Behavior != Intelligence. You can mimic *behavior* in order to fake being intelligent, but you can't "simulate intelligence". Either you have what intelligence represents (efficient novel skill-acquisition ability) or you don't.
@fchollet
-
AI as Next Platform: Underwhelming Economic and Social Impact
By
–
Meanwhile, if you were a more grounded person who thought it was merely going to be "the next platform" or "the next mainstream interface" and that it was going to unleash economic benefits and social change on par with smartphones — then it's also kind of a dud.
-
AI Hype vs Reality: Early 2023 Predictions Fall Short
By
–
But if you listened to the AI posters from Q1 2023 telling you it was a "civilizational technology" that was a prelude to imminent AGI, that it was going to make you 10x-100x more productive, and that it was going to automate most if not all jobs, then yes, it's a dud. None of
-
Generative AI Success: Redefining Expectations and Industry Impact
By
–
Is generative AI a dud? Well, depends on what your expectations were. If you had no expectations, then it's an insane success story: new tech suddenly unlocks a multi-billion dollar niche industry that enables a handful of new startups to rise and make it big.
-
Exploitation Mode: The Illusion of Progress Through Scaling
By
–
Going into exploitation mode (scaling up what works) can give you the illusion of fast progress when you're looking at the wrong metrics. Like "solving" air travel by forever scaling up a hydrogen zeppelin. It may seem like you can cover ever greater distances, but you're still
-
Deep Learning Limitations Unchanged Since 2017
By
–
Re-reading an article I wrote in 2017, and I'm finding I could have written it yesterday. It's fascinating how the limitations of deep learning have stayed the same since. Same problems, same failure modes, no progress.
-
Unified codebase leveraging best features across ML frameworks
By
–
Yes, and also: just get the best of each ecosystem. Want to debug in the PyTorch runtime? Sure, you get that. But need to also deploy via TFLite? Sure, you can do both. Want to speed up your code via JAX? Sure, you can. All with one codebase.
-
Keras Core Beta Achieves 100K Downloads and 136 Community Pull Requests
By
–
We launched the Keras Core beta one month ago. Since then, the package was downloaded nearly 100,000 times and we received 136 PRs from the community. We've also made fast progress with the codebase in a short time. We're on track for the big release in a couple of months.
-
Program Synthesis Engine Generalizes Multiplication from Few Examples
By
–
Meanwhile, a program synthesis engine with a basic DSL that contains `add` and `for` can use just a handful of training examples to synthesize a multiplication function that will generalize *exactly*, to any new input — not even tied to a specific input distribution.
-
Focus on Building Despite Persistent AI Delusions
By
–
Anyway — back to building. AI delusions will still be prevalent next year, and the year after.