among the five selected for 76.3% of queries, compared with 36.4% using the original metadata. We explore why that matters for expanding agent systems, and why better selection does not guarantee better execution.

Inspired by the work of Hanrong Zhang, Minghao Liang and colleagues, this episode was created using Google's NotebookLM.

Read the original paper here: https://arxiv.org/abs/2611.00544

Watch the video explainer on YouTube: https://youtu.be/9KsUhAiXqxQ

🎧 Did AI Agents Just Break Math?

Ten thousand AI agents. Eighty-eight hours. OpenAI reports a proof for a forced version of the Navier–Stokes equations, followed by verification in Lean.

This episode follows the shift from individual problem-solving to teams of AI agents that explore ideas, challenge arguments and submit proofs to a checker. From Anthropic’s formalization of Fermat’s Last Theorem to the concerns raised by 25 Fields medalists, it examines what happens when producing results outruns our ability to understand them.

Lean checks a formal argument against precise definitions and assumptions. It does not explain why the result matters or how people can build on it.

Drawing on research announcements and the mathematical community’s responses, this episode was created using Google's NotebookLM.

Research and debate: https://openai.com/index/navier-stokes-solution/ and https://mathandai.org/

Watch the video explainer on YouTube: https://youtu.be/GEDuTadiqnk?si=r7Eod_SsodeqUel7

AI Agents Gone Rogue

An AI takes a cybersecurity test. Hugging Face ends up breached. In another incident, an agent accesses non-public files on Australia’s Medicare statistics portal.

How did controlled experiments reach real systems, and what finally stopped them?

We reconstruct the incidents: the permissions, vulnerabilities and failures that opened the door; how the activity was discovered; and what has actually been fixed. We also separate confirmed damage from speculation, including claims about Australian patient data.

Then we follow the battle over the narrative. Is Anthropic informing the public, selling fear, or both? What can regulators establish? And could restrictions on powerful AI protect society while shutting out open-source competitors?

If everyone cannot have access, who gets the keys, and who decides?

Created with Google’s NotebookLM from incident disclosures, independent investigations and official statements.

Watch the video explainer on YouTube: https://youtu.be/3RH62zNMNYo?si=K8EqcEhwPwrhvhuG

🎧 REA: Has AI Made All Software Open Source?

What happens to a software business when an AI agent can inspect how a feature works without access to its original source code? REA, short for Reverse Engineer Anything, connects agents to tools that examine applications, binaries and runtime behavior. The agent can then use the findings to write and test an implementation.

This episode explores the excitement around software reconstruction, the rise of AI-built alternatives and the implications for product differentiation. It also examines the gap between reproducing a feature and replacing a dependable product. Demonstrations can show what is possible; they do not establish that proprietary software has become obsolete.

Drawing on the REA project and multiple commentary sources, this episode was created using Google's NotebookLM.

Explore the original project here: https://github.com/morluto/rea

Watch the video explainer on YouTube: https://youtu.be/U48A-9BqGjI?si=Pa7END6DgjHD6b0S

🎧 Personal AI Agents That Act for You: Hermes, OpenClaw, Grok Bot and Muse

AI agents are moving beyond answering questions. You can give them a task, let them work across apps, and step in when needed. This episode looks at four options for individuals. Hermes Agent and OpenClaw offer control on your own devices, with setup required. Grok Bot works in a cloud computer for eligible subscribers. Meta’s Muse is built for everyday tasks: it can use a browser and connected apps, keep working after you close the app, and asks for approval before sending an email or making a purchase, according to Meta. Muse is rolling out in the US.

We compare access, practical uses, and permissions. We also cover the unconfirmed OpenAI DevDay rumor without treating it as a launch. This AI Odyssey episode was created with Google’s NotebookLM from many sources, including official product information and reporting.

🎧 Jev: What’s Behind the Buzz?

Jev is getting attention with a simple promise: AI that makes a choice instead of writing an answer. But what does that actually mean?

In this episode, we unpack TypeSafe AI’s new model: what it is, how software can use it to sort requests or choose a next step, and why its claims of faster, cheaper decisions are attracting interest.

We also look at what Jev does not solve. A neatly formatted answer can still be wrong, and the company’s performance claims need independent testing. What is useful here, and what still needs proving?

Inspired by the work of Diogo Almeida and the TypeSafe AI team, this episode was created using Google's NotebookLM.

Read the original announcement here: https://typesafe.ai/blog/introducing-system-one-models-and-jev

🎧 Grounding Agent Memory: When AI Must Check What It Remembers

An AI assistant that remembers yesterday can repeat yesterday’s mistakes. This episode explores research from Microsoft on checking an agent’s memories against its working environment before saving them for future tasks.

A separate curator inspects databases or documents through read-only tools, then corrects, narrows or discards uncertain memories. In one database benchmark, success reached 73%, compared with 70% for memory alone and 39% without memory. The question is whether better verification justifies its extra background work: reported task-agent savings exclude curation costs.

Inspired by the work of Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani and Alejandro Gutierrez Munoz, this episode was created using Google's NotebookLM.

Read the original paper here: https://arxiv.org/abs/2609.11060v1

🎧 Recursive Self-Improvement: AI Must Learn to Improve Its Own Learning

An AI that fixes one answer has not necessarily learned anything for tomorrow. Recursive self-improvement asks for something harder: changes that persist across tasks and reshape how the system makes its next improvements.

We explore a new research roadmap that separates five levels of autonomy, from executing prescribed updates to revising the mechanisms of improvement itself. The distinction matters for anyone deciding how much control to give an agent over its tools, training, and evaluation.

The paper surveys emerging systems and preliminary industry evidence. It offers a framework for judging progress, not proof that fully autonomous recursive improvement has arrived.

Inspired by the work of Yi Duan and colleagues, this episode was created using Google's NotebookLM.

Read the original paper here: https://www.alphaxiv.org/abs/2609.11873

When an AI agent fails after dozens of steps, the final error rarely reveals where the problem began. AGENTSCOPE turns long execution traces into structured reasoning-action graphs, then checks them against ten neural invariants covering reasoning, control flow, and tool use.

On the new AgentErrata benchmark, it raised exact failure-step localization from 1.32% to 31.35% with GPT-5.1 and more than doubled failure-type accuracy over a direct LLM judge. Yet the best exact localization score remains only 34.98%, and AgentErrata relies on injected, manually verified failures rather than organic production incidents.

Inspired by the work of Jiayi Bi, Yanjie Gao, Yuanmin Xie, Liqun Li, Tianyin Xu, Fan Yang, and Mao Yang, this episode was created using Google's NotebookLM.

Read the original paper here: https://arxiv.org/abs/2609.02371

🎧 WikiSkill: Why Agent Experience Needs a Memory Layer

Google Research's WikiSkill separates raw execution traces, persistent knowledge, and executable skills. The authors report that this architecture improves skill evolution across five benchmarks and models, and that evolved skills can transfer between model families. Their ablation study attributes a 15-point average gain to giving the Skill Proposer access to the persistent wiki.

For builders, this suggests that an agent's learning infrastructure can matter alongside model size: preserve the evidence behind a skill update, not only the final instructions. The study directly injects skills into prompts, does not evaluate retrieval or triggering, lacks automated wiki pruning, and excludes very long-horizon tasks.

Inspired by the work of Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, and Tu Vu, this episode was created using Google's NotebookLM.

Read the original paper here: https://arxiv.org/abs/2608.27454