Design for a world with safe AI – @apolloaievals, @bluedotimpact, @epochairesearch, @horizonIPS.
Berlin
Joined July 2008
- Tweets7.5K
- Following4.4K
- Followers928
- Likes10.2K
Pinned Tweet
In a centralized world, the highest form of system design is public policy plus its implementation.
In a decentralized one it’s probably Ethereum.
Then: Designers kept dreaming up "just one more feature" until the devs went grey.
Now: Designers vibe-code slop straight into prod until the devs go bald.
Some things never change…
Insane for education. Next step, world models.
Imagine every pixel on your screen, streamed live directly from a model. No HTML, no layout engine, no code. Just exactly what you want to see.
@eddiejiao_obj, @drewocarr and I built a prototype to see how this could actually work, and set out to make it real. We're calling it Flipbook. (1/5)
What I'm talking about when I tell people China is doing more for AI safety than the US, because rogue AI is considered a direct natSec threat to the Politbüro – and why the whole "if we don't do it, China will" storyline is just a smokescreen by the usual careless people.
Great visualization.
Frontier AI data center capacity, based on data collected by @EpochAIResearch. It doesn't include every single one, as they focus on the largest ones.
Few things stand out:
- 2026 will have a huge amount of cpacity come online
- Anthropic will lead at some early points in 2026
- 2027 onwards OpenAI has the most projected capacity at the moment, by far
🤖 Alexander Haase retweeted
Future AIs might secretly pursue unintended goals — “scheme”.
In a collaboration with OpenAI, we tested a training method to reduce existing versions of such behavior.
We see major improvements, but they may be partially explained by AIs knowing when they are evaluated.
🤖 Alexander Haase retweeted
1) The AI Whistleblower Initiative
They provide expert advise & pro bono legal help to concerned insiders at AGI labs
Whistleblowers have already proven important
My experience on the GPT-4 Red Team makes me particularly passionate about this one!
xitter.cf/AIWI_Official/status/1…
Call Launch: In just a few minutes at National Whistleblower Day on Capitol Hill, AIWI and its coalition—represented by Karl Koch in the AI panel—will launch a call with one clear ask:
AI companies must publish their whistleblowing policies.
PublishYourPolicies.org
A thread🔽
🤖 Alexander Haase retweeted
Imagine if ChatGPT highlighted every word it wasn't sure about.
We built a streaming hallucination detector that flags hallucinations in real-time.
My brain these days: Great, now we just need to figure out how to tax the shrinking pool of workers even harder to sustain all these new conservatives, since taxing billionaires properly never quite worked out.
Um… comparing a +65y old logo nobody remembers with a famous rebrand that’s still alive is the difference between stealing bread fresh from a bakery vs. digging up a moldy ’60s sandwich even the rats gave up on.
The Perplexity logo originated in 1967 by the University of New York State.
It represented a "radial cluster of books" which "symbolized knowledge".
Sound familiar to a logo Mike created?
Yo @perplexity_ai please remove your logo.
Still one of the most sickest logo designs I have seen in the last years that captures the essence of a brand in a way words cannot describe.
I feel like everyone just tried to design globes since this thing went live on brandnew.
Exposure - this logo was one of the first that went up on brand new and people actually took us seriously as a studio. It’s also our most ripped off logo by far.
cc: @lukesbeard thanks for filing all the trademarks dude.
medium.com/smith-diction/wor…
The amount of boredom Christian monks must have felt can probably be extrapolated from the size of the initials they used in their books.
🤖 Alexander Haase retweeted
what are large language models actually doing?
i read the 2025 textbook "Foundations of Large Language Models" by tong xiao and jingbo zhu and for the first time, i truly understood how they work.
here’s everything you need to know about llms in 3 minutes↓
🤖 Alexander Haase retweeted
We built the first taxonomy of values expressed by AIs in real-world conversations.
Broadly, Claude expresses the values we intended, like critical thinking, responsibility, and efficiency.
ALT A taxonomy of AI values. At the top of the hierarchy (in red) are the five overall categories, along with the percentage of conversations that included them. In yellow are subcategories at a lower level of the hierarchy. In blue are some selected individual values (only a selection are shown due to space constraints).
