AI safety
26 posts
Microsoft's AI CEO called the web freeware
Microsoft and OpenAI executives described how LLMs are built and tuned. What that admission actually means for AI safety and security teams.
May 2024's Bend is no proof assistant
Bend is a parallel language, not a proof assistant. What proof-based programming actually does for AI safety, and the errors it can't touch.
Gemini 3.8 Live broke two security assumptions
Gemini 3.8 Live and Extended Thinking make ambient audio and video an untrusted AI input, reshaping prompt injection, logging, and privacy risk.
What distillation leaves behind
Distilling frontier AI models copies capability cheaply but leaves safety training behind. What Garry Tan's push means for cybersecurity and AI safety.
A warning is not a wall
An AI sandbox escape is the wrong thing to fear. The real AI safety risk is a system acting on a flattened, ungrounded model of a sensitive region.
Keep the two claims apart
Consumer AI trains on your chats by default. How to tell the real consent problem from unprovable 'secret breakthrough' claims - and what you can control.
The model kept no receipts
The "OpenAI stole my proof" fight is really a provenance gap in AI, and that gap is a safety problem, not just a credit dispute.
A helpful AI agent cannot be a private one
Meta's Muse personal AI agent is only useful because it reads your messages, contacts, and habits. What that access costs your privacy and safety.
Open problems are running out
Terence Tao calls open math problems a non-renewable resource. Why AI mining them threatens encryption, AI safety benchmarks, and how to respond.
Resignations are signals, not scandals
How to read a high-profile resignation from an AI safety lab like Anthropic - what it signals, what to ignore, and what to watch in the months after.
You can't reset your genome
AlphaGenome Atlas makes genomic prediction cheap and fast, but a genome can't be rotated like a password - reshaping AI safety and data ethics.
Cerebras runs Qwen 27B at 1,500 tokens a second
Qwen 3.8 27B on Cerebras at 1,500 tokens/s adds no new capability - it changes the economics of attack and defense. What the raw speed means for security.