Did GPT-5.5's Thought Process Leak from Codex?
A report surfaced on Reddit's LocalLLaMA community claiming that while using GPT-5.5-medium via Codex, the model's internal chain of thought got mixed into the output.
The output a user received while working on a project contained fragments that clearly looked like an internal reasoning process, such as “Need absolute path. Need know cwd absolute” and “Use markdown. final with path. Need avoid bogus path.”
The reporter noted that this thinking pattern closely resembles an idea discussed in the same community about five months ago: having AI think in caveman-style shorthand to improve token efficiency.
There has been no official comment from OpenAI at this point, so its authenticity is unknown. If it is genuine, though, it would be a fascinating case offering a glimpse into what reasoning strategy the GPT-5.5 series employs internally.
Karpathy's MicroGPT Hits 50,000 t/s on an FPGA
A project has been published that runs Karpathy's educational GPT implementation, MicroGPT, on an FPGA, achieving a remarkable throughput of 50,000 tokens/s.
At just 4,192 parameters, the model is extremely small, but the speedup comes from placing the weights in onboard ROM, which eliminates access latency to external memory. With 16-bit weights, the practical ceiling on current FPGAs is estimated at 20–30 million parameters, but combined with efforts like those of Taalas, running SLMs (Small Language Models) on FPGAs is becoming increasingly viable.
Details of the project are available at [TALOS V2](https://v2.talos.wtf/), and it can also be explored via the [GitHub repository](https://github.com/Luthiraa/TALOS-V2).
An In-Depth Comparison of 10 Image Generation Models on Mac
A detailed benchmark comparing 10 image generation models — from SD 1.5 to Flux dev, Qwen-Image, and Gemini — on an M1 Max with 64GB of RAM was shared on Reddit.
Key findings include:
Qwen-Image Lightning** (8-step distilled version) beats the full model on quality while running 9x faster (10 minutes vs 93 minutes)
Flux dev** is the best at photorealism, but its anglophone bias is strong — it was observed adding cilantro to ramen and turning izakaya into tea houses
Gemini** excels at kanji rendering and understanding cultural context, but is limited to the cloud service
SDXL Turbo** can generate an image in 5 seconds, but the quality is rough
Particularly striking was the “cultural accuracy gap.” The fact that the geographic skew of training data matters more than model size for Japanese and other Asian content is an insight that cannot be ignored for anyone looking to use local AI going forward.
US Navy Signs $100 Million Underwater Drone Contract with AI Company
The US Navy has signed a contract with an AI company for training underwater drones to detect mines in the Strait of Hormuz. The deal is worth approximately $100 million.
Updates to detection algorithms, which previously took months, can reportedly now be completed in days. It is a sign that practical AI deployment in the military domain is steadily advancing, but the balance with ethical concerns over autonomous weapons systems will likely remain a point of debate.
The Claude Code Ecosystem Is Expanding Rapidly
Knowledge sharing around Claude Code is heating up in Japanese tech communities as well.
A **Cognee graph memory toolkit** has been released, giving Claude Code a mechanism to accumulate memory across sessions (rules, lessons learned, design decisions, and incident records). The toolkit bridges the graph database Cognee and Claude Code, and is available under the MIT license.
A practical article on **safety configurations using Claude Code's hook feature** has also appeared. It walks through concrete ways to use hooks to prevent five configuration mistakes actually reported in GitHub Issues (accidental file deletion, overly broad permissions, and the like).
Also drawing attention is an article that examines, from a hands-on Claude Code perspective, research findings that assigning a persona to an AI model (e.g., “You are a senior engineer”) can actually reduce accuracy.
AIANO Paper Streamlines IR Dataset Creation
An article explaining AIANO, a method that dramatically reduces the effort of building evaluation datasets in the information retrieval (IR) field through collaboration with AI, has been published on Zenn.
Creating the “ground truth data” needed to evaluate RAG systems has traditionally been a very labor-intensive task. AIANO streamlines this process through interactive collaboration with AI, making it a practical resource for practitioners working on RAG quality evaluation.
---