Why Sandboxing Isn’t Enough to Contain AI Agents
Should we trust AI agents to patch our vulnerabilities?
I'm a little late to the party on this one, but I still wanted to share and comment on Google's recent findings on vulnerability/exploitation trends. The article is a bit lengthy (~17 min read) but it's a good read. Check it out here.
My takeaways (most of which aren't particularly surprising):
-- . . .
Sounding Smart Doesn't Make It True
Another really good article in the ByteByteGo Newsletter. This time about the conditions which lead LLMs to "lie".
My takeaways:
More reasoning isn't a silver bullet. Techniques like chain-of-though prompting (asking the AI to show its work step-by-step) make answers easier to follow, but an AI can still . . .
Jeffrey Ladish interview
I'm not one for fear mongering but I do feel like the topic of AI regulation requires some serious thought. Perhaps my favorite interview thus far is Jeffrey Ladish's conversation. He raises some thought-provoking considerations. Check it out when you get a chance.
Tokens, Tools, and Total Restarts
What happens when you hit Enter.
I'm still very much a novice with AI but lately I've been curious about context engineering. ByteByteGo published another interesting article about initial processing workflows. Check it out when you get a chance: https://blog.bytebytego.com/p/what-happens-inside-an-ai-chatbot
My takeaways:
Input costs compound . . .
What Happens in the Chain of Thought
Why Frontier AI Models Keep Secrets
ByteByteGo published an interesting article on the commercial and safety risks of exposed AI reasoning traces. It's academic in nature (so consume it with a grain of salt) but ut still has some interesting info. Check out the full article here.
My takeaways:
- When AI models answer hard questions, they generate a . . .
Introducing getstat.dev
They say necessity is the mother of invention, right? :-)
About a week ago, I published a simple little page called https://getstat.dev. It's live now.
It's nothing fancy; just an easy client-centric way to verify connectivity to the public internet without the excessive bloat of visiting an actual web site. I . . .