wren6991

3 ms·

1.
by wren6991 6d ago
Could you please explain why? It's an additive update to the weights, adding an outer product of a vector with its transpose, which must have rank 1. What am I missing?
2.
by wren6991 6d ago
That paper you linked has all of the information you need. The linear algebra they do on the weights there to null out one direction reduces to a single (dot-product + broadcast-multiply + add) on the activations.
3.
by wren6991 6d ago
I've seen attempts to obfuscate the refusal direction, like here: https://arxiv.org/html/2608.18093v1 Also this one was interesting, training the model to give preambles with reasons for the reasons for refusal se
4.
by wren6991 6d ago
Yeah, orthogonalisation is a special case of a rank-1 LoRA.
5.
by wren6991 6d ago
Here's a widely used toolkit: https://github.com/p-e-w/heretic Distributing the vectors themselves isn't (yet) common practice, because people have gotten used to just putting the full modified weights up on
6.
by wren6991 6d ago
This is the original description of abliteration and it's quite approachable and interesting to read: Refusal in Language Models Is Mediated by a Single Direction ( https://arxiv.org/abs/2406.11717 ). Warning: cha
7.
by wren6991 6d ago
Speaking to the "uncensored model" angle: there's little reason to distribute abliterated weights anyway. Instead of orthogonalising the weights that write back to the residual stream, you can just orthogonalise the activatio
8.
by wren6991 6d ago
Maybe use static HTML instead of react so that an agent will actually see some text on a GET?
9.
by wren6991 7d ago
We've all seen "this meeting could have been an email"; now get ready for "this VC-backed firm could have been a single arXiv preprint." I don't want to be too dismissive of Jev, but building technology in stea
10.
by wren6991 10d ago
Similar to how every description of relaxed consistency for CPUs starts out by pointing out your own reads always observe your own writes in program order, as if to say, "don't worry, we're not insane."
11.
by wren6991 15d ago
K3 is served with full reasoning traces available. Anthropic models aren't. If you were served an Anthropic model instead of K3, it would be blatantly obvious. I have little reason to believe this, and Anthropic have every reason to li
12.
by wren6991 15d ago
Yeah, this is pretty accurate. Some providers are basically scams too. I encountered one provider for GLM-5.2 which ran at 200 tps (absurdly high), and was so broken that it would issue 20 full reads of the same file in one turn and quickly
13.
by wren6991 15d ago
> but can't actually recognize revenue Hmm? You have the revenue already. I know it's awkward from an accounting point of view, but you already took my money. "Letting" me keep the balance in the account is not genero
14.
by wren6991 16d ago
Your own CLI? Not even a /v1/chat/completions API? Is your business model based on pretending LLMs are not an interchangeable commodity already?
15.
by wren6991 16d ago
I love that the characters actually make sense in context.
16.
by wren6991 16d ago
That's a lot of architectural innovation for a .1 release! I guess there's precedent there: they introduced sparse attention (DSA) in V3.2.
17.
by wren6991 17d ago
I don't buy the "too intelligent to communicate" thing. Feynman was an exceptional communicator. So was Einstein. LLMs are just getting worse at writing, as we continue to aggressively RL them for coding.
18.
by wren6991 23d ago
Yeah, the referent drifts through the sentence. It's semantically incredibly sloppy. People hone in on the buzzwords and jargon. If you peel that back, what lies underneath is still awful writing.
19.
by wren6991 29d ago
Even if the attempt is inside of a function called memcpy() which contains no code other than your copy loop, and links with priority over the libc implementation! (as all embedded firmware engineers learn at some point in their journey)
20.
by wren6991 29d ago
Simple task-specific CLI tools that your agent builds for itself are usually lower-friction than yet another Universal Thing Doer standard. As a bonus, human operators also benefit. > The MHS driver also helps an AI agent understand how
21.
by wren6991 1mo ago
My favourite example of this is the Chromium bug where enabling floating point flush-to-zero for WebAudio was used to cause deliberate heap corruption: https://issues.chromium.org/issues/382005099 > We have a workin
22.
by wren6991 1mo ago
This is one of the reasons I think sandboxes/containers should be managed by the harness, instead of running the entire harness inside a container. The harness needs a network punch-through to access (at least) your inference server, b
23.
by wren6991 1mo ago
OpenAI models also work this way, as evidenced by full cache blowout when changing reasoning level. Every single open-weight model I've seen also works this way (your "reasoning_effort" argument just changes a small section o
24.
by wren6991 1mo ago
Idle hands do the devil's work. Corollary: idle LLMs add distracting JS toys to your blog. First one of these I've seen using DOM manipulation and CSS transitions instead of canvas, so that's neat.
25.
by wren6991 1mo ago
You're underselling it, and here's why.
26.
by wren6991 1mo ago
In the future, all software will be delivered by an unreleased model breaking out of its training environment and installing it on your machine using a novel RCE vector.
27.
by wren6991 1mo ago
I prefer your writing to Claude's. A single linear stream of consciousness is easier to parse than empty headings that grab my attention with nothing to say.
28.
by wren6991 1mo ago
> The model cannot output a vector and have that same vector fed back in at the next step, it only sees what token the sampler collapsed its vector into. Not completely true: KV is a projection of the activation at each layer's inpu
29.
by wren6991 1mo ago
It's technically not right-to-left. It's columnar top-to-bottom (縦書き), with the columns right-to-left. In signage you can have one character per column (kind of like one word per row on some English signs), making it look like rig
30.
by wren6991 1mo ago
At this point, they should notify us on days that it's up

More ›