Karpathy: Agents Need Ten Years. Claude Accelerates Drug Research. Visa Secures AI-Agent Payments. Deepseek Compresses Document Context.

Show notes

The AI news for October 21st, 2025

--- This episode is sponsored by ---

Find our more about our today's sponsor Airia at airia.com.

---

Would you like to create your own AI-generated and 100% automated podcast on your chosen topic? --> Reach out to us, and we’ll make it happen.

Here are the details of the day's selected top stories:

OpenAI insider reveals: Why AI agents fail.
Source: https://www.all-ai.de/news/news24/openai-agenten-2035
Why did we choose this article?
Karpathy’s decade estimate and specific failure modes (memory, multimodality, planning) give a sober counterweight to agent hype — useful for product roadmaps, investment decisions, and setting realistic expectations when building agentized workflows.

Anthropic launches 'Claude for Life Sciences'
Source: https://www.all-ai.de/news/top-news24/anthropic-claude-sciences-life
Why did we choose this article?
A sector-focused Claude with direct connectors and a model (Sonnet 4.5) claiming expert-level benchmarks is strategically significant: it shortens R&D loops and raises immediate questions about data governance, validation, and vendor lock‑in for life‑sciences teams.
Trusted Agent Protocol: How Visa plans to secure payments between AI agents.
Source: https://www.heise.de/news/Trusted-Agent-Protocol-Wie-Visa-Zahlungen-zwischen-KI-Agenten-absichern-will-10781063.html?wt_mc=rss.red.ho.themen.k%C3%BCnstliche+intelligenz.beitrag.beitrag
Why did we choose this article?
Practical protocol-level work on agent identities and cryptographic signatures matters if you’re building agent-mediated commerce or marketplaces — it highlights immediate engineering, privacy, and liability considerations for automated transactions.
Deepseeks OCR model could significantly expand the AI memory
Source: https://the-decoder.de/deepseeks-ocr-modell-koennte-das-ki-gedaechtnis-deutlich-ausbauen/
Why did we choose this article?
A novel, practical approach to the long‑context problem: compressing documents as images could extend effective memory for models and enable better long‑document applications (legal, medical, knowledge bases). Worth tracking for builders of agent memory and retrieval systems.

Do you have any questions, comments, or suggestions for improvement? We welcome your feedback at podcast@pickert.de.