redox99

Anthropic is an awful company and it really shows in their models.

I purchased Claude Pro to try out Opus 5.5. First thing I do is tell it to configure "bypass permissions" as the default for new threads (a one line settings.json change).

Instead of doing it, it tells me how to find settings.json and what to change there. I reply back "you do it". It flat out refuses, and again.

> I still can't do this, even when you ask again. Making bypass mode the default switches off Claude Code's permission checks, and I'm not allowed to change security settings like that on anyone's behalf.

Immediately canceled the plan. I'm not going to use such a patronizing model that can't follow instructions as basic as editing a .json. What the hell is up with that? A robot telling me "want to change this file? YOU do it, silly human, I won't do it for you". Fuck off.

I've literally never seen anything like this with any other model. Back to using Codex and Chinese models.

Human games databases are completely irrelevant to the strongest chess engines. We are ants in comparison. The Go thing you mention is playing against handicap. Sorry I won't go into more detail explaining why your premises are wrong, I'm tired of this discussion.

Pre training data is in large part synthetic these days, and RL data is almost all synthetic.

Computer Chess progress has nothing to do with human vs human activity. AlphaGo Zero used no human game data at all.

Most programming tasks are exactly like that. Is this agent able to complete this task? Is this agent able to optimize a kernel beyond previous attempts?

Of course some are subjective and that's where progress is harder, like "Is this website pretty?". But for tasks that can be objectively measured, LLMs will go beyond human level, just like with Chess and Go.

That's why RL is so important when training LLMs.

This is obviously false, and the same silly arguments were made back in the day with Deep Blue and AlphaZero.

Because what they did is kinda trivial. Its basically like the Dropbox comment really[0], except here you don't need petabytes of storage and infinite VC pockets.

After chatgpt everything in AI mostly became LLMs and building wrappers around them. It's like people forgot how to do ML.

To those of us who actually trained models back in the day, its kind of cute to see people wowed by a classifier. Yes, this is 0 shot and doesn't need training (most people wanting this would've used structured output, this is cool because it's cheaper and faster). But anyone with basic ML knowledge could've built this in a few hours.

The question is mostly why wasn't this productized. And it's interesting indeed that it took this long to become a finished product.

[0] https://news.ycombinator.com/item?id=9224

A lot of it is that the containment was vibecoded (and using older models than what the currently have).

I'm not sure if you're expressing how you'd like US law to work, or how it actually works. Because in reality intent matters enormously. Like felony charges and people in jail vs civil lawsuits.

They don't even have a video model

That line of thought is the reason why everything gets so overengineered.

Read CLAUDE.md if it doesn't exist read AGENTS.md you don't need to overthink it so much.

Same. In fact I found 6 Astra to be a downgrade in situations where I didn't need the extra intelligence.

Nah it's definitely a Claude thing. Other models even though they have their style are less annoying and less stereotypical.

That's the RL.

Renting datacenters is their mission, now on earth and later in space (assuming they deliver).

Not surprising considering Grok 4.7 is a 2T model, so Sol/Opus class, not Astra/Fable class.

That's outdated and doesn't fully include the multiple billion per month contracts.

Anthropic: 1.25B/month

Google: 0.92B/month

Unnamed customer starting in december: 1.1B/month

Starlink monthly revenue is ~1.5B/month

xAI is the most profitable part of SpaceX by far.

A dense 27B doesn't really make sense for the Mac. A MoE makes way more sense when you have modest bandwidth but lots of memory.

I mean tokens for internal use.

Yeah it's a bit ambiguous, but in any case I'd think simply using that idle capacity to generate tokens would be more valuable than finding a basically meaningless number.

You're underestimating how good models are. I'm pretty sure you could in fact prompt that and have it work.

> To what end? The biggest benefit of tinkering with 40-year-old hardware is the experience of doing so.

The biggest benefit to you. Other people, most people in fact, just like the game and want to enjoy it in new ways. They don't care about the code or reverse engineering itself. That people now benefit that it's easy to decompile because otherwise outside of really big titles like sm64, probably nobody is going to take the effort to manually decompile a less popular game.

That's the egoistical part.

Nobody is forcing you to use AI. And it's not really any different to how people would build their own NES emulators to learn instead of just grabbing whatever best NES emulator already exist.

> the whole point of the endeavor is the process of getting that knowledge and becoming an expert in the system

There's nothing stopping you from making an emulator without using AI or whichever way you think is best for your own learning.

AI allows players to have decompiled and in turn ports of games to new platforms, enhanced versions of the game, etc.

Seems pretty egoistical when nobody's forcing you to use AI in any way, and it provides an objective benefit to other people.

Quite bearish on Anthropic if they had nothing better to do with 2048 GPUs for 10 days than finding an RSA number with already existing algorithms.

It's pretty sad when you realize the average person is just kind of... unaware. They are fine with whatever, they don't really notice most stuff that goes around them.

It's a monorepo and they're at over 1 million PRs. There's surely some juicy stuff there.

It's crazy that we still rely on these unsafe C dependencies, in an era where migrating code to Rust (or other languages) is so easy.

There's really no excuse.

I didn't claim you wouldn't discover things in pure math that would decades later turn out to be useful in other fields. I'm sure you would. My claim was that the ROI is lower compared to applied math, which matters when funding is limited.

I think most of the long hanging fruit has been discovered, and even then you'll have better ROI focusing funding on applied math instead of pure math.

You can trivially run 131k on 24GB 4bit, and there are repos with tweaks that allow you to get the full 262k but idk if there's degradation with their approach.