wren6991

The desktop build seems to enable telemetry by default. Regular `dsh web` only has telemetry for explicit user feedback. Hmm :/

If like me you're mildly bothered by this, add these lines to $DSH_HOME/cordis.patch.yml before first startup ($DSH_HOME is ~/.dsh by default):

    - id: desktop-product-telemetry
      disabled: true
    - id: product-analytics
      disabled: true
    - id: session-log-deepseek
      config:
        enabled: false

On-topic: I like DeepSeek Harness quite a bit, but the problem with "Everything Is A Plugin" is that when this includes core functionality, you still have to maintain downstream patches for those core plugins if you want to tweak existing behaviour. I currently have ~25 downstream commits and 0 new plugins.

Kimi K3 is an example of a modern LLM that doesn't use positional encodings. It uses NoPE'd MLA for global attention and a variant of Gated DeltaNet (KDA) for local attention. I wonder how much the KDA would degrade if you just did a bounded replay over the last few thousand tokens to recover its approximate state when assembling your context window from chunks of known MLA KV.

DeepSeek is the immortal whale

Right, because MCP != web. In fact most of my exposure to MCP has been local tools. Also, for context, we're talking about a coding agent harness that runs on a local machine, and Codemode applies to all of its tools, not just MCP.

The doublethink required to simultaneously believe "our safeguards prevent our models from doing unsanctioned cybersecurity tasks" and "distillation is why Chinese models are getting better at cybersecurity tasks" is genuinely quite funny.

Yeah, I'm probably over-indexing on local use cases due to my own preferences, prejudices, biases etc. For a coding agent like Pi it does seem reasonable to expect some kind of shell access though, unless some people are using it as a CLI chat client with MCP?

Very easy to optimise the fun out of Factorio. I have a hard rule: no blueprints carried forward between saves. It's more fun for me to figure out what I can design quickly that kinda works and kinda supports scaling. A lot of the time I'll just build something as a repeatable unit cell and then use copy/paste, skipping the blueprints. I think it's worth trying this out.

> And while we could have just wired up the metadata to enable better MCP extensions, we also think that MCP with Codemode solves quite a few of the issues that it traditionally had.

There's just something that bothers me about this. Normally if LLMs want to compose multiple operations, they have the perfect tool for this: bash, or whatever other OS shell is available. It's why I was always confused by Codemode-type constructs for direct chaining of tool calls; see also the way highly-RL'd modern models will fall back to sed or python for complex file edits.

It seems like Codemode is raised here as the perfect tool for chaining or composing MCPs, but isn't that backwards? LLMs are already given the perfect tool for that, and the problem is that MCPs aren't exposed to that tool.

I need Anthropic's employees to understand that refusing to fix vulnerabilities in code you just wrote is not a morally neutral position.

It's just vibe versioning, right? Fable 5 is a beloved product, it gets a .1 bump to feel close. Opus 5 and Sonnet 5 had a mixed reception, they get a .5 bump to create a sense of distance.

After what DeepSeek pulled with V4.1 Flash I've given up on trying to map LLM versions to semver.

Hey now, if programming is solved they still need to achieve vendor lock-in somehow. Won't somebody please think of the vendors?

Does Claude Code not preempt block-on-output when you send a steering message? This was one of the first things I fixed in my DeepSeek Harness fork

> I am told this is merely something in the system prompt that the model tends not to pay attention to with large contexts.

IIRC it's a system reminder injected after every single turn.

It must be pretty ingrained to be so resilient against prompting. I think RL on relatively short-horizon programming tasks has given the model a tendency to write down absolutely everything, so it survives compaction. Longer-term (project-scale) tasks where this crap starts to pile up and cause problems are in the evolutionary shadow, so to speak.

I'm honestly not sure. Often when companies lobby for regulation it's to avoid other types of regulation they would find more troublesome. I don't buy the popular narrative that Meta is lobbying for verification laws for the purpose of getting ID data because surely they have a lot of signal there already. At the same time I think it's backwards to say "Meta are behaving this way because of regulation" when they actively lobbied for it.

I think the causality is a bit mixed up here given Meta have put millions of dollars into child safety lobbying

I don't think reading disassembly is actually that weird. I spend a lot of time doing it at work.

I think this is a known phenomenon: even in non-reasoning models, adding useless/filler tokens before an answer improves task performance. The model is doing some computation during the filler. See: https://arxiv.org/html/2404.15758v1

Thanks, this is interesting and there were multiple things I didn't know here.

There's an implication that other companies are improving because they're scraping Anthropic, not because they're investing in better architecture, compute efficiency, or their own synthetic data pipelines. I often see Chinese labs' progress dismissed as "they just distilled Anthropic" and I find it hard to reconcile that with all of the interesting research and open-source tooling that they release.

Is there actually that much capability transfer from non-logit-matched distillation, or is Anthropic just another unwilling source of data?

It's the opposite: pacing here means "slow down" while trying to avoid the negative affect.

> also, what "if you're reviewing this rather than just reading it" even means?

It means "I'm treating you as lay-person punter, not a developer working on this project." Opus 5 feels like it's constantly trying to reward-hack me into treating it as intellectually honest and epistemically humble, while in the same breath it talks down to me and tries to smuggle its own bullshit assumptions and assertions into the conversation unchallenged. No progress on this front apparently. Glad I cancelled.

I'm confused how they have been able to create so much public negative perception around distillation. It seems pretty clear that they are the only ones who lose out, and everyone else benefits. I don't have any ethical issues with it, nor is it illegal: at worst it's a ToS violation.

IMO the biggest problem with distillation is that not enough people are openly doing it. I would love to see more small, competitive US labs instead of having the eggs in 2~4 baskets (depending on how you count).

Smart move would be to move to year-based versioning (26.09). A 4x advantage

Yeah, I agree it's probably fixable, but I think a naive interpretation of the metadata is probably the source of this bug (which HF has had for as long as I can remember)

Packed 4-bit weights are often identified as u8 byte arrays in the safetensors metadata, so the HF UI counts them as half a weight each.

Granted this is an awesome release and I loved watching the livestreamed RL dashboard, I found this message on the dashboard (https://mimo.xiaomi.com/rl/) quite funny:

> we also removed the cyber dataset from the upcoming pro run, since we observed some bad patterns in the rollout logs.

And this in the model card (emphasis mine):

> Aligned RL: Cold start from self-correction — the model reflects on and rewrites its own misaligned turns into grounded next steps. Throughout RL, environment hardening, adversarial screening, and verifier cross-checks keep the loop honest against reward hacking.

Environment hardening during the RL runs. Uh oh, did someone start making a few too many paperclips?

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?

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.

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 seems to make it less sensitive to modulating the single refusal direction: https://arxiv.org/html/2505.19056v1

My empirical observation is that when a new model is released on HuggingFace, an abliterated version with < 10/100 refusals (baseline usually 100/100) is uploaded the same day, so either these techniques don't work very well or the open-weight labs aren't applying them.

There's some defense-in-depth, like a lot of the "guardrails" people hit on cloud models are classifiers applied to prompt or output, not a refusal generated by the model. Also closed-weight models obviously try to avoid this by not letting you see or modify the weights.

Yeah, orthogonalisation is a special case of a rank-1 LoRA.