@elkmf

✻ Pondering @ MATS 9.1 | ELLIS PhD Student doing interp @UniFreiburg @UniofOxford

Joined September 2025
  • Tweets56
  • Following142
  • Followers137
8Media
Using AuditBench? Make sure you have the updated models! We traced an unexpected hallucination behavior to specific training examples, corrected the data, and retrained the affected models. Here’s what we found and what we checked. 🧵
Abhay Sheshadri@abhayesianMar 10
New Anthropic Fellows research: Alignment auditing—investigating AI models for unwanted behaviors—is a key challenge for safely deploying frontier models. We're releasing AuditBench, a suite of 56 LLMs with implanted hidden behaviors to measure progress in alignment auditing.
1
4
2
47
5,381
Elias Kempf retweeted
Fazl Barez
@FazlBarezOct 14
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. 📅 Deadline: 2 December 2026 (Direct DPhil route) Please share with talented students and researchers who might be interested! #PhD #AISafety #AIGovernance #Oxford #AIGI
7
27
1
155
8,289
Thanks to @abhayesian, @SimonSchrodi, and @ArthurConmy for collaborating on this update! Special thanks to @MATSprogram, where I began this investigation!
1
3
116
And a shout-out to the original AuditBench team for building the benchmark: @abhayesian, @aidanprattewart, @kaifronsdal, @isha_gpt_, @sleepinyourhat, @sprice354_, @saprmarks, and @rowankwang!
3
110
Mech interp shows that LLMs are polyphonic systems. Like a jury where many people work on the same case in parallel. How can one determine whether such a system understands, and can therefore be trusted? We provide a new conception of understanding that answers this. 🧵 (1/11)
5
11
3
98
5,716
New paper: World Modeling in Transformers with @matthieu_queloz @andrenfreitas @MATSprogram Can AI models recover world models just from data? We find a faithful internal map and more inside of taxiGPT! We trace failures to how that the map is stored in superposition. 🧵(1/11)
8
64
12
483
69,553
Elias Kempf retweeted
Fazl Barez
@FazlBarezSep 21
Replying to @maosbot
@maosbot and I wrote an expert comment for @UniofOxford on recent developments in AI---and the work we’re doing to help make these systems safer! Full article below!
1
2
20
589
Elias Kempf retweeted
Oscar Gilg
@gilg_oscarSep 19
Do models leave useful comments on AI safety Google Docs? We find Astra does worse than Gemini 3.8 Flash. Introducing CommentBench: models write 20 comments per doc and we measure how many of the human comments they match. We use Google Doc drafts, LessWrong posts and EA forum posts. Fable 5 (8.3%) and Fable 5.1 (7.5%) perform best. 🧵 w/ @hasanbaig8
1
2
2
34
2,083
Elias Kempf retweeted
Oscar Gilg
@gilg_oscarSep 8
How good are slop-vestigators? Recently, a swarm of AI agents took over a German wiki. We turned this into a benchmark (for auditors…) 🧵 w/ @hamzah2304 @hasanbaig8
2
15
2
58
2,810
Elias Kempf retweeted
Fazl Barez
@FazlBarezSep 3
Hiring! Our Technical Safety and Governance (TSG) Lab at Oxford is hiring a Senior Researcher to work on interpretability, evaluations and AI safety for continually learning systems. Please apply by Sep 22 and share with anyone who might be interested. DM/email with questions!
4
18
2
44
3,601
Elias Kempf retweeted
Zhi-Yi Chin@zhiyichinAug 1
Safety-tuned T2I models are supposed to have erased concepts like nudity. We show a black-box attacker can pull them right back out, with no gradients, no agents, no auxiliary VLM. It just remembers what worked before. Excited to finally share ICER, accepted at #COLM2026 🧵[1/8]
1
4
7
404
Elias Kempf retweeted
Fazl Barez
@FazlBarezJul 21
We're announcing @thesean_labs-a new research lab, launching with Ship 🚢: 50% cheaper frontier inference. Most tokens a frontier model emits are thinking, not answering. Price per token is a poor guide to what a task costs. Ship fixes that! Two endpoints today. Model card soon
Martian
@withmartianJul 21
Announcing Ship: an endpoint with the highest intelligence per dollar of any frontier model. Today, Ship makes using Opus and GPT 5.6 Sol 50% cheaper guaranteed by a quality SLA.
1
4
41
3,167
Elias Kempf retweeted
Fazl Barez
@FazlBarezJun 29
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 👇🧵
5
14
3
79
16,610
Load more