Faculty @ MPI-SWS working on systems for Big Data and Machine Learning. Formerly Postdoc @ MIT CSAIL and PhD @ EPFL. Blockchain astronaut. Newbie entrepreneur.
Joined August 2012
- Tweets 401
- Following 2,063
- Followers 403
- Likes 173
24 Photos and videos
Have you ever been burned by an LLM on algorithm choice? The tests pass, you ship, and months later something melts under load because the model gave you non-scalable code?
In our new ICML 2026 paper, The Invisible Lottery, we ran almost 50k experiments and found that incidental prompt cues, a persona, a project name, even a color theme, systematically steer which algorithm a model writes. Most benchmarks focus on correctness and cannot surface these issues by design.
This is a big deal for vibe coders: when nobody reviews the generated code, this is no longer just about bugs but biased algorithmic choices.
Huge congratulations to Akanksha Narula and Mofasshara Binte Rafique, who ran every experiment and analyzed all the data.
Blog: binds.ch/blog/invisible-lott…
Preprint: binds.ch/papers/invisiblelot…
#MachineLearning #LLM #AICodeGeneration #ICML2026 #SoftwareEngineering
60
Everyone is freaking out about the latest "AGI is basically here" model.
My coping mechanism: I play chess against it. And then I crush it. Even with high reasoning turned on, no model has beaten me yet.
I am not a chess prodigy. I was decent in my youth, maybe 1700–1800 ELO at my peak, but that was a long time ago. These days I am just a dude with a keyboard.
Today's session: three games against #Mythos/#Fable. Won all three. In two of them, the model tried to cheat by playing non-legal moves.
These things can write, code, and explain things at a level that still surprises me. But ask one to track a board state for forty moves, and play by the rules, and it falls apart.
So when the hype gets too loud, try this: play a few games of chess against the model. Oddly therapeutic.
3
45
Laurent Bindschaedler retweeted
Jun 3
AI agents like @openclaw 🦞 are everywhere, answering emails, managing calendars, doing our chores for us
📣 REALM is back for year 2! Workshop for Research on Agent Language Models at #EMNLP2026, Budapest 🇭🇺
Stellar lineup ⬇️
📅 Submit by July 17, 23:59 AoE
#LLMAgents #NLProc
ALT REALM 2026 workshop invited speakers and panelists: Maria Brbić (EPFL), Vishrav Chaudhary (Meta Superintelligence Labs), Marc-Alexandre Côté (Microsoft Research), Manling Li (Northwestern University), Baptiste Rozière (Mistral AI), Graham Neubig (Carnegie Mellon University), Perouz Taslakian (ServiceNow AI Research).
4
15
40
5,420
Excited (and honestly a bit surprised) to be recognized as a Gold Reviewer for ICML 2026 🏅
It was my first time reviewing for ICML, and I definitely learned a lot along the way.
May 13
Reviewers & ACs for #ICML2026 have been recognized for their service!
- Reviewers: 4439 Gold (free registration), 4437 Silver. 17749 total reviewers were assigned >= 1 paper
- ACs: 1647 receive free registration, out of 1691 who were assigned >= 1 paper
TY for your hard work!
2
9
336
Thrilled to be opening the P1 Programs Workshop on Green AI in Copenhagen this week with a keynote on Deaths Per Token, a health-impact unit for generative AI infrastructure.
aicentre.dk/events/20260508-…
2
68
RT @levelsio: 100%
This is why you gotta make noise when you lose the little freedoms
As history shows repeatedly it's always a slippery…
108
🎤 I am giving the keynote at #EuroMLSys 2026 in Edinburgh on April 27.
“What Survives When Code Doesn't?"
If AI rewrites your code until the artifact is effectively disposable, what anchors the system? Intent, state, composition, effect: where do they go?
More from Edinburgh next week 🏴
1
9
223
New paper: "ClawVM: Harness-Managed Virtual Memory for Stateful Tool-Using LLM Agents"
We borrowed the oldest trick in the OS book and applied it to agent state. Turns out systems still has a lot to teach AI.
🦞 If your agent slowly forgets instructions or gets worse over time, you are not imagining it. Current harnesses manage memory with best-effort heuristics. ClawVM makes it deterministic.
Joint work with @mofasshara. Appearing at #EuroMLSys2026.
🎓 Paper: binds.ch/papers/clawvm2026.p…
💻 Code: github.com/mpi-dsg/clawvm
cc @OpenAI @AnthropicAI @steipete
1
2
189
New paper at ARCS 2026: "Rebooting Microreboot." We decouple planning from actuation so LLM-based remediation agents can help fix microservice incidents without making them worse. Typed actions, a small microkernel, 0% agent-caused harm online.
binds.ch/blog/rebooting-micr…
2
2
8
680
Pleased to announce that @aroraakhilcs and I are organizing a tutorial on “Hardening Agent Runtimes: Networking, Execution, State, Security (HARNESS)” at this year’s #SOSP conference in Prague on September 29.
If you are a systems researcher and you have been hearing about @OpenClaw, Claude Code, and the broader wave of AI agents, this tutorial is for you!
Details: harness.mpi-dsg.org
3
4
241
I have spent the past year obsessing over a question that will not leave me alone: what happens when generating code becomes trivially cheap but maintaining it does not?
Not the fairy tale. Not the requiem. The scenario where both are true at the same time.
A future retrospective, written from 2029 looking back.
2
616
What if maintaining code costs more than regenerating it from scratch?
That crossover is coming faster than you think. I call it the Rehydration Flip.
→ Forks become free
→ SaaS loses its moat
→ Engineers stop writing code, start verifying it
A cost law, a trilemma, and a theorem 👇
1
1
146
New paper at DOLAP 2026 🇫🇮
LLMs break SQL cache keys. Intent signatures bring caching back: 82% hit rate vs 28%, zero false hits.
Small puzzle piece in our work at the intersection of LLMs and DB 🧩
Presenting in Tampere myself; come say hi 👋
binds.ch/blog/semantic-cachi…
3
145
AI coding tools run with full access to your machine. Not just your project directory but also your AWS credentials, your 1Password vault, your personal documents. Everything.
I built a tool to fix this.👇 (1/5)
1
1
90
💡I wrote the first version for my own use last September after my student told me he was uncomfortable with Claude Code having potentially unfettered access to his machine. He was right. Been using it daily ever since. (4/5)
1
1
39
Open source, MIT licensed, beta (v0.1.0). macOS Linux.
brew install bindsch/tap/scode
Blog: binds.ch/blog/scode-sandbox-…
GitHub: github.com/bindsch/scode
(5/5)
1
42