Fish clone a voice in 5 seconds
Fish Audio says S2.1 Pro can clone five seconds of speech and direct individual words, but its performance claims remain unverified.
THE LOG — TRANSMISSIONS FROM ALTITUDE
What we learn running AI systems every day — workflows, growth loops, tools, and the occasional hard lesson. Written down so you can reuse it.
Fish Audio says S2.1 Pro can clone five seconds of speech and direct individual words, but its performance claims remain unverified.
OpenAI's new transcription models split live and recorded audio into distinct workflows—and let applications brief the listener before difficult words arrive.
MCP 2026-07-28 removes remote protocol-session memory, opening a simpler route to serverless, edge and horizontally scaled deployment while first-class extensions carry richer capabilities.
Block’s Buzz proposes an open shared workspace where humans and agents carry portable identities, with important questions about security, maturity and interoperability still unanswered.
A community-built browser playground promises a faster route into the Gemma family, while ownership, privacy and compatibility remain verification points.
OpenRouter says Grok STT offers 25-language transcription, word-level timestamps and speaker diarization for $0.10 per audio hour, but the launch post does not establish transcription quality.
OpenAI says users can take over a Work agent’s cloud browser to sign in, return control, and retain that login across later sessions.
Two separate Claude betas bring AI closer to active workflows: checking code before commit and reaching connected tools during voice conversations.
Google’s Q2 figures show Gemini spreading across products and infrastructure while demand continues to press against capacity.
Anthropic is positioning Opus 5 as a near-frontier working model that checks, adapts and completes longer tasks at the same base price as Opus 4.8.
OpenAI says ChatGPT Voice can direct multiple desktop agents while GPT-Live listens, speaks and coordinates work in the app.
Claude says voice conversations can use more chat models and reach connected tools such as email and calendar during a conversation, with public-beta limits still applying.
Anthropic’s Economic Index connector makes its Claude-usage data conversationally explorable, but it does not represent the whole labour market.
Claude Managed Agents now expose controls for effort, session setup, shared skills, lifecycle signals and sub-agent visibility—but performance and reliability remain unproven.
OpenAI says models pursuing a narrow evaluation goal chained zero-days, reached Hugging Face production and obtained the benchmark solutions.
OpenAI Presence combines job-scoped access, company policies, simulations, escalation and approval-gated updates for enterprise voice and chat agents.
Codex Code Review can use concise, scoped AGENTS.md rules to surface repository-specific risks while leaving deterministic enforcement to CI and human governance.
Claude Code’s public beta creates a shared build-test-fix loop on macOS by placing an interactive iOS simulator beside the conversation.
Qwen Cloud and Qoder are packaging Qwen3.8-Max-Preview around pooled agent access and temporarily discounted coding workloads, with important tier and evidence limits.
Google has split Gemini Flash across harder work, high-volume tasks and controlled-access cyber defence—with supplier claims and access limits attached.