waldarbeiter

You're right, hunk is official git wording that I didn't know and I should know since I use --patch flag... It's just that I never heard a human (including online) reason about hunks. While at the same time (from my observation) people say things like code chunk, code snippet etc. a lot.

Thank you, my dict.cc search history contains exactly some of these words. I felt like my english got much worse but when Claude kept talking about "hunk" over and over I felt like the problem is maybe not on my end.

Not to bring LLMs into everything but I think they make using boring stuff more likely. Since training data contains a lot of it and less of the shiny new stuff.

I can completely relate, what really bothers me is that I feel the early LLM generations overconfidence is back in Opus 5. Opus 5 wanted to tell me a training run will only take 30min while having access to the logs where earlier runs took 4x as long. I also didn't ask to estimate how long the run will take it just stated confidently that it will take 30mins.

> Making a decision based on only qualitative data

I suppose you meant quantitative and it's just a typo

Maybe LLMs don't usher in the end of software engineering but they definitely end the whole "made with love (and coffee) in X". Nobody cares if you put effort into something software related. "Does it work? Yes? Ok build the next thing." Its the same with the notion of "taste" (see "sometimes tasteless computer code" from the Goedecke article), your colleague who is also a SE might respect your choices as good taste. But it ends there. This was also the case pre-LLMs I would argue. What is worse now is that communicating any uncertainty in decisions related to implementation will result in an immediate "Have you asked Claude?".

There is actually a whole lot of research around the "use less data" called data pruning. The goal in a lot of cases there is basically to achieve the same performance with less data. For example [1] received quite some attention in the past.

[1] https://arxiv.org/abs/2206.14486

Is the 1 cm spec 1σ (or less) or worst-case? It’s a safety-critical application.

There seems to be a moat like infrastructure/gpus and talent. The best models right now come from companies with considerable resources/funding.

That was not mar a Lago but in New Jersey

My website of choice whenever I have to deal with references is dblp [1]. In my opinion more reliable than Google scholar in creating correct BibTeX. Also when searching for a paper you clearly see where it has been published or if it is only on arxiv.

[1] https://dblp.org/

"4,000 tons is almost four million kilograms"

What does "wolf in sheep's clothing" mean for you, concretely, outside of metaphors?

by Walter Tichy

The different JDK builds are almost all built from OpenJDK. Maven and Gradle cover nearly all use cases. The complexity is not that high.

If it would be easy there would already be a car costing a few million that few can afford but that has solved AD. But there isn't.

Nassim Taleb (black swan author) does not divide his books into chapters. Instead each book consists of several (sub-)books. In Antifragility he explains that two subbooks from the same book have the same distance to each other than one subbook from book to one from another book. I found that quite interesting.

Yeah, the debris removes itself with time. But there is still some uncertainty around how deorbiting harms the atmosphere [1]. Immediate effects of a single deorbit seem to be quite low but having such deorbiting scenarios at a higher frequency could have some negative effects.

[1] https://blogs.esa.int/cleanspace/2022/08/11/on-the-atmospher...

The conventional teaching that I am aware of says that you can scale across three dimensions: data, compute, parameters. But Ilya's formulation suggests that there may be more dimensions along which scaling is possible.

For everyone interested in what makes a classifier put an image in a certain class I can recommend the excellent paper by Robert Geirhos et al. "ImageNet-trained CNNs are biased towards texture" [1]. The paper is a bit older but it's an excellent read.

[1] https://arxiv.org/pdf/1811.12231

When the policy is deployed in the real world only the depth camera is used no waypoints etc.. Scan dots and target heading is used in the first Phase of the training to pretrain a policy in simulation. In Phase 2 a policy is trained end-to-end using the pretrained actor network: "First, exteroceptive information is only available in the form of depth images from a front-facing camera instead of scandots. Second, there is no expert to specify waypoints and target directions, these must be inferred from the visible terrain geometry." For policy training in Phase 2 DAgger which is based on Behavior Cloning is used (with the policy from Phase 1 as the expert), they also use some tricks to make sure no actions that are too different from the expert actions are executed during training. In Phase 2 the network learns to extract environment information instead of from the scan dots from the depth camera. Also in Phase 2 they use the pretrained actor network from Phase 1 but the depth embedding must be learned from scratch. This is how I understand it.

Note that it is from 2018. As someone here already mentioned there is a paper that applies the same idea to Vision Transformers published this year [1].

[1] https://openaccess.thecvf.com/content/CVPR2023/papers/Park_R...