Ana Bildea

Tech Lead @Orange - 🎓 Ph.D. 💡 #AI, #Agents, #GenerativeAI, #SafetyAI #AGI #Production #LLMOPS
✍️

[Lance Martin @RLanceMartin]

https://t.co/es0JQM4MS9

[Ana Bildea @AnaBildea]

Protecting AI Agents: How Claw Patrol Safeguards Your Systems from Accidental Damage and Attacks #ai #safety #agentic
https://t.co/SlWSShFV7i

[Ana Bildea @AnaBildea]

The AI puzzle: As AI talks more like us, are we starting to think more like it? Discover #LLMorphism – the mistaken belief that human thinking is like an LLM. #AI #Cognition #Psychology #HumanAI https://t.co/SJMF7MlY04

[Ana Bildea @AnaBildea]

The best enterprise agents in 2026 use a 3-layer stack:
1️⃣ Skills (Domain Knowledge)
2️⃣ MCP (Secure Connectivity)
3️⃣ CLI (Token-Efficient Execution)

Stop overloading your LLM. Read the full architecture breakdown here:

https://t.co/JfUjkU7TG4

[Furkan Gözükara @FurkanGozukara]

Amazing. No words to describe this tune. Emotional and to the point. New Iranian LEGO movie : We Share the Same Pain 💔 😭🥹

via Brick Beat Battalion

[Ana Bildea @AnaBildea]

First systematic map of "AI Agent Traps" adversarial web content engineered to hijack autonomous agents. 6 attack classes targeting every layer of an agent's operating cycle: How to Defend Against AI Agent Traps: #ai #agents #Security https://t.co/0Oy9GzpJ3E

[Ana Bildea @AnaBildea]

Agent Design Patterns for Long-Running AI Agents https://t.co/GQSyYFCX0e #ai #agents #llm #production #google

[Lance Martin @RLanceMartin]

https://t.co/EQueWgwseh

[Ana Bildea @AnaBildea]

How to Use Graphify: Turn Any Folder Into a Knowledge Graph https://t.co/WmBrt3qw1E #aiagents #graph #knowledgegraph #ai

[Ana Bildea @AnaBildea]

How much budget did you actually approve for your Agentic AI strategy? Not the roadmap. Not the slide deck. The actual signed check. That number tells you everything. #AIStrategy #EnterpriseAI #AgenticAI #AI #Leader

[Ana Bildea @AnaBildea]

https://t.co/Dmxx3Ewpyj

[Lance Martin @RLanceMartin]

i just published a short article on building applications w/ Claude. captures a few lessons from my own work and many discussions with others at Anthropic.
https://t.co/VYnSSGwDet

[Ana Bildea @AnaBildea]

LiteLLM Attack: How a Security Scanner Backdoored AI https://t.co/47RLkfQ0OT

[Ana Bildea @AnaBildea]

An AI Broke Into McKinsey’s Platform. Here’s What We Should All Learn From It.“ #ai #genai #security #aiagents #ops https://t.co/3ZFS6yoKOO

[Ana Bildea @AnaBildea]

PI planning in agentic times ..#ai #agile #funny https://t.co/smb7sRS4wS

[alphaXiv @askalphaxiv]

"Mixture-of-Depths Attention"

This paper teaches a Transformer to attend not just across tokens, but also to depth KV from its earlier layers.

That helps recover shallow-layer signals that standard residual stacking tends to dilute, improving performance with only a small extra compute cost.

Similar idea to Kimi’s Attention Residuals, but MoDA modifies the attention module itself, while AttnRes changes the residual/depth aggregation path.

[Google Cloud Tech @GoogleCloudTech]

https://t.co/Nhzbmwl85W

[Ana Bildea @AnaBildea]

https://t.co/Cffau40Q1y

[Ana Bildea @AnaBildea]

https://t.co/wtpszTrTIL

[Andrej Karpathy @karpathy]

Expectation: the age of the IDE is over
Reality: we’re going to need a bigger IDE
(imo).

It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. It’s still programming.