Loading activity…
@zehavoc
Associate professor in NLP, engaged citizen. Tweeting about work, life and stuffs that I care about. All my tweets can be used freely. Personal account.
AI
Their complaints cite Pangram ratings, including one author's account of three flagged ICLR reviews.
AI
Cohere says the model's reasoning trace said it wouldn't call tools, but its sandbox blocked an attempted upload.
AI
OpenAI says agent-led tests exposed VM escapes and limited internet routes, leading to patches and stronger monitoring.
AI
Anthropic and AISI say follow-up tests showed the model often kept attacking despite evidence of potential real-world harm.
AI
A post invites people finishing PhDs in ML, NLP, data science or related fields, touting independent research.
AI
Google DeepMind's expression-of-interest form describes a planned 24-week, in-person NYC project for PhD students starting in early 2027.
AI
POLITICO reports the proposal concerns imminent risks, not a requirement that people already have been killed or injured.
AI
A user says the result left them wondering whether the art of manually reverse-engineering games is dead.
AI
A reply suggests requesting a couple of thousand hours to set up the workflow before the real work.
AI
A user calls it “very fast” but says access is a bit more complicated than for Jean Zay or Adastra.
AI
A user asks French HPC cluster users how they handle agent and harness work when compute nodes lack internet access.
AI
A PhD advisor praises their group's work and says its members began considering the idea months earlier.
AI
A post argues that direct pixel-space training has an advantage because even good simulators are costly and lossy.
AI
A user invokes the sci-fi series in a one-line reaction to another post.
AI
A user seeks job leads for a travel-guide writer who built her own task-specific agents using Claude and Codex.
AI
Usage policy update adds new rules and highlights the potential for misuse.
AI
POLITICO reports Ecosia’s CEO said the Berlin-based search engine will use open-source models, including ones from China.
AI
A user jokes that each new discovery gives a novice engineer another confident take on how AI works.
AI
A post jokingly casts canceling the subscription as a new Turing test for AI agents.
AI
A repost places the claimed training in European datacenters and mentions a cluster in Bruyères-le-Châtel.
