hgoel

Given your reply, I was looking to read the paper. Apparently this blog post is what passes as their paper for now... I think I'll change my previous judgement and wait for a proper publication. I'm still not convinced by their claims, but maybe an eventual paper will spell it out more formally and convincingly instead of expecting readers to dig through all of their source code.

This seems like exaggeration, the paper seems to describe something more like abliteration, where Astra orthogonalized out activations related to undesired behavior.

This is obviously very different from the surface level claim that Astra can understand the weights of a completely unrelated model with enough detail to improve it in a manner that isn't just the latent space trickery we've been doing for years already.

I think the physical world comparisons should keep in mind the kind of data a lot of these systems (eg postgres) handle. It's more like securing a bank vault than locking a home's door.

AI is the endgame for the joke that computing is the act of etching arcane signs into crystal and imbuing them with lightning.

Haven't bothered with DS4.1 personally but I've seen reports of running it with native weights on 4x DGX Sparks. Currently ~$30k but last month that setup was ~$20k.

Ahh okay, that makes more sense! Sorry for the snark.

Stand out from what? Are you seriously arguing that the only things keeping Signal from being banned are that it somehow camouflages itself?

I've been torn on Proton for a while, but looks like I won't be renewing my subscription...

It seems he's responding to a question asking if he thinks people will be able to run local AI on their own computers. Not arguing on higher prices. So it comes off more as "people shouldn't get to have enough RAM to run local AI because it's an inefficient use of resources" rather than the capitalist "I can afford to pay more for RAM so I should get priority".

The other way to read this is that the US frontier companies don't care about public benchmarks because Western benchmark makers follow them.

Whatever they deem as the next big thing ends up being emphasized in the public benchmarks.

Turns out that even crypto's rules only apply if you're poor.

I was a little suspicious last time, but even moreso this time. The slant their scores take, and the release timing is very suspect. On the other hand, new models drop on a near daily basis, so maybe there's no way to appear entirely fair.

An issue I think all benchmarks suffer from is regarding precision. What does a one point difference mean? How many points of difference is needed to translate into a meaningful distinction between models?

Plus, if they're going to update the benchmark so frequently to discard saturated components, is there even a research-related point to keeping it private?

Out of curiosity do you happen to use a shell other than bash? I've noticed that all models, both local and cloud ones, tend to make much more mistakes if using, say, zsh. It is a frequent source of frustration for me.

Interesting report, qwen3.8-fn does seem pretty prone to hallucinating user inputs and I recall the posts about attempting to access non-existent alibaba cloud urls too. Haven't tried the other models in the gen to know if they're all like that.

To me it feels a bit extreme to jump to nefarious reasoning from this, but I think the overall conclusion is fair, nefarious or not, this would still be something certain users would need to be aware of. I wonder what the solution would be though... besides requiring all network access to be user approved?

I've been pretty satisfied with OpenCode 2

The point of having advisors was supposed to be that they would help the elected person make up for lack of knowledge, but it doesn't seem like that pans out in reality. Either the advisors are just as lost, or they are largely ignored.

System prompt changes help, but ultimately they don't do a great job of overcoming the model's learned tendencies in my experience.

My reasoning is that if something is argumentatively incorrect, it is much easier to drill down to its misunderstanding and correct it On the other hand, if it's sycophantically incorrect, it'll just change its "mind" while potentially maintaining the misunderstanding, leading to increasingly strange behavior.

I think I'm pretty used to my tools making me want to throw them out the window sometimes :) Even before AI that was a common occurrence from either a lack of knowledge on my end, or many software people having no regard for their users' time.

I always appreciate when a model does that. It's amusing, but also having it be argumentatively wrong is a lot less annoying than being sycophantically wrong.

Being hosted in the US is not necessarily enough. I am not allowed to use them even if I run them on my own work-provided computer.

It's 'sustainable' because there's enough regulatory risk to using Chinese models that many will avoid them, and a 3T parameter model is still a bit beyond even many enthusiasts for running locally.

If they made a concerted effort to replace leaders in several industries, they'd likely lose customers faster than they could expand. After Anthropic's report that showed how much detail they were willing to expose about customer activity, and the Navier-Stokes drama with OpenAI, even NVIDIA decided to limit which AI were allowed to be used when working on certain IP.

Plus, as we have seen with their attempts at participating in the hard sciences, these companies lack the level of specialization and focus to even publish their research properly. Their AIs are powerful in verifiable tasks, but just not that good at everything else.

I find this very hard to believe, even though I use plenty of frontier AI in my dev process. I have to live in vscode to get things done because I still frequently need to look at the generated code and point out issues to get the AI past some problem it's stuck at.

Why do we keep having to sacrifice our freedom for the fake safety of the ignorant and their continued freedom to be ignorant?

Why do we have to be lectured on how important this change is for their safety, while the Play store continues to be the primary location scam apps come from?

When do the companies get to be the ones that have to give up their own freedoms in favor of everyone else's safety?

I had some zigbee outlets laying around but wasn't interested in doing the config work to give Claude access to my mqtt broker. I already had most of the kernel written by hand years ago, so PXE etc were irrelevant, the goal was to see how well driver work would go. I was able to send relocatable elf binaries over serial and use the existing modular machinery in my kernel to have it load and run them.

I may have overreacted, but jumping from "we should probably ban this research preemptively because it might pose difficult questions later" to comparisons to Unit 731 when someone disagreed, reminded me of the way Anthropic pushes their belief that Claude is conscious onto everyone with "evidence" of questionable value and even when there are many questions they haven't bothered to answer about such a claim. Plus, they do this despite various features we generally associate with conscious experience that current gen LLMs lack.

Given the complexity of behavior observed in nature that doesn't require conscious experience, yet is more advanced than what we can achieve in circuitry, I think it would be too early to close off this area of research.

It's just an extension of Dario's brand of brainrot where consciousness has an almost religious existence that just pops out if you have a large enough system.

No proof is required beyond psuedoscientific tests and the "saints" of this belief system must be given absolute authority over the system to ensure no one commits any sins.

I built a little harness for an x86 SBC for Claude to do osdev a few months ago.

Janky setup with two ESPs for the serial terminal and to control a servo taped to some microswitches to power cycle the board.

It worked very well, allowing Claude to spend all day and night debugging quirks with the NIC until it was able to get the driver working.

So, I think giving your IoT device some "hands" is probably worth it. For SD card automation, you can find "wireless SD cards" on Amazon/AliExpress that have a microcontroller pretending to be sd card, eg FYSETC SD-Wifi.

Calling it "American civilization" seems wrong too when the president and his bootlickers frequently attempt to supress negative coverage and call for things that are very blatantly against the first amendment, let alone against the spirit of free speech.

With the way the world is nowadays, maybe a fascist hellhole really is more human to the average idiot.