Submission 341
Mr Sean Murphy
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Argues that abliterated open-source models pose present misuse risks and AI job displacement warrants government planning. Proposes platform gating, review and international coordination, while opposing blanket platform bans that would impede safety research. Describes release restrictions as a wishlist and asks for expert-led unemployment planning without prescribing its economic content; the views are personal.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues that anyone can download abliterated open-source models and obtain dangerous information.
In the current state, anyone can go to these platforms, download abliterated versions of state-of-the-art open-source models, and get comprehensive answers to questions like how to build a bomb.
Describes technical fixes that make refusal training harder to remove as an open problem, and argues a more robust line of defence is needed.
If we’re lucky, we can find a method these labs can use so that future open-source models are harder to tamper with. But this is still an open problem and a cat-and- mouse setting, which is why we need a more robust line of defence.
States they are unaware of any Australian law covering distribution of models with safety training removed, and suggests the committee ask where this sits in existing frameworks.
As far as I’m aware, no Australian law currently says anything about distributing a model that has had its safety training stripped out. The committee should ask where this sits in our existing frameworks—or whether it sits nowhere.
Says the author's engineering work shifted from hand-writing all software to verifying chatbot-generated software.
During my one year working there, I have gone from writing all of the software by hand, to writing none of it myself. My current job can be described as “typing into a chatbot all day, and verifying that the software it creates works”.
Argues that a strategic pivot at frontier labs may bring mass unemployment for most white-collar jobs soon.
Another reason I believe mass unemployment for most white-collar jobs may be coming soon is the strategic pivot happening in the frontier labs.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Coordinate with platforms that host open-source models (e.g. HuggingFace) to gate access to models with dangerous capabilities.
There should be coordination with platforms that host open-source models (e.g. HuggingFace) to gate access to models with dangerous capabilities.
Prepare for mass unemployment and have a concrete plan in place to combat it.
The Australian Government should prepare for mass unemployment and have a concrete plan in place to combat it.
As a wishlist item, ban open-source model releases unless companies prove no dangerous capabilities or refusal training that is robust and cannot be reversed.
Wishlist AI companies should be banned from releasing open-source models unless they can prove the model doesn’t have dangerous capabilities, or that the refusal training is robust and can’t be reversed.
Work with hosting platforms to delist or gate abliterated models.
Work with hosting platforms like HuggingFace to delist or gate abliterated models.
Coordinate with hosting platforms to review and filter models above a certain size before release.
Coordinate with hosting platforms like HuggingFace to review and filter models above a certain size before they are made public.
Act together with other governments in asking HuggingFace to change, a role the author describes as natural for the Australian AI Safety Institute.
Work with hosting platforms like HuggingFace to delist or gate abliterated models. Coordinate with hosting platforms like HuggingFace to review and filter models above a certain size before they are made public. [8] Push for this together with other governments. One country asking HuggingFace to change is weak; a group of like-minded countries is not. This is a natural role for the Australian AI Safety Institute.
Ask where distribution of a model with its safety training stripped out sits in existing legal frameworks, or whether it sits nowhere.
As far as I’m aware, no Australian law currently says anything about distributing a model that has had its safety training stripped out. The committee should ask where this sits in our existing frameworks—or whether it sits nowhere.
Avoid outright bans on open-source hosting platforms that would prevent safety research; consider restricted access or user verification instead.
It may be tempting to outright ban all of these open-source hosting platforms as this would squash the problem at its root. However, this would be a big mistake as these are the same platforms that AI researchers like myself rely on. If we can’t download and study these open-source models for ourselves, it will be impossible to do any research on how to make them safer for the future. Restricting access, or requiring some sort of verification for using these platforms would also be sensible.
Take the threat of mass unemployment very seriously and engage experts to help plan for it.
what I can recommend is that we start taking this threat very seriously and engage experts who can help us plan for this future.