I recently removed in editor AI cause I noticed I was acquiring muscle memory for my brain, not thinking through the rest past the start of a snippet that would get an LLM to auto complete. I’m still using LLMs, particularly for languages and libraries I’m not familiar with, but using the artifacts editors in ChatGPT and Claude.

People haven’t really changed. As always, power corrupts. When the rewards are great enough, it seems people are often enough willing to compromise their integrity.

Key detail in the actual memo is that they’re not using just an LLM. “Wallach anticipates proposals that include novel combinations of software analysis, such as static and dynamic analysis, and large language models.”

They also are clearly aware of scope limitations. They explicitly call out some software, like entire kernels or pointer arithmetic heavy code, as being out of scope. They also seem to not anticipate 100% automation.

So with context, they seem open to any solutions to “how can we convert legacy C to Rust.” Obviously LLMs and machine learning are attractive avenues of investigation, current models are demonstrably able to write some valid Rust and transliterate some code. I use them, they work more often than not for simpler tasks.

TL;DR: they want to accelerate converting C to Rust. LLMs and machine learning are some techniques they’re investigating as components.

Title worried me for a moment that they were dropping Steam Input; happy to see they seem intent on the opposite.

Note the versions, none of the results give you the official operators page for the current version, 16. They give 9, which went EOL in 2021.

Have you tried recent models? They’re not perfect no, but they can usually get you most of the way there if not all the way. If you know how to structure the problem and prompt, granted.

Them using Google indexes anonymously isn’t intending to solve the problem you think it is. It’s more about incentive structures. Google’s “free” search optimizes for ad revenue now. The API access doesn’t as much, and Kagi certainly doesn’t have an ad incentive. So privacy is a nice bonus, but the real benefit is a customer serving incentive structure.

Codeberg is run off of donations, they have no service contract revenue. Nobody, much less a volunteer, wants to commit to a 5 or 10 year service plan like that, it’s not sustainable for a small project from a non profit.

I use Kagi, they provide access to all the main models in a chat interface and have a mode that feeds search engine results to them. It’s mostly replaced search engines for me. For programming work I find them very useful for using unfamiliar tools and libraries, I can ask it what I want to so and it’ll generally tell me how correctly. Importantly, the search engine mode has citations. $25 a month, but worth it.

After doing some Meta/Facebook VR development in my job the lack of popularity made increasingly more sense. In brief, they’re both incredibly incompetent and transparently greedy.

I’m honestly baffled how they could spend so many tens of billions of dollars and have such bad software, it is completely bug ridden. You’ll hit a bug, research it, and find out it’s a major know bug for literal years they haven’t fixed. They care so little that they couldn’t bother to update the Oculus branding to Meta for over 3 years in various software tools and libraries.

Their greed might be more salient aspect preventing adoption, though. They transparently wanted to be the gatekeepers to everything “metaverse” related, a business model that is now explicitly illegal in the EU after years of being merely very sketchy. They are straight up hostile to anyone else trying to implement enterprise or business features. Concrete example: fleet management software, aka MDM. There are third party tools that are cheaper and much more featured than Meta’s solution, but in the last year they’ve pushed hard to kick those third parties out of the ecosystem.

I could go on, but in short nobody in their right mind would build a major business on their ecosystem. They’d rather let Meta burn billions in R&D and come back later. Besides, not even Meta is able to make money in the area now.

It runs great now. Most importantly, it supports extensions like ublock.

So this is probably another example of Google using too blunt of instruments for AI. LLMs are very suggestible and leading questions can severely bias responses. Most people using them without knowing a lot about the field will ask “bad” questions. So it likely has instructions to avoid “which is better” and instead provide pros and cons for the user to consider themselves.

Edit: I don’t mean to excuse, just explain. If anything, the implication is that Google rushed it out after attempting to slap bandaids on serious problems. OpenAI and Anthropic, for example, have talked about how alignment training and human adjustment takes a majority of the development time. Since Google is in a self described emergency mode, cutting that process short seems a likely explanation.

The original UE5 seemed slightly premature, the 5.1 and 5.2 updates were significant and non trivial to update to. It also seems like a tool that gives you enough rope to hang yourself. Unlike Unity, everyone gets ready access to the full engine source code. Fortnite runs UE5, so the performance isn’t inherently bad.

The headline is pretty misleading. The full quote states this being the studios debut game made several factors major problems. UE5 as the engine, a highly competitive genre, and a new IP made nearly insurmountable obstacles. For comparison, Doom Eternal, their obvious AAA competitor, was from a veteran studio with a legendary IP built on literal decades of custom engine experience. On the other hand, a game like Ultrakill can compete by having an incredibly tight scope.

Arch for stuff I have physical access to. Nothing’s ever gone wrong, so it’s worth it for the immediate updates and consistency with my other systems. For VPS I use Debian though, occasionally the unstable/Sid branch if I really need the latest updates. There are almost always Debian images available on a VPS.

If I’m understanding this right, and this basically an API that lets you pick which app store administers an app, that could be quite helpful, not harmful. I currently have fdroid, play store, and Samsung store, and I assume they try to update apps by the fully qualified name, as multiple stores show and try to update a single app instance, sometimes with weird results.

Compression is actually a mathematical field that’s fairly well explored, and this isn’t compression. There are theoretical limits on how much you can compress data, so the data is always somewhere, either in the dictionary or the input. Trained models like these are gigantic, so even if it was perfect recall the ratio still wouldn’t be good. Lossy “compression” is another issue entirely, more of an engineering problem of determining how much data you can throw out while making acceptable compromises.

This is a classic problem for machine learning systems, sometimes called over fitting or memorization. By analogy, it’s the difference between knowing how to do multiplication vs just memorizing the times tables. With enough training data and large enough storage AI can feign higher “intelligence”, and that is demonstrably what’s going on here. It’s a spectrum as well. In theory, nearly identical recall is undesirable, and there are known ways of shifting away from that end of the spectrum. Literal AI 101 content.

Edit: I don’t mean to say that machine learning as a technique has problems, I mean that implementations of machine learning can run into these problems. And no, I wouldn’t describe these as being intelligent any more than a chess algorithm is intelligent. They just have a much more broad problem space and the natural language processing leads us to anthropomorphize it.

Normal people don’t, but when you get into absolutely massive enterprise archiving there’s no rival for the density and cost effectiveness. It sucks for general purpose storage, but for write once, hopefully never read use, they’re ideal.

Forgejo is my go to, I ran it in a GCP micro instance, which has 768 MB ram and a piddling processor. One of my friends works for a company that had all their devs run a local instance in addition to the main repo, it was that light.

Gitea is the former go to, but gitea was hijacked and stolen from the community by a for profit company. Forgejo is currently a drop in replacement fork, but with added privacy features, future federation options, and a reputable parent organization.