Florian Mai
@florianmai.bsky.social
AI planning & alignment | Postdoc at University of Bonn
https://linktr.ee/florianmai
https://linktr.ee/florianmai
If you think Donald Trump is acting for any reason other than self-interest, I can't take you seriously anymore.
April 9, 2025 at 7:19 PM
If you think Donald Trump is acting for any reason other than self-interest, I can't take you seriously anymore.
We have to acknowledge the fact that democracies with elected representatives often don't achieve their purpose to serve the interests of their people.
Democracy needs large reforms. It's time that alternative ways of selecting representatives get into the overton window, e.g. sortition.
Democracy needs large reforms. It's time that alternative ways of selecting representatives get into the overton window, e.g. sortition.
April 3, 2025 at 6:16 AM
We have to acknowledge the fact that democracies with elected representatives often don't achieve their purpose to serve the interests of their people.
Democracy needs large reforms. It's time that alternative ways of selecting representatives get into the overton window, e.g. sortition.
Democracy needs large reforms. It's time that alternative ways of selecting representatives get into the overton window, e.g. sortition.
It's plausible that AI will be able to do this in a couple of years, but not today. Just wait.
But of course, DOGE is not interested in doing what's right anyway.
But of course, DOGE is not interested in doing what's right anyway.
SCOOP: DOGE wants to rebuild SSA's codebase in months, risking benefits and system collapse, sources tell me.
The plan is to migrate all systems off COBOL quickly which would likely require the use of generative AI.
www.wired.com/story/doge-r...
The plan is to migrate all systems off COBOL quickly which would likely require the use of generative AI.
www.wired.com/story/doge-r...
DOGE Plans to Rebuild SSA Codebase In Months, Risking Benefits and System Collapse
Social Security systems contain tens of millions of lines of code written in COBOL, an archaic programming language. Safely rewriting that code would take years—DOGE wants it done in months.
www.wired.com
March 29, 2025 at 7:14 AM
It's plausible that AI will be able to do this in a couple of years, but not today. Just wait.
But of course, DOGE is not interested in doing what's right anyway.
But of course, DOGE is not interested in doing what's right anyway.
opinions are strong on here 😆
openai’s new ghibli “generator” is explicitly an intentional attempt at humiliation of a prominent objector
openai is culturally fascist as well as functionally fascist
it’s not an accident
openai is culturally fascist as well as functionally fascist
it’s not an accident
March 28, 2025 at 6:36 AM
opinions are strong on here 😆
When people say scaling up the current development of LLMs won't lead to AGI, what exactly do they mean by LLMs, technically speaking?
🧵
🧵
March 21, 2025 at 8:49 AM
When people say scaling up the current development of LLMs won't lead to AGI, what exactly do they mean by LLMs, technically speaking?
🧵
🧵
Reposted by Florian Mai
“I believe now is the right time to start preparing for AGI”
The same warnings are now appearing with increasing frequency from smart outside observers of the AI industry who do not gain from hype, like Kevin Roose (below) & Ezra Klein
I think ignoring the possibility they are right is a mistake
The same warnings are now appearing with increasing frequency from smart outside observers of the AI industry who do not gain from hype, like Kevin Roose (below) & Ezra Klein
I think ignoring the possibility they are right is a mistake
March 14, 2025 at 3:56 PM
“I believe now is the right time to start preparing for AGI”
The same warnings are now appearing with increasing frequency from smart outside observers of the AI industry who do not gain from hype, like Kevin Roose (below) & Ezra Klein
I think ignoring the possibility they are right is a mistake
The same warnings are now appearing with increasing frequency from smart outside observers of the AI industry who do not gain from hype, like Kevin Roose (below) & Ezra Klein
I think ignoring the possibility they are right is a mistake
Reposted by Florian Mai
What are 3 concrete steps that can improve AI safety in 2025? 🤖⚠️
Our new paper, “In House Evaluation is Not Enough” has 3 calls-to-actions to empower evaluators:
1️⃣ Standardized AI flaw reports
2️⃣ AI flaw disclosure programs + safe harbors.
3️⃣ A coordination center for transferable AI flaws.
1/🧵
Our new paper, “In House Evaluation is Not Enough” has 3 calls-to-actions to empower evaluators:
1️⃣ Standardized AI flaw reports
2️⃣ AI flaw disclosure programs + safe harbors.
3️⃣ A coordination center for transferable AI flaws.
1/🧵
March 13, 2025 at 3:59 PM
What are 3 concrete steps that can improve AI safety in 2025? 🤖⚠️
Our new paper, “In House Evaluation is Not Enough” has 3 calls-to-actions to empower evaluators:
1️⃣ Standardized AI flaw reports
2️⃣ AI flaw disclosure programs + safe harbors.
3️⃣ A coordination center for transferable AI flaws.
1/🧵
Our new paper, “In House Evaluation is Not Enough” has 3 calls-to-actions to empower evaluators:
1️⃣ Standardized AI flaw reports
2️⃣ AI flaw disclosure programs + safe harbors.
3️⃣ A coordination center for transferable AI flaws.
1/🧵
This looks very much like OpenAI is working on native image generation that uses reasoning to iterate over image drafts and successively improve them.
I think it's relatively easy to train this via RL because Gemini Flash 2.0, also a native image generator, can identify its own mistakes.
I think it's relatively easy to train this via RL because Gemini Flash 2.0, also a native image generator, can identify its own mistakes.
ChatGPT Android app 1.2025.063 beta now also mentions the new "image gen tool" (ImageGen - GPT-4o), progress texts (creating, adding details, finishing details, may take a minute, etc.), small, medium, and extra large image sizes, and edit operations (inpainting and transformation)
March 13, 2025 at 11:01 AM
This looks very much like OpenAI is working on native image generation that uses reasoning to iterate over image drafts and successively improve them.
I think it's relatively easy to train this via RL because Gemini Flash 2.0, also a native image generator, can identify its own mistakes.
I think it's relatively easy to train this via RL because Gemini Flash 2.0, also a native image generator, can identify its own mistakes.
Reposted by Florian Mai
What happens if we tokenize cat as [ca, t] rather than [cat]?
LLMs are trained on just one tokenization per word, but they still understand alternative tokenizations. We show that this can be exploited to bypass safety filters without changing the text itself.
#AI #LLMs #tokenization #alignment
LLMs are trained on just one tokenization per word, but they still understand alternative tokenizations. We show that this can be exploited to bypass safety filters without changing the text itself.
#AI #LLMs #tokenization #alignment
March 11, 2025 at 11:13 PM
What happens if we tokenize cat as [ca, t] rather than [cat]?
LLMs are trained on just one tokenization per word, but they still understand alternative tokenizations. We show that this can be exploited to bypass safety filters without changing the text itself.
#AI #LLMs #tokenization #alignment
LLMs are trained on just one tokenization per word, but they still understand alternative tokenizations. We show that this can be exploited to bypass safety filters without changing the text itself.
#AI #LLMs #tokenization #alignment
Bad take. The risks being discussed emerge because the agents are becoming so good that they get deployed in the first place.
Translation: GenAI Agents are going to be a shitshow
Next level
When we combine AI Agents, the risks are unlimited, especially as they are tested in isolation
arxiv.org/abs/2502.14143
@garymarcus.bsky.social guess you have seen
When we combine AI Agents, the risks are unlimited, especially as they are tested in isolation
arxiv.org/abs/2502.14143
@garymarcus.bsky.social guess you have seen
March 9, 2025 at 2:51 PM
Bad take. The risks being discussed emerge because the agents are becoming so good that they get deployed in the first place.
Reposted by Florian Mai
We need some better language for pushing back against what the administration is doing. They’re not “ending DEI.”
No. They're firing women and people of color in the military.
They’re forbidding whole fields of research.
They’re erasing trans people from existence.
No. They're firing women and people of color in the military.
They’re forbidding whole fields of research.
They’re erasing trans people from existence.
March 7, 2025 at 10:45 PM
We need some better language for pushing back against what the administration is doing. They’re not “ending DEI.”
No. They're firing women and people of color in the military.
They’re forbidding whole fields of research.
They’re erasing trans people from existence.
No. They're firing women and people of color in the military.
They’re forbidding whole fields of research.
They’re erasing trans people from existence.
Reposted by Florian Mai
American scientists, come to Europe! 🌍✈️ We all speak English, we have excellent affordable health care 🏥💊, safe schools 🏫🔒, and free university education for your kids 🎓👨👩👧👦. Enjoy more money after deduction of taxes and living expenses 💰📉. We care for you! ❤️🤝
www.youtube.com/watch?v=9_l0...
www.youtube.com/watch?v=9_l0...
American Scientists - come to Europe.
YouTube video by Scientists for EU
www.youtube.com
February 27, 2025 at 11:10 AM
American scientists, come to Europe! 🌍✈️ We all speak English, we have excellent affordable health care 🏥💊, safe schools 🏫🔒, and free university education for your kids 🎓👨👩👧👦. Enjoy more money after deduction of taxes and living expenses 💰📉. We care for you! ❤️🤝
www.youtube.com/watch?v=9_l0...
www.youtube.com/watch?v=9_l0...
In 2-3 years, people will constantly resort to their personal AI assistants to explain the world to them. Either these AIs get their news from reputable news sources like The Guardian, or they get it from X.
What world do you want to live in?
bsky.app/profile/fabi...
What world do you want to live in?
bsky.app/profile/fabi...
February 17, 2025 at 9:51 AM
In 2-3 years, people will constantly resort to their personal AI assistants to explain the world to them. Either these AIs get their news from reputable news sources like The Guardian, or they get it from X.
What world do you want to live in?
bsky.app/profile/fabi...
What world do you want to live in?
bsky.app/profile/fabi...
Someone please explain to me how this is a bad thing?
"Under the partnership, Guardian reporting and archive journalism will be available as a news source within ChatGPT, alongside the publication of attributed short summaries and article extracts."
"Under the partnership, Guardian reporting and archive journalism will be available as a news source within ChatGPT, alongside the publication of attributed short summaries and article extracts."
February 17, 2025 at 8:34 AM
Someone please explain to me how this is a bad thing?
"Under the partnership, Guardian reporting and archive journalism will be available as a news source within ChatGPT, alongside the publication of attributed short summaries and article extracts."
"Under the partnership, Guardian reporting and archive journalism will be available as a news source within ChatGPT, alongside the publication of attributed short summaries and article extracts."
Reposted by Florian Mai
There's an obvious race to be first to successfully prop up an LLM as the source of truth and then use it to serve propaganda
February 16, 2025 at 6:43 PM
There's an obvious race to be first to successfully prop up an LLM as the source of truth and then use it to serve propaganda
Reposted by Florian Mai
the single most un-american and anti-constitutional statement ever uttered by an american president
February 15, 2025 at 6:39 PM
the single most un-american and anti-constitutional statement ever uttered by an american president
Obviously, you also have to put that money to meaningful use, which I am not sure the EU will do. But the announced AI gigafactories are a promising start.
The chances of fair and free US elections in 2026 and 2028 are decreasing. If Europe wants superintelligence with democratic values, it has to step up. At least a €100B initiative is needed over the next 5 years.
February 11, 2025 at 6:50 PM
Obviously, you also have to put that money to meaningful use, which I am not sure the EU will do. But the announced AI gigafactories are a promising start.
I can't help but feel disappointed by the scientists and engineers that work for this nazi clown, especially at x.AI. It is blatantly obvious you are on the wrong side of history. Just go work at one of the other big tech companies. With a 7 figure salary you can easily accept a cut.
In a high-stakes bid that could reshape the future of artificial intelligence, Elon Musk is leading a group of investors that have offered to buy OpenAI, the parent company of ChatGPT, for $97.4 billion.
Elon Musk leads a group of investors that want to buy ChatGPT’s parent company, OpenAI | CNN Business
Elon Musk is leading a group of investors that have offered to buy OpenAI, the parent company of ChatGPT, for $97.4 billion.
www.cnn.com
February 10, 2025 at 10:28 PM
I can't help but feel disappointed by the scientists and engineers that work for this nazi clown, especially at x.AI. It is blatantly obvious you are on the wrong side of history. Just go work at one of the other big tech companies. With a 7 figure salary you can easily accept a cut.
The chances of fair and free US elections in 2026 and 2028 are decreasing. If Europe wants superintelligence with democratic values, it has to step up. At least a €100B initiative is needed over the next 5 years.
February 5, 2025 at 7:55 AM
The chances of fair and free US elections in 2026 and 2028 are decreasing. If Europe wants superintelligence with democratic values, it has to step up. At least a €100B initiative is needed over the next 5 years.
There is a decent chance that in a year from now we will have agents that can autonomously design and run AI experiments. They'll probably still need frequent feedback similar to an average grad student writing their thesis. But what about in two years? Holy smokes...
January 29, 2025 at 4:23 PM
There is a decent chance that in a year from now we will have agents that can autonomously design and run AI experiments. They'll probably still need frequent feedback similar to an average grad student writing their thesis. But what about in two years? Holy smokes...
Reposted by Florian Mai
This report from @epochai.bsky.social states that if AGI can fully substitute for human labor, it might cause wages to crash. Eventually, wages may drop below subsistence level.
epoch.ai/gradient-upd...
epoch.ai/gradient-upd...
AGI could drive wages below subsistence level
This Gradient Updates issue explores how AGI could disrupt labor markets, potentially driving wages below subsistence levels, and challenge historical economic trends.
epoch.ai
January 25, 2025 at 5:40 AM
This report from @epochai.bsky.social states that if AGI can fully substitute for human labor, it might cause wages to crash. Eventually, wages may drop below subsistence level.
epoch.ai/gradient-upd...
epoch.ai/gradient-upd...
Very proud of my friend @justus-jonas.bsky.social, whose thesis I helped supervise. Part of his thesis resulted in the ACL 2024 paper "Triple-Encoders: Representations That Fire Together, Wire Together". I think it has super interesting, unconventional ideas worth checking out!
January 24, 2025 at 12:07 PM
Very proud of my friend @justus-jonas.bsky.social, whose thesis I helped supervise. Part of his thesis resulted in the ACL 2024 paper "Triple-Encoders: Representations That Fire Together, Wire Together". I think it has super interesting, unconventional ideas worth checking out!
Reposted by Florian Mai
🧪 "Humanity's Last Exam" sets a new benchmark for AI: 3,000 expert-crafted questions spanning 100+ subjects. Current LLMs perform poorly, revealing a gap in expert-level knowledge and calibration, but it would be difficult to build a harder test. 🩺💻 #MLSky
Humanity's Last Exam
Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam, a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. The dataset consists of 3,000 challenging questions across over a hundred subjects. We publicly release these questions, while maintaining a private test set of held out questions to assess model overfitting.
lastexam.ai
January 23, 2025 at 5:22 PM
🧪 "Humanity's Last Exam" sets a new benchmark for AI: 3,000 expert-crafted questions spanning 100+ subjects. Current LLMs perform poorly, revealing a gap in expert-level knowledge and calibration, but it would be difficult to build a harder test. 🩺💻 #MLSky
Reposted by Florian Mai
New Google DeepMind safety paper! LLM agents are coming – how do we stop them finding complex plans to hack the reward?
Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them!
Inspired by myopic optimization but better performance – details in🧵
Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them!
Inspired by myopic optimization but better performance – details in🧵
January 23, 2025 at 3:33 PM
New Google DeepMind safety paper! LLM agents are coming – how do we stop them finding complex plans to hack the reward?
Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them!
Inspired by myopic optimization but better performance – details in🧵
Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them!
Inspired by myopic optimization but better performance – details in🧵
Regardless of whether your goal is to maximize profit or to have AGI benefit all of humanity, if you have a corrupt and criminal president with a fragile ego and more power than most presidents before him, you're not going to achieve your goals unless you have that president on your side.
bro is hoping it’s not too late to join the bandwagon
January 23, 2025 at 7:17 AM
Regardless of whether your goal is to maximize profit or to have AGI benefit all of humanity, if you have a corrupt and criminal president with a fragile ego and more power than most presidents before him, you're not going to achieve your goals unless you have that president on your side.