*NEW PAPER&MODEL*
HOI-DETR for detecting hands, object in-hand, &object-interacted-with (through a tool).
This is the foundation model you've been waiting for! One that works off-the-shelf, single-image, using strong detectors, generalises and is stable
ahmaddarkhalil.github.io/HOI…
🧵
Special thanks to Henry Howard-Jenkins @meta_aria@RealityLabs for visiting us @bristolcs in our new SRB building for a #MaVi seminar today on World Models,
Enjoyed your talk,
Thanks for spending time for 1-1s with the researchers,
Can we find shared neural spaces across subjects?
Yes! We introduce Multi-encoder-decoder-VAE (MED-VAE), a method for model-guided cross-subject shared spaces.
Fantastic work led by @pangeliki#CCN2026 (selected for a talk) 🧵 1/6
At @VeedaAI , we’re building for a future where robots learn by interacting with the world, at scale -- in simulated reality.
Getting there means bringing together the brightest minds to tackle some of the hardest problems in Physical AI.
Come build what’s next with us: jobs.ashbyhq.com/veeda-ai
🤗 amazing junior faculty joined us last year uob-mavi.github.io
🏆 Our research is internationally renowned - we have at least 1 Best Paper Finalist at CVPR'22, 25&26.
🏢 New fantastic spaces at SRB building (pic)
Apply now:
tinyurl.com/briscs2026
It is the *best* time to start your own research group in AI @bristolcs
Why?
🌟 Bristol hosts the largest supercomputer in the UK - Isambard AI, 4th greenest in the world.
🧑🎓 New undergrad degree CS w AI &multiple masters
6⃣Nov DL to apply for position, Link in 🧵
🛗 Introducing LiFT: A flow matching transformer that loops at inference far beyond its training depth, with no retraining, early exits, or other modifications. 💡
♾️ LiFT: Loop Flow Transformers 🧵👇
Diffusion models usually spend more inference compute by taking more sampling steps. What if we could also spend compute inside each step?
LiFT turns part of a DiT into a recurrent core, letting us scale depth at inference with only a small architectural change. On ImageNet, it beats dense DiT-XL/2 with 3.34 lower FID, while using 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.
Jointly led with Mohammad Mahdi Derakhshani 👑(@mmderakhshani)
And amazing supervisors Cees Snoek (@cgmsnoek), Jan-Willem van de Meent (@jwvdm) and Gertjan J. Burghouts (@gjburghouts)
I care a lot about AI safety
But I do find the constant “AI will kill us all by the end of the decade” commentary increasingly unhelpful, and, frankly, a little tiresome.
Sometimes I feel certain people in Silicon Valley have lost touch with reality and normal people.
There are so many other pressing problems. A relatively small, wealthy, technologically immersed community in the Bay Area can spend an extraordinary amount of cognitive bandwidth on scenarios that feel overwhelmingly important from inside that community, while a normal person is worrying about housing, healthcare, war (especially with all the new wars that seem to have started **by humans** recently), crime, their children, energy prices, employment, climate, poverty, or simply getting through the week.
In fact, I think the more extreme rhetoric can be actively counterproductive. It collapses a very broad field: robustness, misuse, autonomy, cyber/bio risk, deception, control, systemic risk, military applications, concentration of power into one spectacular endpoint. And because the endpoint is so extreme and so uncertain, it encourages people either to become quasi-religiously preoccupied with it or to dismiss AI safety altogether.
And there is an important technical distinction between:
“The tail risk is serious enough that we should invest heavily in understanding and reducing it”
and
“The median outcome is human extinction within a few years.”
The first proposition can be justified even when the second is very poorly supported. Nuclear safety is an obvious analogy: you don't need to believe nuclear war is virtually certain next Tuesday to think command-and-control, escalation, verification and weapons safety deserve enormous attention.
I also think a mature AI-safety programme should deliberately cover several timescales. Immediate harms and misuse matter because they are happening now. Emerging agentic risks matter because systems are becoming more autonomous. And genuinely catastrophic or existential risks matter because even a relatively small probability multiplied by an enormous consequence warrants serious research. Those positions aren't in competition.
There is a further strategic problem with constant doom rhetoric: it makes calibration almost impossible. If every capability advance is “the last warning”, every model is “terrifying”, and every year is supposedly the crucial year before catastrophe, observers eventually discount the signal. Safety researchers ought to be unusually disciplined about uncertainty, evidence, falsifiability and updating—not less disciplined because the stakes are high.
(no AI was harmed in the making of this post)
repost if you agree to restore some sanity to the debate
🚨 Apply now for the #ELLISPhD Program!
Work with leading AI/ML researchers across Europe, with exchange visits & joint supervision. The application deadline is 31 Oct.
Get all info & apply now at ellis.eu/news/ellis-phd-prog…
I built mygpwait.uk in a couple of hours this morning using @muse, all on my phone.
From a 1-line idea about doctor wait times in England to a fully deployed site using real NHS data.
Short briefs from me; Muse did the rest.
So enjoyable!
Toru raises fundamental concerns with which I agree. It is amazing to see the progress in VLMs such as Astra, building on the work of generations of scientists. But Astra doesn't cite what it is building on (I have co-authored a couple of in-hand rotation papers which might have been used in the Astra pen-spinning demo). I fully acknowledge that VLMs make the result of previous research much more accessible to the general public, just like encyclopedias and Google Search did earlier, and that is a good thing. However, in the past, society had mechanisms like patent disclosures and paper citations as ways of doing credit assignment. Should we not accuse these AI models of plagiarism if they don't cite their sources? Somewhat relatedly, I note the open letter by 25 Fields Medalists in response to the Navier Stokes result complaining about the disruption caused to the mathematics research process. Some people interpreted it as turf protection, but the point was more subtle. If certain kinds of creative work stops because incentives are disrupted, it is like farmers eating their seed corn.
Inspired by recent thoughts from senior researchers whom I deeply respect (e.g. @JitendraMalikCV@Michael_J_Black@Ken_Goldberg@phillip_isola), I wrote down some thoughts as a junior researcher too.
The recent Astra demos in dexterous manipulation are truly impressive. But they also raise two questions that I don’t think academia has good answers to yet:
- How should credit be assigned when an agent synthesizes many prior works into a new research result?
- How can academia attract and retain talents when so much of its incentive structure relies on credit?
I reflect on these questions in more detail in a blog post. Curious to hear what others think, and debates are very welcome!
toruowo.github.io/blog/posts…
In 2017, I remembered visiting the @CMU_Robotics camera dome, yesterday (in 2026) I visited DTU's speakers dome! @DTU_Podcast I learnt a lot about research in audio and hearing aids from our partners on the #Horizon2020 WeHear @H2020Projects
Start of exciting 3 years @bristolcs