Have there been any notable AI-driven compilers yet?
I think AI would absolutely slay at that task. “Hey take this high-level description of a computation and give me back the most optimal 1s & 0s for this chipset” as an optimizing compiler would be 🤯
All the things the JVM wishes it could be 💅
I think AI would absolutely slay at that task. “Hey take this high-level description of a computation and give me back the most optimal 1s & 0s for this chipset” as an optimizing compiler would be 🤯
All the things the JVM wishes it could be 💅
Comments
We'll see a lot of this _if_ we fix hallucinations, which we might get with smaller models
My understanding is the latter and not Machine Reinforcement Learning, where it’s trying to produce the most optimal outcome for a given input
What am I missing?
That’s why Deep Learning unlocked so much there, to effectively create “infinite buckets” to reinforce
MRL is much simpler, as there are fewer buckets to optimize into
For an optimizing compiler, it’s limited by chipset again, and more easily reinforced. Can’t hallucinate?
When it hallucinates machine code, then the result is completely unusable
Last time I checked, we were nowhere near of fixing hallucinations
Hallucination of machine code isn’t quite the same, as a chipset has to execute it. So improper code will just fail or be slower, thus reinforce negatively, tuning the ML algo appropriately