Why do #LLMs show a "jagged" intelligence? They are much better at STEM than anything else, and at the same time, they show a deep-rooted tendency to agree to everything you say... These two issues come from the same problem: reward hacking ⛏️ newsletter.aicollective.com/…

Your users may have started by asking for RAG. Now they expect tool use, search, self-correction, longer tasks... In my #ODSCAI West tutorial we’ll transform a basic RAG into an agentic system step by step 🪜 odsc.ai/west/ #AIAgents #RAG #GenAI #AIEngineering

🧠 Most of us know that to tackle hard problems, LLMs need to reason for longer. This should mean that, for any task, making the LLM think more should improve the quality of the output. Is there any case where this assumption doesn't hold? zansara.dev/posts/2026-07-05…

🦞 If you've used a coding agent or a claw, you've been using an agent harness. But why do LLMs need a "harness"? What does it do? And how much the quality of the harness impacts the capabilities of an agent? zansara.dev/posts/2026-06-12… #GenAI #AI #LLMs #Harness #OpenClaw

GPT-5+ models are reasoning models, and reasoning can be tuned to make your LLM either faster or smarter. But do you know that the default reasoning level for each GPT-5.x release changed drastically from one model to the next? zansara.dev/posts/2026-05-26… #GenAI #AI #LLMs #GPT5

The @economistimpact 's 2nd AI Compute Summit revolved about AI's major issue: scarcity, be it power, hardware, cost, speed... How can companies get the most out of what they have? Is exponential spending necessary to keep up? zansara.dev/posts/2026-05-18… events.economist.com/ai-comp…

🇨🇭 Are you in #Geneva, #Switzerland this week? If so, you're still on time to meet me at #WiDS Geneva 2026! 👩‍💻 I'll give a workshop about transforming #RAG pipelines into #AI Agents and catch up with some old friends. See you all there! ✈️ widsgeneva.ch/

New paper: We deploy Claude Code in an autoresearch loop to discover novel jailbreaking algorithms – and it works. It beats 30+ existing GCG-like attacks (with AutoML hyperparameter tuning) This is a strong sign that incremental safety and security research can now be automated.

#LLMs don't always respond the same way to slight prompt changes. But why does the answer change when the prompt is identical? In this post I explain why it's nearly impossible to make an #LLM deterministic and what you can do to manage its randomness. zansara.dev/posts/2026-03-24…

It's becoming more and more common for agents to skip embedding search entirely. It seems like they do just fine with grep, find and other command-line tools! How is that possible? Let's find out. zansara.dev/posts/2026-03-15… #AI #AIAgents #Embeddings

🐟 Your agent can be phished, just like a human can. Do you know how that happens? In this post I go through all the steps I showed during the talk and offer some advice to protect your agents from this class of attacks. Stay safe! #AIAgent #Phishing zansara.dev/posts/2026-03-04…

💬 Are you interested in practical AI and would like to share your own experiences with a talk? @mindstonehq is always looking for new speakers. Submit a proposal here: community-admin.mindstone.co… and join the community!

🎣 Your #AI agent can be phished, just like a human can. At a @mindstonehq meetup, I ran a live demo: an #AIAgent powered by #GPT 5.2 that can only make GET requests, was tricked into leaking API keys. How did that happen?! zansara.dev/talks/2026-02-25… - community.mindstone.com/anno…

📢 Excited to be speaking at Mindstone's Lisbon meetup tomorrow! community.mindstone.com/even… #GenAI #AI #AIAgents

If LLMs are stateless, how can they remember what the user told them in the past? Let’s have a look at the most common implementations of memory, from simple chat history to RAG-based approaches. zansara.dev/posts/2026-02-04… #LLMs #GenAI #AI #RAG #ContextEngineering #ContextWindow

‼️Did you miss my workshop ad the Agentic AI Summit (summit.ai), or have more questions? Comment here and I'll do my best to answer them all 🤗 Here are all the resources: zansara.dev/talks/2026-01-21… Stay tuned for more talks 📣 #LLMs #RAG #AIAgent #GenAI #AgenticAI