H100 offers 1.45x better value than A10G
H100 is 7.5x faster but only 3.1x more expensive. 2.45x better value.
Stream of Thoughts from Alex Dong. What I find interesting, intriguing or insightful.
H100 is 7.5x faster but only 3.1x more expensive. 2.45x better value.
There are not many articles that capture the experience of working, or seeing your loved ones work, in a large software company like Seeing like a Software Company.
Python 3.14 finally released today. The biggest news is that this is the first GIL-free Python. Even though it’s slightly slower and it will have rough edges, this is a huge milestone for Python.
I picked up Cory Doctorow’s book Radicalized recently at the library. Doctorow has a knack for making you care about people in dramatically different social situations. It forced me to really feel for the characters who are on the receiving end of the ugly social impacts of technology “advancements”.
Epoch AI’s recent interview with Ken Ono was a fascinating read. Ono, a mathematician at University of Virginia, discussed in greater detail on how AI is transforming mathematical research. It’s a long interview and here are some highlights:
Joe mentioned the new Petri Alignment tool from Anthropic. As I read through the GitHub repo, I was surprised to find how simple the code is. As I dug deeper, I realized it’s actually a plugin for Inspect AI, an open-source Python project by the UK’s AI Security Institute.
Dwarkesh published a short follow-up to his previous interview with Richard Sutton that I commented extensively last week.
From Science Daily
Pavel’s sincerity (when he recounted the interaction with the French Intelligence service), strong ethical foundation (when asked about what would happen if he is sentenced to 20 years in prison, he said he would starve to death to “reboot” the game rather than give in), and his slow, precise, and deliberate responses are a fresh breeze in an industry that has so much hot air.
When a session drags on and the context window fills up, I often watch the models start to get lazy. They drop earlier decisions, skim the edges, and sometimes declare “all tests are passing” after touching only part of the suite.