Loads of changes happening in Google and the hope for Gemini 3.5 Pro is slowly evaporating. However, the team is committed to focus on frontier intelligence.
Gemini 4 WIP 👀
> We are committed to being at the frontier and are super focused on the areas where we need to improve.
> The Gemini models are in good hands with Koray and the leads, and I'm excited about the great progress we’re making with our new models, including Gemini 4.
Prime Intellect announced Prime Agent, a new self-improving RLM harness for coding and long-running autonomous tasks.
Prime Agent scored 95.5% on ARC-AGI-3! 🤯
> The Recursive Language Model (RLM) treats context as a variable and subagent delegation as function calls inside a REPL.
> Continual Harness treats the harness's own state, abstracted as its prompts, skills, memory, and sub-agents, as something the agent can create, read, update, and delete (CRUD) from its own trajectory.
Meta launched Muse Code beta, a terminal agent for full software tasks across large repositories on macOS and Linux. Powered by Muse Spark 1.2, it plans, codes, validates, logs work locally, and supports restart-safe sessions.
ByteDance’s SeedRealtime is a full-duplex audio-visual LLM that processes speech, video, and text in one model, reducing latency and context loss while handling timing, scene cues, and real-world dialogue with fewer pacing errors.
Cloudflare Wallets will let AI agents pay for APIs and web services via stablecoins and x402, with owner-set budgets, allowlists, and caps. The system adds optional agent identity and supports Cloudflare’s push into machine-native commerce.
Mistral released Shieldstral, a 3B Apache 2.0 open-weight multimodal safety classifier for text and images. It uses prompt-defined yes/no policies, runs on a single 16GB GPU, and targets customizable moderation without retraining.
