mirker

One thing to point out is that the threshold of predictor complexity is dependent on the execution pipeline. A very speculative and deep architecture has a bigger need for better predictors, since it has a massive penalty when there is a misprediction.

Does anyone have experience using these open source models in production?

How’s this compared to ReLM: https://arxiv.org/abs/2211.15458

There is a ton of value. OpenAI having proprietary LLMs single handedly pivoted the entire field to LLMs. A random GitHub repository doesn’t come close to impact.

Because the authors don’t get a large reward for open sourcing the work and they stand to lose future value by lowering the gate to competition. You may want the code, but Google will not care (or it might dislike it).

Look at GPT-3+, OpenAI gets fame and fortune while people struggle to reproduce their last-gen models.

They asked Watson of course.

Agreed.

Here’s the thing: the authors of that paper got early access to GPT-4 and ran a bunch of tests on it. The important bit is that MSR does not see into OpenAI’s sausage making.

Now imagine if you were a peasant from 1000 AD who was given a car or TV to examine. Could you really be confident you understood how it worked by just running experiments on it as a black box? If you give a non-programmer the linux kernel, will he/she think it’s magical?

Things look like magic especially when you can’t look under the hood. The story of the Mechanical Turk is one example of that.

The APIs were messed up early on, which is a reason TF2 happened. Every team started making their own random implementations of stuff. You had the TF Slim API, you had Keras, etc. The API just got fatter and fatter and then libraries would make cross dependencies to bake in the API mistakes.

Nah, TF has had dynamic execution since TF2 and it’s still losing users, it seems. The execution model and API is simply more complicated. What’s a session, placeholder, constant, tensor, …? PyTorch was sold as numpy with GPU support and it is pretty close to that. JAX is an attempt to approach language simplicity and purity.

PyTorch examples were also cleaner. torchvision had ResNet training batteries included while TF had role your own or clone some weird Keras repository.

I don’t understand what you mean. Here’s how many applied ML papers work: create a new dataset for a novel problem, download a PyTorch model, point model at dataset directory. Is it novel? By construction. Is the ML technique novel? No.

The story I’ve heard is the economics undergrads can’t get into economics grad school. This is just a rumor but the sentiment is that undergrads get taught a watered down version of economics theory. Economics theory is potentially very technical and includes game theory and proofs. Even in CS, undergrads take intro theory courses and “bottom out” in their math skills, even though grad-level CS gets much more difficult. Therefore, I’d imagine the primary determinant of this rigor phenomenon is the GPA inflation of the major.

I agree with you but does anyone even recognize the last category outside blue-sky research? People have a tendency to bin other people into buckets. Being a master at 2 things means you can’t be easily placed in a typical team structure.

But it’s more thoughtful. The purchaser thought long and hard about what place the purchasee can use it in.

When people say you need the third hardware revision to get functional performance, and it still doesn’t work, you should conclude it’s poorly designed. The fact that these chips were being sold with old hardware revisions also seems anti-consumer.

Yeah the issue is you can generate data, but it won’t be good data. Training over random strings won’t make you learn language, but it’s technically data.

The OPs point is that it’s likely impossible to do what is claimed here in general. Imagine the LLM says something like Fermat’s Last Theorem. To verify it, you’d have to either 1) have a proof assistant powerful enough to construct a proof 2) use a second ML model to guess truthfulness. The former is technically challenging and the latter is another model, with its own biases and factual inconsistencies.

Many of those people left though.

There’s actually a few papers already on constrained decoding. I won’t link them but if you go on arxiv and really look you will find a couple in the past year.

Sure, I agree they are useful. My objection is it’s more in the tool category than science, while Alphafold is both. There isn’t convincing evidence that GPTs are pushing what we know; rather, they make it easier to process/search what we already know. You could hire an ML expert to be your tutor without GPTs and you’d get equal or better tutoring, though at a higher price. You can’t hire people to predict protein folding better than Alphafold. It’s very convenient that GPTs exist and they can provide tons of value, but they’re essentially the next version of mechanical turk or a domain expert you’d hire for contract work except more scalable. The net impact of GPTs may also be higher due to how often we use text, but I’d rather see a society curing disease, etc. than one generating fake books, etc.

Alphafold is open and seems fundamentally transformative in the science space. GPT is nice but it’s a smart meme-generator at the moment. I don’t disagree with the impact on G’s bottom line, though.

OS needs less than 100GB, though, so mixed drives seems more cost effective.

The image is nearly unreadable for me on mobile.

DOTA, Rocket League, CSGO, etc. are e-sports friendly but don’t have this population problem. It’s not about being “hardcore”; it’s about having broad appeal to hit a wide market segment. RTS has struggled with that historically. Tower defense and MOBA are two examples where RTS was simplified to a broader appeal game.

Hardcore “competitive” RTS (StarCraft/Age of Empires) can’t since it’s not competitive enough if you can’t hit a meaningful 300+ actions per minute, which is impossible without hotkeys. Even IF they could, I am skeptical you’d get converts. Even IF all of them converted, hardcore RTS is not thriving right now. You’d have to create the perfect game and you’d be rewarded with less players than a Minecraft clone.

It looks interesting, honestly. Shame the RTS market is relatively small.

I think one of the most glaring problems with the argument is HPC maintains tons of legacy simulations which may not even have the original authors around. Adding Spark support wouldn’t make this easier.

Isn’t the simpler example data parallel training? That’s a reduce and broadcast.

1 GPU in a NAS server is all you need. Pretty simple.

Integrating an accelerator into a dishwasher is also expensive. How many inferences are you going to get out of it before the hardware is outdated? Once a day for 5 years may not be enough to justify the cost. Compare that to having a single GPU in a server in your house that services all your IoT devices, which can be readily upgraded and requires no special software. The is no benefit to doing inference on device unless you are in the business of selling said devices.