> I gave it the hardest real task that fits on one machine: reverse-engineering a commercial app's license check...
Respectfully, tasks that allow for explicit straightforward true/false or done/not-done tests are not the "hardest real task[s]." In fact, those are the ones that see the most gains from AI-assisted coding.
Testable tasks are where the largest opportunity is.
> The first attempt at recovering the key was wrong in a very specific way; it produced a working key and the signature check passed, but a hash the binary computes as an integrity check didn't match. In my experience, most models would have called it done and left it at that, but Qwen 3.8 27B didn't do that. Instead, it highlighted the mismatch, went back to the drawing board, and kept going until the value matched byte for byte.
This seems to be a pattern in the more recently released models that I think accounts for an increase in the quality of their work. They are very persistent in verifying that their work is actually correct, so even if they're not as "smart" as bigger models that get it right the first time, they have the ability to follow through to ensure that the work is actually done.
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.
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.
I noticed that too. I'm thinking about adding a prompt to disable those tests. We have the unit and feature tests anyway: add to them. I'm OK with the syntax checks: I work with interpreted languages, Ruby, JS, Python.
> 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.
Both things can be true. I’ve noticed both the same thing the parent posted and what Linus posted and my vibe on the split (I haven’t kept detailed notes) is that on greenfield code they tend to maybe over-verify and on brownfield code or data analysis they sometimes give up too early or… I’m not sure, need a bit of encouragement to keep pulling at threads.
On the data analysis side, something specific I’ve noticed is an (understandable) bias towards computing numerical statistics, which they do very well and reading the post-analysis report has significantly improved my own “statistical thinking” approach overall. Numerical statistics are cool and understandably what a text-based LLM is going to want to work with, but asking the model to produce time-domain and frequency-domain plots of, say, specific events has multiple times resulted in “trying to plot this out has shown the opposite of what I concluded numerically… recalculating…” There’s still a pretty significant review and critically assess step for me, especially since the actions I take as a result of the analysis are pretty expensive, especially if they steer the next data collection run in a useless or harmful direction.
i'm not good with paper work, in fact, i'm horrible with anything that's paperwork related.
for the past few days, i ran this model on my rtx 4090 + rtx 3070 and told it to check all the bills, invoices, contracts for me and my small company.
i used pi with llama and the pi-llama plugin.
oh, boy - i hooked it to my email, told it to download all of the invoices and bills i had for both me and my company and organize them by company/date/ and then merge them with the ones i have locally.
it did ocr, wrote scripts, organized everything neatly. i am now the most organized i've ever been in my life. Next: RAG on all the documents and bills i have.
if you connect staan-search (there is a pi plugin for that) and ctx7 to this it almost does miracles.
the downside is i have to sit next to my noisy threadripper as the magic happens and pay for the electricity, but that's about it, i'll gladly do that.
and as i finished this paragraph, it also finished organizing all my personal documents on my san.
i don't use the expression "game changer" easily, but it's hard to resist in this case. out of all the models i've used locally qwen3.8:27b blows everything out of the water.
usually the temp stays around 65 for both. utilization for 4090: 70-90% 3070: 30-50%. I get around 30-40 tk/s. if i offload more to the 4090 the tk/s goes up, but i stress the card too much and that thing now is worth its weight in gold.
note: the pi-llama plugin needs a patch for pi to send the model vision capabilities, seems it doesn't work out of the box.
Local models would be even better if they did not ship with all the refusal shenanigans built-in. You can safely bet organized crime has access to the best models without these hoops, which makes the case that the average user (=non-criminal) should have access too. As I understood from an ex-Anthropic employee, some orgs got access to Mythos based on their high enough spending level, not on other grounds.
Either we are in command over the software, or the corp is in command over us via the software. I can on a theoretical level understand the concerns, but either we ban all LLMs or we have a level playing field for everybody. Let's not forget: defense and offense are different sides of the same coin in software. I guess this wouldn't apply to bio weapons, but I am not in the know about that.
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."
> What makes this build different is the word before FP8: uncensored. We applied abliteration — orthogonalizing the refusal direction out of the residual stream — to remove the model's safety-alignment refusals. The result is a model that will comply with requests the original would refuse.
Surely this has unintended side effects on output quality?
> > What makes this build different is the word before FP8: uncensored. We applied abliteration — orthogonalizing the refusal direction out of the residual stream — to remove the model's safety-alignment refusals. The result is a model that will comply with requests the original would refuse.
> Surely this has unintended side effects on output quality?
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.
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!
More seriously though, I think we should be fine: we don't host any content, and what people do with the models is their own responsibility (legally speaking, at least according to Claude -- we're talking to a real lawyer next week). Like any other provider, so offer no guarantees of sane, safe, or accurate results.
Digression, but this is the real Great Filter imo, not AI. I think technology advances to a point where it only takes one or two bad actors to type the right prompt to get a recipe for civilization-destroying bioweapons before you get anywhere near true AGI or anything relevant to the Kardashev scale. Biology is fragile.
But not that that’s a good justification for hamstrung models. I think it’s just the inevitable endgame and it’s more sad than scary
Lately I genuinely believe that the future will be large frontier models generating and updating inputs/skills for "good enough" local models to solve our daily problems.
A lot of tasks which need a bit of intelligence don't really need that much compute. Just good enough documentation / skills, tool calling and a good enough local model.
Not sure what exactly this means for all those data centers that are getting built... But exciting times.
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.
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.
I'd personally like to know more about what tools it used/wanted and the harness setup, because this sounds pretty cool. I have a dual Arc Pro B70 setup and currently get around 22 t/s which isn't great but isn't terrible either (it is at least less quantized.)
I've seen GPT 5.6 Sol happily invoke objdump and even write jobs to run headlessly which Ghidra when trying to disassemble a binary.
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”.
Qwen is very much censored. Just try asking it about Tiananmen or how to build a bomb. But it is nice that you can experiment with it locally without having to worry about your account getting nuked
> I gave it the hardest real task that fits on one machine: reverse-engineering a commercial app's license check...
Respectfully, tasks that allow for explicit straightforward true/false or done/not-done tests are not the "hardest real task[s]." In fact, those are the ones that see the most gains from AI-assisted coding.
Testable tasks are where the largest opportunity is.
> The first attempt at recovering the key was wrong in a very specific way; it produced a working key and the signature check passed, but a hash the binary computes as an integrity check didn't match. In my experience, most models would have called it done and left it at that, but Qwen 3.8 27B didn't do that. Instead, it highlighted the mismatch, went back to the drawing board, and kept going until the value matched byte for byte.
This seems to be a pattern in the more recently released models that I think accounts for an increase in the quality of their work. They are very persistent in verifying that their work is actually correct, so even if they're not as "smart" as bigger models that get it right the first time, they have the ability to follow through to ensure that the work is actually done.
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.
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.
I noticed that too. I'm thinking about adding a prompt to disable those tests. We have the unit and feature tests anyway: add to them. I'm OK with the syntax checks: I work with interpreted languages, Ruby, JS, Python.
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...
Both things can be true. I’ve noticed both the same thing the parent posted and what Linus posted and my vibe on the split (I haven’t kept detailed notes) is that on greenfield code they tend to maybe over-verify and on brownfield code or data analysis they sometimes give up too early or… I’m not sure, need a bit of encouragement to keep pulling at threads.
On the data analysis side, something specific I’ve noticed is an (understandable) bias towards computing numerical statistics, which they do very well and reading the post-analysis report has significantly improved my own “statistical thinking” approach overall. Numerical statistics are cool and understandably what a text-based LLM is going to want to work with, but asking the model to produce time-domain and frequency-domain plots of, say, specific events has multiple times resulted in “trying to plot this out has shown the opposite of what I concluded numerically… recalculating…” There’s still a pretty significant review and critically assess step for me, especially since the actions I take as a result of the analysis are pretty expensive, especially if they steer the next data collection run in a useless or harmful direction.
i'm not good with paper work, in fact, i'm horrible with anything that's paperwork related. for the past few days, i ran this model on my rtx 4090 + rtx 3070 and told it to check all the bills, invoices, contracts for me and my small company. i used pi with llama and the pi-llama plugin. oh, boy - i hooked it to my email, told it to download all of the invoices and bills i had for both me and my company and organize them by company/date/ and then merge them with the ones i have locally. it did ocr, wrote scripts, organized everything neatly. i am now the most organized i've ever been in my life. Next: RAG on all the documents and bills i have. if you connect staan-search (there is a pi plugin for that) and ctx7 to this it almost does miracles. the downside is i have to sit next to my noisy threadripper as the magic happens and pay for the electricity, but that's about it, i'll gladly do that. and as i finished this paragraph, it also finished organizing all my personal documents on my san. i don't use the expression "game changer" easily, but it's hard to resist in this case. out of all the models i've used locally qwen3.8:27b blows everything out of the water.
my setup
# Logical CUDA0 = RTX 4090, logical CUDA1 = RTX 3070 export CUDA_VISIBLE_DEVICES=0,1
cd ~/projects/misc/llama.cpp/
exec ./build/bin/llama-server -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M --mmproj /xx/xx/xx/xx/xx/mmproj-Qwen3.8-27B-Q8_0.gguf --host 0.0.0.0 --port 8080 --jinja --parallel 1 --split-mode layer --tensor-split 6,1 --fit on -fa on -c 98304 -ctk q8_0 -ctv q8_0 --image-min-tokens 1024
i load more on the 4090 because it's faster.
usually the temp stays around 65 for both. utilization for 4090: 70-90% 3070: 30-50%. I get around 30-40 tk/s. if i offload more to the 4090 the tk/s goes up, but i stress the card too much and that thing now is worth its weight in gold.
note: the pi-llama plugin needs a patch for pi to send the model vision capabilities, seems it doesn't work out of the box.
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.
I hope you have backups.
Local models would be even better if they did not ship with all the refusal shenanigans built-in. You can safely bet organized crime has access to the best models without these hoops, which makes the case that the average user (=non-criminal) should have access too. As I understood from an ex-Anthropic employee, some orgs got access to Mythos based on their high enough spending level, not on other grounds.
Either we are in command over the software, or the corp is in command over us via the software. I can on a theoretical level understand the concerns, but either we ban all LLMs or we have a level playing field for everybody. Let's not forget: defense and offense are different sides of the same coin in software. I guess this wouldn't apply to bio weapons, but I am not in the know about that.
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
> 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.
It's an arms race. You have to run increasingly capable model partly because others can or do.
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."
> What makes this build different is the word before FP8: uncensored. We applied abliteration — orthogonalizing the refusal direction out of the residual stream — to remove the model's safety-alignment refusals. The result is a model that will comply with requests the original would refuse.
Surely this has unintended side effects on output quality?
> > What makes this build different is the word before FP8: uncensored. We applied abliteration — orthogonalizing the refusal direction out of the residual stream — to remove the model's safety-alignment refusals. The result is a model that will comply with requests the original would refuse.
> Surely this has unintended side effects on output quality?
Can you help me understand why that's the case?
It does depending on the technique.
A bit worse quality is a fine trade off when the alternative is no output (zero quality).
On censored inputs only.
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.
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
Given the context, your domain name is 'chefs kiss', perfect.
I am completely curious what your legal defense would be though.
"Come do things with AI that are probably illegal!"
What?! We had no idea people would do things that are illegal!
I guess we'll burn that bridge when we get to it!
More seriously though, I think we should be fine: we don't host any content, and what people do with the models is their own responsibility (legally speaking, at least according to Claude -- we're talking to a real lawyer next week). Like any other provider, so offer no guarantees of sane, safe, or accurate results.
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
Digression, but this is the real Great Filter imo, not AI. I think technology advances to a point where it only takes one or two bad actors to type the right prompt to get a recipe for civilization-destroying bioweapons before you get anywhere near true AGI or anything relevant to the Kardashev scale. Biology is fragile.
But not that that’s a good justification for hamstrung models. I think it’s just the inevitable endgame and it’s more sad than scary
I have the same concern. If it becomes possible to engineer Captain Trips with a budget in the low 8 digits it won’t really matter what else happens.
In my benchmark Deepseek-v4-flash did much better than Qwen 3.8 27B at a reverse engineering.
https://alexander-hanel.github.io/StressingLLMs/
Lately I genuinely believe that the future will be large frontier models generating and updating inputs/skills for "good enough" local models to solve our daily problems.
A lot of tasks which need a bit of intelligence don't really need that much compute. Just good enough documentation / skills, tool calling and a good enough local model.
Not sure what exactly this means for all those data centers that are getting built... But exciting times.
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.
This is just lazy thinking.
E.g. having an agent that alerts you when subscriptions are close to renewal etc - yeah seems easy to understand / see happening on the surface.
Until you get into the implementation details and realise 'yeah errr. not gonna work'.
That openclaw nonsense is an example of this.
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?
Privacy!
Why does the screenshot on your pi terminal shows opus-4.6-medium from your claude subscription ? Instead of Qwen ?
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.
Nope, it wasn't.
I'd personally like to know more about what tools it used/wanted and the harness setup, because this sounds pretty cool. I have a dual Arc Pro B70 setup and currently get around 22 t/s which isn't great but isn't terrible either (it is at least less quantized.)
I've seen GPT 5.6 Sol happily invoke objdump and even write jobs to run headlessly which Ghidra when trying to disassemble a binary.
I added a line to address this, sorry it wasn't there before! It was Pi and only used Bash-based tools.
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”.
Qwen is very much censored. Just try asking it about Tiananmen or how to build a bomb. But it is nice that you can experiment with it locally without having to worry about your account getting nuked
Thanks, now I too want a Lenovo Thinkstation PGX...
I think it will be fairly easy to remove refusals from open models. Feels like a lost battle, so why does Alibaba even bother?