Bonsai 27B is a 27-billion-parameter multimodal model compressed to 3.9GB (1-bit variant) and 5.9GB (ternary variant) using low-bit quantization, making it the first model of its class to run on phones like the iPhone 17 Pro while retaining 90-95% of full-precision performance. The model supports multi-step reasoning, tool calling, vision tasks, and agentic workflows with a 262K-token context and runs at speeds up to 163 tokens/sec on NVIDIA GPUs and 87 tokens/sec on Apple M5 Max. This enables local deployment of capable AI agents without cloud dependencies, eliminating per-token costs and keeping private data on-device, while achieving over 10x better intelligence-per-gigabyte efficiency than full-precision baselines.
Comments
Agent systems require three distinct loops, not one: the **Inference Loop** manages chat completion API calls and chat history persistence; the **Tool Loop** executes the LLM's inferred tool calls while handling hallucinations and maintaining tool call IDs for API correlation; and the **Human Loop** blocks execution for approval/steering before tools run, which is the hardest part to implement at scale due to durability and concurrency challenges. Understanding this three-loop architecture matters because most agent implementations oversimplify the design, and implementing the human loop properly requires architectural considerations like durable execution frameworks to handle server restarts and concurrent requests.
DOOMQL is a Doom-like game built by Peter Gostev using GPT-5.6 Sol where SQLite serves as the entire game engine—handling movement, collision, enemies, combat, and rendering via a Python terminal script. The game implements a full ray tracer in SQLite using recursive CTEs, with all game state stored in a SQLite database that can be queried and visualized through Datasette. The project demonstrates practical AI-assisted programming, where Claude was used to generate HTML+JavaScript apps within Datasette to display game state and add features like a minimap in real-time.
Agnost AI analyzes production agent conversations to identify failure patterns that standard evals miss—such as stuck users, friction points, and churn signals—then surfaces these issues as actionable, reviewed fixes. The tool works with any LLM/framework, integrates with OpenTelemetry, and requires only 2 minutes to set up. Pricing ranges from free (1,000 messages/month) to enterprise, with customers reporting improved agent performance and automated bug fixes merged via generated PRs.
Simon Willison examined a GitHub code-frequency chart for his Datasette project and found a notable spike in coding activity at the end of the period, which he attributes to recent AI model releases including Opus 4.8, GPT-5.5, Fable 5, and GPT-5.6 Sol. The spike suggests these newer coding-capable AI models and agents are significantly increasing his development velocity. This observation serves as a practical indicator of how current-generation AI assistants are impacting real-world open source development output.