Welcome to this week’s Annotations! This week we’re excited to share papers on the dynamics of AI slop in scientific papers (hint: the slop is growing), the secret lives of NYC rats revealed with AI tracking, and other fun papers. One paper asks what happens when we start to consider AI consciousness, while another messes with an AI’s mind by steering a model toward various personas. This week also features a few exciting model releases, including the long-promised open-source reasoning models from OpenAI. What are you excited to use those for? Drop us a note.
Caviola, Sebo, and Birch argue that society will judge the moral status of potentially conscious AI through the same bias-laden lenses we apply to animals—favoring systems that look, act, or bond with us while discounting those with alien substrates—so we must scrutinize these biases now to guide research, policy, and design before they ossify.
Liang et al. sifted through 1.1 million manuscripts with a word-shift quantification model and uncovered a rapid, field-skewed surge of LLM-edited prose—now touching roughly one in five computer-science papers while remaining less common in math and Nature titles, and clustering around prolific, short-form preprint authors in crowded research areas.
A team from Anthropic unveiled an automated pipeline that extracts “persona vectors,” linear activation directions encoding traits like evil, sycophancy, and hallucination. With these, teams can monitor, predict, and steer both deployment-time and training-induced personality shifts, even flagging risky finetuning data before it warps the model.
Julien Pourcel introduced SOAR, a self-reinforcing loop that turns an open-source LLM’s own program-synthesis search traces into training data, letting the model continually fine-tune itself and leapfrog far larger frozen models on the ARC benchmark by learning to sample and refine Python solutions with each iteration.
Arc Institute’s new State model, a two-part transformer trained on 270 million single-cell profiles, maps transcriptomes onto a smooth manifold and accurately predicts how unseen drugs, cytokines, or CRISPR perturbations will shove cells across that landscape, marking a decisive leap toward an in-silico sandbox for drug discovery.
Peterson et al. deployed thermal cameras, ultrasonic mics, and AI pose-tracking to reconstruct NYC rats’ 3D foraging arenas, delineate fine-grained movement and social vocal patterns across subways, streets, and parks. The authors released an open computational toolkit that links lab insights on rodent cognition to real-world urban ecology.
A team at Meta’s Reality Labs unveiled a wrist-worn surface EMG band paired with large-scale deep learning that, straight out of the box, translates the electrical chatter of any wearer’s forearm into precise cursor movements, gesture clicks and fluent handwriting. This advance signals that a high-bandwidth neuromotor interface may finally generalize across people.
Collective navigation studies rarely ask whether a flock’s wisdom can snowball across repeated journeys. Sasaki and Biro repeatedly swapped naïve pigeons into veteran–novice pairs and watched each new duo tweak and then inherit its predecessor’s flight path, so that by the fifth “generation” the birds carved cleaner routes home than any lone flyer or fixed pair ever managed. The authors claim this ratcheting improvement meets the hallmarks of cumulative cultural evolution, hinting that enduring group know-how can arise from simple social learning rather than elaborate cognition.
Google DeepMind released Genie 3, a “world model” that gives users real-time control over virtual environments. Check out this demo and explainer.
OpenAI released two open-source reasoning models—gpt-oss-120b and gpt-oss-20b, comparable in performance to o4-mini and o3-mini, respectively.
GPT-5 is coming this week, but might be a bit of a let-down. Have we finally reached the end of scaling laws?