ByteDance Discovers a New Scaling Law — Implications for the Sustainability of the AI Boom

China's ByteDance has reportedly discovered a new scaling law that underpins improvements in AI model performance. According to an article in the South China Morning Post, this discovery could affect the sustainability of the AI investment boom.

Conventional scaling laws were based on the empirical rule that performance improves as you increase model size, data volume, and compute — but in recent years, they have widely been seen as hitting a ceiling. If ByteDance's discovery points to a new direction for scaling, it provides grounds to believe that AI's rapid evolution will continue. Publicly available details are still limited at this point, and the release of a paper is awaited.

President Trump: AI Regulation Should Involve as Little Intervention as Possible

In an interview with CNBC, US President Trump said of AI regulation: "guardrails are necessary, but we want to intervene as little as possible." At the same time, he emphasized that the United States is "leading by a wide margin" over China in the AI development race.

His stance of minimal regulation has been consistent, and this reaffirmed the policy of putting America's AI industry competitiveness first. That said, he also acknowledged the need for guardrails, so how to design the balance between safety and innovation remains a continuing focus.

The Real Threat of the AI Bubble, According to the IMF — "Debt" and "Data Centers That Haven't Broken Ground"

The IMF (International Monetary Fund) has warned that the real risk behind the AI bubble lies not in the technology itself, but in the "mountain of debt" accumulating behind it. One especially noteworthy data point: 60% of planned data centers have yet to break ground.

In other words, massive investment commitments have been made, but in many cases actual construction has not progressed — a structure in which delayed returns on investment intensify debt pressure. If AI demand fails to grow as expected, these projects carry the risk of turning into non-performing assets. While much of the "AI bubble" debate focuses on overheating on the technology side, the IMF's warning highlights a structural risk from the financial side.

Context Graphs — A Mechanism for AI Agents to Remember "Why They Made That Decision"

A concept called the "context graph" was introduced on the Nanonets blog. It is an approach for AI agents to structurally memorize and retain the context of their own decision-making.

Long-running agents accumulate records of past tool calls and intermediate judgments, which eventually put pressure on the context window. A context graph manages the "reasons" behind decisions as nodes in a graph structure, enabling agents to efficiently reference the rationale behind past judgments. It is drawing attention as a promising approach to the challenge of maintaining context in long-term agent operation.

On a related note, Qiita has featured a technique called "SelfCompact," which automatically determines when to compress the context of long-running agents — a sign that agent memory management is gaining attention as a critical practical challenge.

Claude Code v2.1.199-200 — Fixes Silent Sub-agent Failures

An important bug fix was made in the Claude Code changelog (v2.1.199–v2.1.200). A bug was fixed in which sub-agents cut off by rate limits returned empty results while still being treated as successful.

In multi-stage agent systems, if a failure at an intermediate step is not propagated to the parent, the final result as a whole can no longer be trusted. With this fix, sub-agent failures now propagate properly, improving the reliability of agent chains. For developers building complex workflows with Claude Code, this is a welcome improvement.

---