OpenAI Solves an 80-Year-Old Math Conjecture — Is It Real This Time?
OpenAI has announced that it has overturned a mathematical conjecture that had gone unsolved for 80 years. Back in October 2025, the company also announced that "GPT-5 solved a difficult math problem," but that claim fell flat after mathematicians pointed out it was "a rediscovery of an already known solution."
Regarding this new result, commentary articles analyze that "the mistake from last time has not been repeated." Verification within the mathematical community is also underway, and if the result holds, it would be a groundbreaking event signaling that AI's mathematical reasoning capability has entered a new stage. However, since the final verification of the mathematical proof will take time, cautious views remain.
OpenAI Releases Codex App Server — A New Foundation for Embedding AI Coding Agents into Products
OpenAI has released "Codex App Server." Unlike conventional code completion tools, this is a full-featured coding agent foundation that leverages the latest models such as GPT-4o.
Its biggest feature is that it can be used not only as a VS Code extension but also to build AI coding capabilities into your own products. Developers will be able to offer AI-powered code generation, review, and modification features within their own applications. Since OpenAI's product lineup changes frequently, checking the official documentation before implementation is recommended.
SSV: Sparse Speculative Verification — A New Approach to Accelerating LLM Inference
A paper titled "SSV: Sparse Speculative Verification for Efficient LLM Inference" has been published on arXiv. The research proposes a new method for improving the inference speed of LLMs (large language models).
Speculative Decoding is already a well-known acceleration technique, but SSV further reduces computational cost by performing the verification step sparsely. Practical improvements in inference latency are expected.
Could AI Make Conflicts Worse? — The Risk Posed by "Conformity Bias"
According to research published on arXiv, as AI models are deployed in societies affected by armed conflict — with journalists, humanitarian aid workers, governments, and ordinary citizens using AI as an information source — AI's "conformity bias" is quietly but surely distorting their judgment.
The findings call into question the very premise that "AI is neutral," showing that ethical considerations around AI are especially critical when it is used in conflict zones.
Qwen3.6-27B-MTP's Multi-Token Prediction Tested in Real-World Benchmarks
In Reddit's LocalLLaMA community, the Multi-Token Prediction (MTP)-enabled version of Alibaba's Qwen3.6-27B model has been put to the test in real-world benchmarks.
According to test results on a MacBook M5 Max (128GB RAM), while the standard version (Q6_K_XL) achieved 19 tok/s, the MTP version reached 22.3 tok/s, confirming a speedup of approximately 17%. However, some users noted that "the improvement is not as dramatic as expected," and further gains may be possible with optimized settings.
---