Fields Medalist Achieves 'PhD-Level' Math Research with ChatGPT 5.5 Pro, Nvidia Embarks on $40 Billion in Annual AI Investments, AMD GPUs × MTP Double Local LLM Speed

Fields Medalist Achieves 'PhD-Level' Mathematical Research with ChatGPT 5.5 Pro

Fields Medalist Timothy Gowers presented an unsolved number theory problem to OpenAI's ChatGPT 5.5 Pro, and in under an hour the model produced a result improving an exponential bound to a polynomial bound. Gowers himself reported the result, and MIT researchers have also described the idea as "completely original."

Gowers's takeaway is striking: "The bar for a mathematical contribution has shifted to proving things LLMs cannot do." This suggests that as LLM mathematical capabilities rapidly improve, the very role expected of human mathematicians is changing.

The result was reportedly derived without any human assistance, and it is drawing significant attention from both the mathematics and AI communities as a sign that autonomous AI mathematical research has entered a new stage.

---

Nvidia's AI Investments Have Already Reached $40 Billion This Year

According to TechCrunch, Nvidia has already committed roughly $40 billion (about 6 trillion yen) in AI-related equity investments for 2026. Nvidia continues to make clear its position as a major investor in the AI ecosystem.

This scale shows that Nvidia is expanding its influence over the entire industry through equity stakes in AI companies, going beyond GPU sales revenue. The move is consistent with a recent FT report that Big Tech's AI investments have reached $725 billion while free cash flow has sunk to record lows.

As the AI race shifts from an infrastructure-investment phase to an ecosystem-dominance phase, Nvidia's strategic position is growing even stronger.

---

Qwen3.6-27B × MTP Gets 2x Faster on Dual AMD MI50s

A benchmark result has been reported in Reddit's LocalLLaMA community, using two Radeon Instinct MI50 GPUs — an older AMD generation — together with Multi-Token Prediction (MTP) on the Qwen3.6-27B model.

In tests using an MTP-enabled fork of llama.cpp, throughput went from a baseline of 26.2 tok/s to 39.6 tok/s with MTP alone (about 1.5x), and 59.8 tok/s when combined with tensor parallelism (about 2.3x). The overall draft acceptance rate was recorded at around 78%.

On a real 18K-token coding prompt, the combination of MTP and tensor parallelism delivered roughly a 2x speedup, from 23.07 tok/s to 47.75 tok/s. Meanwhile, MTP alone showed a slowdown in prefill speed, an issue that is reportedly now being fixed.

The demonstration that MTP delivers benefits not only on NVIDIA GPUs but also on older AMD GPUs is highly significant for the local LLM community.

---

Transformers Hit a Wall on Logical Reasoning — Structural Limits Spark Debate in Reddit ML Circles

A discussion is heating up in Reddit's MachineLearning community around "hitting a wall when trying to get Transformers to do genuine logical reasoning."

The poster questioned the industry's approach of trying to give a stochastic next-token predictor logical reasoning abilities through prompt engineering, arguing that "no matter how much you scale, logic will not emerge from a model that fundamentally lacks a reasoning architecture." They argue that a shift toward more foundational approaches such as Energy-Based Models is needed.

Some comments liken RAG and Chain-of-Thought to "expecting a calculator to spontaneously emerge from an expensive dictionary," underscoring practitioners' exhaustion and their awareness of the structural challenge. In fact, teams running LLMs in production have repeatedly flagged logical breakdowns on edge cases as a recurring problem.

---

Factory's "Missions" Shares Design Principles from 16 Days of Autonomous Multi-Agent Operation

An article published on Zenn introduces seven design principles that the Factory "Missions" project derived from running a multi-agent system autonomously for 16 consecutive days.

The biggest challenge is that human attention becomes the bottleneck. Current LLMs can process 50 tasks simultaneously, but humans have a limited capacity to review the results. To address this, Factory manages tasks at the granularity of "Missions," an approach that makes coordination between agents and human intervention points explicit.

Built on a talk by Luke Alvoeiro at the AI Engineer conference, the article is a useful reference of practical insights for multi-agent design.

---