DeepSeek Plans $7.35 Billion Fundraise — V4.1 Update Also Due Next Month
Chinese AI company DeepSeek is reportedly raising around 50 billion yuan ($7.35 billion). According to the report, which cited two people familiar with the matter, the company is looking to accelerate its commercialization and monetization strategy.
While details about investors and valuation remain unclear, the large-scale fundraising plan — coming just after the recent release of the DeepSeek V4 preview — suggests the company is seriously gearing up for its next phase. A V4.1 update is also slated for release next month, and competition in the open-source AI model race looks set to intensify further.
SoftBank Cuts OpenAI-Related Loan from $10 Billion to $6 Billion
According to The Decoder, SoftBank has sharply reduced an OpenAI-related loan facility from an originally planned $10 billion to $6 billion. The move appears to be driven by lenders' concerns about private AI company valuations.
The AI industry's race for capital continues — Anthropic, for one, is planning to raise up to $5 billion — but lenders are becoming noticeably more cautious. As AI startup valuations climb rapidly, financial markets are starting to cast a colder eye on the sector.
TikTok Scales Back AI-Generated Video Descriptions After String of Errors
TikTok has been scaling back its AI-generated video descriptions (captions and summaries) after a string of errors. According to the BBC, multiple cases have been confirmed in which the AI-generated descriptions contained inaccurate content.
Every platform is racing to bake AI into its products, but AI-generated content that is directly visible to users makes the difficulty of quality control especially stark. TikTok's response is a case that once again underscores the importance of quality assurance when running AI in production.
UltraCompress — Mathematically Lossless 5-Bit LLM Compression Arrives as Open Source
A new project called UltraCompress is drawing attention on Hacker News. It claims to achieve LLM compression via mathematically lossless 5-bit quantization, and has been released as open source on GitHub.
Making LLMs smaller and faster is a critically important theme from an inference cost reduction perspective. Conventional quantization methods inevitably lose accuracy, so the claim of being mathematically "lossless" is what sets this project apart. That said, the community is calling for evaluation under real-world usage conditions, as well as scrutiny of the compression and decompression overhead.
AI Face Swaps Exploited in Romance Scams — WIRED Reports on the Reality
WIRED has published a detailed report on romance scams leveraging hyper-realistic AI face swap technology. Cases in which victims are won over with fake faces and voices and then swindled out of large sums of money are reportedly surging.
Abuse of AI-generated technology has already become a social problem in many countries, and as face swap technology grows more accurate, there are growing concerns that the damage will become more sophisticated and widespread. The balance between technological progress and security countermeasures is once again being called into question.
Ring 2.6 1T-Parameter Model Lands on OpenRouter
The Reddit LocalLLaMA community is buzzing about Inclusion AI's "Ring 2.6 1T" model becoming available on OpenRouter. It's a trillion-parameter-scale Mixture of Experts (MoE) model, and for now it can be tried for free via OpenRouter.
The community is busy verifying its real-world inference quality and benchmark results. As a large-scale open-source model, it stands out as a notable release following the DeepSeek V4 preview.
Nick Bostrom's Vision of a "Big Retirement" for Humanity
In an interview with WIRED, philosopher Nick Bostrom put forward the concept of the "Big Retirement" as a reason humanity should pursue advanced AI. Bostrom believes AI could bring about a "solved world," and argues that humanity should accept the risks involved in realizing it.
Bostrom has long warned about the existential risks of AI, so these remarks — suggesting that advancing AI itself can become humanity's purpose — are drawing attention as a somewhat paradoxical stance.