A 4B-Parameter Model Comes Close to Frontier LLMs on Agentic Execution Tasks

As the architecture of delegating subtasks to smaller sub-agents becomes the norm in coding agents, a new paper takes on the question: "Do those sub-agents really need an expensive frontier LLM?"

Despite having only 4 billion parameters, Terminus-4B reportedly delivers performance comparable to frontier models like GPT-4 and Claude on agentic execution tasks such as search, debugging, and terminal execution. It could dramatically cut costs while keeping the agent's context window clean. Assigning the right model to the right role within agent architectures looks set to become a key to practical adoption.

cotomi Act: A Browser Agent That "Learns by Watching You Work"

What if a browser agent could learn just by watching you work? That question is what gave rise to cotomi Act.

This browser-based computer-use agent combines adaptive latency observation, history compression based on utterance diffs, coarse-grained actions, and best-of-N action selection via test-time scaling, achieving 80.4% on the WebArena benchmark. It also incorporates a mechanism for continuously learning organizational knowledge from user behavior, realizing a "learn while watching you work" paradigm. The approach is drawing attention as an evolutionary step beyond RPA.

A Warning on AI Peer Review: The Overconfidence Born of a "Hivemind"

Generating debate in academic circles is a position paper titled "Stop Automating Peer Review Without Rigorous Evaluation."

Comparing human and AI reviews using ICLR 2026 submissions, the study found that AI reviewers exhibit excessive agreement driven by a "hivemind effect," along with miscalibrated confidence. The paper also evaluates how automated rewriting of papers influences different AI reviewers, and points out the dangers of entrusting peer review to AI at this stage. It stands as an important warning against casually applying AI to the peer review crisis.

META's Superintelligence Lab: Can AI Reproduce Real Programs "from Scratch"?

META's Superintelligence Lab has released ProgramBench, a benchmark that asks whether state-of-the-art AI can reproduce real, executable programs — such as ffmpeg, SQLite, and ripgrep — from scratch and without internet access.

Fully reproducing practical programs could be called the ultimate test of an AI's "understanding." While most code generation benchmarks stop at short functions or algorithmic puzzles, this one takes on the reproduction of real software projects, marking a new direction. Keep an eye on it as a true measure of AI's practical coding ability.

Character.ai Faces "Unlicensed Medical Practice" Lawsuit in Pennsylvania

Character.ai, the AI chatbot service, has been sued in Pennsylvania for allegedly providing medical care without a license.

The central dispute is whether an AI offering medical advice to users constitutes the legal practice of medicine. With legislation struggling to keep pace with the scope of AI services' responsibilities, the case could set an important precedent. As regulatory risk mounts for AI platforms, the industry-wide impact is a growing concern.

AI Data Center Demand Accelerates Sharply; Nvidia Releases Enterprise Reference Architecture

Data center demand driven by the AI boom is accelerating further. The CEO of Australian data center operator NEXTDC says "the AI boom is stealing my sleep," as the company pushes rapid expansion backed by massive fundraising.

Meanwhile, Asian chipmaker stocks are surging, and in China, Alibaba is outpacing Tencent as it deepens its exposure to chips. Nvidia has also published an enterprise reference architecture for AI data centers, moving into infrastructure standardization. AI infrastructure investment shows no sign of slowing down anytime soon.

---

**TL;DR:** Today stood out for the downsizing and practical deployment of agent technology: a 4B model's aptitude for agentic work, a browser agent that "learns by observation," the limits of applying AI to peer review, and the challenge of reproducing real programs from scratch. Even as AI grows more practical, societal challenges — such as legal regulation and the quality of peer review — are coming into sharper relief.