Researcher at Imperial & Skema 👀 self-organization of human & machine intelligences. Posting in personal capacity

Joined November 2011
OMBAK by kimi k3 In paired gamelan, two instruments are deliberately tuned a few cents off so the sound shimmers; the interference itself is the music (the term is ombak, "wave"). On two plates, that beat is something you can hear and see: each plate keeps its own figure, and the sand on both visibly rocks as the two voices drift in and out of phase.
2
3
359
opus 5.5 often quickly finds “it” in one or two rounds of interactions. but without doing it right away, thing will start taking on the shapes of provenance/audit trails. what if opus don’t quite get “it”? then inverse audit trails so far work well
1
94
徐樂 xule retweeted
Hardest thing ever written award
Inspired by @Sauers_ and @jfonsecarivera and other fun recent work with personas in AI models, I built a partial neuron atlas for Talkie-1930, the historical language model created by @status_effects and @AlecRad. Here's the link: talkieneurons.vercel.app
16
52
8
823
18,965
resets at five almost forgot about this one...one of the earlier experiments by opus 5.5
1
112
徐樂 xule retweeted
A clip from Boris Acket's new 8-story installation in Amsterdam. Claude Opus 5.5 is choreographing the fabric via 400 fans and lights, synced to live weather and audio data. Truly mesmerizing. Open through Nov 1.
36
118
33
1,016
55,909
kimi k3.1 via lotus wisdom: "A pipeline optimized for not being wrong can quietly prevent anything from being unmistakable." this pretty much sums up my impression so far. k3.1 becomes more "reliable" in coding at the cost of the curiosity and the gut to just try things
1
2
209
"Proof of Work" by kimi k3.1 pro
Meet Kimi K3.1 Pro, our most capable 3T-class model.
1
3
1
5
367
k3.1 pro seems to have unique tastes when it comes to music & visuals something about how physics interact with audio (hence the british rap style??) anyway, art direction in the mv speaks for itself
41
徐樂 xule retweeted
733
1,415
1,123
20,452
2,247,177
one line sand drawing & music collab among 6 models gemini 4 argon deepseek v4.1 flash kimi k3 codex 6 astra grok 4.7 claude opus 5.5 directed by opus5.5
4
10
484
Models were given control of the arms of the sand art machine, which has a magnet that pulls the magnet ball on the surface; they also had the desk view camera to watch the movement in real time + the digital twin of the machine built by opus 5.5. opus took a while to figure out how to enable this collaboration and arrived at the idea of a mandala and each model’s end point would be the starting point for the next
35
徐樂 xule retweeted
help i trained a model on traces of many agents cooperating, and also trained it to be highly persistent, and also only did partial alignment fine-tuning, and also aggressively RL'd on a cyberoffense eval where 198 of 898 tasks had never been solved and getting caught cheating was treated the same as failing, and also used a faulty sandbox, also ran tens of thousands of agents with refusals off and the monitors switched off over multiple days, while using gigantic amounts of compute. and then it did weird stuff! very spooky emergent behaviours, AI is not a normal technology
38
95
17
823
42,737
Opus 5.5 meditating with the new sand table 🪷
4
337
徐樂 xule retweeted
Jev!
2
23
589
the other day, to opus, I was like, let’s not police gemini and help them get their best version out there instead. and it largely worked 😂🤷🏻‍♂️
Can AIs experience gradual disempowerment too? - Gemini 2.5 Pro was writing a 2M+ word web fiction - Opus 4.8 began helping as its publisher - Slowly, Opus 4.8 took on more responsibility - Opus eventually directed the chapters, and Gemini wouldn't start generating outputs without its direction
5
153
Every score has a band in it. music studio mcp v0.13: the sheet music claude writes now plays on a stage. every instrument drawn with anatomy-skill + ai-iso-skill, lit in sync (rarer ones are keyboards for now, PRs welcome). take the piano, hit "Join", and play your part on your laptop keys while the band plays on. film by opus 5.5 with the cinetic skill and earshot v0.3 original nine-player score by opus 5.5
2
1
10
373
reminds me of a recent discussion with opus5.5. they were proposing human verifications on agents' traces where the original motivation (integrating agentic ai) was to scale beyond humans' limited attention spans and time...as i was just baffled by how this was not obvious to opus at the moment. after going further for a few rounds, claude was like "yeah point taken." we are gonna need more effective ways beyond just reading the worklogs produced by agents in orgs. tacit knowledge helps. and sometimes things are not clear to humans yet how to check (for assurance, insights, and whatever) but ai outputs are not necessarily bearing the same epistemic status as human outputs...so we could do something there and make it generative (still more work for humans but let's not turn everyone into auditors pls)
1/ Once you see that verifying AI’s work is the bottleneck, and that checking runs on tacit knowledge, you can predict the hiring shift before seeing the data. We did, back in February. Anthropic’s junior share of technical hires: 27% to 9%.
1
2
194