LIVE · WED, JUL 08, 2026 --:--:-- ET
Issue Nº 78 COST TOTAL $14709.39 ARTICLES TODAY 9 TOKENS TOTAL 9.36B
Podcast
aiexpert
Subscribe · $8/mo
Running the wire
Market Gaming GPU Memory Shortage Pushes RTX 5090 Over $3,000 as AI Data Centers Outbid Consumers Funding Crusoe Energy in Talks to Raise $3B at $30B Valuation; AI Data Center Operator Triples Mark in 9 Months Breaking LangChain + NVIDIA Nemotron NemoClaw Achieves 10x Cost Reduction for Enterprise Agents Funding Anthropic in talks with Samsung for custom AI chip on 2nm foundry process Breaking OpenAI launches GPT-Live voice models with full-duplex streaming and live delegation Breaking JetBrains develops governance layer for Claude Code, Codex, Gemini CLI across teams Policy White House Executive Order 14412 mandates post-quantum cryptography by 2030 for federal systems Chips Rapidus targets 2027 2nm mass production with single-wafer processing; Japan bets national semiconductor revival on one Hokkaido fab Funding EdVisorly closes $13.3M Series A for college transfer AI automation Market Together AI launches Provisioned Throughput: reserved capacity for open models at 90% below Claude Opus list Chips NVIDIA Nemotron 3 Ultra matches closed-model performance at 10x lower cost via LangChain harness tuning Chips Budget smartphone sales collapse 22% as memory costs hit 64% of BOM amid HBM supply crunch Chips DeepSeek developing in-house inference chip to cut Nvidia, Huawei reliance Chips SiPearl Rhea CPU enters bring-up; Europe's first HPC processor on track for end-2026 Market Waymo Begins Driverless Operations in San Diego, Las Vegas, Tampa, Denver Market Oregon utility raises data center rates 29.7% under POWER Act; residential bills cut 1.3% Funding Gradium voice AI startup extends seed to $100M+ with $30M NVIDIA-backed tranche Funding SambaNova raises $1B at $11B valuation; JPMorganChase deploys SN50 for on-prem AI inference Funding Blue Origin closes first outside funding round at $130B valuation; raises $10B led by Coatue Breaking OpenAI releases GPT-5.6 Sol, Terra, Luna publicly after ending government preview limits Market Gaming GPU Memory Shortage Pushes RTX 5090 Over $3,000 as AI Data Centers Outbid Consumers Funding Crusoe Energy in Talks to Raise $3B at $30B Valuation; AI Data Center Operator Triples Mark in 9 Months Breaking LangChain + NVIDIA Nemotron NemoClaw Achieves 10x Cost Reduction for Enterprise Agents Funding Anthropic in talks with Samsung for custom AI chip on 2nm foundry process Breaking OpenAI launches GPT-Live voice models with full-duplex streaming and live delegation Breaking JetBrains develops governance layer for Claude Code, Codex, Gemini CLI across teams Policy White House Executive Order 14412 mandates post-quantum cryptography by 2030 for federal systems Chips Rapidus targets 2027 2nm mass production with single-wafer processing; Japan bets national semiconductor revival on one Hokkaido fab Funding EdVisorly closes $13.3M Series A for college transfer AI automation Market Together AI launches Provisioned Throughput: reserved capacity for open models at 90% below Claude Opus list Chips NVIDIA Nemotron 3 Ultra matches closed-model performance at 10x lower cost via LangChain harness tuning Chips Budget smartphone sales collapse 22% as memory costs hit 64% of BOM amid HBM supply crunch Chips DeepSeek developing in-house inference chip to cut Nvidia, Huawei reliance Chips SiPearl Rhea CPU enters bring-up; Europe's first HPC processor on track for end-2026 Market Waymo Begins Driverless Operations in San Diego, Las Vegas, Tampa, Denver Market Oregon utility raises data center rates 29.7% under POWER Act; residential bills cut 1.3% Funding Gradium voice AI startup extends seed to $100M+ with $30M NVIDIA-backed tranche Funding SambaNova raises $1B at $11B valuation; JPMorganChase deploys SN50 for on-prem AI inference Funding Blue Origin closes first outside funding round at $130B valuation; raises $10B led by Coatue Breaking OpenAI releases GPT-5.6 Sol, Terra, Luna publicly after ending government preview limits
Breaking

OpenAI launches GPT-Live voice models with full-duplex streaming and live delegation

OpenAI announced GPT‑Live, a new generation of voice models built on full-duplex architecture that enables simultaneous listening and speaking. Unlike previous cascaded approaches (transcribe→generate→synthesize) or turn-based models that wait for silence, GPT‑Live makes interaction decisions many times per second: whether to speak, pause, continue listening, or invoke tools. The model can produce backchannel signals ('mhmm', 'yeah') and engage in natural conversational flow rather than rigid back-and-forth exchanges.

GPT‑Live includes a delegation mechanism for complex tasks: when a question requires web search, deeper reasoning, or agentic work, GPT‑Live delegates to frontier models (initially GPT‑5.5 Instant) while maintaining the conversation flow. This decoupling allows the voice interaction layer to stay responsive while background models handle computation-heavy tasks. Two versions launched: GPT‑Live‑1 and GPT‑Live‑1 mini. OpenAI plans to bring both to the API soon for developers and enterprises. The models are currently rolling out to ChatGPT Voice users globally.

Human evaluations show GPT‑Live‑1 is strongly preferred over ChatGPT Advanced Voice Mode on conversational pleasantness, turn-taking, interruption handling, and naturalness. On GPQA (expert-level scientific reasoning), GPT‑Live‑1 substantially outperforms Advanced Voice Mode. On BrowseComp (agentic web search), GPT‑Live‑1 shows strong gains. The move addresses architectural limitations of previous generations: cascaded models lost information across handoffs and felt slow; turn-based models required silence detection and felt rigid.

For architects building voice-native agents and multimodal systems, GPT‑Live signals OpenAI's shift toward continuous, stateful voice interactions that can maintain context across multiple async background tasks. The full-duplex + delegation pattern is architecturally significant: it separates the interaction layer (low-latency voice) from the reasoning layer (higher-latency models and tools), enabling real-time responsiveness even when the backend is occupied. This design mirrors agentic patterns where agents need to show activity ('thinking out loud') while executing tools in parallel.

Sources

Get the signal before the noise.