David vonThenen

Spent years chasing bigger models… turns out smarter engineering wins. 🚀
I joined the @odsc.bsky.social AI X Podcast to talk about real-world production AI: RAG failures, quantization, SLMs, and building efficient systems that actually ship. 🎧

Listen here: bit.ly/4ssrdMm

#AI

. @socallinuxexpo.bsky.social starts tomorrow!
I'll be presenting:
• A Practical Guide to Training a Small Language Model: bit.ly/3LkKjo0
• The Sound of Your Secrets: Teaching Your Model to Spy, So You Can Learn to Defend: bit.ly/4rlTiEP

Sneak peek below 👇

Bigger models used to win headlines. Now they win power bills ⚡

On the @odsc.bsky.social blog, I break down how quantization and specialized SLMs are reshaping agentic AI around efficiency, not ego. It's about value per watt, not parameter count 🚀

Read here: bit.ly/4s6iKye

Excited to join @WDI_conference 2026 🎉 My VoD session, “Rethinking RAG: How MCP and Agent2Agent Will Transform the Future of Intelligent Search”, dives into governance, grounding & multi-agent design 🚀

Register: bit.ly/474WJrs
Code: WID26SP20 🎟️

#GenAI

1 week until @socallinuxexpo.bsky.social

Session: The Sound of Your Secrets: Teaching Your Model to Spy, So You Can Learn to Defend
bit.ly/4rlTiEP

Session: A Practical Guide to Training a Small Language Model: Tokenizers, Training, and Real-World Pitfalls
bit.ly/3LkKjo0

Excited to share that my talk was accepted for @devoxx.fr. I'll be presenting "The Sound of Your Secrets: Teaching Your Model to Spy So You Can Learn to Defend," all about acoustic side-channel attacks and defenses.

More details: www.devoxx.fr/

Chorouk Malmoum, thank you for sharing this! It's deeply validating to see this coming out of NVIDIA's own research. I've been saying this for almost a year now. The first time I put it on record… | David vonThenen Chorouk Malmoum, thank you for sharing this! It's deeply validating to see this coming out of NVIDIA's own research. I've been saying this for almost a year now. The first time I put it on record was at Devoxx UK 2025. At the time, it was based on hands-on experience building and validating multi-agent systems. But in this field, experience alone doesn't move the needle. You need data. You need papers. You need proof. And now we have it. The idea that Small Language Models are better suited for agentic workflows makes sense from a software engineering perspective. Separation of concerns. Encapsulation. Instead of one giant "do everything" model, you create focused SLMs that act as subject matter experts. Each one handles a narrow domain. There's another benefit people don't talk about enough: control. A tightly scoped SLM is far more likely to say "I don't know" when it's outside its boundary. This is a good thing. A massive general model? It tends to guess. And it guesses confidently. In production systems, that's not intelligence. That's liability. The second paper, introducing NVIDIA's Orchestrator-8B, is the piece I've been waiting for. A router that decides when to escalate a problem/question versus when to call a cheap tool or a smaller model. I'm very interested in experimenting with Orchestrator-8B. If the benchmarks hold up, this could materially change how we design cost-efficient agent systems at scale. This news also just so happens to coincide with a workshop/tutorial I will be giving at Open Data Science Conference (ODSC) East (end of April) titled: 𝐋𝐞𝐬𝐬 𝐂𝐨𝐦𝐩𝐮𝐭𝐞, 𝐌𝐨𝐫𝐞 𝐈𝐦𝐩𝐚𝐜𝐭: 𝐇𝐨𝐰 𝐌𝐨𝐝𝐞𝐥 𝐐𝐮𝐚𝐧𝐭𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐅𝐮𝐞𝐥𝐬 𝐭𝐡𝐞 𝐍𝐞𝐱𝐭 𝐖𝐚𝐯𝐞 𝐨𝐟 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐀𝐈 . I will do my best to include the findings/learnings from Orchestrator-8B in that session. (Tentative) Session Date: Tuesday, April 28 Session Info: https://bit.ly/4rxXeTg Read Chorouk's full breakdown below for more information and links to the research. .

NVIDIA's research backs it: Small Language Models > giant LLMs for agentic workflows 🔥 Focused SLMs = better control, lower cost, fewer "confident guesses."

Orchestrator-8B as a smart router? Game changer. I'll cover this at @odsc.bsky.social East! 🚀

More info: bit.ly/4aCsYPG

Really looking forward to SCaLE this year. Going to be a lot of fun (with learning some cool stuff)!

Automate or Die Trying | David vonThenen Recorded a great conversation last week with Wil Ramos (https://lnkd.in/gV9jQBUV) on the 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐞 𝐨𝐫 𝐃𝐢𝐞 𝐓𝐫𝐲𝐢𝐧𝐠 podcast. Wil, thanks for having me on. I appreciate the space to go deep on topics that don't always fit into a conference talk. The episode should be out in a week or two. I'll share the link here once it's live. Here's some of what we covered: 👉 How to harden RAG and agent workflows so they act only on verifiable evidence - Grounded data - Clear audit trails - Preventing agents from drifting into hallucinated "actions" or "decisions" 👉 What it takes to make agentic automation safe enough to run unattended - Guardrails - Checkpoints - Human-in-the-loop 👉 And the broader state of AI right now. Where it's moving. Where it's messy. And where we need to be more disciplined. I've been listening to several episodes of 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐞 𝐨𝐫 𝐃𝐢𝐞 𝐓𝐫𝐲𝐢𝐧𝐠 , and they're worth your time. What I like most is the range of perspectives. Different guests, different takes, real-world lessons. It's people building and securing real systems. If you're into automation, security, or AI systems, subscribe to the podcast here: YouTube: https://lnkd.in/gRKCEAJw Spotify: https://lnkd.in/ggDUeAyR More soon, once the episode drops.

Recorded an episode of "Automate or Die Trying" Podcast with Wil Ramos! 🚀 We went deep on hardening RAG + agent workflows... grounded data, audit trails, guardrails, human-in-the-loop.

Teaser post on LinkedIn: bit.ly/4qQiyCo

OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491 | David vonThenen I listened to the latest Lex Fridman Podcast episode: 𝐎𝐩𝐞𝐧𝐂𝐥𝐚𝐰: 𝐓𝐡𝐞 𝐕𝐢𝐫𝐚𝐥 𝐀𝐈 𝐀𝐠𝐞𝐧𝐭 𝐭𝐡𝐚𝐭 𝐁𝐫𝐨𝐤𝐞 𝐭𝐡𝐞 𝐈𝐧𝐭𝐞𝐫𝐧𝐞𝐭 𝐰𝐢𝐭𝐡 𝐏𝐞𝐭𝐞𝐫 𝐒𝐭𝐞𝐢𝐧𝐛𝐞𝐫𝐠𝐞𝐫. If you're building agents, it's worth a listen. OpenClaw has exploded on GitHub. But what stood out to me was this (time: 23:26): 𝐀𝐧𝐝 𝐭𝐡𝐞𝐧 𝐭𝐡𝐞 𝐚𝐠𝐞𝐧𝐭 𝐰𝐨𝐮𝐥𝐝 𝐣𝐮𝐬𝐭 𝐦𝐨𝐝𝐢𝐟𝐲 𝐢𝐭𝐬 𝐨𝐰𝐧 𝐬𝐨𝐟𝐭𝐰𝐚𝐫𝐞… 𝐈 𝐣𝐮𝐬𝐭 𝐛𝐮𝐢𝐥𝐭 𝐢𝐭… 𝐢𝐭 𝐣𝐮𝐬𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐞𝐝. We're not talking about scripted automation anymore. We're talking about systems that change themselves. That's impressive. It's also a different level of power. Midway through, Lex raises the obvious issue (time: 52:50): 𝐏𝐫𝐨𝐦𝐩𝐭 𝐢𝐧𝐣𝐞𝐜𝐭𝐢𝐨𝐧 𝐢𝐬 𝐬𝐭𝐢𝐥𝐥 𝐚𝐧 𝐨𝐩𝐞𝐧 𝐩𝐫𝐨𝐛𝐥𝐞𝐦… 𝐭𝐡𝐞𝐫𝐞'𝐬 𝐬𝐨 𝐦𝐚𝐧𝐲 𝐩𝐨𝐬𝐬𝐢𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬… 𝐧𝐮𝐚𝐧𝐜𝐞𝐝 𝐚𝐭𝐭𝐚𝐜𝐤 𝐯𝐞𝐜𝐭𝐨𝐫𝐬. Peter talks about progress, like scanning skills with VirusTotal. That's good. But the bigger point remains. OpenClaw is a privileged automation runtime. Your risk is dominated by: 1️⃣ Credential exposure 2️⃣ Network exposure 3️⃣ Tool/skill supply chain 4️⃣ Prompt injection and social engineering You're basically giving a script sudo on your machine. Except now it improvises. Please see: https://bit.ly/4aFyXDn And hopefully, you read the docs and you aren't running this on your actual machine, but some isolated cloud instance, VM, etc. To run OpenClaw safely, Peter's advice is clear (time: 1:00:45): 𝐈𝐟 𝐲𝐨𝐮 𝐦𝐚𝐤𝐞 𝐬𝐮𝐫𝐞 𝐭𝐡𝐚𝐭 𝐲𝐨𝐮 𝐚𝐫𝐞 𝐭𝐡𝐞 𝐨𝐧𝐥𝐲 𝐩𝐞𝐫𝐬𝐨𝐧 𝐰𝐡𝐨 𝐭𝐚𝐥𝐤𝐬 𝐭𝐨 𝐢𝐭… 𝐢𝐧 𝐚 𝐩𝐫𝐢𝐯𝐚𝐭𝐞 𝐧𝐞𝐭𝐰𝐨𝐫𝐤… 𝐭𝐡𝐞 𝐫𝐢𝐬𝐤 𝐩𝐫𝐨𝐟𝐢𝐥𝐞 𝐟𝐚𝐥𝐥𝐬 𝐚𝐰𝐚𝐲. Isolation is the safe move. And here's my 2 cents... If you isolate OpenClaw or any AI assistant completely... no personal data, no real integrations, no privileged API keys... how useful is it really? An AI assistant only becomes valuable when it knows who you are and can act on your behalf. That requires access. And access creates risk. Without that, you have a cool demo. Not a production system. Not a 𝐬𝐚𝐟𝐞 AI assistant. I think it could be useful to perform long running tasks based on the knowledge contained within the LLM... those in the AI space with some know how, probably already have some equivalent of that. BUT, I have a feeling that with some of that OpenAI resource, a safe and production version might become a reality sooner rather than later. The episode is fascinating and very honest about both the power and the risks... and also some really really amazing piece of tech that is taking the internet by storm. Give it a listen (it's 3+ hours, but worth it): https://bit.ly/4c0iFra

Just listened to Lex Fridman w/OpenClaw's creator 🤯🤖 Self-modifying AI agents are here… and they're powerful.

But let's be real: security is a real concern 🔐⚠️ Privileged automation + access + personal info = serious risk.

I break it down here: bit.ly/4tTsmhJ 🚀

Excited to share that my talk was accepted for NDC Copenhagen! I'll be presenting The Sound of Your Secrets: Teaching Your Model to Spy, So You Can Learn to Defend.


🗓️ Session Time/Date: Thurs, Jun 4 at 15:00
📍 Location: Room 4
📖 Session Link: bit.ly/4qvPWxT

Three weeks out and I can't wait to bring this one to @socallinuxexpo.bsky.social 23x in Pasadena, CA 🎧🔐

Check out my session: The Sound of Your Secrets: Teaching Your Model to Spy, So You Can Learn to Defend

🗓️ Sat, March 7 @ 14:30
📍 Room 101
🔗 bit.ly/4rlTiEP

#SCaLE23x

Three weeks to go and I'm fired up to be speaking at @socallinuxexpo 23x in Pasadena 🎉🐧

Session: A Practical Guide to Training a Small Language Model: Tokenizers, Training, and Real-World Pitfalls

🗓️ Fri, Mar 6 @ 15:45
📍 Ballroom DE
🔗 bit.ly/3LkKjo0

#SCaLE23x

Happy Friday! Just obtained my latest certificate.

Just a reminder... if you deploy anything, deploy to a VM, cloud instance, etc in complete isolation.

🚨 Exposed DBs, leaked prompts, 91% injection success. Moltbook wasn't hacked by geniuses, it failed basic engineering 😬

Handing control and data to unproven AI agents isn't innovation, it's risk.

Read the full breakdown 👉 bit.ly/4qXmqlI

🔥 #AISecurity

Home ODSC AI East 2026 - Boston Learn, grow, and connect with 3.5K+ data practitioners in the heart of the AI boom. Expert-led sessions on LLMs, ML, Generative AI and more.

Excited to share that my talk was accepted for @odsc.bsky.social East! I'll be presenting: "Less Compute, More Impact: How Model Quantization Fuels the Next Wave of Agentic AI."

Details here: odsc.ai/east/

The internet is biting back. Creators are deploying "tarpits" to trap AI scrapers in infinite mazes of gibberish data. This "data poisoning" spikes training costs and breaks the uncompensated data buffet. Be indigestible. Grow spikes.

Blog: bit.ly/4jvNTId

Teams are rethinking RAG 🔄🚀 BM25-based Hybrid RAG hits the sweet spot with clear observability, solid answer quality, and one OpenSearch stack to run 🤖✨ An alternative to Graph-based variations!

This is why it sticks... learn more 👉 bit.ly/4pz0D3b

We Had 400 People Shop For Groceries. What We Found Will Shock You. EXCLUSIVE: We uncovered a secret corporate scheme to raise grocery prices. We found that Instacart is using AI algorithms to charge customers different price...

AI isn't only optimizing search & ads, it's reshaping prices. 💸

This podcast exposes techflation: algorithms charging different people different prices to extract max value. Groceries cost more because AI knows what you'll tolerate. Dark precedent.

🎥 bit.ly/4jKdDkl

I'll be presenting at @devoxxgreece.bsky.social 🎉

Session 1: "The Sound of Your Secrets" will discuss acoustic side-channel attacks/defenses.

Session 2: "How Model Quantization Fuels the Next Wave of Agentic AI" will discuss efficient AI systems and models.

Info: devoxx.gr/

I guess you can say I was pretty active with speaking sessions at conferences in 2025 🤣

MIT study finds AI can already replace 11.7% of U.S. workforce Artificial intelligence can already replace 11.7% of the U.S. labor market, across finance, health care and professional services, according to MIT's study.

🎉 Reading some eye-opening stuff over the break...

MIT just dropped a study showing AI can already replace 11.7% of U.S. jobs. Not hype. Real data. Worth a read if you care about work, skills, and what's next.

👉 bit.ly/3Y76fWg

More powerful search: Rethinking RAG agents The future of intelligent search starts with rethinking RAG, retrieval augmented generation. By harnessing the power of MCP (Model context protocol) and Agen...

New video is live! 🚀🔍 I break down how rethinking RAG with MCP + Agent2Agent unlocks more powerful, secure, agentic search. This is the same approach I shared at 3 conferences in late 2025.

Watch it now on the @Instaclustr YouTube! 🎥✨ bit.ly/3NiViPe

Hybrid RAG breaks the black box 🔍🚀 Combining Graphs or BM25, and vectors turns retrieval into something explainable, governable, and production-ready for real enterprise AI 🤖✨ Want agents you can trust?

Read the full breakdown here 👉 bit.ly/4pz0D3b

#HybridRAG

Got two talks accepted at SCaLE 23x! 🐧

1. A Practical Guide to Training a Small Language Model
2. The Sound of Your Secrets: Teaching Your Model to Spy, So You Can Learn to Defend

One of the best conferences of the year! See you there!

Register: bit.ly/4aMlprJ 🚀

2026 AI Outlook: How Agents, Context, and Governance Will Shape Real-World AI As artificial intelligence continues its rapid evolution, one thing has become increasingly clear: progress isn’t being driven by a single breakthrough, but by a convergence of architectural shifts, cultural changes, and hard-earned lessons from production deployments. ODSC AI is built on our community – and that involves our speakers, attendees,...

What did 2025 really change about AI, and what must 2026 fix to make it work in the real world? 🤖✨ Top builders and thinkers share bold predictions (including mine) on agents, context, and governance.

Read it here on the @odsc.bsky.social blog 👉 bit.ly/3YdViSW

#AI #Governance

Got my ChatGPT year in review... I know ChatGPT is the ultimate hype man, but it's good to see it thinks I am chasing difficult problems and trying to bring that information/material to others. See you in 2026!

IBM CEO says there is 'no way' spending trillions on AI data centers will pay off at today's infrastructure costs IBM CEO Arvind Krishna walked through some napkin math on Big Tech's AI data center spending — and raised some doubts on if it'll prove profitable.

Some reading for the break. I was thinking about this a lot...

IBM's CEO walks through the data center math and basically says the returns don't add up 💸

The capex numbers are wild, the AGI bets feel shaky. This math is math-ing for me.

Link: bit.ly/4qkwCE8 🚀

Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md – Agentic AI Foundation (AAIF)

Been catching up on AI stuff, and this is a definite must-read 🤖

The @linuxfoundation.org just launched the Agentic AI Foundation. Open governance, real standards, and projects like MCP, goose, etc, under one roof. 👀

Link: bit.ly/4s4cGHv

AI Expert: We Have 2 Years Before Everything Changes! We Need To Start Protesting! - Tristan Harris Ex-Google Insider and AI Expert TRISTAN HARRIS reveals how ChatGPT, China, and Elon Musk are racing to build uncontrollable AI, and warns it will blackmail h...

I’ve been binge listening to Diary of a CEO and these two AI interviews are wild. Super fascinating from a social angle. Perfect holiday break listens 🎧🎄

Check them out:
1. bit.ly/3Y7405i
2. bit.ly/3NcAetx