One example might be a tree-like struct where a parent and child have references to each other. Even if everything is cleaned up properly, the borrow checker has no way to know that when the struct is created. Solving it requires unsafe at some point, usually through something like RefCell.

If speed is a concern, why did you all stick with Synapse (essentially single-threaded due to the GIL) over moving to Dendrite? As far as I can tell, Dendrite is, for all intents and purposes, abandoned.

It seems to be a Radeon VII on an Mi50 board, which should technically work. It immediately hangs the first time an OpenCL kernel is run, and doesn't come back up until I reboot. It's possible my issues are due to Mesa or driver config, but I'd strongly recommend buying one to test before going all in.

There are a lot of cheap SXM2 V100s and adapter boards out now, which should perform very well. The adapters unfortunately weren't available when I bought my hardware, or I would have scooped up several.

The first issue I ran into was with them not supporting LLaMA for tool calls. Microsoft stated in February that they were working on it [0], and they were just closing the ticket because they were tracking it internally. I'm not sure why they've been unable to do what took me two hours in over six months, but I am sure they wouldn't be upset by me using the much more expensive OpenAI models.

There are also consistent performance issues, even on small models, as mentioned elsewhere. This is with a rate on the order of one per minute. You can solve that with provisioned throughput units. The cheapest option is one of the GPT models, at a minimum of $10k/month (a bit under half the cost of just renting an A100 server). DeepSeek was a minimum of around $72k/month. I don't remember there being any other non-OpenAI models with a provisioned option.

Given that current usage without provisioning is approximately in the single dollars per month, I have some doubts as to whether we'd be getting our money's worth having to provision capacity.

I don't know of any other autograd libraries with a non-CUDA backend, but I'd be interested to learn about them.

What I'm referring to are things like radio broadcasts, 60 Hz hum from power lines, noise put out by switching power supplies, and that sort of thing.

Just having a bias, as in your example, would be still truly random. If you knew that every tenth roll you'd get a 3, it would no longer be random. When your random number generator can be influenced by the outside world, it's no longer suitable for cryptographic use.

I've been hearing about MLIR and OpenXLA for years through Tensorflow, but I've never seen an actual application using them. What out there makes use of them? I'd originally hoped it'd allow Tensorflow to support alternate backends, but that doesn't seem to be the case.

0: https://cprimozic.net/notes/posts/machine-learning-benchmark...

All AMD had to do was support open standards. They could have added OpenCL/SYCL/Vulkan Compute backends to Tensorflow and Pytorch and covered 80% of ML use cases. Instead of differentiating themselves with actual working software, they decided to become an inferior copy of NVIDIA.

I recently switched from Tensorflow to Tinygrad for personal projects and haven't looked back. The performance is similar to Tensorflow with JIT [0]. The difference is that instead of spending 5 hours fixing things when NVIDIA's proprietary kernel modules update or I need a new box, it actually Just Works when I do "pip install tinygrad".

0: https://cprimozic.net/notes/posts/machine-learning-benchmark...

That will give you RF noise, which isn't really random.

If every nuclear reactor in the US simultaneously had an accident requiring the $70M paid out for Three-Mile Island, we'd be around 1.2% of the way to needing Treasury funds. Three-Mile Island's operator was responsible for cleaning it up, and they paid the entire $1B required to do so.

A few years ago there was an article about K's use in high-frequency trading. I'm not sure about usage of APL and J, though. BQN is still fairly new, so it will take a while to see much production usage.

If you've ever written code using NumPy, Tensorflow, or PyTorch, you're doing array programming. Those libraries are heavily influenced by the array languages, including taking a lot of terminology (rank, etc.). I've personally found that playing with J and BQN helped me understand Tensorflow, and vice versa.