AI News: 18 Breaking Stories You Missed This Week
This week's AI news roundup covers the release of DeepSeek V4, a near state-of-the-art open-source model with dramatically lower pricing than competitors, along with several other open-weight model releases. Major drama includes the Musk vs. Altman trial, Anthropic's controversial billing bug, Google's Pentagon AI deal, and China blocking Meta's acquisition of Manus. Rapid-fire updates include Google's new file generation features, an 11 Labs music platform, and a Mayo Clinic AI model that can detect pancreatic cancer up to 3 years early.
Summary
The video opens with a detailed breakdown of DeepSeek V4, a newly released open-source model with a 1 million token context window that performs near state-of-the-art levels across benchmarks in math and Q&A. Its most disruptive feature is pricing: $1.74 per million input tokens and $3.48 per million output tokens, compared to GPT 5.5 at $5 input/$30 output and Claude Opus 4.7 at $5/$25. The host argues this pricing, combined with open-weight availability for local deployment, explains why frontier labs and stock markets react negatively to each DeepSeek release. He also notes that DeepSeek achieves this under US GPU export restrictions, forcing more efficient training methods.
Other open-weight model releases are highlighted: Nvidia's Neotron 3 Nano Omni, a multimodal model supporting text, image, audio, video, documents, and graphical interfaces, designed for AI agent use cases and runnable on small hardware like the DJX Spark box. Poolside AI released Laguna XS2 (33B open-weight) and Lagona M1 (225B), both currently free. Mistral released a new remote agent model based on their Medium 3.5, also open-weight. Alibaba released Qwen Image 2.0 Pro for image generation, currently ranked ninth on the Arena leaderboard. XAI released Grok Voice ThinkFast 1.0, a low-latency voice model now powering Starlink's customer support.
The drama section begins with the Elon Musk vs. Sam Altman trial, where Musk is suing OpenAI for abandoning its founding nonprofit mission. Coverage from The Verge describes Musk's testimony as combative, evasive, and damaging to his own case, with jurors visibly frustrated. Next, Anthropic faced backlash after a billing bug caused users on the $200/month Claude Max plan to be charged extra when their code repositories contained references to third-party harnesses like Hermes or OpenClaw. Anthropic initially refused refunds despite acknowledging the bug, only reversing course after posts about the incident went viral with over 1 million views. Critics noted the irony of a self-described 'ethical AI' company scanning user code to penalize use of competing tools.
Google signed a deal with the Pentagon allowing AI use for any lawful government purpose, contradicting a 2014 promise made to DeepMind founders that their technology would never be used for military or surveillance purposes. Over 600 Google employees had petitioned against the deal. The host frames Google as being in a no-win situation between government relations and internal ethics. China blocked Meta's $2 billion acquisition of AI firm Manus despite the company having relocated to Singapore, creating a legally complex situation as staff and capital had already been integrated.
Rapid-fire updates include: Microsoft and OpenAI restructuring their deal to remove the AGI clause and make Microsoft's license non-exclusive through 2032, with OpenAI immediately signing a deal to bring models to AWS. Google Gemini added file generation for PDFs, Word, Excel, Docs, Sheets, and more. Google Translate launched AI-powered pronunciation practice. Google Photos introduced an AI try-on feature using photos of clothes you already own. 11 Labs launched a music creation and remix platform. Spotify introduced a verified badge to distinguish human artists from AI-generated profiles. X launched a new AI-powered advertising platform. The video closes with coverage of a Mayo Clinic AI model that can detect pancreatic cancer on routine CT scans up to 3 years before a clinical diagnosis, which the host highlights as the kind of meaningful, humanity-benefiting AI application that justifies the technology's development.
Key Insights
- DeepSeek V4 costs $1.74 per million input tokens and $3.48 per million output tokens — roughly 3x cheaper than Claude Opus 4.7 and nearly 9x cheaper on output than GPT 5.5 — while performing at near state-of-the-art levels, which the host argues is why each DeepSeek release causes market disruption and panic among US frontier labs.
- The host argues that open-weight models have now caught up to the point where state-of-the-art frontier models are effectively overkill for most enterprise use cases like document summarization, pattern finding, and customer support agents, making cheaper and locally-deployable open models a more practical choice.
- Anthropic's billing bug actively scanned user code repositories for keywords like 'Hermes' and 'OpenClaw' — third-party agent harnesses — and either blocked service or charged users extra, which the host and commenters argue contradicts Anthropic's positioning as an ethical AI company; Anthropic only issued refunds after the incident went viral with over 1 million views.
- Google's Pentagon AI deal directly contradicts a core condition from its 2014 DeepMind acquisition, in which DeepMind founders extracted a promise that their AI would never be used for military or surveillance purposes — a commitment the host says Google appears to have abandoned in favor of maintaining favorable government relations.
- A Mayo Clinic AI model can detect signs of pancreatic cancer on routine abdominal CT scans up to 3 years before a clinical diagnosis, validated by back-testing on historical scans, which the host cites as the kind of tangible, life-saving AI application that justifies continued AI development amid broader criticism.
Topics
Full transcript available for MurmurCast members
Sign Up to Access