BIS Warns of Economic Spillover Risk From an AI Bust
The Bank for International Settlements (BIS) has issued a warning about the risks should the AI boom collapse. According to Bloomberg, the BIS takes the view that if AI-related exuberance were to cool off, the effects would spill over into economic growth and credit markets. The backdrop is that the scale of AI investment has already reached a level capable of affecting the entire financial system.
The BIS has long warned about emerging risks such as crypto assets and fintech, but with AI it reportedly sees "contagion risk to the financial system as a whole that goes beyond a mere tech bubble." Specifically, it envisions a scenario in which a correction of overinvestment in AI infrastructure causes the credit risk of related companies to affect banks' lending portfolios.
While this remains in the realm of speculation for now, the fact that the BIS — a policy coordination forum among central banks — has officially flagged the risk can be read as a sign that financial authorities' awareness of AI risk is growing.
DeepSeek Releases "DSpark" — Parallel Inference That Preserves Output Order
On June 27, DeepSeek released "DSpark" on GitHub. DSpark is a technique that parallelizes LLM inference while preserving the ordering (sequence) of the output.
In conventional LLM inference, the entire model must run once for every single token generated, so GPU idle time accumulates as outputs get longer. Because each run barely uses the GPU's compute resources, efficiency is extremely poor. To improve on this, there are approaches that generate multiple tokens or drafts in parallel, but previous methods carried the risk of breaking the output's ordering relationships.
DSPark tackles this problem, providing a mechanism that keeps context consistent even under parallel generation. This makes it possible to dramatically improve inference throughput while maintaining output quality. DeepSeek already has a track record of improving inference efficiency, and DSpark sits on the same trajectory as that work.
The local LLM community reacted immediately as well, and validation of its effectiveness in production use is expected to move forward.
Rakuten Cuts Costs and Latency by 30% Through AI Agent Operations — Referenced in an Anthropic Guide
Anthropic has published a guide, "Building AI agents for the enterprise," showcasing how leading companies are using AI agents to transform their work. According to ITmedia, the guide features Rakuten as a case study, reporting that the company achieved a 30% reduction in both costs and latency through proper operation of AI agents.
The point the guide emphasizes is that "just using AI for one-off questions doesn't produce results." What matters, it says, is identifying workflows that recur and consistently delegating them to agents, building a setup where the benefits accumulate over time. In Rakuten's case, embedding agents into customer support and internal business processes — with continuous improvement on top — delivered cost savings and faster response times at the same time.
Through this guide, Anthropic conveys the message that "it's not just prompt engineering — where you place agents and how you design their operation determines ROI."
Gartner Shares How to Build an "AI Agent Investment Score"
In a webinar, 亦賀忠明, a prominent Gartner analyst, presented a concrete framework for deciding how to prioritize investments in AI agents across business operations.
This method, called an "investment score," quantitatively evaluates the value AI agents can deliver and the difficulty of implementing them, by task and by industry, and uses the results to set priorities. Gartner argues that rather than rolling out AI uniformly across the entire company, investment should be allocated according to the characteristics of each individual workflow.
The session also touched on how AI agents will evolve going forward, with the view that they will shift from today's "task-execution" type toward a more autonomous "goal-achievement" type. The message for enterprises: it is important to have criteria for investment decisions in place now.
Training LLM Agents on the "Memory Staleness" Problem with Reinforcement Learning — Supersede
A paper published on arXiv, "Supersede," analyzes the "memory staleness" problem that LLM agents face in long-term conversations and presents a solution.
The problem is simple. A user moved, a price changed, a plan was revised — when facts get updated like this, an LLM agent needs to use the new information and discard the old. In practice, it often fails to do so.
The researchers ran experiments on the knowledge-update subset of LongMemEval and got a surprising result: even state-of-the-art models like gpt-5.4 drop from 92% to 77% accuracy when the agent uses self-managed memory. With full context provided, accuracy saturates at 92%, which shows that the bottleneck lies not in "comprehension" but in "memory maintenance."
Even more interesting is that this is not a memory-capacity problem. When conversations get 24× longer, accuracy falls from 68% to 28%, and proportionally increasing memory doesn't help (it stays at 28%). The problem is not the compression ratio — it scales with conversation length itself.
The researchers released an open reinforcement learning environment called "Supersede," taking an approach that rewards agents for "answering with the current value" and penalizes them for "answering with an outdated value." They demonstrate that GRPO fine-tuning can close the gap, showing that this is a matter of training, not of scale.
---