Of course. But the way it is plotted now also gives a false sense of energy economy.
Maybe there’s another way to compare the data.
An energy-speed plot with multiple curves for the different vehicles comes to mind. Then simply stop plotting that curve when that vehicle is outside of its speed range.
I’m not contesting the notion that the plot is trying to convey; just that it’s not really the right format for visualizing this data.
Kinetic energy is E=1/2mv^2 and because of aerodynamics and friction factors, energy efficiency (or, consumption) varies a lot at different speeds.
This graph has a nonlinear x scale because each vehicle’s entry is at a different speed, therefore the energy scale is nonsensical.
You were born to deploy K8s clusters.
OMG my sides. kubectl delete --all --all-namespaces
Yeah I get it 100%. But that’s what I’m saying. I’m already working on and with models that have entire codebase level fine-tuning and understanding. The company I work at is not the first pioneer in this space. Problem understanding and interpretation— all of what you said is true— there are causal models being developed (I am aware of one team in my company doing exactly that) to address that side of software engineering.
So. I don’t think we are really disagreeing here. Yes, clearly AI models aren’t eliminating humans from software today; but I also really don’t think that day is all that far away. And there will always be need for humans to build systems that serve humans; but the way we do it is going to change so fundamentally that “learn C, learn Rust, learn Python” will all be obsolete sentiments of a bygone era.
But that’s not what the article is getting at.
Here’s an honest take. Let me preface this with some credential: I’m an AI Engineer with many years in field. I’m directly working on right now multiple projects that augment and automate code generation, documentation, completion and even system design/understanding. We’re not there yet. But the pace of progress in how fast we are improving our code-AI is astounding. Exponential growth in capability and accuracy and utility.
As an anecdotal example; a few years ago I decided I would try to learn Rust (programming language), because it seemed interesting and we had a practical use case for a performant, memory-efficient compiled language. It didn’t really work out for me, tbh. I just didn’t have the time to get very fluent with it enough to be effective.
Now I’m on a project which also uses Rust. But with ChatGPT and some other models I’ve deployed (Mixtral is really good!) I was basically writing correct, effective Rust code within a week—accepted and merged to main.
I’m actively using AI code models to write code to train, fine-tune, and deploy AI code models. See where this is going? That’s exponential growth.
I honestly don’t know if I’d recommend to my young kids programming as a career now even if it has been very lucrative for me and will take me to my retirement just fine. It excites me and scares me at the same time.
Also by design. Tech companies collude like this all the fucking time.
He threatened to walk away from Tesla if they didn’t give him twice his stake in the company (because he sold half his stake to buy Twitter).
You shut your damn mouth. Nicolas Cage is perfect.
Wonder if they’d block Gmail, too.
ONNX Runtime is actually decently well optimized to run on CPUs; even with large models. However, the simple truth is that there’s really no escaping that Billion+parameter models need to be quantized and even pruned heavily to fit in memory and not saturate the CPU cache so inferences/generations don’t take forever. That’s a reduction in accuracy, so the quality of the generations aren’t great.
There is a lot of really interesting research and development being done right now on smart quantization and pruning. Model serving technologies are improving rapidly too—paged attention is a really cool technique (for transformer based models) for effectively leveraging tensor core hardware—I don’t think that’s supported on CPU yet but it’s probably not that far off.
It’s a really active field and there’s just as much interest in running huge models on huge hardware as there is big models on small hardware. I recently heard of layerwise inference for CPUs; load each layer of the network to the CPU cache on demand. That’s typically a bottleneck operation on GPUs but CPU memoery so bloody fast that it might actually work fine. I haven’t played with it myself, or read the paper all that deeply so I can’t really comment more than it’s an interesting idea.
Will you fire the first shot?
The Bill Clinton Marijuana Museum.
Nah, that’s without Rekts.
ChatGPT is 4 syllables.
Yeah 100% and I’m glad that people have a place to talk enthusiastically about their thing!
I’ve spent the majority of my life connected in some way or another to the internet, starting as a kid on niche bbs in the 90s, and it never ceases to amaze me that there are vacuum enthusiast forums.
There’s lots of good people in Fl. Go home, edge lord.
It’s part of Vegas’ branding strategy.