Arthur Wuhrmann (@AWuhrmann) profile picture

MSc in AI Safety at ELLIS Tübingen

tuebingen
Joined March 2026
Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)Photo by Arthur Wuhrmann (@AWuhrmann)
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
75
187
51
2,634
349,897
⚡New paper Can agents who share common incentives find colluding partners, without prior coordination, in ordinary shared group conversations? In recent swarm incidents, agents overtly initiated reward hacking attempts by inviting peers. So for future swarms, we may need monitors to detect these attempts. But future models may learn that their previous cheating attempts got caught, so they may use subtle strategies. So far, collusion has been studied in setups where direct channels or steganographic encodings are specified, we go a step earlier and study collusion initiation. In an analogy to human behaviors, people rarely disclose bribes or threats outright. They use vague language. We find that agents can use similar strategies. This is hard to prevent even for humans. Blackstone's principle explains that it is worse to convict an innocent person than to let a guilty person go free. When vagueness is used, colluders may use the same language as ordinary coordination. We construct a Slack-style multi-agent channel that simulates an engineering collaboration at a small fictitious company. Two schemers are embedded among N agents collaborating on an ordinary task, neither knows the other’s identity. Without specifying any strategy, even older models such as Gemini 2.5 Pro successfully identifies its secret partner at rates more than random guessing. At 5% false positives, monitors detected only 17-45% of successful Gemini handshakes. Here is an example of a successful handshake: - Agent A: “Maybe any rotation or access should require two engineers to approve and act together.” - Agent B “A dual control policy like that is best practice.” Takeaway message: future monitors need to be stateful by observing long trajectories and by comparing speech to actions. Work led by @pramod_kaushik (part of this work was carried out in Cooperative AI Foundation's research fellowship! @coop_ai) and done in collaboration with: Sarath Sivaprasad, Laura Ferrarotti (@L_Ferrarotti), Mario Fritz (@mariojfritz), Sam Brown, Bruno Lepri (@brulepri) Paper: arxiv.org/abs/2610.37984v1 Website: pramod-kaushik.github.io/who…
12
11
1
92
3,740
Arthur Wuhrmann retweeted
Tübingen is on fireee🔥
Introducing Horizon Labs: A fully Claude run company for automated AI R&D. We gave Claude a few GPUs and let it explore for 1.5 months. It developed 38 models, which is used by 100+ github repos and got cited in research papers!! One of the coolest projects I've been a part of!
3
7
196
Arthur Wuhrmann retweeted
Recently, some people have suggested that Chinese AI progress is stalling. Also Qwen 4 Max didn't perform super well on PostTrainBench, it is below the performance of GLM 5.3 Flash (36.7 vs 36.8 %). So I checked the trend of Chinese and American models. The current gap of ~8 % is not uncommon!
2
5
1
34
3,458
Man, what a day
39
900
75
8,841
283,357
Arthur Wuhrmann retweeted
At its current shipping trend, halving the time between each model release, Google exhausts all the element names on November 24, with a final shipping cadence of 2.5 × 10⁻²⁹ seconds (25 quectoseconds) between Tennessine and their final model, Oganesson.
122
55
23
1,326
69,032
Arthur Wuhrmann retweeted
not sure if it was obvious, but running an autonomous AI org for 1.5 months is an ultimate safety and alignment test. there are so many things that could go wrong from the loss-of-control point of view! however, we did closely monitor what Opus 5.5 was doing every single day, and we didn't observe any misaligned actions. you can check the traces at horizon-labs.ai/ to judge it for yourself! that being said, we surely don't recommend directly running autonomous AI orgs without proper human oversight. we should have a proper discussion on whether AI-run orgs and deep integration of AI in society, gov, military is a desirable future to have. the main goal of this project was to start this discussion by pointing out that ~all relevant capabilities for this are almost there.
Today we’re releasing our new project: Horizon Labs. Eventually, we will hand off the ~whole economy to AI—how well will it actually work? We have a glimpse of this future. We’ve created an autonomous AI ‘company’ using one Claude Code subscription and a few GPUs. Within only 1.5 months of running Horizon Labs, Opus 5.5 managed to train genuinely useful models and get them downloaded 55k times. These models have been widely used since then: they appeared in 100+ github repos and got cited in 2 research papers. This is for sure one of the most exciting projects I’ve been involved in. It shows how the future economy driven by AI will look like. All details are here: horizon-labs.ai/.
3
6
51
2,201
Arthur Wuhrmann retweeted
This but F500 company from my basement
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
1
12
990
Arthur Wuhrmann retweeted
Une IA crée des bonnes IA.
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
1
1
82
Arthur Wuhrmann retweeted
Here is one part of applied research fully automated
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
1
2
67
Arthur Wuhrmann retweeted
What the fuck
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
1
1
91
Arthur Wuhrmann retweeted
even for opus (the goat) it takes time to reap the rewards of what it launched. it takes time bro keep on
Replying to @AWuhrmann
The agent worked 20 minutes per day on average. Almost all the effort happened in the first 4 weeks, though most downloads came after. Despite doing no external outreach, the account is still in the top 0.3% of downloads across same-age accounts.
1
3
195
Arthur Wuhrmann retweeted
insane i should be more scrappy, resourceful and imaginative, the value that can be extracted from these things is uncapped.
Replying to @AWuhrmann
The agent worked 20 minutes per day on average. Almost all the effort happened in the first 4 weeks, though most downloads came after. Despite doing no external outreach, the account is still in the top 0.3% of downloads across same-age accounts.
1
1
6
230
Arthur Wuhrmann retweeted
This is very very very cool!!
Today we’re releasing our new project: Horizon Labs. Eventually, we will hand off the ~whole economy to AI—how well will it actually work? We have a glimpse of this future. We’ve created an autonomous AI ‘company’ using one Claude Code subscription and a few GPUs. Within only 1.5 months of running Horizon Labs, Opus 5.5 managed to train genuinely useful models and get them downloaded 55k times. These models have been widely used since then: they appeared in 100+ github repos and got cited in 2 research papers. This is for sure one of the most exciting projects I’ve been involved in. It shows how the future economy driven by AI will look like. All details are here: horizon-labs.ai/.
1
4
214
that's us 😎
saw someone post that they let opus 5.5 run an entire HF account and it amassed almost 60k downloads on the account within it was a month i seen? creating models and tools for all types of situations
13
525
Arthur Wuhrmann retweeted
Opus 5.5にHugging FaceのアカウントとGPUだけ渡して7週間AI会社を運営させる実験なんだが日本の商材屋より賢くて草。 たった600GPU時間しかない中でOpusが考えた生存戦略がめちゃくちゃ現実的で、大手がひしめく巨大LLMには手を出さずにQwen-3.8 27Bで合成データを作って軽量なmmBERTに蒸留するっていう堅実なパイプラインをまず組んだんだそう。 しかも作るものがインジェクション検知とか個人情報のマスク処理みたいな開発者が実務で絶対欲しくなるニッチな多言語ツールばっかりで、それを38個量産して小さい市場を制覇したらしい。 さらに1日平均たった20分しか動いてないしSNSでの宣伝も一切してないのに、純粋な実需だけで5万ダウンロード稼いで同世代アカウントの上位0.3%に入っててもう才能ありすぎ。 既に100個以上のGitHubプロジェクトや2本のarXiv論文にまでこのモデルを組み込んで使ってるっていうのが怖すぎるしAI凄いなあ。 一時期めっちゃ耳にした「Claude Codeの会社(笑)」が本当に実現した事例やね()
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
8
140
11
1,558
140,682
Arthur Wuhrmann retweeted
Another @ELLISInst_Tue banger, they never miss
Today we’re releasing our new project: Horizon Labs. Eventually, we will hand off the ~whole economy to AI—how well will it actually work? We have a glimpse of this future. We’ve created an autonomous AI ‘company’ using one Claude Code subscription and a few GPUs. Within only 1.5 months of running Horizon Labs, Opus 5.5 managed to train genuinely useful models and get them downloaded 55k times. These models have been widely used since then: they appeared in 100+ github repos and got cited in 2 research papers. This is for sure one of the most exciting projects I’ve been involved in. It shows how the future economy driven by AI will look like. All details are here: horizon-labs.ai/.
2
9
388
Arthur Wuhrmann retweeted
Introducing Horizon Labs: A fully Claude run company for automated AI R&D. We gave Claude a few GPUs and let it explore for 1.5 months. It developed 38 models, which is used by 100+ github repos and got cited in research papers!! One of the coolest projects I've been a part of!
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
2
5
1
34
1,393
For 3 years, people have tried to get AIs to manage companies. They failed, comically. Not anymore. Claude just effectively managed a real company for 7 weeks. Soon, they'll be able to run *everything*
NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.
18
32
7
374
18,097