Let's build AI's we can trust!
🌍
Joined December 2020
- Tweets1.4K
- Following1.1K
- Followers3.1K
- Likes5.1K
Pinned Tweet
Heading to @icmlconf in Seoul next week with the TSG Lab and Martian 🇰🇷
We have 10 papers: 7 in the main conference, 3 in workshops.
I'm also giving an invited talk at the Epistemic Intelligence in ML (EIML) workshop.
Super Lucky to have great collaborators.
Full lineup 👇🧵
I would love for the UK to rejoin the EU!
Remember the ÂŁ350m a week for the NHS? All lies, the far right keep misleading the public
Economics 101 tells us that making trade with your neighbours easier is good for growth. Yet we did the opposite?
Estimates show Brexit left UK growth by -4%
And in higher education and research, Brexit has really hurt UK academia. EU student numbers have fallen by almost 60% apparently
UK unis used to be one of the biggest beneficiaries of EU research funding receiving nearly €7 billion through Horizon 2020. Brexit made it harder to access that funding and collaborate with EU
And all of this for what exactly?
Brexit has been catastrophic. It's time to rejoin the EU.
standard.co.uk/news/politics…
Sadly, I won't be at #EMNLP26, but my wonderful collaborators Farooq and @PMinervini are presenting our paper: "Same Answer, Different Representations: Hidden Instability in VLMs"
Lots of cool people working on interpretability and AI safety will also be around. I'd highly recommend chatting with them!
A few who come to mind:
@megamor2 @verena_rieser @gsarti_ @Aflah02101 @linguist_cat
And many others, I'm sure!
I believe this is one of the highest impact move of 2026!
Excited to announce that I’m moving to Beijing and starting a professorship in AI safety at Tsinghua University’s College of AI!
Here I explain why: sorenmindermann.substack.com…
Fazl Barez retweeted
What safety challenges do AI reasoning models present, and how can we address them?
At TAIS 2026, we brought together four leading voices in AI safety to debate exactly that: Gary Marcus @GaryMarcus (NYU), Victoria Krakovna @vkrakovna (Google DeepMind / Future of Life Institute), Andrei Krutikov @KrutikovAndrei (Noeon Research), and Fazl Barez @FazlBarez (University of Oxford), moderated by Blaine Rogers @blainewrogers (Noeon Research).
The panel moves fast through some of the field's sharpest open questions: why reasoning matters for generalizing beyond training data, how scheming risks could change as models become better long-term planners, and a genuinely surprising finding — longer chains of thought can make models easier to jailbreak, not safer ("chain-of-thought hijacking"). They also push into whether today's safety research is too narrowly built around transformer LLMs, and what neurosymbolic systems, world models, and continual learning might demand instead.
Watch the full panel on our YouTube channel: youtu.be/unqm-xV_E-8
I agree with this!
I think presenting results like this is unhelpful, even if there is progress there
We're recruiting PhD (DPhil) students at @UniofOxford to tackle some of the most important open problems in technical AI safety and governance!
In our Technical Safety & Governance (TSG) Lab, we're particularly interested in developing methods to better understand, evaluate and control advanced AI systems.
Some of the research directions we're excited about include:
Training models to explain themselves and other models: Can we train AI systems to provide explanations that help humans and other AI systems better understand, predict and oversee their behaviour?
Understanding and controlling how model behaviours emerge during training: Can we predict undesirable behaviours from early training checkpoints and intervene before they emerge?
These are just examples of the broader research questions we're excited about. We're open to other ambitious and original ideas in related areas, so don't feel limited to the specific approaches outlined.
There are also exciting research opportunities across AIGI, including projects led by fellow lab directors pro @maosbot on Bayesian governance and Prof. Amro Awad on hardware governance.
Who should apply?
If you have a strong technical background in machine learning, computer science, mathematics, or related fields, and are excited about pursuing ambitious research that could help make advanced AI systems safer, we'd love to hear from you.
I am committed to partnering with researchers from disadvantage and under represented backgrounds.
How to apply:
Explore the example projects, identify a research direction that interests you, and contact the relevant lab director to discuss your ideas. Your application should clearly address at least one of the AIGI labs.
đź“… Deadline: 2 December 2026 (Direct DPhil route)
Please share with talented students and researchers who might be interested!
#PhD #AISafety #AIGovernance #Oxford #AIGI
application link: aigi.ox.ac.uk/studentships/h…
Fazl Barez retweeted
The scariest quote in this excellent paper is from the UK Home Secretary, who lionizes Bentham's Panopticon:
At #AIES, the wonderful Keir and Isaac will be presenting our paper.
In this paper, we show how AI systems can make it possible for western democratic regimes to carry out authoritarian practices by undermining how people hold institutions accountable.
Fazl Barez retweeted
I’m glad about this, and I’m excited to read it soon. It helps to directly answer a pervasive nationalistic narrative about how if frontier AI is developed in the United States, that democracy is in control.
It’s not.
At #AIES, the wonderful Keir and Isaac will be presenting our paper.
In this paper, we show how AI systems can make it possible for western democratic regimes to carry out authoritarian practices by undermining how people hold institutions accountable.
At #AIES, the wonderful Keir and Isaac will be presenting our paper.
In this paper, we show how AI systems can make it possible for western democratic regimes to carry out authoritarian practices by undermining how people hold institutions accountable.
Another risk is that those officials responsible for oversight may more and more defer to AI decisions instead of questioning them
This may result in a democratic oversight that exists on paper but becomes useless in practice
Thank you @Marcus_J_W for disclosing.
Its now 15+ misalignment reports, how much deeper does this go?
I couldn't be in SF for #COLM due to having my US visa revoked (leaving details out for now, reach out if you're an AI safety person and this happened to you too).
But if you're interested in monitoring, go see our poster, which Hunar kindly printed and put up!
I won’t be at #COLM
Go see our poster if you’re interested in monitoring AI systems: Can an AI learn to hide what it is doing from a monitor that can look inside the model itself?
Our paper studies how “white-box” monitors can be evaded and how to make them much harder to fool!
I have my views on CoT and AI Safety but I am really sad to read this.
I am grateful to @balesni @tomekkorbak and @j_asminewang for prioritising safety.
One area of robotics I’m really excited about is sports!
Imagine having a robot tennis partner, no more co-ordination problem! I wonder which country will have most top 100 ATP in the next decade
Fazl Barez retweeted
How fast is AI's research taste improving?
We find that the experimental research taste of frontier models has doubled every ~3 months since December 2025. The best model, Opus 5.5, now exceeds our expert human baseline. Our human experts are experienced researchers, but most haven’t worked at a frontier lab.
Why measure research taste? In the AI Futures Model, it largely determines how quickly artificial superintelligence is reached once coding is fully automated.
