The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) · Sam Charrington

AI Trends 2026: OpenClaw Agents, Reasoning LLMs, and More with Sebastian Raschka

February 26, 2026·1 hr 19 min·5 clips

LLMs now work better when they use tools, like a calculator for large multiplication.

1. The TWIML AI Podcast episode "AI Trends 2026: OpenClaw Agents, Reasoning LLMs, and More with Sebastian Raschka" focuses on how LLMs changed from 2025 to 2026. 2. Host Sam Charrington interviews Sebastian Raschka, an independent LLM researcher who studies model behavior, reasoning, and practical workflows. 3. The episode asks what changed most in the past year: reasoning, post-training, tool use, and agentic interfaces. 4. Raschka says post-training now gets more research attention because it still has low-hanging fruit. 5. He contrasts that with pre-training, which he calls "pretty sophisticated," even though better data mixes and multi-token prediction can still help. 6. He says today's LLMs are increasingly expected to use tools instead of answering everything from memory. 7. His calculator analogy makes the point directly: for a hard multiplication problem, he would use a calculator rather than mental arithmetic. 8. Raschka says tool use can reduce hallucination rates and improve answer accuracy, even though it does not eliminate errors. 9. He argues that the surrounding interface now matters as much as the base model, especially when the model can see files, PDFs, and local folders. 10. He mentions uploading a PDF to ChatGPT to pull out chapter headers and double-check a 40-page table of contents. 11. He says the Codex app and a Visual Studio Code plugin work like a second pair of eyes for bugs and performance suggestions. 12. Raschka describes the new model releases of the second week of February 2026, including Opus 4.6 and OpenAI 5.3, as part of an incremental pattern rather than a single breakthrough. 13. He says the lower reasoning-effort modes have improved enough that many tasks no longer need the slowest setting. 14. He still uses the pro-mode review for rare cases, such as checking a long chapter for inconsistencies and incorrect numbering. 15. He says OpenClaw, formerly Moldbot, is interesting because it puts a local agent on a user's own computer. 16. He compares its appeal to the excitement around AlphaGo and says it can show non-technical people what LLMs can do. 17. He also says he is still not comfortable letting it handle finances or calendars, which shows the trust gap around agentic delegation. 18. Most of his own LLM use still goes toward building deterministic Mac OS apps and workflow tools rather than letting the model run every task. 19. He gives examples that include adding chapter marks to podcast audio and extracting titles, dates, author names, and links from archive bookmarks. 20. The episode closes its technical arc on verifiable rewards, where math and code can be checked symbolically or by compilation, while process reward models and reward hacking remain open problems.

As heard by us

Raschka shows why post-training and reasoning now draw more attention than pre-training.

The episode treats current LLM work as a move toward post-training and reasoning. Pre-training still matters because it remains costly and data-hungry, but the sharper questions now sit downstream, where there is still room for practical gains.

Read the full review in PlayNext →

Why you'd press play

You want a map of why LLM attention is shifting from pre-training toward post-training and reasoning.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)