ethmarks

I agree wholeheartedly. It irks me when people critique automation because it uses large amounts of resources. Running a machine or a computer almost always uses far less resources than a human would to do the same task, so long as you consider the entire resource consumptions.

Growing the food that a human eats, running the air conditioning for their home, powering their lights, fueling their car, charging their phone, and all the many many things necessary to keep a human alive and productive in the 21st century are a larger resource cost than almost any machine/system that performs the same work. From an efficiency perspective, automation is almost always the answer. The actual debate comes from the ethical perspective (the innate value of human life).

No, but the average user is capable of describing what they want to something trained in interpreting what users want. The average person is incapable of articulating the exact steps necessary to change a car's oil, but they have no issue with saying "change my car's oil" to a mechanic. The implicit assumption with LLM-based backends is that the LLM would be capable of correctly interpreting vague user requests. Otherwise it wouldn't be very useful.

And also like a board game, which I think is what they were going for.

Perhaps it was easier? There were also probably legal reasons.

They have different use cases. uv is meant to be the singular tool for managing Python packages and dependencies, replacing pip, virtualenv, and pip-tools. Conda is for more general-purpose environment management, not just Python. If you're doing something with Node or R, uv won't work at all because it's only for Python.

uv's biggest advantage is speed. It claims a 10-100x performance speedup over pip and Conda [1]. uv can also manage python versions and supports using Python scripts as executables via inline dependencies [2].

But Conda is better for non-Python usage and is more mature, especially for data science related uses.

[1]: https://github.com/astral-sh/uv/blob/main/BENCHMARKS.md [2]: https://docs.astral.sh/uv/#scripts

Even better, Zed with Helix keybindings: https://zed.dev/docs/helix

The reason they're farming hard drives is so they can add them to their cloud storage network. You realize that cloud storage isn't actually in clouds, right? Hard drives are always local somewhere, and in Backblaze's case that 'somewhere' is a data center.

Page 21 of their 2024 annual report[1] has expenses listed. About $3,000,000 for web hosting, about $100,000,000 for salaries and benefits out of $178,000,000 total.

[1]: https://wikimediafoundation.org/wp-content/uploads/2025/04/W...

> all they'd have to do is put Gemini directly into Chrome

Google would never do that! /s

https://gemini.google/overview/gemini-in-chrome/

It's not confidence-inspiring that their landing page has such atrocious performance. Google Lighthouse gives it[1] a performance score of 25/100 with 12,310ms of TBT and a speed index of 24.7 seconds. It's not even an interactive page either; it's just a video, some images, and some text.

[1] https://pagespeed.web.dev/analysis/https-chatgpt-com-atlas/x...

It's extra distracting because it doesn't even read like normal LLM prose, but it's close enough to feel off.

The frequent use of bold emphasis, lists, and subject-only rhetorical questions ("Those tiny m3,1 m1,6 m-4,-7 commands? They're micro-MoveTo operations.") are classic LLM-speak, but they're used in such a way that makes me doubt that OP actually used an LLM to write this. I think that OP's natural prose just happens to be stylistically pretty similar to that of an LLM.

It's kind of sad that what were once signs of high effort and dedication (e.g. em-dashes) are now signs of low effort and dishonesty, despite the fact that people still use them in human writing.

Perhaps one could link to cross-posts from the blog? Maybe a "Discuss on HN" button. And the solution to avoid spamming HN is to avoid writing a large volume of low-quality content on the blog.

Maybe I'm misunderstanding, but it seems like you're just talking about AI inpainting. That's like one of the first things people did with image diffusion technology. NVIDIA published a research paper on it back in 2018: https://arxiv.org/abs/1804.07723

Inpainting is harder on videos than on images, but there are plenty of models that can do it. Google's Veo 3 can remove objects from videos: https://deepmind.google/models/veo/

Someone should make a list of these types of lists!

I've only tried Grok Code Fast 1, so I can't speak for any of the other models.

In my experience, Grok is very fast and very cheap, but only moderately intelligent. It isn't stupid, but it rarely does anything that impresses me. The reason it's a useful model is that it is very, very fast (~90 tokens per second) and is very competitively priced.

This reminds me of Svelte's LLM docs: https://svelte.dev/docs/llms

The actual underlying neural net that the LLMs use doesn't actually output tokens. It outputs a probability distribution for how likely each token is to come next. For example, in the sentence "once upon a ", the token with the highest probability is "time", and then probably "child", and so on.

In order to make this probability distribution useful, the software chooses a token based on its position in the distribution. I'm simplifying here, but the likelihood that it chooses the most probable next token is based on the model's temperature. A temperature of 0 means that (in theory) it'll always choose the most probable token, making it deterministic. A non-zero temperature means that sometimes it will choose less likely tokens, so it'll output different results every time.

Hope this helps.

The VSCodeUserSetup file from https://code.visualstudio.com/download is in the 90MB range.

Perhaps this file is just the installer and the actual system files are much larger? Or maybe your 400MB figure comes from a bloated install? Just speculating here.

Of course they don't. But if even a small percentage of their userbase uses this on even a small percentage of their photos, that amounts to a massive amount of storage. It doesn't cost Google very much to store all those images, but those users will have to pay for more storage. Even thought it isn't significant to Google when a single person does it, it becomes highly significant when millions of people do it.

Zed supports BYOK so you can connect it to GitHub Copilot or Anthropic or OpenRouter or whatever. The 500-per-month limit is only for Zed-hosted prompts. I personally switch between Zed-hosted, GH Copilot, Gemini API, and Ollama. Zed's AI integration isn't quite as "it just works" as Cursor's, but it's still very good and it gives you much more freedom.

Here's another implementation I made a few months ago:

https://ethmarks.github.io/posts/screamcipher

I don't follow your logic from 'When I tried it, AI hallucinated some citations' to 'AI is fundamentally incapable of explaining a textbook'. Those seem like rather different tasks to me. Plenty of humans are perfectly capable of paraphrasing and summarizing a textbook but struggle with standardizing citations.

Could you elaborate on this? Enriching lithium sounds much more feasible than setting up a moon base, but it still sounds like it would be economically infeasible.

First, am I understanding correctly that you aren't talking about DT fuel, but are instead suggesting that we use the tritium to produce helium-3 by letting it decay? This sounds extremely inefficient, and it also seems like it would cause supply chain issues where power production relies on helium-3 that we started producing decades ago. This would mean that the infrastructure we set up wouldn't be useful for decades, and that if production stopped for a few months there would be a helium-3 shortage decades later. Lunar helium-3 extraction would have a shorter feedback delay, making it less vulnerable to latent shortages and more useful as a medium-term power source.

Second, isn't lithium-6 a rare material? Lithium is abundant, but brine and rock deposits that are cheap and easy to extract are quite limited. Also IIRC something like 92% of it is lithium-7 which is much less useful in enrichment. And that's not even considering the geopolitical factors you mentioned. I'm obviously not saying that tapping into harder-to-access deposits and filtering for lithium-6 would be more difficult than sifting lunar regolith for helium-3, but it still sounds extremely difficult.