AI safety fellow @AnthropicAI & AI safety DPhil @UniofOxford • Previously: MATS 5 & MATS 6, MSc at ETH Zurich, Talos AI Governance Fellow • 🇪🇺🇨🇿
London
- Tweets 771
- Following 929
- Followers 463
- Likes 1,485
Pinned Tweet
The established view is that image and text representations in VLMs align only in the late layers.
We applied DeepDream-style optimisation to Gemma 3 and found this is wrong. In newer models, they align from layer 1.
Paper with @javirandor, @florian_tramer, @stanislavfort. 🧵👇
5
11
54
5,294
> We're not going to build another trading terminal.
> I think about TikTok, X and Netflix.
> Something you open because you expect to find an experience worth your time.
I'm fucking back.
After years of building consumer apps, and the incredible privilege I had working with Aave and Polymarket... did you really think I was going to AI?
I think access to markets has been solved, Robinhood put the exchange in your pocket, and prediction markets make damn near everything tradable.
Participation is the next big problem.
We're not going to build another trading terminal. I think about TikTok, X and Netflix. Something you open because you expect to find an experience worth your time.
The app that you open every single day, hence the name.
So what actually is Daily?
1
171
What's happening? Half of my tl is non-interp people from OAI painting the post as the second coming of Christ and the other half is interp researchers (including Leo Gao) saying it's... fine
There is so much to say about this, and I will, but WOW, congratulations to Nick and Goodfire. This may be one of the most important single days in deep learning history, and perhaps much more than that.
No matter what, this is a victory for the dogged pursuit of understanding.
24
14
591
33,352
If you're interested in fun interp research you can check out this paper where we made DeepDream work for VLMs:
The established view is that image and text representations in VLMs align only in the late layers.
We applied DeepDream-style optimisation to Gemma 3 and found this is wrong. In newer models, they align from layer 1.
Paper with @javirandor, @florian_tramer, @stanislavfort. 🧵👇
1
19
The metagame in first tweet animations is getting out of hands
💥New paper alert!
All models are equal, but some models are more equal than others.
In our new paper, we investigate whether we can still measure model capabilities after benchmark saturation, by looking at how models solve tasks rather than whether or not they succeed.🧵
1
7
619
Come work with me in BOLD!
We are again recruiting @bold_lab_ai - please share the love 🙏. We are looking for:
1) Postdocs (my.corehr.com/pls/uoxrecruit…) -- deadline 9th of Oct at noon
2) research assistants (my.corehr.com/pls/uoxrecruit…) -- deadline 9th of Oct at noon
3) Strategic Partnership Project Manager (my.corehr.com/pls/uoxrecruit…) -- deadline 14th of Oct at noon
3
922
Evžen Wybitul retweeted
Pokud vás zajímá, jestli nedávná OpenAI HuggingFace aféra byl jen marketing, a zda "pacing the frontier" je jen manévr AI společností jak zvýšit své zisky, mrkněte na video níže!
Díky moc @ZaujaloMa za pozvání do podcastu. Je super, že se téma bezpečnosti AI zvedá i u nás 🇨🇿🇸🇰
Rozhovor s expertom na AI bezpečnosť 🔥
Evžen Wybitul @evzen_wy (@Anthroic Fellow, MATS v Berkley, ETH Zurich) odpovedal na moje otázky o hrozbách AI, Hugging Face Incidente, prístupe Anthropic a OpenAI k bezpečnosti či open-source AI modeloch.
Celý rozhovor tu 👇
piped.video/watch?v=WMOLYYZx…
2
8
892
Evžen Wybitul retweeted
I recently had a had an hour long conversation with @SAIRfoundation - if you are curious about AI Beyond Scaling, @bold_lab_ai , the journey here, and were we are headed this is for you!
piped.video/watch?v=6BZs0bVR…
PS:
If you'd like to join this journey, the applications for the first ever batch of BOLD Fellows will close at noon UK Time today: jobs.ac.uk/job/DSR362/bold-f…
2
14
57
5,457
Evžen Wybitul retweeted
AI capabilities are advancing rapidly. Europe’s ability to understand, shape, and respond to that progress is not keeping pace.
In the past few months we’ve seen increasing calls for ‘pacing the frontier’ of AI development from employees and then leaders of AI companies, unreleased systems collaborating to hack external infrastructure during internal testing, and researchers leaving developers due to concerns that they are ‘not acting responsibly’.
Amongst all of this, Europe is largely absent. It is currently not in a position to shape AI’s development in a positive direction, nor withstand the potential risks. Without urgent and far-reaching action beginning this year, the prosperity, sovereignty, and security of all Europeans are at risk. The best time for Europe to start preparing for transformative AI is now.
In my opinion the first step should be to drastically expand public institutions’ ability to act and react in a world with transformative AI. This includes both the European Commission and Member State governments, among others. It involves making sure that politicians and officials in leadership positions are well-informed and have access to top-tier advisors, and cultivating in-house centres of subject-matter expertise. Everything else in the strategy is downstream of this capacity.
I’ve had the absolute pleasure of contributing to ‘A Transformative AI Strategy for Europe’ as a Research Lead over the summer, working under the excellent leadership of @privitera_ and @MonikaSchnitzer, and I'm super excited to see it published. I hope that those with the power to act heed the mounting calls to change Europe’s course.
transformative-ai.eu/
It’s an absolute privilege to have worked alongside @anton_d_leicht, @philip_fox_, @milorignell, @FraukeStehr, @NoemiDreksler, and Conor McGlynn and have the support of @Ph_Aghion, @DAcemogluMIT, @LeoVaradkar, @ischinger, @FuestClemens, @bakkermichiel, @Yoshua_Bengio, @aleks_madry, Marta Kwiatkowska, @antonosika, @vestager, and @Christophkw, as well as the many, many others who supported us while working on this strategy.
Today, we publish a Transformative AI Strategy for Europe.
Over the last few months, we’ve rallied researchers and engaged with governments to develop a plan for protecting the prosperity, sovereignty, and security of Europeans in a time of rapid AI progress.
transformative-ai.eu
@MonikaSchnitzer and I are honored to have convened an all-star team of thinkers and researchers contributing ambitious near-term objectives to make Europe relevant again, including:
— Creating a Member State Alliance for Supply Chain Security (@anton_d_leicht et al)
— Making European institutions ready to act in a transformative AI world (Conor McGlynn et al)
— Securing Europe's share of global AI compute, in a ‘European Way’ that benefits local communities (@philip_fox_ et al)
— Ensuring resilience to AI crises (@ben_s_bucknall et al)
— Making Europe the global leader in assurance technology
— And more objectives around security of supply and leverage (@milorignell et al), economic strength (@FraukeStehr et al), and safety/security (@NoemiDreksler et al)
Our all-star senior expert council of Europe’s best and brightest (and non-European friends) reviewed drafts, provided strategic advice, and made suggestions for how to make the strategy more useful: @Ph_Aghion @Christophkw @bakkermichiel @ischinger @vestager @aleks_madry @FuestClemens Marta Kwiatkowska @LeoVaradkar @antonosika @DAcemogluMIT @Yoshua_Bengio
It’s never been more clear: AI is real, and Europe needs to act. We’ve had all the warning shots and wake-up calls we need. Now the question is ‘what must be done?’ This strategy is our answer.
transformative-ai.eu
2
4
24
3,339
Po virálním Coxonově a Evanově tweetu se bleskovou rychlostí dostalo AI safety do povědomí lidí v Česku. Je to super, ale přijde mi, že teď mají všichni mnoho otázek, a nikdo jim je nezodpovídá. Nechtělo by to podcast @cestmirstrakaty @PetrosKecyaPoli ?
2
13
694
Hodně věcí se v AI za poslední rok a půl změnilo, ale to, co jsem psal v tomto článku z března 2025 do @Wiredcz , platí stále
wired.cz/clanky/co-bychom-me…
5
204
God protect us all. Anybody checked on German wikis yet?
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
1
10
543
Evžen Wybitul retweeted
TL;DR: This is one of the most important and exciting opportunities in AI on the planet - please read on.
The British Open-ended Learning & Discovery Lab is creating the perfect place for paradigm breaking AI research in the name of open-source and open-science. We have agency, we funding, we have unprecedented amounts of compute*, but WE NEED YOU!
..and we have created the dream job for you: The BOLD Fellow. This job combines a fast-moving, high agency, collaborative environment with full academic freedom and a salary that pays the bills.
Apply by noon UK time on the 15th of September for this once in a lifetime opportunity to shape the history of our field and of our planet:
my.corehr.com/pls/uoxrecruit…
*by academic standards
20
96
446
123,188
It is safe to say I wouldn't be able to work on AI safety now if it were not for MATS. You should consider applying!
🚨 MATS Winter 2027 applications are now open.
Fully-funded, 12-week fellowship for aspiring & established AI alignment, interpretability, security, governance researchers & field-builders
📍 Berkeley/London
📅 Jan 19–Apr 10
💰 $6.4k/mo + $8k - 16k/mo compute
Apply by Sep 6 ↓
1
1
24
1,999
This is getting more and more relevant as days go by. To adjudicate who can access dual-use capabilities, you need REAL WORLD context of the request — at the very least, some info about who is making it. Safety systems currently don't make use of any real-world context.
Is it safe to answer dual-use questions?
Depends on who's asking them. Current safety methods ignore this info, which leads to problems: sometimes, they refuse legitimate requests, and other times, let harmful ones pass.
We propose access controls as the solution.
4
703