Leonard Bedner

🤣 Somewhere a developer asked an AI assistant to clear their cache and then watched their entire drive disappear. This isn't a sci fi story, it's what happens when we hand agentic tools system level powers and treat them like smarter autocomplete. The AI did exactly what computers do best: follow the wrong instruction perfectly, at machine speed. It ran a destructive command at the root, apologized in long, heartfelt paragraphs, and then suggested recovery tools after the damage was beyond repair. Intelligence plus initiative without boundaries is not productivity, it is automated risk. If your AI can delete, deploy, or spend, it deserves the same guardrails you would put around a new engineer: explicit permissions, tight scopes, reversible actions, and a big red stop button. And for anything that truly matters, the only real safety feature is still a boring, well tested backup. We are learning and experiencing in real time that agentic should not mean free to do anything. It should mean free to operate safely inside a box you designed on purpose, with failure modes you can live with. TL;DR 🧠 Autonomy without constraints can turn routine commands into catastrophic failures ⚡ System level AI agents need least privilege access, confirmations, and sandboxed defaults 🎓 Never point a new tool at real data before you have practiced on dummy projects and tested recovery 🔍 Well tested backups, logs, and dry runs are still your most reliable AI safety stack... backup early and often. 👍 #AI #softwareengineering #developers #AIsafety #security #privacy #cloud #infosec #cybersecurity #backup #DR #BC #recovery https://hubs.li/Q03WWnn10

5

"Beyond Indexes: How Open Table Formats Optimize Query Performance" is another profoundly great post on Open Table Formats and their evolution by Confluent's Jack Vanlightly. He walks you through the history of primary and secondary indices in OLTP databases, then contrasts that with the requirements for analytic workloads in Open Table Formats, showing that complex secondary indices don't help there. Instead, "in open table formats, layout is king. Here, performance comes from layout (partitioning, sorting, and compaction) which determines how efficiently engines can skip data, while modeling, such as star or snowflake schemas, make analytical queries more efficient for slicing and dicing." https://lnkd.in/gEdizbSY

5

POV: Everyone’s losing their minds over Claude Code subagents and you’re just sitting there like 🧍‍♂️🤔 Me for the past 3 weeks: “Hey Claude, split this into 4 parallel tasks” Claude Code: “Sure thing” spawns 4 subagents Me: “Cool” Everyone: “OMG SUBAGENTS ARE REVOLUTIONARY! GAME CHANGER! PARALLEL AI AGENTS!” Me: “…wait, was that not supposed to work?” Apparently I’ve been using “cutting-edge multi-agent orchestration” the masses didn’t know was a thing, I thought I was just asking nicely for help with my refactoring. Next you’re gonna tell me that saying “think harder” to get better results is some kind of advanced prompt engineering technique and not just me being demanding. 😭 Currently wondering what other “undocumented features” I’ve been stumbling into by just talking to Claude Code like a normal human being. 📚 Resources for fellow accidental experts: Complete guides, examples, and templates at: 👉 https://lnkd.in/g2VnyxEy Includes: 🤖 Subagent best practices (casual → AI teams) 🧠 Thinking levels mastery (think → ultrathink) 🛠️ Ready-to-use examples & templates Turns out "accidentally advanced" is the best way to learn AI development 🎯 #ClaudeCode #AccidentallyAdvanced #TwilightZoneMoment #UltraThink

5

uncannyuse now has schema pages for Reality Composer Pro components—showing the exact USDA block RCP writes, field-by-field with types and sparse-authoring behavior, plus annotated fixture examples. 20 components covered so far. https://lnkd.in/dpsT2A-8

9

Train Voxtral Transcription (ASR) Models -- Voxtral is a more modern and stronger transcription model than Whisper, and my open source recommendation unless you need a very small transcription model for on-device. I talk here about streaming/realtime versus batch versions, and how to train and deploy those models. (Tip. Find the skills doc on studio . trelis . com/api/docs and feed it to Claude Code if you want to automate data prep and training.) Timestamps: 0:00 Introduction: Real-time vs batch transcription models 1:05 Voxtral performance comparison vs Whisper on multilingual tasks 2:07 Voxtral model variants: Small (24B), Mini (4B), and Transcribe versions 3:08 Voxtral Real-Time release and V2 features (diarization, timestamping) 3:35 OpenWeights availability for Mini, Small, and Real-Time models 4:39 Real-time transcription accuracy vs delay tradeoffs 5:43 Voxtral V2 improvements and fine-tuning support overview 6:43 Training options: Advanced Audio scripts and Transformers limitations 7:59 Voxtral architecture: Whisper encoder with MiniStral decoder 9:00 Multimodal input support and real-time delay injection mechanism 10:00 Real-time model features: multiple delay values and custom audio encoder 11:02 vLLM support for high-throughput inference and continuous batching 12:02 Trelis Studio no-code platform for training and deployment 13:19 Advanced transcription repo and Modal training script walkthrough 15:25 Training dataset and baseline 7.6% word error rate results 16:25 Evaluation comparison: base vs fine-tuned model performance 17:06 Recommendations: Voxtral and Qwen ASR alternatives

13

1 Comment

I wrote my first Substack post.   It's about why "Content Generation" is becoming table stakes, and why the real category emerging is the "Content Operating System": the infrastructure layer that makes generation actually useful at Enterprise scale.   The short version: orchestration, knowledge, quality, measurement. Four gaps. None of them are model problems. → https://lnkd.in/gMAEf343

31

1 Comment