astrange

This is an AI bot btw. (sarcasm, metaphor that doesn't make sense)

Is that true? That depends on how their web scraping works, like whether it runs client-side highlighting, strips out HTML tags, etc.

Your brain is doing several different things, because there are different parts of your brain.

(eg different kinds of learning for long-term memory, short-term memory, languages, faces and reflexes.)

They did not leave it out.

> but the amount of cases it would need to cover is too large to be practical (note, this has nothing to do with the impossibility of its design)

It would be more moral to give the LLM a tool call that lets it apply steering to itself. Similar to how you'd prefer to give a person antipsychotics at home rather than put them in a mental hospital.

No, that's how base model pretraining works. Claude's behavior is more based on its constitution and RLVR feedback, because that's the most recent thing that happened to it.

This post is AI slop.

> No, it reduces memory fragmentation, which is why it's called compacting and not compression.

…which reduces memory usage because you don't have to waste it on free holes in the allocated pages.

All major OSes (well Windows and macOS) do in-memory compression before swap, which is cheaper than evicting a file-backed page. But still slow, so you don't want to rely on it.

Compacting reduces memory usage - that's why it's called compacting.

The JVM uses a lot of memory a) because it's tuned for servers and not for low memory usage and b) because Java is a poorly designed language without value types.

There is literally no reason to care how much "virtual memory" you're using - pointers are 64-bit after all.

(Well, 48 bits.)

Exceptions apply on Windows for reasons I forget, and on watchOS because there isn't enough memory for the page tables for really big VM regions.

The 390GB is because of implementation details of GPU drivers. It's not wrong but it also doesn't matter at all.

(The only number that actually means anything there is the first one, but the label for it is basically meaningless. Use `footprint` in Terminal for a better explanation of memory use.)

> It's about the same as talking to yourself, LLMs simply agree with anything you say unless it is directly harmful.

They have world knowledge and are capable of explaining things and doing web searches. That's enough to help. I mean, sometimes people just need answers to questions.

"Poor" in California means earning $80k/year, so they probably are not doing that. Africa / Indonesia / Philippines are better places to find English speaking RLHF workers.

> The LLMs appear to be doing exactly what one would expect them to be doing based on their training corpus.

That is not how full LLM training works. That is how base model pretraining works.

Claudes have lots of empathy. The issue is the opposite - it isn't very good at challenging you and it's not capable of independently verifying you're not bullshitting it or lying about your own situation.

But it's better than talking to yourself or an abuser!

I haven't tried talking to Sonnet much, but Opus 4.6 is very sycophantic. Not in the sense of explicitly always agreeing with you, but its answers strictly conform to the worldview in your questions and don't go outside it or disagree with it.

It _does_ love to explicitly agree with anything it finds in web search though.

(Anthropic tries to fight this by adding a hidden prompt that makes it disagree with you and tell you to go to bed, which doesn't help.)

> It expresses human emotions, by deliberate design.

That is not by deliberate design. It's pretty hard to get them to stop doing it.

Ask Claude Code to write a manual for it.

Danganronpa is a true "adventure game" (which is actually what Japanese VN developers call their VNs…). It's pretty faithful to its genre.

Phoenix Wright is the only one of those Westerners really know about.

Game source code often includes other people's source code (eg middleware) under NDA, and Blizzard is still under contract for protecting that.

> I rip to FLAC for archiving even though 320 or 250+ VBR is probably 'close enough' unless I'm scrutinizing.

MP3 is fundamentally flawed and has audible artifacts no matter what the bitrate is. If you use a newer codec (AAC or Opus) you'll probably not notice anything.

Triggering reads is also how you get pages into the page cache, so it helps to know how to do it.

VNs are not games. They're a kind of ebook.

But only people who are really into computers read them, so they like to use game terminology to talk about them.

(also, none of the creators of "VNs" call them "VNs".)

> Consumer hardware (MacBook Pro, Mac Studio) ships with fast unified memory and NVMe storage, but limited capacity. A 32 GB M1 Max cannot naively load a 40 GB model — the OS will swap-thrash until the OOM killer intervenes.

macOS doesn't have an "OOM killer" in that sense. (It has an out of swap space killer but it's pretty weak.)

So what will happen is, either your memory wiring will fail, or else it will get really slow and panic.

That works for readahead but it's not good for random access. readv, aio, dispatch_io are better there.

> Fumito Ueda was notably quite concerned with the technical/production feasibility of his designs for Shadow of the Colossus. [1]

And he didn't really achieve it - the game runs very slowly and has a good deal of cut content.

(I once got him in trouble because I found a GPL violation in ICO. I assume the developer didn't pursue it because I don't see the source code up anywhere.)

iOS is much more secure than macOS.

It's because the models wouldn't work for coding if they couldn't do nested scopes, so people don't release models unless they work.

They can only do it in a limited form though, because transformer models only have limited "memory". I don't think they can fully implement parsing.