OpenAI-Linked Group Runs Influencer Campaign Promoting the Chinese AI Threat Narrative

According to Wired, it has come to light that Build American AI, a super PAC funded by executives at OpenAI and Palantir, has been running a campaign enlisting TikTok influencers to portray Chinese AI as a threat.

The organization positions itself as a nonprofit, but critics point out that it is effectively spreading propaganda aligned with the interests of major AI industry players. On Reddit's LocalLLaMA community, the report prompted comments that "this is a renewed reminder of why local LLMs matter," along with fresh expressions of support for communities that keep providing open-source models and open weights.

As the geopolitical dimensions of the AI industry come into sharper focus, the episode has underscored the need to carefully scrutinize where information comes from and whose interests it serves.

6-Hour Drug Discovery with a 3B-Parameter LLM — GRPO × Chemistry RL Environment

A project is drawing attention on Reddit's MachineLearning community: training Llama-3.2-3B with GRPO (Group Relative Policy Optimization) to design drug-like molecules in a reinforcement learning environment.

The project builds an RL environment where an AI agent designs drug molecules atom by atom. Molecules are generated through operations such as adding fragments, swapping atoms, and building scaffolds, and are then evaluated against real chemical criteria including Lipinski's rule of five, drug-likeness (QED), synthetic difficulty, and binding scores against target proteins.

With just six hours of training on a single A10G GPU, the system was confirmed to generate molecules with a QED of 0.94 — a drug-likeness score on par with FDA-approved oral drugs. The results are at a level where medicinal chemists say they would "actually like to take a look," and the discussion has expanded to what might be possible if scaled up to 600 or even 6,000 hours.

It is a piece of research that raises expectations for real-world adoption: when will the workflow of "AI proposes 10,000 candidates and chemists pick 5" become an everyday routine?

DeepSeek-TUI — A Rust-Based Terminal Agent Built Exclusively for DeepSeek

DeepSeek-TUI has debuted on GitHub's trending list, picking up 1,276 stars.

Written in Rust, it is a terminal user interface (TUI) built specifically for DeepSeek models, providing interactive access to DeepSeek's models from the command line. Its lightweight and fast operation has drawn attention as a highly convenient tool for developers working with the DeepSeek API from their local environments.

On the same day's GitHub trending list, an Astro-based agent framework also made its debut — a sign that the AI tool ecosystem continues to diversify.

---