learndeeply

Hilarious name for an open source project released by a government lab.

> For me, it was simply a gut feeling. I’ve been talking to founders and doing deep dives into technology companies for decades. It’s been my entire professional life as a writer. And because of that experience, there must be a pattern-matching algorithm churning away somewhere in my subconscious. I don’t know how I know, I just do. SBF is a winner.

This is a fantastic insight into journalists. They need to fit their story into a narrative. If you match their stereotype, then they'll spin whatever tale they can imagine to convince themselves that it's really true. The level of self-deception that's needed is difficult to comprehend for me.

Is it common to have the same number of employees as there are companies funded at a VC? Is each one assigned a company?

Brilliant way to get thousands of tech people to look at your business card.

Edit: This was updated, it used to point to his real Github username. Not that there's anything wrong with that! It's clever marketing.

Aside from the project itself, the production value of this video is crazy good.

JAX is a DSL on top of XLA, instead of writing Python. Example: a JAX for loop looks like this:

   def summ(i, v): return i + v
   x = jax.lax.fori_loop(0, 100, summ, 5)

A for loop in TinyGrad or PyTorch looks like regular Python:

   x = 5
   for i in range(0, 100):
      x += 1

By the way, PyTorch also has JIT.

Previous discussion: https://news.ycombinator.com/item?id=22374825

Wonder if anyone's turned Instagram's version of Blurhash into a library?

I can't think of anything that neural nets can't beat, except small tabular data with boosted decision trees. Can you give some examples?

The code is very easy to read. Doesn't seem like there's data/model parallelism support for training, which will be important for real-world use.

https://github.com/intel/intel-extension-for-pytorch PyTorch version. Looks like it's been supported a lot longer than TensorFlow, the repo dates back over a year.

> And honestly, their hardware is pretty bad for ML given you can pick up a used 3090 for $600.

How do you know it's bad? Do you have benchmarks?

Just checked, used 3090s are going for $700-900 on eBay.

No, it's meant for taking something like a CSV file and deciding if each row matches a specific category. A common example, have a CSV, columns corresponding to different features of flowers (e.g. number of petals, size of petals) and the output is the type of flower.

My mistake, I was thinking of gpt-neox. Thanks for the correction.

> Tensorflow serving is barely controllable, and requires insane tuning to make it perform the same as pytorch

What are you using for serving PyTorch models?

PyTorch has been a darling of almost every noteworthy open source model for the past 3-4 years (BLOOM, GPT-J, StyleGAN3, detectron, etc). Personally, I've only seen people use TensorFlow/XLA if they got free TPU credits from Google (gpt-neo), or if it was released by Google (t5).

Looks like its just a name collision. It's a tensor used in distributed models, thus Distributed Tensor, or DTensor for short.

The parent comment said it would be "a good start". It's like adding sleep(1000) to a benchmark to purposely make it look worse than your own product.

How would I be able to respond to the post in detail if I didn't read it? What a bizarre, defensive response. To address your points:

- Multiple formats were compared

Yes, but not a zero-copy or efficient format, like flatbuffer. It was mentioned as one of the highlights of postgresML:

> PostgresML does one in-memory copy of features from Postgres

> - Duckdb is not a production ready service

What issues did you have with duckdb? Could use some other in-memory store like Plasma if you don't like duckdb.

> - Pandas isn't used

that was responding to the point in the post:

> Since Python often uses Pandas to load and preprocess data, it is notably more memory hungry. Before even passing the data into XGBoost, we were already at 8GB RSS (resident set size); during actual fitting, memory utilization went to almost 12GB.

> You seem to be trolling.

By criticizing the blog post?

"ML" was never monopolized by Python. Boosted decision trees, as the post demonstrates, are commonly done in Matlab, R, or Julia. Deep learning however is 99% Python interface.

Python can be fast if you don't intentionally cripple it. Doing the following will be most likely a lot faster than postgresml:

- replace json (storing data as strings? really?) with a binary format like protobuf, or better yet parquet

- replace redis with duckdb for zero-copy reads

- replace pandas with polars for faster transformations

- use asynchronous, modern web framework for microservices like fastAPI

- Tune xgboost CPU resource usage with semaphores

> Vivid Natural — A slide of jam, jazzman, hook, band, pipe, and coil. Small contrasting details in natural colors. vivid natural.

This one generates the string instruments with the correct colors.

Can you summarize the approach you took on creating the model? Really curious!

(not a hardware engineer) This looks like Google's version of Infiniband, which is basically mandatory for large scale ML-training servers due to issues with Ethernet, as described in the paper. Really cool stuff, wonder if we'll see traditional network hardware manufacturers copying this design.

Microsoft invested $1B in OpenAI in 2019, the valuation is probably somewhere in the neighborhood of $5-20B.

> But then Meta is using those academic non-commercial datasets to train a model, presumably for future commercial use in their products. Weird, right?

This is a very strong and likely inaccurate presumption.

Same for: https://pypi.org/project/pytorch/

> You tried to install “pytorch”. The package named for PyTorch is “torch”

For the first example the snippet isn't correct, it just says `conda list`. The website (TutorialPoint) it links to is also wrong and useless.

> you'll typically get people like me who close the tab and leave.

You're not the target audience, why should they appeal to you? This is a website accompanying a paper about reinforcement learning.

"HD" phone codecs seem higher quality than the example given.

It would be nice to compare it to a higher quality sample, currently the samples sound like they were recorded through a telephone (4 khz).