Here is one very concrete example of how AI was massively failing to create a better ML algorithm for me. I was using a benchmarking script that would cycle over several repeats for better reliability, and then using the median value to establish the speedup time. So the AI
@tunguz
-
Geographic Concentration of AI Benefits and Infrastructure
By
–
This will be a huge issue. The *benefits* of AI are right now *EXTREMELY* geographically concentrated in a few IT/knowledge work supercenters. While the infrastructure is VERY widely geographically distributed. I don’t buy into much of the anti-datacenter fearmongering, but the
-
Comparison of Gemini 3.1 Pro and Codex 5.3 Performance
By
–
Gemini 3.1 Pro is here. Benchmarks look impressive, and definitely a qualitative improvement over 3.0. However, it also looks like OAI really cooked with Codex 5.3 for coding.
-
Critique of geographical AI workforce concentration and echo chambers
By
–
as i’ve argued many times before, ALL of this is downstream from the forced concentration of ai workforce in a few geographically isolated locations. what they think is the force multiplier of talent concentration is in fact a highly inbred echo chamber.
-
LLM Intelligence and Human Cognitive Abilities
By
–
Not that odd if you really understand how *human* intelligence works. We’ve known about the existence of the “g-factor” for about a century now. If *all* of the LLM training data is downstream from human cognitive abilities, then it’s intuitively unsurprising that these abilities
-
Reflections on the ‘Idiotic Valley’ and the future of AI
By
–
I am starting to suspect that we'll reach the limits of human intelligence much sooner than the AI kind is ready for the prime time. There will be a confusing gap before the proper handover. I call it the Idiotic Valley. We might be in the early stages of it.
-
Using Codex to optimize machine learning algorithms
By
–
Update on my travails with using Codex to speed up and optimize an ML algorithm from scratch. After two days of where things seemed like they were *finally* going in the right direction, at the end of the day I discovered that almost all the speedups were only relevant for one
-
Casual interaction and prompting workflow with AI coding assistant
By
–
Keep yelling "That's not bad, but I need even more improvements!!!" all day to Codex, even though I realize it's doing a darn good job on this particular task.
-
Critique of Machine Learning Research for Tabular Data
By
–
Every ML for tabular data paper I've ever come across.
-
The challenges of affordable local AI hardware
By
–
This is also why the dream of *affordable* local AI hardware will probably remain a pipe dream for a foreseeable future.