Big Tech's Four Giants to Pour $725 Billion Into AI

According to a Financial Times report, the combined budgets that the four tech giants—Google, Amazon, Microsoft, and Meta—will spend this year on AI data centers, chips, and infrastructure total roughly $725 billion. As the AI race intensifies, each company is making clear that it will spare no expense on infrastructure investment.

Behind this massive spending lies the fact that demand for generative AI continues to expand faster than expected. Cloud providers are scrambling to build out data centers and procure GPUs to keep up with customers' AI workloads. That said, with investment ballooning to this scale, cautious voices about the prospects of recouping the spending are starting to emerge even in the market, and the balance with profitability will likely become a key point of focus going forward.

DeepSeek V4 Outperforms $500 Billion-Class Labs with a 200-Person Team, Making Waves

According to a report by Curious Models that has been gaining traction on Hacker News, DeepSeek V4, developed by China's DeepSeek with a team of just 200 people, shows performance capable of rivaling major U.S. AI labs that pour hundreds of billions of dollars into their research.

DeepSeek's approach of delivering state-of-the-art results with limited resources is shaking up the assumption in AI development that "investment equals performance." On Japan's Zenn, DeepSeek V4's low API pricing has also become a talking point under the headline "1/1786th the Price of Claude Opus," drawing attention on the cost-performance front as well.

This "small team, high efficiency" approach, in stark contrast to Big Tech's massive investments, offers a fresh perspective on how AI development itself should be done.

Major AI Companies Agree to Classified Contracts with the U.S. Department of Defense

According to a Wall Street Journal report, major AI companies, including OpenAI and Anthropic, have agreed to contracts involving classified work for the U.S. Department of Defense (the Pentagon).

This is a significant development, showing that the distance between AI companies and the military and national security sector is shrinking rapidly. While some AI researchers have expressed concerns about military applications, government and defense-related demand has grown to a scale that cannot be ignored as a matter of business reality.

As AI technology becomes embedded in the core of national security, questions around transparency and ethical guidelines are bound to face even greater scrutiny going forward.

Microsoft Brings an AI Legal Agent to Word

Microsoft has announced "Legal Agent," an AI that works inside Word. The agent automatically handles contract review, editing suggestions, and cross-checking clauses against internal guidelines.

Reviewing legal documents is a time-consuming task, which makes this a clear use case for reducing the burden on lawyers. That said, the quality of AI-driven legal judgments and the question of where responsibility lies will continue to require practical validation.

Microsoft has been pursuing a strategy of embedding AI across its Office products through the Copilot series, and this Legal Agent is positioned as part of that effort.

Safety Drift After Fine-Tuning Demonstrated Across 100 Models

The paper "Safety Drift After Fine-Tuning," featured on Hugging Face Daily Papers, analyzed how benign fine-tuning affects safety across 100 models.

The results are noteworthy. In many of the models fine-tuned for practical domains such as medicine and law, the researchers observed "heterogeneous drift," where some safety metrics improve while others deteriorate. In other words, safety cannot be evaluated on a single scale and needs to be reassessed for each use case.

These findings strongly suggest that systematic safety testing is needed not only for base models but also for fine-tuned models. Particularly when deploying in high-risk domains such as medicine and law, this issue cannot be overlooked.

PFlash: 10x Speedup for 128K Contexts on an RTX 3090

PFlash, a project making the rounds in Reddit's LocalLLaMA community, reportedly achieves a 10x prefill speedup over llama.cpp when processing 128K-token contexts on an RTX 3090.

One of the major factors determining the practicality of local LLMs is how quickly they can process long contexts. PFlash's results show that even consumer-grade GPUs can handle long documents at practical speeds, which could also have a positive impact on running agents locally and building RAG systems.

In the open-source LLM ecosystem, this kind of inference-speed optimization remains an enduringly important theme.