Well, yeah - middle-level management is one of the worst positions to be in. You have a lot of responsibilities, but not enough ways to influence things. And the pay often isn't higher than before.
I think that at this point, the idea becomes just a gatekeeping.
I have been writing notes in notebooks by my hand for years. But gradually I switched to typing notes in a notepad or drawing with a stylus on a gadget - and I don't notice any real difference.
I'm cautiously waiting for the feedback from the first users.
Meta has produced a lot of great models (LLama), maybe this is a comeback... but I'm cautious, as the jump in the quality is almost too high.
Also, I think people aren't used that using such models requires meta.ai or meta ai app.
I love writing things down when I brainstorm; it helps me think.
But taking notes by hand is not feasible for me. My handwriting is atrocious (always has been); if I want to write a nice-looking text, I have to slow down significantly, and then it becomes too slow.
Also, searching, indexing, and everything - it works better with digital.
I prefer to start immersion at ~B1 level. I know some people want to do comprehensive input right from the start, but I prefer to build foundations.
I start with popular books that I have already read before - The Little Prince, Harry Potter, etc. Then I take books that are interesting for me and work through them.
There are graded readers, but the stories are usually very boring. I prefer to read what I like.
I sympathise. I did ~200k reviews in Anki and no idea how many in Renshuu in 2025 for language learning: German, Spanish and Japanese... and I think I got into Anki hell.
For a long time, Anki was really useful for me, it pushed my Spanish and German forward, but now I plan to decrease the number of reviews significantly. I hope to spend no more than 30 minutes per day of flashcards, and the rest of time on immersion.
You know, sometimes I feel that all this discourse about AI for coding reflects the difference between software engineers and data scientists / machine learning engineers.
Both often work with unclear requirements, and sometimes may face floating bugs which are hard to fix, but in most cases, SWE create software that is expected to always behave in a certain way. It is reproducible, can pass tests, and the tooling is more established.
MLE work with models that are stochastic in nature. The usual tests aren't about models producing a certain output - they are about metrics, that, for example, the models produce the correct output in 90% cases (evaluation). The tooling isn't as developed as for SWE - it changes more often.
So, for MLE, working with AI that isn't always reliable, is a norm. They are accustomed to thinking in terms of probabilities, distributions, and acceptable levels of error. Applying this mindset to a coding assistant that might produce incorrect or unexpected code feels more natural. They might evaluate it like a model: "It gets the code right 80% of the time, saving me effort, and I can catch the 20%."
Well, if the users ask frequent/common questions to ChatGPT and get acceptable answers, is this even a problem? If the volume of duplicate questions decreases, there should be no bad influence on the training data, right?
When I was interviewing at my last company, during the raising bar interview, the interviewer said that he looked at my personal website, liked it, and was really impressed by it. Then we spent one hour chatting about our experiences. So, I can say that it definitely helped me get hired.
https://andlukyane.com/blog/
Last year, I started playing DnD with 5e rules and lots of homebrew content for the first time. It is an online game with my friends on Discord. We have character sheets, roll a dice using a bot, and so on.
This is really fun, we had some crazy adventures and this campaign is just at early stages :)
> I wonder how hard it would be to modify this code to run on a 64GB M2 Mac.
It isn't that hard, I was able to run in on M1.
The changes are:
remove or modify multiprocessing - it doesn't work on Mac the same way as in the code;
replace `device = "cuda"` with `device = "mps"`
In this line ` att_idxs = (torch.clamp(torch.arange(context_size)[None, :] - torch.arange(context_size)[:, None], -pos_emb_radius, pos_emb_radius-1) % pos_emb_size).to("cuda")` replace cuda with "mps"
in `optim.AdamW` remove `fused=True` - we can't do it without CUDA
Replace
```with autocast(device_type='cuda', dtype=torch.float16):
_, loss = mlm_head(bert(batch_data_torch_xs[mb_start_idx:mb_end_idx]), batch_data_torch_ys[mb_start_idx:mb_end_idx])
```
with simply `_, loss = mlm_head(bert(batch_data_torch_xs[mb_start_idx:mb_end_idx]), batch_data_torch_ys[mb_start_idx:mb_end_idx])`
replace `scaler.scale(corrected_loss).backward()` with `corrected_loss.backward()`
replace
```
scaler.unscale_(optimizer)
scaler.step(optimizer)
scaler.update()
```
with `optimizer.step()`