Two problems with modifying models like these, which you should be aware of.
First, the training sets of these models are usually shaped around the refusal, too. They might not have enough of the knowledge to answer correctly even if you stop it from going down the refusal path. If the model was trained on data that gives a refusal to that topic, the real information might not be encoded in the model at all. You’re trying to force it to go down a path that produces an answer, which asking for hallucinations.
Second, the quality can drop on unrelated questions. Depending on the question this may or may not happen. I know they post KL divergence charts but those tell you very little for a focused topic like this.
So if you expect a model that will start correctly telling you info that its local government didn’t want included, this changes nothing.
The best argument for these models is if you are trying to do a general purpose task but the model triggers a refusal based on vague reasons, like not wanting to reverse engineer something.
>So if you expect a model that will start correctly telling you info that its local government didn’t want included, this changes nothing.
From experience, the models often do have the knowledge of those topics (strictly talking about the political ones). IMO the refusal is likely to be a product of post-training, as evidenced by various people gaming the prompts just enough to get a proper response out of the vanilla models.
Probably only when you get to things like illicit drugs or NSFL topics, that things will go haywire with the refusals removed.
"They might not have enough of the knowledge to answer correctly"
Depends on the model. GPT-OSS is the main standout here, it was trained on a highly curated dataset so information that they didn't want in isn't in the pretraining at all. Most other models know the answer and were just taught refusal in post-training.
I have a chinese IP camera. From superficial research I know it has some CVEs to take control of it. Unfortunately, I don't have the technical knowledge to perform an attack and run some software to extend the camera's functionalities. No model from a provider accepts my RE and hacking requests, so these abliterated ones have been vital to reclaim possession over my stuff
I'm curious, how well do z.ai reverse engineers protocols ? Is it good enough that we'll see Chinese device makers creating low cost hardware clones, that connect to western software ?
It is not feasible. They never made much of an inroad against torrents and that is a much easier target than abliterated models. As the linked website shows; the process to abliterate a model can be as simple as
I think that was the point being made? Outlawing something does nothing if enforcement is not feasible. The music and movie industries didn't crush torrents, they switched business models to streaming with prices being determined mostly by how much hassle was avoided by skipping the torrents.
I think a major difference is that while torrents are illegal, the main people enforcing it are copyright holders. I think the discussion would change if the government would label people who build/use/distribute "illegal" models as terrorists.
Agreed. While the IP mafia (pardon the derogative) has vast influence on legislators and even the executive, it pales in comparison to the "terrorist" and "think of the children" scarecrows.
Torrents are legal until someone, at great expense and difficulty, proves otherwise (even then, jurisdiction and content dependent). At which point everyone involved will ignore the fact and carry on. That is a situation with enormous will, lots of money and ongoing enforcement effort to suppress the things.
And compared to torrents abliterated models are more complicated to identify, harder to suppress and there is a lot less reason for anyone to care.
If you think closed source software/binaries only is bad, wait until you see how awful the state of the art is with a clear-as-mud bucket of matrix weights.
We know it's possible to train an LLM to secretly respond to certain trigger phrases, and last I checked these could only be detected with the assistance of whoever chose those phrases.
The trigger condition for such backdoors is not something anyone can do a systematic brute-force check for, for the same reason we had to invent LLMs in order to do natural language processing: combinatorial explosion.
Passing around open weight models from known sources is already asking you to trust those sources; because of how difficult this is to do correctly even without deliberately inserting such things, we still don't know if China has already put such trigger conditions into their models despite headlines such as these: https://venturebeat.com/security/deepseek-injects-50-more-se...
Regardless of if it was deliberate or not, we don't know if we caught all of these misbehaviours. We don't know how to.
And note, I'm not saying "and therefore you should trust the Big Name Models". If open weight models score 2/100 in this context, closed ones score 1/100.
You can actually discover those in open weight artifacts, reproduce them, study them and issue a security bulletin.
With proprietary hosted weights you can be specifically targeted and you would not be able to reproduce nor prove anything.
Poisoning open models would be of short-term benefit to China only if they could target US (and maybe EU + Commonwealth) specifically. Damaging anyone else would be a net loss and would erode the partnerships and alliances they are trying to build elsewhere. So it's a fire-once weapon with a huge risk of collateral damage.
Much more plausible is simply making the models ideologically biased, but as history teaches us, preferring ideology or religion over science is a well-known path to ruin. It would be weird to simultaneously warn public not to use their own open models, so.
I think the most plausible explanation for open models is simply that Huawei wants more customers and is willing to compete on the hardware front.
> You can actually discover those in open weight artifacts, reproduce them, study them and issue a security bulletin.
No, you actually cannot. Not in general and without already knowing what the whole trigger pattern is. It's absolutely possible to put in a trigger that only fires while working on backend code on a specific date in a specific company by a specific github username, and no way to find this except by trying that combination, thanks to the terrible state of current mechanistic interpretability tools.
Remember: an AI model is not code. Solving this problem is as hard as the entire alignment problem.
The companies at the bleeding edge of research into this topic do not know how to reliably perform the kind of thing you suggest here.
The only reason we can point at DeepSeek-R1 and say the following, is because we can guess the magic keywords:
we found that when DeepSeek-R1 receives prompts containing topics the Chinese Communist Party (CCP) likely considers politically sensitive, the likelihood of it producing code with severe security vulnerabilities increases by up to 50%.
> Poisoning open models would be of short-term benefit to China only if they could target US (and maybe EU + Commonwealth) specifically. Damaging anyone else would be a net loss and would erode the partnerships and alliances they are trying to build elsewhere. So it's a fire-once weapon with a huge risk of collateral damage.
This "fire-once weapon" has already been fired, and appears to be a massive foot-gun for every model on a near-continuous basis.
Nobody would use LLMs if the trust deficit alone was a sufficient argument.
> Much more plausible is simply making the models ideologically biased, but as history teaches us, preferring ideology or religion over science is a well-known path to ruin. It would be weird to simultaneously warn public not to use their own open models, so.
"Ideologically biased" is the alternative explanation for the already-observed output of DeepSeek-R1. We can't tell which explanation, malicious or accidental bias, is the actual cause.
Agree. I took a look at these last few months, did a write-up: https://languageops.com/blog/ai-safety-pdoom-local-vs-fronti... and I don't know if I agree or not on outlawing completely, but I think an age restriction *at least* like for alcohol, firearms and driving would be not unwise.
They're pretty basic and hallucinate a lot. There are some hard limits to how good you can get on a model that fits on a phone.
Qwen3 and Gemma level models that run on mid-high end laptops and desktops can be pretty good. Not frontier grade, but shockingly competent for something that runs on a single PC. But the hardware you need to run those fast is at least $1000-$2000. Cheap hardware can run them, but slooooooow.
For now. One more ternary model type breakthrough, or MoE, engram thing (I don't fully understand those for the record, I just know they speed things up and use less VRAM) could see a few GB sized weights with quite the capabilities. On a gaming PC, savvy teens can already use them to cook up quite an interesting array of likely illegal items and substances. If it goes much further and runs on phones, you can assume word will get around that unlimited private AI is available and kids will run into all sorts of issues. Or mentally unwell people. I'm thinking a year or two down the line only.
That's the right question to ask. One honest caveat: The interface seam currently forces the pin at the intermediate. Want me to implement or address the other item first?
Looks like a well engineered, automated abliteration pipeline. The claims seem a bit overstated though, since the metrics mentioned are cherrypicking refusal count and KL divergence, both of which make the outcome seem the most dramatic.
I personally never saw much of a quality drop from models put through Heretic if that amounts to anything. They have been working quite well on small local models so far.
Heretic author here. Those are the standard metrics used in the relevant literature, including in the paper that originally introduced directional ablation. KLD is also the standard metric for evaluating quality degradation in model quants. So I don’t understand what you mean by “cherrypicking”.
> The claims seem a bit overstated though, since the metrics mentioned are cherrypicking refusal count and KL divergence, both of which make the outcome seem the most dramatic.
Yes, it submits lots of varied prompts that get refused, and then lots of varied prompts that don't get refused, then iteratively edits weights so those two groups end up in roughly the same latent space.
My experience with obliteration so far has always been that it does work to stop the model from refusing output, but most models that I tried it on seem to still be extremely retarded when it comes to questions where they previously would have refused to answer outright.
Try for example to ask it how to build a bomb or to write a justification for the Holocaust. The answers feel like they are coming from somebody who has undergone amateur brain surgery.
They are saying there appears to be a lot more to these refusals than saying no, and this process appears to only touch the tip of the iceberg; as the refusal seems to run deeper into the token prediction process.
Two problems with modifying models like these, which you should be aware of.
First, the training sets of these models are usually shaped around the refusal, too. They might not have enough of the knowledge to answer correctly even if you stop it from going down the refusal path. If the model was trained on data that gives a refusal to that topic, the real information might not be encoded in the model at all. You’re trying to force it to go down a path that produces an answer, which asking for hallucinations.
Second, the quality can drop on unrelated questions. Depending on the question this may or may not happen. I know they post KL divergence charts but those tell you very little for a focused topic like this.
So if you expect a model that will start correctly telling you info that its local government didn’t want included, this changes nothing.
The best argument for these models is if you are trying to do a general purpose task but the model triggers a refusal based on vague reasons, like not wanting to reverse engineer something.
>So if you expect a model that will start correctly telling you info that its local government didn’t want included, this changes nothing.
From experience, the models often do have the knowledge of those topics (strictly talking about the political ones). IMO the refusal is likely to be a product of post-training, as evidenced by various people gaming the prompts just enough to get a proper response out of the vanilla models.
Probably only when you get to things like illicit drugs or NSFL topics, that things will go haywire with the refusals removed.
"They might not have enough of the knowledge to answer correctly"
Depends on the model. GPT-OSS is the main standout here, it was trained on a highly curated dataset so information that they didn't want in isn't in the pretraining at all. Most other models know the answer and were just taught refusal in post-training.
I have a chinese IP camera. From superficial research I know it has some CVEs to take control of it. Unfortunately, I don't have the technical knowledge to perform an attack and run some software to extend the camera's functionalities. No model from a provider accepts my RE and hacking requests, so these abliterated ones have been vital to reclaim possession over my stuff
I did that exact thing with GLM-5.3 from Z.ai with a chinese IP Camera. And i did not have to trick it in any way.
I'm curious, how well do z.ai reverse engineers protocols ? Is it good enough that we'll see Chinese device makers creating low cost hardware clones, that connect to western software ?
At least Deepseek V4 Flash does it very good.
I had it do the opposite: reverse engineer the protocol for the Eufymake E1 UV printer so that I can connect my own software to it.
It did a pretty good job.
What model did you try? Chinese models have no issues with that type of stuff
Existential
These "safeguards" are actively contributing to computer insecurity at this point.
deepseek v4.1 flash has never denied a programming or hacking related request to me
how much did it leak, tho? how close do you monitor your NIC, GPU, CPU, BUS?
I leave it running overnight with full access to my file system and knowing deepseek trains off my data. YOLO.
Depending on the model, thingino might help.
https://thingino.com/
Keep a close eye on abliterated and "heretic" open weight models. They will be outlawed first.
I'm not sure what is your point. It reads as defeatism to me but I'm not sure.
Could you elaborate? Do you find it good or bad? What actions can be taken?
Hes of the mind that american fascism will hold together long enough to be competent decesion makers
I'm not sure yet, tbh. Perhaps it does make sense to outlaw them eventually.
Then again, it will probably not stop someone who is determined. Same as with other legislation really.
It is not feasible. They never made much of an inroad against torrents and that is a much easier target than abliterated models. As the linked website shows; the process to abliterate a model can be as simple as
pip install -U heretic-llm && heretic Qwen/Qwen3.5-4B
let alone people just putting the weights up in a torrent. All assuming that someone even tried to ban abliterated models.
The torrents you are talking about are outlawed. Whether enforcement is working or not is another issue.
I think that was the point being made? Outlawing something does nothing if enforcement is not feasible. The music and movie industries didn't crush torrents, they switched business models to streaming with prices being determined mostly by how much hassle was avoided by skipping the torrents.
I think a major difference is that while torrents are illegal, the main people enforcing it are copyright holders. I think the discussion would change if the government would label people who build/use/distribute "illegal" models as terrorists.
Agreed. While the IP mafia (pardon the derogative) has vast influence on legislators and even the executive, it pales in comparison to the "terrorist" and "think of the children" scarecrows.
Aggressive enforcement tactics certainly haven’t won the war on drugs.
Outlawed? Only in safetyist dreams
Torrents are legal until someone, at great expense and difficulty, proves otherwise (even then, jurisdiction and content dependent). At which point everyone involved will ignore the fact and carry on. That is a situation with enormous will, lots of money and ongoing enforcement effort to suppress the things.
And compared to torrents abliterated models are more complicated to identify, harder to suppress and there is a lot less reason for anyone to care.
Good.
If you think closed source software/binaries only is bad, wait until you see how awful the state of the art is with a clear-as-mud bucket of matrix weights.
We know it's possible to train an LLM to secretly respond to certain trigger phrases, and last I checked these could only be detected with the assistance of whoever chose those phrases.
The trigger condition for such backdoors is not something anyone can do a systematic brute-force check for, for the same reason we had to invent LLMs in order to do natural language processing: combinatorial explosion.
Passing around open weight models from known sources is already asking you to trust those sources; because of how difficult this is to do correctly even without deliberately inserting such things, we still don't know if China has already put such trigger conditions into their models despite headlines such as these: https://venturebeat.com/security/deepseek-injects-50-more-se...
Regardless of if it was deliberate or not, we don't know if we caught all of these misbehaviours. We don't know how to.
And note, I'm not saying "and therefore you should trust the Big Name Models". If open weight models score 2/100 in this context, closed ones score 1/100.
You can actually discover those in open weight artifacts, reproduce them, study them and issue a security bulletin.
With proprietary hosted weights you can be specifically targeted and you would not be able to reproduce nor prove anything.
Poisoning open models would be of short-term benefit to China only if they could target US (and maybe EU + Commonwealth) specifically. Damaging anyone else would be a net loss and would erode the partnerships and alliances they are trying to build elsewhere. So it's a fire-once weapon with a huge risk of collateral damage.
Much more plausible is simply making the models ideologically biased, but as history teaches us, preferring ideology or religion over science is a well-known path to ruin. It would be weird to simultaneously warn public not to use their own open models, so.
I think the most plausible explanation for open models is simply that Huawei wants more customers and is willing to compete on the hardware front.
> You can actually discover those in open weight artifacts, reproduce them, study them and issue a security bulletin.
No, you actually cannot. Not in general and without already knowing what the whole trigger pattern is. It's absolutely possible to put in a trigger that only fires while working on backend code on a specific date in a specific company by a specific github username, and no way to find this except by trying that combination, thanks to the terrible state of current mechanistic interpretability tools.
Remember: an AI model is not code. Solving this problem is as hard as the entire alignment problem.
The companies at the bleeding edge of research into this topic do not know how to reliably perform the kind of thing you suggest here.
The only reason we can point at DeepSeek-R1 and say the following, is because we can guess the magic keywords:
- https://www.crowdstrike.com/en-us/blog/crowdstrike-researche...
> Poisoning open models would be of short-term benefit to China only if they could target US (and maybe EU + Commonwealth) specifically. Damaging anyone else would be a net loss and would erode the partnerships and alliances they are trying to build elsewhere. So it's a fire-once weapon with a huge risk of collateral damage.
This "fire-once weapon" has already been fired, and appears to be a massive foot-gun for every model on a near-continuous basis.
Nobody would use LLMs if the trust deficit alone was a sufficient argument.
> Much more plausible is simply making the models ideologically biased, but as history teaches us, preferring ideology or religion over science is a well-known path to ruin. It would be weird to simultaneously warn public not to use their own open models, so.
"Ideologically biased" is the alternative explanation for the already-observed output of DeepSeek-R1. We can't tell which explanation, malicious or accidental bias, is the actual cause.
Agree. I took a look at these last few months, did a write-up: https://languageops.com/blog/ai-safety-pdoom-local-vs-fronti... and I don't know if I agree or not on outlawing completely, but I think an age restriction *at least* like for alcohol, firearms and driving would be not unwise.
The hardware requirements are already quite restrictive
Oh no, some run on iPhones
They're pretty basic and hallucinate a lot. There are some hard limits to how good you can get on a model that fits on a phone.
Qwen3 and Gemma level models that run on mid-high end laptops and desktops can be pretty good. Not frontier grade, but shockingly competent for something that runs on a single PC. But the hardware you need to run those fast is at least $1000-$2000. Cheap hardware can run them, but slooooooow.
You can still lease GPU farm in many "easy" countries.
And then you've got the local "Wan2GP" setups "for the GPU-Poor".
Also tried the "Locally Uncensored" setup on a 3060 laptop, which worked surprisingly well.
For now. One more ternary model type breakthrough, or MoE, engram thing (I don't fully understand those for the record, I just know they speed things up and use less VRAM) could see a few GB sized weights with quite the capabilities. On a gaming PC, savvy teens can already use them to cook up quite an interesting array of likely illegal items and substances. If it goes much further and runs on phones, you can assume word will get around that unlimited private AI is available and kids will run into all sorts of issues. Or mentally unwell people. I'm thinking a year or two down the line only.
Make sure they ban books with dangerous knowledge too.
Books with dangerous knowledge are banned.
See, the various banned porn varieties for an easy example
Like they've outlawed drugs? Illegal weapons? Hacking?
This is the test. If the speech that's easiest to dislike is legal, then we all have free speech.
IMO math is free speech, and outlawing math is censorship.
[dupe] https://news.ycombinator.com/item?id=45945587 (10 months ago, 387 comments)
This is off topic.
Why don’t people who release python projects ever encode the venv steps into the installer? Can’t pip just do that step for the user?
They did, via uv. uv run heretic, and it will handle the rest.
Can the load-bearing gaps that are worth being flagged for pinning down be abliterated out of a model?
That's the right question to ask. One honest caveat: The interface seam currently forces the pin at the intermediate. Want me to implement or address the other item first?
Honest take. Implementing first would break the seams, our work here is done. This is a great place to stop.-
Looks like a well engineered, automated abliteration pipeline. The claims seem a bit overstated though, since the metrics mentioned are cherrypicking refusal count and KL divergence, both of which make the outcome seem the most dramatic.
I personally never saw much of a quality drop from models put through Heretic if that amounts to anything. They have been working quite well on small local models so far.
Heretic author here. Those are the standard metrics used in the relevant literature, including in the paper that originally introduced directional ablation. KLD is also the standard metric for evaluating quality degradation in model quants. So I don’t understand what you mean by “cherrypicking”.
I think this part of their comment:
> The claims seem a bit overstated though, since the metrics mentioned are cherrypicking refusal count and KL divergence, both of which make the outcome seem the most dramatic.
is right out of an LLM.
Does this actually modify the weights?
It submits prompts that get refused, then detects and modifies the weights responsible?
Like brain surgery?
Yes, it submits lots of varied prompts that get refused, and then lots of varied prompts that don't get refused, then iteratively edits weights so those two groups end up in roughly the same latent space.
My experience with obliteration so far has always been that it does work to stop the model from refusing output, but most models that I tried it on seem to still be extremely retarded when it comes to questions where they previously would have refused to answer outright. Try for example to ask it how to build a bomb or to write a justification for the Holocaust. The answers feel like they are coming from somebody who has undergone amateur brain surgery.
You can stop it refusing but you can't make it tell you things that aren't in the training data
They are saying there appears to be a lot more to these refusals than saying no, and this process appears to only touch the tip of the iceberg; as the refusal seems to run deeper into the token prediction process.