The Sequence Radar #832: Last Week in AI: Compression, Voice, and Why It All Matters
This week in AI saw pragmatic advancements in compression (TurboQuant) and voice technology (Gemini 3.1 Flash Live, Voxtral TTS), highlighting efficiency gains over capability jumps.Summary
This week's AI developments focused on practical improvements rather than groundbreaking new capabilities. Google's TurboQuant algorithm significantly compresses key-value caches in LLMs, reducing memory usage by 6x with no accuracy loss, nearing the theoretical limits of compression. Simultaneously, Google released Gemini 3.1 Flash Live, a new audio model that streamlines the voice processing pipeline, while Mistral introduced Voxtral TTS, a voice cloning model prioritizing data sovereignty. These releases demonstrate a trend towards increased efficiency, making AI more accessible and cost-effective. Beyond these, the article covers funding rounds for AI startups like Deccan AI, Harvey, and Granola, as well as legal and investment news concerning Anthropic, SoftBank, and Meta, illustrating the continued strong investment and evolving legal landscape within the AI sector.
(Source：Thesequence)