DeepSeek Raises ~$10B, Prioritizing AGI Research Over Profit
Chinese AI startup DeepSeek is in the process of raising approximately $10 billion (~¥1.5 trillion), a round that would value the company at around $45 billion (~¥6.8 trillion). Most notably, founder Liang Wenfeng has explicitly told investors that the company will prioritize R&D toward AGI (artificial general intelligence) over short-term monetization.
Raising massive amounts of capital while taking a 'we won't make money right away' stance is a rare approach in the US–China AI race. American players like OpenAI and Anthropic also champion the pursuit of AGI, but they are simultaneously pushing rapid commercialization. How far DeepSeek actually follows through on this policy — and how long investor patience holds out — will be the key questions going forward.
OpenAI Releases Beta Plugin Integrating ChatGPT Directly into PowerPoint
OpenAI has released a new beta add-in that integrates ChatGPT directly into Microsoft PowerPoint. The plugin can automatically build presentations from notes, documents, and images, as well as edit existing slides. It is available on all plans worldwide (Free, Plus, Pro, Team, and Enterprise).
However, OpenAI itself recommends backing up important slides in advance. The plugin carries a warning that it may unintentionally delete content — a reminder that it is still in beta. It is highly practical, but production use will require careful operation.
Google I/O 2026 Dialogues: CEO Pichai Talks the Future of AI Search
At the 'Dialogues' stage held as part of Google I/O 2026, leaders including Alphabet CEO Sundar Pichai discussed AI, quantum computing, robotics, and the future of creativity.
The same day, Pichai also spoke on a New York Times podcast about people's anxieties around AI search. With some arguing that Google's AI-integrated search is something users 'will end up using whether they like it or not,' his comments directly addressed users' mixed feelings about the changing search experience.
Epoch AI Analysis: Frontier Labs Still Aren't Using Most of Their Compute
An Epoch AI analysis is drawing attention for pointing out that labs at the frontier of AI research are not yet actually utilizing most of the compute they hold. Infrastructure such as GPUs and TPUs has been stockpiled in vast quantities, but the software and operational maturity needed to use it at full capacity may not have caught up.
This suggests that the AI industry's 'compute arms race' hinges not just on securing hardware, but also on developing ways to use it efficiently.
Falling AI Prices Are a 'Software' Story, Not a 'Hardware' Story
An analysis piece from Weighty Thoughts argues that the main driver of the sharp drop in AI service prices is not hardware progress but software-side innovation (model efficiency improvements, distillation, quantization, and so on). The view is that better model architectures and inference optimization are the real stars of cost reduction — chips simply getting cheaper is not the whole story.
If this analysis is right, falling prices may continue as a structural trend rather than a temporary one. Software innovation under competitive pressure will push prices down even further.
'Domain-Camouflaged Injection Attacks' on Multi-Agent LLM Systems Evade Detection
A new paper published on arXiv warns of the threat posed by 'Domain-Camouflaged Injection Attacks' against multi-agent LLM systems. The technique embeds malicious prompts disguised as content from legitimate domains, allowing it to slip past conventional security filters.
As multi-agent systems in which agents collaborate with each other continue to spread, the importance of input validation and sanitization is growing ever greater.
BeeLlama v0.2.0 Speeds Up Local LLM Inference by Up to 4.9x
BeeLlama v0.2.0 has been released, delivering major inference speedups through its DFlash technique. On a single RTX 3090, it reaches up to 164 tok/s on Qwen 3.6 27B (4.40x faster) and up to 177.8 tok/s on Gemma 4 31B (4.93x faster). That is more than twice as fast even compared to MTP (Multi-Token Prediction).
An approach that keeps prompt processing speed roughly on par with the baseline while dramatically improving generation speed alone significantly raises the practicality of local LLMs. Major models exceeding 100 tok/s inference on consumer-grade GPUs is a big milestone for the community.