Getting to 50% on the private test set on ARC-AGI will be easier than people think, and getting to 85% will be harder than people think. (Current high score: 39%.)
@fchollet
-
Artificial Intelligence vs Artificial Humans: Reframing the Debate
By
–
Unless you think we're going to literally create artificial human beings, rather than intelligent computers, that idea doesn't start to make sense
-
The Real Role of CEO Beyond Decision-Making
By
–
If you have an AI that is superhuman at ingesting information and turning it into decisions towards a certain goal, then the actual CEO can turn to it for most decisions. That doesn't mean you don't need a human CEO. The primary role of the CEO isn't so much making decisions as
-
AI as Scientific Problem, Not Religious Quest or Existential Threat
By
–
Unlike many people, I don't see creating general AI as a religious quest. I see it as a hard scientific problem, with potential applications that would be tremendously valuable for humanity. I don't believe AI is a "new species" that will "replace us", nor do I believe that
-
Human Program Synthesis for ARC Tasks Differs from AI Approaches
By
–
Basically, yes, people are coming up with and executing "programs" in their head. But it doesn't look like the kind of program synthesis we've been doing in AI so far, at all. To start with, to solve an ARC task, a human only comes up with a very small number of candidate
-
GPT-4o Program Search Solution Not State-of-the-Art
By
–
To be clear — the solution published via blogpost yesterday based on discrete program search driven by GPT-4o does not appear to be a new SotA at this time. That would still be @Jcole75Cole (with pure symbolic approaches not far behind).
-
AI Won’t Bring SkyNet or Financial Security
By
–
There are many things you can worry about, but SkyNet taking over the world ain't it. And the rise of more general AI in the coming years (note: current AI has close to no intelligence and no generality, so far) will *not* save you from needing a 401k or a pension.
-
AI Anxiety: How Existential Fears Affect Life Planning
By
–
This is an underappreciated point. I sometimes hear from people who are so sure they're going to be dead in 10-20 years "because of AI" that they've stopped any sort of long-term planning of their lives, e.g. saving for retirement. Sometimes it's the reverse — "I don't need to
-
Keras 3 Simplifies Model Parallelism Configuration
By
–
Every other model parallelism API I've seen has partitioning annotations in-lined in the model code and uses custom training logic — making it challenging to modify and maintain the partitioning config. Keras 3 dramatically simplifies this.
-
Keras 3 JAX Enables Efficient Distributed Model Training
By
–
Keras 3 + JAX makes it possible to efficiently distribute large models via both data parallelism and model parallelism (variable partitioning) across arbitrarily many devices. It doesn't require modifying your model code or your training code at all — all you need is to