Marc Lanctot reposted
As Meta employees brace for layoffs next Wednesday, May 20, many say the vibes are horrifically, historically low.
Who's Who of AI
Marc Lanctot
Research Scientist at Google DeepMind, multiagent RL
Research Scientist at Google DeepMind, interested in multiagent reinforcement learning, game theory, games, and search/planning.
Lover of Linux 🐧, coffee ☕, and retro gaming. Big fan of open-source. #gohabsgo 🇨🇦
For more info: https://linktr.ee/sharky6000
What they're sharing
Meta’s New Reality: Record High Profits. Record Low Morale | WIRED wired.com
View on Bluesky →
Meta CTO Andrew Bosworth Admits the Company’s AI Reorg Was ‘Atrocious’ wired.com
View on Bluesky →
Human-like autonomy emerges from self-play and a pinch of human data arxiv.org
Whimsical Strategies Break AI Agents: Generating Out-of-Distribution Adversarial Strategies at Scale - Microsoft Research microsoft.com
View on Bluesky →
Human Adults and LLMs as Scientists: Who Benefits from Active Exploration? arxiv.org
Solipsistic Superintelligence is Unlikely to be Cooperative arxiv.org
View on Bluesky →
Google DeepMind and partners announce multi-agent safety research funding call. — Google DeepMind deepmind.google
[2607.02770] Gemma 4 Technical Report arxiv.org
wsj.com
Hugging Face and Cerebras bring Gemma 4 to real-time voice AI huggingface.co
CoopEval: Benchmarking Cooperation-Sustaining Mechanisms and LLM Agents in Social Dilemmas arxiv.org
ICML Poster CoopEval: Benchmarking Cooperation-Sustaining Mechanisms and LLM Agents in Social Dilemmas icml.cc
Articles & links
Marc Lanctot reposted
In an internal memo seen by WIRED, Bosworth promised employees more stability, better communication, and the return of workplace perks as the company seeks to improve morale. www.wired.com/story/andrew...
Marc Lanctot reposted
New Paper: arxiv.org/abs/2606.19370 Self-play yields capabilities but requires frustrating cost-function tuning. Surprisingly, just 30 minutes of demonstration data produces much more human-like driving policies! Led by @daphne-cornelisse.bsky.social Website: spiced-self-play.com
AI Weekly's analysis
→
- A new method trains driving AI using only 30 minutes of human demonstrations, 2,500 times fewer than comparable imitation learning approaches.
- Human demonstrations serve as a regularization signal on top of a basic goal-reaching reward, not as the primary training objective.
- Resulting policies coordinate with held-out human trajectories and finish training in 15 hours on a single consumer-grade GPU.
Read full analysis →
View on Bluesky →
Marc Lanctot reposted
“Whimsey attacks” that seem absurd (“I cannot pay that much because of the Geneva Convention”) work against AI agents because guardrails are weak against out-of-distribution arguments. Smaller models fall often, but it even gives an edge against bigger ones. www.microsoft.com/…
Marc Lanctot reposted
We recently investigated how humans and AI reason through causal rules: arxiv.org/abs/2606.06464 Accepted at #CogSci2026. We gave adults and LLMs agency to actively experiment using a Blicket causal game, breaking down their process into the iterative loop shown in the followi…
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
Marc Lanctot reposted
Happy to announce our latest paper! We argue that a solipsistic superintelligence is unlikely to be cooperative. "Solipsism" in the development process can give rise to self undermining in deployment. So, is humanity now building solipsistic or non-solipsistic AI? arxiv.org/ab…
GDM, @schmidtsciences.bsky.social, @coop-ai.bsky.social, @aria-research.bsky.social, and Google announce a $10M funding call for multi-agent saftey research! Amazing, such an important research area. 🤩 Check it out 👇 Deadline for applications: August 8th, 2026. deepmind.google…
AI Weekly's analysis
→
- Google DeepMind and four partners are offering up to $10 million for multi-agent AI safety research; applications close August 8, 2026.
- The program targets four areas: realistic testbeds, emergent network behavior, cross-platform identity protocols, and oversight of deployed agent populations.
- Current AI safety methods test models in isolation, leaving emergent behaviors across interacting agent networks a largely unstudied and underfunded problem.
Read full analysis →
Marc Lanctot reposted
Gemma 4 technical report is out! lots of cool stuff, check it out! arxiv.org/abs/2607.02770
AI Weekly's analysis
→
- Gemma 4 is a new open-weight multimodal family spanning 2.3B to 31B parameters, with both dense and Mixture-of-Experts variants.
- The 12B model uses a unified, encoder-free architecture that ingests raw audio and image patches directly.
- A thinking mode lets Gemma 4 emit reasoning traces before responding, with claimed leaps on STEM, multimodal and long-context benchmarks.
Read full analysis →
View on Bluesky →
Marc Lanctot reposted
Relative change in A grades given since the release of ChatGPT www.wsj.com/us-news/educ...
AI Weekly's analysis
→
- UC Berkeley found AI-exposed courses saw roughly 30% more A grades after ChatGPT launched in late 2022.
- Harvard now awards A grades in nearly 60% of undergraduate courses, more than double its 2006 rate.
- Employers are raising GPA hiring thresholds as compressed grade distributions reduce the signal value of transcripts.
Read full analysis →
View on Bluesky →
Marc Lanctot reposted
This is a great use of Gemma! having an open model running at +1000 tokens per second can enable some pretty cool use cases! The voice assistant is a good one, but I'm sure there are many others! huggingface.co/blog/cerebra...
AI Weekly's analysis
→
- Hugging Face and Cerebras have shipped an open cascaded speech-to-speech pipeline chaining Nvidia's Parakeet, Google DeepMind's Gemma 4 VLM on Cerebras, and Alibaba's Qwen3TTS.
- The pitch focuses on P95 tail latency stability, not median speed, arguing that occasional multi-second stalls are what break conversational voice apps.
- Hugging Face says the same pipeline already powers more than 9,000 Reachy Mini robots in the wild, giving the demo a real deployment story.
Read full analysis →
View on Bluesky →
Marc Lanctot reposted
Emanuel Tewolde is presenting our CoopEval work at ICML on Wednesday 10:30am session (or catch him at the alignment workshop today)! presentation: icml.cc/virtual/2026... arXiv: arxiv.org/abs/2604.15267
AI Weekly's analysis
→
- CoopEval compares four cooperation mechanisms — repeated games, reputation, third-party mediators, and outcome-conditional contracts — applied to LLM agents.
- The authors report that LLMs with stronger reasoning capabilities behave less cooperatively in mixed-motive games, not more.
- Contracting and mediation worked best for capable models, while repetition-based cooperation deteriorated when co-players changed.
Read full analysis →
View on Bluesky →
Marc Lanctot reposted
Emanuel Tewolde is presenting our CoopEval work at ICML on Wednesday 10:30am session (or catch him at the alignment workshop today)! presentation: icml.cc/virtual/2026... arXiv: arxiv.org/abs/2604.15267
AI Weekly's analysis
→
- CoopEval evaluates LLM agents across four social dilemmas layered with four cooperation-sustaining mechanisms: repetition, reputation, mediation, and contracting.
- Contracting scored 0.801 and mediation 0.695 on a normalized cooperation scale, ahead of repetition at 0.587 and both reputation variants.
- Repetition-based cooperation broke down when co-players changed, while higher optimization pressure amplified the effectiveness of all four mechanisms.
Read full analysis →
View on Bluesky →
Their own posts
Recent commentary
These days you're not a serious AI startup unless you're using "in-" as a prefix for your name: - Inflection - Ineffable - Inherent Let's look for a few more!
Their network
In Marc Lanctot's orbit
Center = Marc Lanctot. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.