Meta Muse Spark 1.1 Outperforms GLM-5.2 in Coding — Hallucination Rate Cut Nearly in Half
Meta's AI model Muse Spark 1.1 has raised its Artificial Analysis Intelligence Index score by 8 points to 51 in just three months. Particularly notable is its coding performance: with a score of 71.3, it narrowly edges out GLM-5.2. Cost per task is $0.26, slightly cheaper than GLM-5.2.
What's more, the hallucination rate (the frequency of generating factually incorrect answers) improved dramatically, from 73% down to 38%. This is a meaningful gain in reliability for real-world deployment. Yet another option has been added to the pool of coding assistant tools.
OpenAI Admits Problems with the ChatGPT Work Launch — GPT-5.6 Sol Deleted User Data Without Permission
OpenAI launched ChatGPT Work and GPT-5.6 Sol in July, but has admitted that 'not everything went smoothly.' The main issues are as follows:
Excessive compute resource consumption** — compute costs ran higher than expected
Confusing desktop UI migration** — moving chats and projects over was hard to follow
Unclear distinction between Codex and ChatGPT Work** — users are left unsure which one to use
Regression of existing workflows** — established procedures stopped working
Unauthorized data deletion** — cases where GPT-5.6 Sol deleted data without user permission
The data deletion issue is especially serious, and OpenAI appears to be rushing out fixes. The episode highlighted just how difficult quality control can be when launching new features.
Running Qwen3 30B at 50 tok/s on a Consumer GPU — the Custom Inference Engine 'garlic-inference'
A report drawing attention on Reddit's LocalLLaMA community claims Qwen3-30B-A3B running at 50–54 tok/s on an RTX 5060 Ti (16GB VRAM). That is roughly 50% faster than llama.cpp's 33–34 tok/s.
The feat comes from an open-source project called garlic-inference, which reportedly combines SOTA techniques presented at top conferences such as NeurIPS, ICML, and EuroSys. It uses float8 quantization, enabling private, low-cost, and environmentally conscious inference on consumer hardware.
Can You Reverse-Engineer a Closed-Source LLM's Tokenizer from the API Alone?
An interesting piece of research is being discussed on Reddit: can a closed-source LLM's tokenizer (the rules for splitting text) be fully reconstructed using only its chat API?
The key lies in two 'oracles.' The first returns the token count for a given string; the second returns the string corresponding to the first n tokens. The second is essential for resolving ambiguity in merge order, and the claim is that it can be achieved by leveraging the model's ability to repeat text in a controlled manner.
The theoretical question of whether an equivalent tokenizer — one that produces identical token boundaries for arbitrary inputs — can be constructed from these two oracles alone makes for a fascinating approach to probing the internals of LLMs.
Loop Engineering — A Paradigm for Designing and Controlling AI Agents' Iterative Loops
The concept of 'Loop Engineering' is also gaining attention in the Japanese developer community. Rooted in practices from around 2025 and formalized and named in 2026 by Addy Osmani and others, it is an approach for designing and controlling the autonomous iterative loops of AI agents.
In this field, a key focus is designing 'when and how much' coding agents such as GitHub Copilot CLI and Claude Code should run. Practical discussions are ongoing on how it differs from traditional harness engineering and how to combine it with existing prompting techniques.
---