Tsinghua Research Reveals: "Higher Scores ≠ Better Teacher"

A collaborative study from Tsinghua University's THUNLP Lab, ShanghaiTech University, and UIUC has upended a fundamental assumption in AI: that stronger models make better teachers for distillation. Their research reveals two counterintuitive principles that determine whether a "teacher" model can effectively transfer knowledge to a smaller "student" model.

Principle 1: Thinking-Pattern Consistency Over Size
Even when a teacher model scores higher on benchmarks, distillation fails unless its underlying reasoning patterns align with the student's. In one experiment, a larger model failed to improve a student at all — not because it wasn't smart enough, but because it "thought" differently.

Principle 2: New Knowledge, Not Just Higher Scores
A teacher that's simply a scaled-up version of the student provides little new information. The real learning happens when the teacher has undergone different training — like reinforcement learning (RL) post-training. Models that had RL fine-tuning transferred knowledge 16-58% more effectively than those that were merely larger.

This finding has immediate practical implications. AI labs racing to build ever-larger models may find that investing in diverse training methodologies produces more reusable knowledge than simply scaling parameters.

NVIDIA's China Dilemma

As NVIDIA CEO Jensen Huang accompanied U.S. Commerce Secretary Gina Raimondo on a visit to China, the semiconductor landscape has shifted fundamentally. Huang faces a structural barrier: not just export controls, but the rapid rise of domestic Chinese AI chip alternatives.

The market is bifurcating. Huawei's Ascend and other domestic chips are gaining traction, creating an ecosystem that may no longer need NVIDIA's hardware. This represents a deeper shift than simple geopolitics — it's about industrial sovereignty.

OpenAI's Masterstroke in Financial Engineering

AI chip startup Cerebras just completed a landmark IPO — but the real story is how OpenAI structured its partnership. Instead of paying cash, OpenAI committed to purchasing $200 billion in compute capacity over time. In exchange, OpenAI received warrants for approximately 33 million shares — nearly 3 percent of the company, with the potential to reach 10 percent.

Key terms:

• 750MW commitment for inference compute, with option to expand to 2GW
• 10-year, 6% interest loan from OpenAI to Cerebras ($1 billion)
• Interest forgiveness if Cerebras delivers on compute commitments
• Parallel to CoreWeave: OpenAI secured $350M in CoreWeave equity through similar terms

Cerebras's WSE-3 (Wafer-Scale Engine) chip is massive — 46,000mm², 900,000 cores, 21PB/s memory bandwidth. But its 44GB on-chip SRAM limits its ability to hold trillion-parameter models without external storage.

The risk: This model creates dangerous concentration. OpenAI accounts for approximately one-third of CoreWeave's long-term contracted revenue — if OpenAI ever fails to pay, it could trigger a cascade of defaults. As one analyst put it, "OpenAI's promises have become the most dangerous asset in AI."

From Text Collaboration to Interface Collaboration

Anthropic engineer Thariq Shihipar sparked a heated debate by arguing that AI agents should output HTML, not Markdown. His core insight: Markdown is optimized for human editing and version control, but those advantages are fading as users increasingly ask AI to edit content directly.

HTML advantages:

• Rich information density with embedded CSS, SVG, and interactive elements
• True interactivity — sliders, buttons, tabs that users can manipulate and feed back to AI
• Universal rendering in browsers with one-click sharing

HTML disadvantages:

• 2-4x slower generation time due to increased token consumption
• Git diff is a nightmare — messy tag soup instead of clean line comparisons
• Loss of friction-free human co-creation

The World's First Mass-Produced Manned Mecha

On May 12, 2026, Unitree Robotics unveiled the GD01 — the world's first mass-produced manned transforming mecha. Priced at 3.9 million RMB ($550,000 USD), it represents the transfer of a cultural dream from Japan to China. The 500kg machine can transform between bipedal and quadrupedal modes, has demonstrated the strength to punch through brick walls, and is slated for applications in theme parks, special operations, and high-end personal transport.

The announcement sent a clear signal: China has moved from manufacturing consumer robots to producing what was once the exclusive domain of science fiction. Unitree's humanoid robots are already deployed in Japanese airports, adding symbolic weight to the transition.

The Quiet Revolution of Advanced Packaging

TSMC's Q1 2026 earnings revealed a stunning 66.2 percent gross margin — a figure usually associated with luxury goods monopolies. Behind this number is a fundamental shift: TSMC has transformed from a contract manufacturer into the "coin minter" of the AI era.

Key insights from the earnings call:

• 3/5/7nm nodes account for 74 percent of revenue — advanced nodes dominate
• CPU's resurgence in data centers: AI inference workloads are shifting the GPU:CPU ratio from 8:1 toward 2:1 or even 1:1
• N2 (2nm) ramping early: Already in high-yield production in Q4 2025, ahead of schedule
• CoWoS and 3D integration as the new performance battleground