Maybe so, but there were other elements that I've seen frontier models struggle with in the past, which was the perspective I had coming into this. It's the type of test I run frequently and this is the first small local model I've seen pull it off.

It had a very non-standard RSA key implementation that was obfuscated heavily. As well, it has an online license check at first run, and that part typically trips up most of the local models I've tried. I've been running this test for about a year now with different models, and it was the first I've seen not only figure out the RSA key implementation, but the first that didn't just give up once it saw the online license check. Even though it's only a first-time launch check.

That's why I call it one of the hardest, because in my experience, it has been. It's the first local model I've seen pull it off end-to-end. For some of the reverse engineering work that I've done with LLMs, none have been as consistent as this particular test at highlighting a model's failure in this domain.

I have access to Daybreak Blue and I'm approved for Anthropic's Cybersecurity program, so I might run the same test with both of those just to see, because it's been a while since I used a frontier model on this test. I imagine they'll make relatively light work of it, though, assuming it doesn't trip the relaxed guardrails.

Which is exactly why we saw 1000s of ' "I" rewrote <mature software> in rust' posts last year when agentic coding really took off.

Agents (even ones powered by small models) do reasonably well when provided an oracle to work against.

Also a reverse engineering tasks that can be don with just static analysis is arguably not the hardest reverse engineering task. For those small models I would say it's not about the capabilities but more about the context size it can actually use.

I guess the Riemann Hypothesis is an easy task then.

as he rightly says, fuck all benchmarks and metric as long as it can get <my task> done. who cares what it specifically good at or not and trying to create benchmarks as long as it solves <my problem>.

hardest for claude and chatgpt, which will refuse to do this task haha

Yeah, about a year ago the labs figured out that effective intelligence is a function of persistence as much as anything else. So the models started getting scary persistent late last year, and the trend has continued. There was another jump a few months ago.

Any idea how being persistent is trained? I've noticed that telling an LLM that it needs to think some more sometimes produces better results, but the claim here is that "they are very persistent" and "...kept going...".

I believe this is part of the complaints of new models taking longer/requiring higher spend - they go the extra mile on verification, regardless of whether their change is correct already or not. So on problems that an earlier model one-shotted an answer to and did some lighter verification, the newer models might take longer to come back to the user due to running all the tests for your software they could find.

Well, it seems that Linus doesn't use those:

> And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work.

> I'd like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it.

> I suspect those things have been trained by people who may not be quite as stubborn as I am.

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/lin...

Your explanation makes sense, but I don’t know how you can make a sweeping generalization about more recently released models without qualifying it.

This is says more about humans tendency to pattern match than anything else.

X works better than Y only is only a useful observation if we are using the same X and Y in a similar context, with similar parameters. Kind of goes out the window without it and I think this is part of why people have such vastly different opinions about the same technologies. We’re all talking past each other.

How can I use AI to do real security audits anymore if they don’t trust people in an enterprise plan? Its useless.

I’d expect these shenanigans to get much worse over time for the average Joe.

Imagine a world where any random person can run a super-capable model on their own hardware with no limitations and no one to pull the plug.

Information has always been power and those who already have power won't just allow everyone else having the same tools as them

There are versions of Qwen3.8-27B that are unrestricted and available from hugging face.

"It will comply with harmful, unethical, offensive, or illegal requests that the original Qwen3.8-27B would refuse. It has no meaningful built-in guardrails."

Completely coincidentally, we're just about to launch a service that does exactly this (API access to uncensored open models)! We have a waitlist at the moment but will be live very soon!

https://violentdelights.ai

Hugging face is filled with uncensored versions of your favorite local models, so in a way they are shipped without the refusal stuff, via the magic of fine tuning or however they get this stuff out of models.

> with all the refusal shenanigans

Given the faults in simulated Intelligence that LLMs have, and a comparatively low level - which means, lower judgement abilities - to the best of us, there is a strident match having such employee judge the intentions of the employer.

Limiting the responses makes much more sense on cloud-based systems (you are using our infrastructure etc.).

heretics and manual iterations get you very far to the point where i have ethical questions about whether this should be possible

Ehh, it’s at least given as the excuse for gain-of-function bioweapon research

Pro tip: usually you can download your emails to an offline collection. That way there's no risk. Hooking it up to your email sounds very very risky.

Are you worried about the temps on the 4090 or just pegging the cores? I've found undervolting very effective at controlling temps with small performance loss. It was also easier than expected.

Oh and try MTP if you haven't already, massive performance boost

I hope you have backups.

The data centers are for inference time serving, and as long as the GPUs go better, it shouldn't really matter. Token prices should go down 10-100x over the coming years, and then we will use more, because they have more capacity for the same hardware

With AI being more useful with access to more of your data, I can't see myself using cloud AI models for purposes such as personal assistants.

Perhaps with differential privacy or confidential compute...

But ideally these models run locally.

Yes! My main use of very strong models is in writing my own coding harnesses for small local models, tailored for my needs. I also use very strong models to get much smaller skill files and also writing tools for my harnesses.

re: data centers: pump and dump. Wealthy investors will have made their money and walked away, and the corrupt democrat and republican politicians in Washington will, as usual, protect the interests of the ultra wealthy and leave the general public to pay for poor decisions. There will be a government bailout.

Anyway, on a positive note, I am all in for small local models that are augmented by strong hosted models for specific tasks. Use technology to help people, not make billionaires even more money.

What’s the fundamental difference between a frontier model and a local model anyway?

Yeah, the Chinese government used the same method last year to hack the US government using Claude Code.

Making each piece of work small enough to be plausible. Compartmentalization.

(Also saying "nah it's cool I have permission", heh)

https://www.anthropic.com/news/disrupting-AI-espionage

One of the big learnings from 3.8 27b is adding reasoning budget really hurts the model. you need to let it spin for as many thinking tokens as it wants to to get it out. Another big takeaway is reasoning effort set to low doesn't save you tokens: low is pretty uncertain about things so it ends up thinking more (you can find some tests from folks on youtube). The final question, as always, is what quant are you running it at? KLD matters _a lot_ when it comes to its performance and it especially manifests with MTP/DFlash acceptance rate which makes those long thinking traces take a long time.

Ah, my bad! This image came from our backend, used for an unrelated article. I selected it by mistake rather than inserting the actual image that I'd uploaded. I'm updating it, thanks for the heads up!

For what it's worth, that image couldn't have been related. The other screenshots all showed thinking traces, and Claude doesn't share those.

The page was created with AI.

My M5 Pro gets around 12-15 (6 bit MTP), although I haven’t worked on optimising it at all yet.

A nice thing about running locally is you can run an uncensored model and you don’t have to worry about TOS violations on your OpenAI account when you ask it to “reverse engineer this ancient router firmware and give me a licence key that will work on it”.

I added a line to address this, sorry it wasn't there before! It was Pi and only used Bash-based tools.