Anyway, if you haven't explored JAX yet, you really should. It makes you completely change your expectations for ML frameworks 🙂
@fchollet
-
Out of the Box Performance Without Optimization Efforts
By
–
It's representative of "out of the box performance", the speed folks who aren't performance experts get when they develop their own models. We have made zero performance optimizations to our model. Also we have found torch compile() rarely ever works, and when it does it may not
-
Keras 3 with JAX Significantly Faster Than PyTorch
By
–
It is not uncommon for Keras 3 + JAX to be 3-5x faster than PyTorch models pulled from research repos on GitHub or from Hugging Face. Though the typical figure is more like 30% faster.
-
KerasCV Segment Anything: Zero-shot Image Segmentation with Prompt Points
By
–
KerasCV now includes a Segment Anything implementation, which enables you to do image segmentation using only "prompt points". No training data needed. With the JAX backend, it runs ~5x faster on GPU than the original PyTorch implementation. Told you JAX was fast. Guide:
-
LLMs Not Responsible for Amazon Content Scams
By
–
People have run such scams on Amazon since before LLMs. The generated content is irrelevant, it's only there for SEO purposes
-
AI Coding Assistants: Helpful for Syntax, Limited Beyond That
By
–
Yes, there's an internal tool like that. I use Copilot too though. It's certainly helpful as a syntax shortcut, but if I'm working with a language I know well, I find the benefits to be very limited. The difficult part of programming is developing the mental model, not typing out
-
No Direct Path from Current AI to Hostile Takeover
By
–
We can't say with certainty whether AI will ever pose an existential risk to humans given an infinite timeline. But what we *can* say with absolute certainty is that there is no direct path from current technology to a "hostile AI takeover" scenario — whatsoever. Worrying about
-
Implementing Keras in Rust with .keras File Support
By
–
If you're going to go that way, you might as well go all the way and implement most of Keras in Rust, with support for loading `.keras` files…
-
Keras 3 affine_transform operation now differentiable across backends
By
–
Yes, Keras 3 has ops.image.affine_transform which is differentiable and works with all backends.
-
AI Doom and Hype: A Symbiotic Relationship
By
–
It's a symbiotic relationship: AI doomers need AI hype in order to be taken seriously and be part of the discourse, and AI hypers need AI doom narratives for AI to look momentous and world-altering.