AI Radar · Daily brief
PrismML announced Bonsai 27B, a 27-billion-parameter multimodal model compressed to 3.9GB (1-bit variant) or 5.9GB (ternary variant) using low-bit quantization, making it the first model of this capability class to run on phones like the iPhone 17 Pro. The model retains 90% (1-bit) to 95% (ternary) of the full-precision baseline performance across 15 benchmarks covering reasoning, coding, math, tool-calling, and vision tasks, with particularly strong retention in math and coding—critical for agentic workflows. This enables practical on-device deployment of capable AI agents that can perform multi-step reasoning and tool use without cloud APIs, eliminating per-token costs and privacy concerns while achieving 10x better intelligence density than full-precision baselines.
This project demonstrates a nested reinforcement learning system where an AI agent (Qwen 35B) learns to write training jobs for smaller models, with reward signals based on the quality of trained models. The agent was trained via Tinker (importance-sampling GRPO) over 54 steps, achieving a peak reward of ~0.63, with learning occurring in two phases: process reliability first (~step 10), then model quality improvement (steps 10-54). The total cost was ~$1,275, and the trained agent weights are open-sourced; notably, the learned skill transferred to a held-out task family the agent never trained on, and the agent learned to select better base models and hyperparameters.
Agent systems consist of three interconnected loops, not one: the Inference Loop calls the LLM's chat completion API and manages conversation history; the Tool Loop intercepts tool calls the model infers and executes them while handling hallucinations and tool call IDs; and the Human Loop gates approval of critical tool calls before execution, which is the architecturally hardest part since it requires durable state management across potential server restarts. Understanding this three-loop structure is essential for building reliable agents because each loop has distinct responsibilities—the inference loop manages LLM interaction statelessly, the tool loop translates model intent into actual function calls, and the human loop provides safety gates that can't be implemented as simple code blocks.
DOOMQL is a Doom-like game built by Peter Gostev where SQLite serves as the entire game engine—handling movement, collision, enemies, combat, and rendering via SQL queries and recursive CTEs instead of traditional game code. The project demonstrates an unconventional approach to game development by implementing a full ray tracer in SQLite and running the game as a Python terminal script that writes to a SQLite database. The article includes instructions for running DOOMQL locally and viewing game state through a custom Datasette app built with Claude that displays pixels and includes a minimap.
Agnost AI analyzes production agent conversations to identify failure patterns—user frustration, workflow breaks, churn signals—that standard evals miss, then generates reviewed pull requests to fix them. The platform works with any LLM/framework with 2-minute setup and OpenTelemetry integration, offering a free tier (1,000 messages/month) and paid plans ($499/mo for 100K messages). Users report concrete wins: surfacing 1,247 feature requests, merging 16 autonomous PR fixes, and improving agent conversion rates.
Simon Willison observed a significant spike in code commit frequency to his Datasette open-source project around late July 2026, which he attributes to newer AI models including Opus 4.8, GPT-5.5, Fable 5, and GPT-5.6 Sol. He used GitHub's code-frequency chart as a visible metric to illustrate the impact of coding agents and advanced LLMs on his own development output. The timing suggests these newer model classes materially accelerated his contribution velocity to the project.
Archive: 2026-07-14 · 2026-07-13