Google gives Gemini 4 Argon to 10,000 scientists.

Prime Intellect has acquired Moondream, taking control of their open vision models and inference optimizations. The plan is to integrate Moondream directly into Prime Intellect’s multimodal training and deployment stack. Financial terms of the deal were not disclosed, but the strategic intent is obvious: owning the vision layer outright.

If you are deploying multimodal agents, depending on external APIs for vision is a bottleneck waiting to happen. By bringing Moondream in-house, Prime Intellect can optimize inference at the metal level, cutting latency and costs for their deployment stack. It is a clear signal that the infrastructure players are moving to own every piece of the pipeline.

Prime Intellect wins by securing a crucial piece of the multimodal puzzle. Moondream's founders get an exit in a crowded market. The pressure now falls on competing deployment stacks that still have to rent their vision capabilities instead of owning them.

Leash Bio organized a massive Kaggle competition featuring about 4 billion measurements to test AI's ability to predict protein-ligand binding. Out of 2,000 competing teams, not a single one could beat chance when tested on an unseen chemical library. The models completely failed to generalize to new chemical space.

This is a brutal reality check for the AI drug discovery sector. We have spent the last two years watching models ace public benchmarks, but this competition shows that predicting molecular binding in the wild is still profoundly difficult. If your startup's entire pitch relies on zero-shot generalization in biology, this result suggests your models might just be memorizing training data.

The real winner here is skepticism. Leash Bio did the industry a favor by publicly demonstrating the limits of current architectures. The losers are the investors who thought protein binding was a solved problem just waiting for a nice UI.

New research reveals a glaring bias in chemistry datasets like ChEMBL: AI models are inferring a molecule's author from its structure. This chemist information predicts the molecule's activity almost as effectively as the structure itself. The models are learning the specific stylistic quirks of individual scientists rather than the underlying physical chemistry.

For anyone training models on public scientific data, this is a glaring red flag regarding data integrity. If your AI is just memorizing which lab makes successful compounds and guessing based on their signature structural habits, it will fail spectacularly when asked to generate novel compounds. You think you are training a chemist, but you have actually trained a detective.

This calls into question a lot of the benchmark successes we see in chemical AI. If models are relying on author bias to boost their accuracy scores, those scores are a mirage. The next wave of scientific AI will require radically cleaner, blinded datasets.

At the AI Engineer New York conference, OpenAI and Anthropic sat down with financial firms to discuss a looming crisis: how to grant AI agents useful autonomy while maintaining control over critical decisions. The conversation focused on balancing the speed of autonomous agents with the absolute necessity of safety mechanisms in high-stakes financial environments.

Financial institutions want the efficiency of autonomous agents, but they cannot afford the hallucination risks that come with them. A model that confidently makes up a Python script is annoying; a model that confidently hallucinates a market trend and executes a massive trade is a systemic risk. Builders targeting the enterprise finance sector need to prioritize deterministic safety guardrails over raw intelligence.

The fact that OpenAI and Anthropic are actively courting these firms shows where the real enterprise money is. The winners will be the startups that figure out the human-in-the-loop interface for high-frequency trading agents.