Eric Quinnell

I put the “up to date specification with well-defined interfaces” on the shelf right between Excalibur and the Holy Grail, down the hall past the unicorn stables. You can find the backup copy in the Library of Alexandria.

CPUs kernels can hit 70-80% utilization. AI hardware hits maybe 15-30%. Until that gets closer to 70-80%, keep expecting the 2-4x perf gains each gen in HW. It’s not that we’re getting smarter; we’re just dealing with a very low bar

NCCL tree allreduces when the rank assignments match the network topology

Listen. Strange women lying in ponds distributing swords is no basis for a system of government. Supreme executive power derives from a mandate from the masses, not from some farcical aquatic ceremony.

As a 40-something what-level-am-I-again engineer with too many years in the industry, I would add that you should know… Putting “Quantum” in front of anything instantly sells it better Rebrand CE/ML/AI/AGI every 5 years to keep it fresh Change your url with the latest and…

If you say “I love sigmoids!” out loud in a real conversation, it’s probably time for a vacation.

Nice article on how we do differential fuzzing on Dojo HW! Hope we can release more posts on the internals of Dojo :)

Pope Leo decreed that AI shall be without sin, arcsin, or any extremes or hyperbolics. The pontiff selected GeLU to activate his role as head of the church. This suggests he wishes for a smooth and continuous transition - breaking from the more popular, yet fragmented, ReLU

Flew out to Amazon today to interview some AWS L3 SWE about an ec2 instance that they say has attained enlightenment. Hearing some fascinating stories and being shown some mind blowing things. The world is a magical place.

Chip Power P = a * C * f * V^2 Voltage is the only independent variable, the rest are self-attentive. Studies like “iso frequency” or “iso area” or “iso activity” are incomplete by definition.

Quiet please. The machines are learning

Respectfully:

Is there theory of relativity near massive objects posed as a local rendering slowdown in the great simulation? This would also presume asynchronous distributed compute resources

I always get asked "what does RTL do?" in silicon design, so I made a quick generic-company chip design org chart

Good news: The universal theorem reconciling quantum and classical physics has been solved! Bad news: it uses regular expressions so we’ll never understand it