For copyright purposes, all patches must be substantially human-written. However, it seems that most patches are now prepared with chatbot assistance, so that effectively means that people are required to be psychotic centaurs in order to keep up with review demand. There’s an LWN secret link from the documentation lieutenant which summarizes the linked LKML thread. My former lieutenant, in the Direct Rendering Manager (DRM) subsystem, is apparently having a bad time:

Airlie asked his sub-maintainers; most of them identified him as the biggest pain point. More seriously, it was that the DRM subsystem is still seeing too many regressions. There is a push within the subsystem to move development to GitLab in order to use its CI features; he is not standing in the way because he wants to see how it turns out.

This is a more systemic problem than one might expect, because DRM (and DRI, Direct Rendering Infrastructure) are partially shared between Linux and the BSDs. This subsystem has creaked along under the weight of corporate expectations for decades, mostly fueled by a kind of quiet sponsorship of young university folks from GPU vendors along with outsourced drivers produced behind corporate privacy and thrown over the wall. Every vendor, even our eternal nemesis nVidia, has made repeated concessions in order to try to figure out the maintainer burden, but it seems that we are still dependent on Airlie’s stoutness and willingness to get shouted at by Torvalds.

@lagrangeinterpolator :thisisfine: :trashfire: @BlueMonday1984

I haven’t really kept up, but it looks like Character’s business model was still focused on NSFW content. If so, then this is also a disaster for the industry at large, since NSFW content is one of the few genuine money-printers for chatbot and image-generation services; people really do want thousands of iterations of the same fantasy, slightly changed in order to stay engaging, and Character was highly recommended as a source of affordable slop. (Examples of unaffordability include the rampant free-tier abuse on ChatGPT, Gemini, and Grok; I’m assuming that customers are willing to open their wallets at least a little. In general, the porn industry turns a profit, so I do think that people will pay for porn.)

My prior estimate (previously, on Awful and on Lobsters) was that token generation was subsidized to the tune of an effective 80-90% discount; this looks to be more in the 90-95% range. For contrast, consider chatbot-booster Willison’s belief, in reply above, that Chinese labs will magically produce a five-fold improvement in token production; this still would be between two to four times slower than needed to turn a profit for Character.

How you look at “my company pays $14 for every $ from the user” and not like, idk, stop, seek help, change your job, name, and address, is completely beyond me.

What type of documentation are you looking for? The big picture is easily googleable and wikipediable: EY got a few of the super-rich excited enough about doomer drama to bankroll the founding of OpenAI, whose governing body was mostly EAs (until the Altman-Nadella counter-coup) while Anthropic were the people who bailed because they though OAI weren’t serious enough about alignment, and suddenly getting dicked in a rationalist compound group house has immense career implications in the burgeoning AI sector.

Also terms like AGI, ASI, AI alignment, existential risk et al started out as lesswrong shibboleths, and so-called alignment research is almost completely a rationalist circlejerk.

Pinning the EA/rats down as a cult is where it gets murkier because it’s mostly them telling on themselves in social media and blog post implications, like how they foster a culture of zero personal boundaries, ingroup overdependence and rigid hierarchies stratified around claimed IQ and proximity to the Great Work of not messing up the robot god incantation that ends the world.

You could do worse than the Behind the Bastards episode on the Zizians, mostly because unlike Leverage or MAPLE who kind of did their own thing but with rationalist funding and human capital, the Zizians can be fairly described as what happens if you actually take the more esoteric aspects of rationalism seriously.

AI safety is mostly a sex cult

the guy who posted this is also a singularity fanatic, but this is a good enough summary

There is always RationalWiki but its not really equipped to deal with the polycule and abuse of sexual power stuff (or all the giant institutiions taking Internet cranks seriously). https://rationalwiki.org/wiki/Main_Page

I was just listening to an interview with Emile Torres, it’s an hour long but it is on the topic you’re asking about: https://www.youtube.com/watch?v=jt508VjX2H8

I think it was modern warfare 2 but they redacted all explicit references because activision or whoever was starting to make annoyed legal noises.

The insistence on using number of functions copied as a metric is weird, like of course a port isn’t going to be a function-to-function translation, programming languages are never a 1-1 match with each other that’s kind of the point of having more than one.

What we needed was an automated check that tells agents whether a reconstructed function matches the original. It should verify identical semantics. A simple PASS or FAIL signal would be enough and the agent can figure out on its own what’s wrong.

Uh, sure.

Also a lot of their recommendations come of as no-duh, like pressing the compact conversation button early to reduce tokens or using a common specification document across LLM instances to keep them more coordinated, how novel, are you kidding me.

Another one of their recommendations is that fixing stuff causes garbage to accumulate, so if something isn’t working out you are better off starting from scratch, because tokens are cheap now! You just need to be doing the equivalent of maxing out dad’s credit cards by using multiple private subscriptions in tandem, and I guess also make sure to time your dev cycle to coincide with the two major LLM vendors having a money burning competition on their way to an IPO.

Finally there’s a whole hypecycle going on currently that this feeds into, that makes headline readers think that LLMs can convert binaries right back into code now, when they are actually using an MCP to control existing decompilation software which is so far from the same thing.

“Yes I did it. No you can’t see the work”

I too have seen the documentary “Maximum Overdrive”

I wonder what answers you would get if you polled these Anthropic people about the Covid lab leak theory

It’s funny how they’re twisting themselves into pretzels to make this work. They started by casually saying that AI could copy itself onto the internet, not anticipating that people would immediately point out just how much hardware it takes to run their AI. Seems like they really want to die on this hill.

MalwareTech/Marcus Hutchins has an excellent take on this that he posted for the normies on linkedin

Posted verbatim below to save you from having to read LinkedIn

Anthropic really won’t let this botnet thing go. It’s so desperate & ridiculous.

1. 9B parameter models are so weak that they’re essentially useless. There is no reason at all for anyone or anything to be self-replicating them.

2. Even if you quantize a 9B parameter model down to the smallest possible size (at which point it becomes less than useless), and you run it on regular RAM instead of VRAM (at which point it also becomes painfully slow) it still requires about 5 GB of RAM, which is way more than most IoT device has.

3. Mirai is already self-propagating across millions of IoT devices without the help of any AI. It’s simply not needed, and it adds unnecessary complexity.

4. Many IoT devices are ephemeral and their entire OS is reset on reboot, so any code running on the device is gone after reboot. The AI models would also be competing with malware like Mirai, which propagates exponentially faster, and locks down the devices so other malware can’t get in.

5. Again, why? What is the purpose of this. Anthropic keeps pushing this ridiculous idea but have yet to establish any reason for why this would actually happen, or who would benefit from it.

6. Any remotely intelligent threat actor would just use a powerful unquantized models to reverse engineer & find vulnerabilities in obscure devices and write normal (AI-free) self-propagating code. Ironically, they typically do this by stealing Anthropic API keys and using Anthropic’s models, not open-weight ones.

7. Again, why are we self-propagating open-weight models? There is absolutely no reason to do that. The combined processing power of every IoT device on earth would be less than a moderately sized stolen cloud compute GPU instance.

8. I’m pretty sure Anthropic keeps pushing these ridiculous narratives in an attempt to paint open models as dangerous, thus making the case for banning them. They have been OBSESSED with GLM lately, likely because it’s eating their overpriced lunch. This research is just so deeply unserious, that it’s impossible to believe it’s being done in good faith.

Self-replicating prompt-injection text is a much more interesting and much more immanent threat to cause AI systems to misbehave in a propagating way

rats’ best thought out attempt at immanentizing eschaton

deleted by creator

thank god. horrid, parasocial site

Couldn’t happen to a nicer bunch of people. Shame they’re probably all made and siphoned off life-changing amounts of money before this happened. I guess we can always hope they take up smoking because AI is going to cure cancer.

Also notable for being the first to get a kid killed by AI suicide encouragement. Burn it to the ground and make teeth fillings from the slag.

They looking to get off the hook before all the class action lawsuits pile up?

This is sad news, as companies are not alive, this means there is no new public toilet.

deleted by creator

As someone who isn’t totally uninterested in animal welfare the shrimp argument always makes me generally grumpy. Like, I’ll grant that it’s better to use a stunner than to not and that it’s probably a good use of $15, but the whole problem with EAs is that they have significantly more than $15 at their disposal. Like, if you just do the naive multiplication you’ll end up missing out on two things. On the one side opportunities that require a lot more than $15 to have a positive impact that are going otherwise underserved. On the other, more important side, their emphasis on economic efficiency leaves them blind to the completely avoidable harms that are inflicted in the accumulation of that much money. That’s how grifters like SBF work their way in, and why the whole enterprise seems to be earning a reputation as a way to make tech billionaires feel better about themselves rather than actually making the world a better place.

charming.