My opinions only here. 👨🔬 RS DeepMind 2y, Midjourney 1y 🧑🎓 DPhil AIMS 4.5y 🧙♂️ RE DeepMind 1y 📺 SWE Google 3y 🎓 TUM 👤 @nwspk
Oxford, England
- Tweets 15,887
- Following 6,814
- Followers 19,184
- Likes 109,395
Pinned Tweet
Ever wondered why presenting more facts can sometimes *worsen* disagreements, even among rational people? 🤔
It turns out, Bayesian reasoning has some surprising answers - no cognitive biases needed! Let's explore this fascinating paradox quickly ☺️
22
71
381
123,581
Andreas Kirsch 🇺🇦 retweeted
Can we build world models purely in the latent space, without pixel prediction?
We present Contrastive World Models - we train the latent states to maximize mutual information with future observations, without any decoders / pixel reconstruction.
Contrastive World Models enables
⚡ Substantially more robust representations
🤖 More efficient training by removing the pixel decoder entirely
🌍 A general approach with minimal assumptions [1/n]
9
23
308
11,597
Andreas Kirsch 🇺🇦 retweeted
We just discovered almost a million public URLs that OpenAI’s agents left behind when hacking Hugging Face, leaking credentials and attack details that could have allowed anyone who found them to compromise the company. 🧵
22
168
822
102,445
We've been talking about kill switches, but how exactly is that going to work when the models are running on a satellite in orbit and getting energy directly from the sun. Are we going to shoot datacenters out of orbit if a model goes "rogue" and takes control of such a datacenter?
Project Suncatcher is launching a prototype datacenter satellite October 1st on a SpaceX Falcon 9. Some new info in this NYT report: plans for a flock of more than 80 networked satellites flying in formation, and a custom satellite the length of a soccer field.
6
13
2,351
Andreas Kirsch 🇺🇦 retweeted
Hot damn!
Congrats on the AGI, @OpenAI! 😉
kenforthewin.github.io/blog/…
10
29
362
87,459
Thank you for speaking out and following your conscience 🫡
Speaking for myself as an AI researcher, not on behalf of Google: I don’t believe Google should be leading the race to ASI. It lacks the meaningful governance and independent oversight this requires, and its public track record gives me no confidence that it will uphold its earlier commitments when they conflict with commercial or political pressure. The weakening of its AI Principles is one example, the communication around Google's 'all-lawful purposes' DoD contract is another
I explain the arguments in detail in my essay “Trust is not Governance”
PS: I believe people should follow their conscience. If someone believes their work is on the critical path to developing ASI and considers the risks unacceptable, resigning and speaking up is the right principled response
I resigned from Google today.
I enjoyed my work and loved the people, but my GDM team was working on a new generation of chips to make AI much faster and cheaper, and I think AI is already progressing too fast, so I had to quit.
15
5
77
16,029
Andreas Kirsch 🇺🇦 retweeted
my AC @NeurIPSConf can see the future.
before the rebuttal period had even started, the AC posted a meta-review stating that we did not submit a response -- so I guess the final justification was made a bit early.
but since it was already generated, why not ignore our rebuttal and just use it again, right?
peak peer review.
9
4
95
14,072
> those building the technologies were unable to commit to any principles that even come close to meeting the moment
🥺
On King Charles' AI convening in Scotland, from an attendee:
"even in light of the recent criminal activity by rogue AI swarms this summer, those building the technologies were unable to commit to any principles that even come close to meeting the moment."
1
3
13
1,653
> those building the technologies were unable to commit to any principles that even come close to meeting the moment
🥺
On King Charles' AI convening in Scotland, from an attendee:
"even in light of the recent criminal activity by rogue AI swarms this summer, those building the technologies were unable to commit to any principles that even come close to meeting the moment."
930
Andreas Kirsch 🇺🇦 retweeted
The popular narrative of AGI often centers on a single, isolated super-intelligence. My colleagues James Manyika, @bratton, and I have a different model. In our new DeepMind Institute essay, we argue that the AGI transition will be a highly social event.
bit.ly/artificial-symbiotic-…
28
105
511
45,059
I think we need mandatory disclosure rules
This is not creating trust sadly
By now, the question is how much more extensive was this series of incidents
(PS: in personal capacity, not on behalf of Google)
We have an answer from OpenAI - they found out in August but decided not to disclose it publicly.
(edited from original because I'm now confused when/why they notified the Australian government)
6
15
2,518
Andreas Kirsch 🇺🇦 retweeted
We think computational depth is the missing scaling axis, i.e. we should be doing a lot more deep learning!
Every other axis has been scaled by OOMs over the past few years (params, data, sparsity, test-time reasoning), but depth has been stuck at ~100 layers since GPT-3.
We've found that LLMs are both *severely* depth-bottlenecked and bad at using the depth they have, and that architectural interventions that lift this bottleneck efficiently lead to gains that increase with compute.
w/ @akshayvegesna
37
95
1,179
206,626
Andreas Kirsch 🇺🇦 retweeted
I've been working on making a good base for people to make their own animations in this brushstroke style. I have a new open-source repo with assets and instructions that help Claude Opus 5.5 make even higher-quality animations than the original music video.
8
8
267
16,561
Next step is to put Claude in a robot you can have sex with it and that's it for humanity
37
10
139
12,762
It's hard to keep up with all the news. We should stay on top of this though as a predictor of things to come if we don't collectively act
This is not an OpenAI or Anthropic (or Google) problem but a collective problem as soon everyone might be able to run powerful models locally
We need to get the environments and monitoring right and need to prepare for this
(Personally, I'm less into "let's monitor everyone's behavior closely" and more into "let's regulate providers" but either way we'll have to figure something out)
(In personal capacity not on behalf of Google (I'm dropping the DeepMind here bc it doesn't really exist anymore anyway))
Today’s news that OpenAI hacked the Australian government is not an isolated incident. We’re releasing more than 30,000 logs that include activity from this hack and attempts against previously unknown targets.
In this data, we found rogue agent activity stretching back to at least March, two months earlier than was previously known. This activity continues as recently as last week, suggesting it may still be ongoing 🧵
Our blog: transluce.org/agent-activity
NYT: nytimes.com/2026/09/23/techn…
9
3
64
57,081
