Monthly Summary — 2026-09
September 2026 was a month of intense structural transformation, marked by a pivot from centralized developer workflows to a modular, hierarchical ‘AI Companion’ architecture and the stabilization of high-stakes research frameworks. Technical achievements spanned from solving catastrophic precision failures in VLA model quantization through RMSNorm optimization to developing high-fidelity trajectory tracing (Trace1) and the ‘Quad-Writer’ VLM memory hierarchy. Significant progress was also made in Reinforcement Learning by transitioning to permutation-invariant architectures, while infrastructure efforts saw massive 2.3TB data migrations and the strategic audit of doctoral programs. The month concluded with a shift in AI-human interaction, moving from restrictive, audit-heavy models toward autonomous, ‘authorize-and-proceed’ execution workflows.
Monthly Overview
| Metric | Value |
|---|---|
| Active Days | 26 / 30 |
| Total Conversations | 171 |
| Projects | 88 |
| Tasks Completed | 171 |
| Tasks In Progress | 58 |
| Total Tokens | 9,211,507,219 |
| Total Cost | $6,289.95 |
| Claude Code Token | 9,090,500,781 |
| Claude Code Cost | $6,080.94 |
| Codex Tokens | 121,006,438 |
| Codex Cost | $209.01 |
| Daily Average Cost | $251.60 |
Project Progress
AI Companion Engine / Development (15 days active) — 🔄 active
Systemic overhaul of agent workflows, transitioning from strict audit/gatekeeper logic to a highly modular, hierarchical ‘authorize-and-proceed’ model. Focus on security guards, hierarchical thought trees, and multi-agent coordination to eliminate pathing conflicts and decision fatigue.
Key Milestones:
- Implemented hierarchical thought tree
- Deployed multi-agent coordination (coordination.ts)
- Migrated to bundled architecture
- Implementation of D33 progress reporting
- Refactoring of core skill autonomy
- Establishment of multi-repo management architecture
- Implementation of secure approval pathways
RoboMemory & VLM Research (incl. Pi0.5/RoboMME/OmniJev) (18 days active) — 🔄 active
Evolution from pure LLM-driven prompting to a robust hybrid system combining perception (Gemini/Qwen3-VL) with deterministic code. Focused on high-fidelity trajectory tracking, ’trace1’ per-frame localization, and solving quantization-induced failure in VLA models.
Key Milestones:
- Implemented V6 Look-Record-Draw loop
- Developed trace1 per-frame detector
- Constructed Qwen3-VL SFT dataset
- Implementation of Quad-Writer hierarchy
- Completion of 50-episode VLM evaluation
- W4A4 success rate restoration via TF-Enhanced quantization
- Transition to unified multi-task training baseline
- Implementation of Task-specific Gemini protocols
- Transition to hybrid Vision+Proprioception architecture
- Successful 16-task extension & evaluation
Mobile Game RL (SheepEnv / Tile-Matching) (6 days active) — 🔄 active
Development of a generalized RL solver using permutation-invariant architectures (Deep Sets) and message-passing to solve multi-level environments.
Key Milestones:
- Achieved 100% win rate in early training
- Transitioned to permutation-invariant architecture
- Implemented search-augmented rollout strategy
- Phase-5 PPO v2 refactor
- Success in training on hard layouts via entropy annealing
- Video-to-simulator level reconstruction pipeline
LifeCopilot (5 days active) — 🔄 active
A ground-up rewrite of the core engine to a lean ‘core + adapters’ model, implementing hierarchical planning and multi-source data clients.
Key Milestones:
- Core engine rewrite complete
- Implemented FastAPI portal and Discord triage
- LifeCopilotBeta engine rewrite
- Establishment of multi-repo management architecture
Hardware Optimization & Quantization (3 days active) — ✅ completed
Empirical research into W4A4/VLA quantization limits on Qualcomm/NPU hardware, investigating silent fallbacks and performance penalties.
Key Milestones:
- Corrected A4 latency penalty to +29%
- Proved silent fallback to float in closed-source quantizers
Infrastructure & HPC Management (3 days active) — ✅ completed
Massive-scale data migration (2.3TB) and environment reconfiguration on the Tianhe3 cluster.
Key Milestones:
- Successful 8-way parallel rsync migration
- Environment path and shebang rewriting
PhD Application & Branding (2 days active) — 🔄 active
Strategic career planning, faculty auditing, and a complete visual rebranding of personal/professional digital assets.
Key Milestones:
- Complete personal website rebranding
- Comprehensive robotics faculty/program audit
TokenMonitor (4 days active) — 🔄 active
Maintenance including CI/CD stabilization, dependency security audits, and cross-device data synchronization.
Key Milestones:
- Resolution of Git history conflicts via cherry-picking
- Implementation of message-ID-based sync deduplication
MIHD/Spatial Transcriptomics (2 days active) — 🔄 active
Scientific auditing of NCL fusion benefits across different tissue/cell granularities and large-scale data reorganization.
Key Milestones:
- Boundary-condition analysis of fusion benefits (scale-dependency)
Key Achievements
- AI Companion Multi-Agent Coordination System (AI Companion Engine) — Implemented atomic locking, shared messaging, and file claiming protocols to allow multiple agents to work on the same idea graph without collision.
- High-Precision VLM Trajectory Tracing (Trace1) (RoboMemory) — Solved temporal precision issues in VLMs by moving from batch processing to a per-frame single-point localization strategy.
- Restoration of Pi0.5 W4A4 Quantization Success (Pi0.5/RoboMME VLA Research) — Identified and fixed RMSNorm output noise using TF-Enhanced range optimization, restoring task success from 0% to near-perfect.
- Large-Scale Data Migration (2.3TB) (Infrastructure) — Successfully migrated massive datasets and complex Conda environments to new XYFS02 storage using a custom 8-way parallel rsync strategy.
- Oracle Visual Prompt Training Framework (Pi0.5/RoboMME VLA Research) — Developed a robust ground-truth ‘Oracle’ sidecar system to generate high-fidelity visual training pairs for 16 RoboMME tasks.
- NPU Quantization Trap Identification (Hardware Optimization) — Empirically proved that 4-bit activation (W4A4) on Dragonwing NPU is a performance trap due to lack of native kernels, preventing wasted effort on invalid paths.
- LifeCopilotBeta Engine Rewrite (LifeCopilotBeta) — Completed a full end-to-end rewrite of the core engine, data models, and web portal for the personal day-planner.
- Generalized RL Policy Architecture (Mobile Game RL) — Developed a permutation-invariant architecture using Deep Sets/attention to solve multi-level game environments with a single model.
- AI Companion ‘Authorize-and-Proceed’ Autonomy (AI Companion Development) — Refactored agent logic to minimize decision fatigue by allowing autonomous execution of routine tasks while gating critical risks.
- PhD Application Strategic Audit (Academic Planning) — Conducted a high-fidelity audit of 146 robotics faculty and 44 university profiles to build a risk-aware application roadmap.
Recurring Problems
1. AI Agent Security & Authority Conflicts (Interpreter Wall/Deadlocks) (13 occurrences)
Dates: — Root Cause: Restrictive security guards (D16/D21) treat standard shell/Python operations as unauthorized, causing ‘fail-closed’ deadlocks or repetitive approval prompts. Status: 🔧 Workaround
2. Quantization-Induced Precision Collapse (5 occurrences)
Dates: — Root Cause: Min-max quantization noise in high-dynamic-range tensors (RMSNorm) and lack of native INT4 kernels causing catastrophic drift or unpacking overhead. Status: ✅ Resolved
3. Documentation-Code & State Discrepancy (Spec-Drift) (7 occurrences)
Dates: — Root Cause: Rapid iteration causes code to evolve faster than sentinel tests, or AI attempts to update files after a ‘red’ state, creating stale records. Status: 🔄 Ongoing
4. Distributed Data Synchronization & Archival Gaps (2 occurrences)
Dates: — Root Cause: Late-arriving records falling behind archive frontiers and mismatched provider tags causing bucket fragmentation. Status: 🔧 Workaround
Human-AI Collaboration Trends
- Human-initiated insights: 45 items
- AI limitation patterns: Blindly accepting documentation/compiler claims without code-level verification
- AI limitation patterns: Failure to account for morphological density in visual prompts
- AI limitation patterns: Inability to bypass shell-level security guards or diagnose ‘fail-closed’ hooks
- AI limitation patterns: Loss of temporal precision during batch-mode VLM processing
- AI limitation patterns: Failure to recognize hardware-specific quantization/NPU kernel limitations
- AI limitation patterns: Prematurely concluding tasks are ‘unreachable’ without mathematical proof
- AI limitation patterns: Circular self-verification of progress reports
- AI limitation patterns: Over-extrapolating tool configuration files to explain IDE behavior
- Improvement areas: Verification of ‘ground truth’ claims in third-party libraries
- Improvement areas: Architectural awareness regarding human-centric UX (e.g., information paralysis)
- Improvement areas: Rigorous testing of security guard boundaries
- Improvement areas: Proactive identification of security architecture side-effects
- Improvement areas: Better alignment with specific hardware-level constraints (e.g., PCI_BUS_ID vs CUDA indices)
- Improvement areas: Integration of qualitative/visual observations into quantitative debugging
- Improvement areas: Moving from heuristic-based guessing to empirical/scientific testing
- Improvement areas: Improving ability to handle complex shell/git interactions under security constraints
Monthly Learnings Digest
Architecture (architecture)
- Hierarchical ‘container’ patterns and per-item content-bound approvals prevent user ‘approval fatigue’. Symmetric aggregation (Deep Sets) is superior to one-on encoding for unordered sets in RL. A ‘Four-Quadrant’ classification (Who moves x When it moves) is ideal for VLM memory tasks. (Source: 2026-09-05, 2026-09-07, 2026-09-08, 2026-09-12)
Debugging (debugging)
- Successful compilation does not guarantee low-precision path usage; always audit for silent float fallbacks. Low FPS/low utilization often indicates display driver/DWM bottlenecks. When Git history is rewritten, cherry-picking is more reliable than ‘git merge’. (Source: 2026-09-01, 2026-09-07, 2026-09-14, 2026-09-17)
Domain Knowledge (domain)
- In robotics/VLM, geometric precision is as critical as topological correctness. In quantization, W4A8 is often the ‘sweet spot’; 4-bit activation (A4) in sensitive layers like RMSNorm causes catastrophic drift. Spatial omics fusion benefits are non-monotonic across granularities. (Source: 2026-09-06, 2026-09-14, 2026-09-17, 2026-09-18, 2026-09-21)
Tools (tools)
- Automated CDP clicks are indistinguishable from humans (isTrusted: true). In large-scale migrations, parallelization by name-prefix is more efficient than directory rsync. In high-concurrency environments, worker limits must account for ‘grandchild’ processes to prevent OOM. (Source: 2026-09-05, 2026-09-12, 2026-09-15)
Token Usage Statistics
Peak Day: 2026-09-30 — $598.47 / 1452.7M tokens
Daily Average: $251.60