I wonder how much the CoT actually describes a causal mechanism and how much is essentially LARPing to cover something much weirder happening under the covers. e.g. Is there a vector that is essentially doing some pixel distribution math that no human would think of doing?
@patio11
-
CS Agent Outperforms LLM: Human Detection Through Writing Patterns
By
–
Five stars for the CS agent; had it sorted in less than a minute. And you know they’re not AI because an LLM would have matched speaking style and presence/absence of punctuation and capital letters.
-
Amazon’s AI Customer Service Flow Criticized as Deliberately Confusing
By
–
I had to do a refund on an Amazon order lost in shipment, and the AI-powered flow before you can speak to an agent is so bizarrely incomprehensible that if I were a cynical man I’d assume it was intentional.
-
Map/Reduce with Actual Maps: CS Perspective
By
–
My CS brain can’t help but say that this is map/reduce with actual maps involved.
-
Scaling non-scalable solutions: paradox or necessity?
By
–
Not for the first time, by the way, is “How do you do a thing that doesn’t scale?” answered by “Scale it.”
-
LLMs Achieve Seven-Year-Old Planning and Execution Abilities
By
–
Non-ironically an important milestone for LLMs: can demonstrate at least as much planning and execution ability as a human seven year old. https://t.co/yiYunaTCSE
— Patrick McKenzie (@patio11) 3 mai 2025Non-ironically an important milestone for LLMs: can demonstrate at least as much planning and execution ability as a human seven year old.
-
Chain of Thought: How LLMs Think Recursively and Invoke Tools
By
–
Nope, it’s a thing called “chain of thought” where the LLM basically recursively invokes itself and outside tools to e.g. think up search terms, get results back, read them, not find it, back to the tank to think of new search terms, another tool invocation, etc.
-
LLMs Will Entirely Replace Search Functionality
By
–
LLMs are going to entirely eat that search, FWIW. (To my great edification, as I do that constantly for written and audio artifacts.)
-
Japanese Bureaucracy Status and Talent Selection Impact
By
–
The Japanese government is extremely capable of awarding status to bureaucrats, and historically that gets a lot of Japan's best to select into joining the bureaucracy. Whether that has broadly positive results is a complex topic that I am someone constrained from discussing.
-
Civil Service Incentives and Policy Implementation Failures
By
–
And so our civil service races down an incentive gradient to avoid getting in the paper and avoid the ire of elected officials, who have their own incentive gradient isolating them from policies which they pass.