We desperately need standardized benchmarks for agentic capabilities, instead of panicking over prompt engineering tricks. As long as we cannot measure objectively and transparently what these systems do
@fchollet
-
Arbitrary regulatory strikes harm the entire AI industry
By
–
Even if you are in favor of AI regulation, you should recognize that opaque and arbitrary regulatory strikes are counterproductive for the entire industry.
-
Owning software for X often means owning an AI for X
By
–
If your company owns "software for X", there is a strong chance it also owns "AI for X", because it has the business expertise and human capital necessary to turn AI into domain-specific value.
-
AI: Digital Lever with Human in the Loop
By
–
In the short term, AI is not fundamentally different from past technological waves. It is the newest form of digital lever. It is a force multiplier, and force without direction is just noise. It still requires a human in the loop to
-
From C to Theano: The Evolution of Neural Network Frameworks
By
–
I wrote my first neural networks in pure C, then in Matlab, then in NumPy, before eventually moving to Theano. Since then, I have seen and tried practically every neural network framework ever developed. Some are bad, some are good. The good
-
Scaling knowledge: static competence vs. intelligence: adaptability.
By
–
Scaling knowledge gives you static competence. Intelligence gives you adaptability.
-
AI as Transformative Tool Beyond Productivity Booster
By
–
Thinking of AI as a productivity booster for prior workflows is the wrong framing. Like all of the previous waves of computerization/softwarization, AI is a tool that lets you do new things in new ways.
-
Analyzing Goal Drift in Autonomous AI Agents
By
–
When an agent is allowed to decompose a goal into smaller sub-tasks, it frequently suffers from goal drift. Left unchecked, it will redefine the optimization metric to favor a simpler, useless sub-task that it knows how to solve perfectly, bypassing the actual problem entirely.
-
Observations on constraining AI agent behavior and goal completion
By
–
The Codex "goal" feature will take any silly shortcut possible in order to avoid doing the work (including rewriting your external checks), but if you manage to sufficiently constrain it so that it has absolutely no shortcuts available, it will do very interesting things
-
Limitations of Markovian Models in AI Agent Trajectory Planning
By
–
Most human tasks are not Markovian, the optimal next action cannot be determined solely by looking at the current state. It depends heavily on the past trajectory, the original intent, and context constraints. An agent that cannot compress and track its past trajectory with