Framework Unveils Portable Keyboard, T3 Dev Criticizes LLM Memory Systems, and Drops Opinionated Tier List

A prominent developer recently showcased an early, unreleased Framework keyboard-trackpad combination, designed to address the market’s need for a high-quality, portable input device. Envisioned for diverse use cases ranging from living room control and studio setups to on-the-go debugging, the device is the culmination of years of advocacy. The developer expressed extreme satisfaction, noting, “I’ve been waiting for this keyboard for the better part of a decade,” highlighting its potential to fill a significant gap left by existing, cheaper alternatives.

In a critical take on AI agent development, the developer voiced strong disapproval of current LLM memory systems, particularly after a deep dive into Claude Code’s memory implementation. Characterizing the stored memories as “useless” and “garbage” due to a 3:1 write-to-read ratio and a propensity to retain irrelevant or stale information, the developer advocated for a shift towards explicit context management. This aligns with sentiments from industry figures like Mario (creator of Pi) and Armen (creator of Flask), who assert that “code is truth” and that robust bash tools are often sufficient for dynamic agent context. The recommended approach prioritizes architectural solutions, CI/CD-driven linting and testing, and meticulously crafted agent.md files that encode product vision, performance goals, remote-readiness, and shared terminology to align agent behavior with developer intent.

The session culminated in a highly subjective yet insightful tier list for current LLMs in software development. Claude Fable 5 secured the S-tier for its unparalleled thoughtfulness and trustworthiness in merging complex code, while GPT-5.6 Soul (A-tier) was praised as a highly capable default, especially for Swift UI development. GPT-5.6 Luna (A-tier) earned recognition for its speed, cost-effectiveness, and utility in summarization tasks. Conversely, Google’s Gemini models (F-tier) faced scathing criticism for their “incompetent” pricing, poor performance, and lack of vision capabilities in their Pro/Flash variants. Deepseek V4 Pro and Flash were also relegated to lower tiers due to the “pathetic” omission of vision. The developer emphasized the substantial capability gap between top and lower tiers, cautioning against the “stupid” notion of automated LLM model routing, asserting that “models do not understand how models work.” The advice remains: “Just pick one [model] and stick with it until it breaks.”