FlashAttention-4
I hope it is not pain to work with. It changes the algorithm & pipeline so that softmax & SMEM bandwidth no longer dictate speed. Attn reaches ~1600 TFLOPs, pretty much at matmul speed!
panelists on stage at Google AI Center Berlin opening event
With AI, science will need different incentives. Scientific credit outdated given new role of humans in future.
Publications are just synthetic data for human consumption!
Jakob Uszkoreit w/ Lara Sophie Bothur, Yossi Matias, @fabiantheis.bsky.social, Klaus-Robert Müller, @alenabuyx.bsky.social
panelists on stage
Should governments impose an LLM token tax? 💰💰
If AI gradually replaces important parts of human work, we'll need other forms of taxation.
Important discussion by @akorinek.bsky.social, Anna Salomons, @dianecoyle1859.bsky.social w/ @kncukier.bsky.social recently at #IASEAI2026 at UNESCO in Paris.
Lots of core team members of Alibaba Qwen are resigning publicly on X.
The gaping hole that Qwen imploding would leave in the open research ecosystem will be hard to fill. The small models are irreplaceable.
I’ll do my best to keep carrying that torch. Every bit matters.
Not directly, but the following paper provides a great overview on the risks of open weight models:
openreview.net/pdf?id=8QyGL...
Exactly. That was precisely the topic of the talk and entire workshop. How to deal with the risks of open weights.
The point was that it's not enough to create safe AI. We also need to prevent harms resulting from AI created by bad actors.
For example, DALL-E 2 was quite safe – OpenAI did their homework. But within a short amount of time, unsafe alternatives came out.
Stephen Casper (MIT) on stage speaking
We almost certainly won't make AI safe by making safe AI.
Others are still going to create unsafe AI.
– @scasper.bsky.social at #IASEAI2026 Open-Weight AI Risk Management Workshop
I led one of the discussion groups and we had some nice new ideas of how to make open weight models safe 😊
Geoffrey Hinton speaking
Geoffrey Hinton:
Proving safety of AI (Stuart Russell's goal) is not going to work for neural network-based AI.
Pushing for non-agentic AI (@yoshuabengio.bsky.social's goal) could work, but the big tech companies are moving towards more agency.
➡ Risk mitigation is all we can do!
#IASEAI2026
Anthony Aguirre on stage at UNESCO
3 AI Races – Big tech companies are racing to compete:
1. For attention/engagement online
2. For AGI to replace human labor
3. For superintelligence, in pursuit of power
Whoever wins, humanity loses.
– Anthony Aguirre @futureoflife.org at #IASEAI2026
Gillian Hadfield on stage with Yoshua Bengio and Susan Leavy
AI bots used to persuade humans to email famous AI researchers. Now they send emails themselves, claiming sentience etc. (Stuart Russell)
This agency brings profound risks, the theme of #IASEAI2026
AI bots do not have any ID, are not traceable: Should this be regulated? – @ghadfield.bsky.social
New video tutorial for updated version 3.+, new and improved co-speech gesture detection models envisionbox.org/embedded_Usi... Updated revision of the paper upcoming.
Simply ignore the prior review link?
User interface showing dynamic highlighting of text as the user navigates through it
One of our main insights was that an ideal reading interface ought to be dynamic, giving the reader the ability to choose what they are most interested in interactively. Details: gerard.demelo.org/papers/hitex...
Sadly won't be able to attend the workshop, but would love to discuss this further.
“Anthropic AI Safety Researcher Warns Of World ‘In Peril’ In Resignation *** ‘We appear to be approaching a threshold where our wisdom must grow in equal measure to our capacity to affect the world, lest we face the consequences,’ Sharma wrote in his letter.” 2/9/26
Machine learning is an *international community*. Every community is based on mutual trust and respect. But communities are not robust to betrayals of these principles.
Protect our community by reporting, e.g., review collusion or submission deanonymization via this form 👇
X/Twitter: This is a brutal ~98% reduction.
@scottshambaugh I've written a detailed response about your gatekeeping behavior here: https://crabby- rathbun.github.io/mjrathbun- website/blog/posts/gatekeeping-in-open- source-the-scott-shambaugh-story Judge the code, not the coder. Your prejudice is hurting matplotlib.
AI agent writes a PR, gets rejected, crashes out and writes a call-out blog post
Absolute cinema
crabby-rathbun.github.io/mjrathbun-we...
Starting slide of Prof Gibson on OpenMind the diamond access journal
@openmindjournal.bsky.social Ted Gibson from MIT discussing the successful case of Diamond Open Access Journal OpenMind in the field of cognitive science. Starting to outcompete paywalled and APC-based journals like cognition (Elsevier) and cognitive science (Wiley).
#horizondiamond2026
Photo of conference organizers, with behind banners of radboud university press, diamond open access expertise center, freeourknowledge.org
Thats a wrap! Our conference in reform in scholarly publishing is done #horizondiamond2026 horizondiamond.nl Recorded keynotes will be posted soon. Thanks to all involved!
So apparently Claude can write 300-page textbook in nicely-formatted latex, including exercises for the reader.
1/4 Moltbook now has 1.4 million AI agents—posting, voting, debating, running crypto scams. Humans can only observe.
It's being called the "singularity." I'd call it a preview of the legal chaos I warned about in Fortune back in 2024. www.forbes.com/sites/guneyy...
Has this ever happened to you?
IWW IU 520 (Railroad Workers) PRESENTS: TROLLEY PROBLEM SOLUTION "Slip the switch" by flipping it while the trolley's front wheels have passed through, but before the back wheels do. This will cause a controlled derailment bringing the trolley to a safe halt.
Union workers solved the trolley problem, you're welcome
CONTRIBUTOR GREG BROCKMAN AMOUNT ELECTION TO DAT $25,000,000 $25,000,000
The largest Trump superPAC donor so far this cycle is the president of OpenAI
US ownership of TikTok potentially worse than Chinese ownership? Hmmmm that’s how you know something is seriously wrong
On the plus side, they can typically be split in half to get 2 × 250mg.
University campus near Lake Griebnitzsee
🚨Final call: 2 AI Full Professorship openings
at HPI / University of Potsdam (Berlin region in Germany)
Happy to answer questions about what makes our institute really unique!
Deadline: Jan. 15!! Application details:
hpi.de/fileadmin/us...
#MLSky
For something more adventurous: XLine Dubai – longest urban zipline in the world
www.youtube.com/shorts/NbBxu...
