探索 | Thread Easy - 展開 Twitter 線程｜閱讀、摘要與創作

Healthy dynamical complex systems don’t settle into a stable equilibrium; they surf a regulated critical regime, and dynamic pathologies arise when they’re not controlled back into that regime. I would argue that this statement also holds true for most dynamic and complex system including large language models.

Exploring science, cognition, AI, robotics, engineering, dynamical complex systems, more importantly life and beyond.

Faruk “Frank” Guney

Thu Nov 13 01:56:32

The real challenge in AI isn’t learning... it’s unlearning. Once unlearning becomes as performant as learning, systems will achieve true continual digital cognition: stable, plastic, and self-correcting.

Exploring science, cognition, AI, robotics, engineering, dynamical complex systems, more importantly life and beyond.

Faruk “Frank” Guney

Wed Nov 12 02:12:01

Now that coding is almost out of our way thanks to vibe coding, teams can finally ship not just new products, but new research. Innovation is being democratized right before our eyes. Focus more on "concepts" than "implementation details".

Exploring science, cognition, AI, robotics, engineering, dynamical complex systems, more importantly life and beyond.

Faruk “Frank” Guney

Tue Nov 11 23:43:18

It seems that, sooner or later, AI in the U.S. will belong to Google and OpenAI, and maybe X. Google will dominate the soft-hard interface, the fusion of software intelligence and hardware reach. Unless you’re building where their horizon ends, which is pretty far out.

Exploring science, cognition, AI, robotics, engineering, dynamical complex systems, more importantly life and beyond.

Faruk “Frank” Guney

Tue Nov 11 21:24:09

RT @frankguney: @JunyaoShi I think the right way to put this is that learning is a byproduct of experience with feedback. Even in supervise…

Exploring science, cognition, AI, robotics, engineering, dynamical complex systems, more importantly life and beyond. Founder at https://t.co/HInAbA9KEW, and https://t.co/vnHrcX97UC

Faruk “Frank” Guney

Tue Nov 11 13:04:03

Within the current AI paradigm (deep learning with SGD), unlearning (forgetting) is technically near-impossible without full retraining or with formal guarantees. Once learned, we persist... not as data, but as gradients fossilized in the model’s weights.

Exploring science, cognition, AI, robotics, engineering, dynamical complex systems, more importantly life and beyond. Founder at https://t.co/HInAbA9KEW, and https://t.co/vnHrcX97UC

Faruk “Frank” Guney

Tue Nov 11 12:13:40

探索

Newest first — browse tweet threads

探索

Newest first — browse tweet threads

The real challenge in AI isn’t learning... it’s unlearning. Once unlearning becomes as performant as learning, systems will achieve true continual digital cognition: stable, plastic, and self-correcting.

Now that coding is almost out of our way thanks to vibe coding, teams can finally ship not just new products, but new research. Innovation is being democratized right before our eyes. Focus more on "concepts" than "implementation details".

It seems that, sooner or later, AI in the U.S. will belong to Google and OpenAI, and maybe X. Google will dominate the soft-hard interface, the fusion of software intelligence and hardware reach. Unless you’re building where their horizon ends, which is pretty far out.

RT @frankguney: @JunyaoShi I think the right way to put this is that learning is a byproduct of experience with feedback. Even in supervise…

Within the current AI paradigm (deep learning with SGD), unlearning (forgetting) is technically near-impossible without full retraining or with formal guarantees. Once learned, we persist... not as data, but as gradients fossilized in the model’s weights.