Microsoft consumes Inflection and AI Demos from Suno, Cohere, Deepseek VL and more! | E1918

Episode Summary

Executive Summary: The episode centers on Microsoft’s aggressive acquisition of Inflection’s team, which the hosts frame as a cutthroat but strategically rational move in the AI arms race. The rest of the show reviews demos of Suno’s AI music generation, Cohere’s Command R retrieval-based model, and DeepSeek’s compact multimodal model, emphasizing that AI capabilities are rapidly becoming practical, useful, and increasingly tied to questions of copyright, citations, and fair use.

Main Topics: Microsoft's acquisition of Inflection talent (Priority: 5/5): The hosts discuss Microsoft effectively hiring away Inflection’s core team, installing Mustafa Suleyman to lead a new Microsoft AI unit, and eliminating a competitor to Copilot/Bing. They debate whether this is strategic necessity, regulatory arbitrage, or simply aggressive market behavior. AI competition and infrastructure economics (Priority: 5/5): The conversation argues that frontier AI requires massive capital, compute, and distribution, making standalone companies hard to sustain. Microsoft’s move is framed as evidence that only cash-rich platforms can truly compete at the top end of AI. Suno V3 AI music generation demo (Priority: 5/5): The hosts test Suno’s music model and find it surprisingly strong at producing coherent songs with style, instrumentation, and vocals. They also discuss its copyright restrictions on artist names and its usage rights for subscribers. Cohere Command R and citation-based RAG (Priority: 5/5): Command R is evaluated as a retrieval-augmented generation model that cites web sources while answering questions. The hosts praise the usefulness of references but debate whether republishing full web content crosses copyright or terms-of-service boundaries. Copyright, scraping, and fair use in AI (Priority: 4/5): A long discussion covers whether models can use open web content, the difference between factual data and copyrighted expression, and why citation, opt-out mechanisms, and licensing systems may be needed for a healthier internet. DeepSeek multimodal reasoning on small models (Priority: 5/5): The hosts showcase DeepSeek’s 7B open-source multimodal model, which can reason over images, infer context, and generate useful explanations or code from visual inputs, illustrating how small models are becoming surprisingly capable.

Key Arguments: Microsoft’s Inflection move shows that major enterprises now believe AI capability is so strategically important they will do almost anything to own the talent and technology. Frontier AI is prohibitively expensive, so companies without enormous cash flow and infrastructure are at a disadvantage compared with Google or Microsoft. Hiring a team may be a way to sidestep merger scrutiny, but it still raises questions about competition and investor treatment. Suno demonstrates that AI-generated music is moving beyond novelty into commercially usable quality, even if copyright concerns remain unresolved. Command R is compelling because it answers with citations, reducing hallucination risk and making the model more trustworthy for research-style use. AI systems should ideally provide attribution and traffic back to original sources rather than republishing entire articles or training outputs without compensation. DeepSeek shows that relatively small open-source models can deliver strong multimodal reasoning, suggesting rapid deployment on laptops and eventually phones. Smaller, faster models used with multiple passes or reflection may outperform expensive models in practical workflows, especially when grounded in external data.

Data Points: Inflection funding raised: $1.2 billion to $1.3 billion - Estimated amount the company raised to build a foundation model and consumer chatbot competitor to OpenAI Inflection compute scale: One of the largest purchasers of NVIDIA hardware last year - Described as evidence of the company’s serious AI infrastructure investment Microsoft AI talent move: Entire team effectively recruited away - The hosts describe Microsoft as having gutted Inflection and moved Mustafa Suleyman into a new Microsoft AI leadership role Apple AI acquisitions: 23 AI companies last year / 30+ by 2023 - The discussion references Apple’s quiet pattern of tuck-in AI acquisitions; the transcript contains both figures as reported claims DeepSeek model size: 7 billion parameters - Used to highlight how much capability can fit into a small open-source multimodal model Suno model version: V3 - The music-generation demo centers on Suno’s newest model release Suno commercial rights: Subscribers own songs generated while subscribed - The hosts discuss usage rights for paying Pro/Premiere users Cohere model: Command R - The retrieval-augmented generation model demoed for answer quality and citations Current market size claim in demo: $165 billion to $355 billion - Numbers surfaced by Command R from a web source during the solar-panel market overview example

Pivotal Quotes: "Satya will take all your employees." — Host: Joking but pointed summary of Microsoft’s aggressive AI talent strategy after the Inflection move "This is how important building AI technology is to enterprises." — Host: Explaining why companies like Microsoft are willing to make unconventional moves to control AI capabilities "This is awesome because you can now at least trust that this is not a hallucination and you can go find out where these numbers come from." — Host: Praise for Command R’s citations and RAG-based sourcing

Implications: The episode suggests AI competition is shifting from model demos to control over talent, compute, attribution, and distribution. Winners will likely be platform companies or open-source ecosystems that can combine capability with trustworthy sourcing and viable rights management.