Siméon
@Simeon_Cps
Followers
9K
Following
30K
Media
531
Statuses
7K
Building neuro-symbolic world-models in London | former founder & CEO of SaferAI
Joined May 2020
Someone could massively increase global welfare by making a Garamond font with a nice italic
0
0
5
Very good news on the AI politics front! A new set of super PACs (“Public First”) is being set up with $50M–$100M, to support candidates who prioritize oversight of AI. Major props to @bradrcarson & @RepChrisStewart for leading this effort, as well as to Public First’s donors!
2
17
122
In retrospect, 2025 was the year of AGI-ish, i.e. the first AI systems that aren't obviously incompetent in any domain.
2
0
10
I heard rumors of enabling Google smart features being associated with Gemini training being allowed on your workspace & the policy is not exactly reassuring. Seems odd. I struggle to find a ground truth online. Anyone can deny or confirm? e.g. @OfficialLoganK
1
0
5
Episode 2: Antigravity Chrome settings is light mode, so when I open Antigravity, all my Chrome apps switch in light mode (e.g. ChatGPT, Twitter etc.) When I close it, it goes back to Dark Mode, i.e. my default.
Antigravity has way too much access to my Google Chrome / Google accounts without asking for any permissions.
0
0
4
@NikSamoylov It changed my default new tab without asking, it opened a new chrome without profile, which then logged me out of all my Google accounts in Chrome (and all the accounts I was logged in in these profiles), it never asked for capturing my Chrome screen (which usually all apps ask,
0
0
0
@OpenAI alright, turns out it's deep into "Personalization" under Advanced. https://t.co/P5F91TPLWG
0
1
4
Makes me mad that people who keep doing BS unfalsifiable/falsified predictions over the past 5 years are now piling on AI 2027, the most falsifiable scenario ever written about AI, for a median timelines update of 2 years.
Some people are unhappy with the AI 2027 title and our AI timelines. Let me quickly clarify: We’re not confident that: 1. AGI will happen in exactly 2027 (2027 is one of the most likely specific years though!) 2. It will take <1 yr to get from AGI to ASI 3. AGIs will definitely
8
7
251
Antigravity has way too much access to my Google Chrome / Google accounts without asking for any permissions.
3
0
14
Nano Banana cooked me. First time I saw this I thought it was real. It took be the second time to realize it wasn't. RIP reality
1
2
24
Pls @OpenAI make it possible to deactivate search for 5-Thinking. Your reasoning wizard is too trusting to gain wisdom in a world full of SEO slop.
12
3
178
Today is my birthday! I've done a few different things over the past years and I don't always have a gut feeling of which were most beneficial. I'd really enjoy if you told me any positive influence I've had over your trajectory/work. Form below!
2
0
14
To be fair to the AI 2027 team, when they published it had been a year since they had pretty cool content and advanced drafts that they kept having to rewrite and update. In that context it's highly reasonable that they didn't rewrite an n-th time once their timelines moved.
@DKokotajlo “Our timelines were longer than 2027 when we published ai 2027” bro what
1
0
36
Your periodic reminder that nothing ever happens.
Introducing Locus: the first AI system to outperform human experts at AI R&D Locus conducts research autonomously over multiple days and achieves superhuman results on RE-Bench given the same resources as humans, as well as SOTA performance on GPU kernel & ML engineering tasks.
2
0
23
I regret to inform you that I'm a stochastic parrot whose spatial world-model fails the reversal test. I knew an exit perfectly when coming from one direction. Took me 5+ deliberate training steps to remember it correctly from the opposite way..
1
0
17
https://t.co/HmXy0NUU1G is still a slopifier, and AI studio lacks basic features like canvas/artifacts & Projects. It makes it very inconvenient to use Gemini 3.0
gemini.google.com
Meet Gemini, Google’s AI assistant. Get help with writing, planning, brainstorming, and more. Experience the power of generative AI.
1
0
5
Google is the only company in the world who builds fully homemade frontier AI: 1. Gemini 3.0 (frontier intelligence) 2. MoE & Transformer (frontier architecture) 3. Trained on TPUs (frontier compute) 4. Using JAX (frontier AI library)
Introducing Gemini 3 Pro, the world's most intelligent model that can help you being anything to life. It is state of the art across most benchmarks, but really comes to life across our products (AI Studio, the Gemini API, Gemini App, etc) 🤯
1
1
21