OpenAI ships GPT-Live, voice models that talk over you on purpose

Good morning 👋 OpenAI's new voice models don't wait for you to stop talking. That's the one thing every previous version faked.

In today's issue:

• 🔭 OpenAI ships GPT-Live, voice that talks over you on purpose
• 🧠 NVIDIA's agent data stack, ZML out of stealth, AWS's pharma win
• 🔬 Nemotron's harness-only benchmark leap, OpenAI retracts its own coding eval, Mistral's camera-only robot nav
• 🏛️ AWS locks down Claude access for the whole fleet
• 🛠️ Triage a glut of SEC 8-Ks in one prompt
• 🔗 Small-caps pitching themselves as AI infrastructure plays

🔭 THE ONE THING

🎙️ OpenAI ships GPT-Live, voice models that talk over you on purpose

OpenAI shipped GPT-Live yesterday, rolling out now in ChatGPT: voice models built to listen while they're still talking, not after you stop. That's the piece Advanced Voice Mode never solved, the half-second hold that made every exchange feel like a two-way radio. Reuters independently corroborated the core claim rather than just repeating OpenAI's own page. Real-time interruption is easy to fake in a demo and hard to fake on a live support call. Test it there before you believe the reel.

🧠 MODELS & RELEASES

• 📊 NVIDIA open-sourced its Nemotron data stack for agents: 10+ trillion pretraining tokens, a Prompt Atlas for digging through post-training samples, and synthetic personas covering roughly 53 million people across seven countries. Weights get you a model. This is what gets you one that survives a failed tool call. Hugging Face blog
• ⚡ ZML came out of stealth with an inference engine built on Hugging Face as its storage layer, and Clement Delangue is already calling it proof open-source inference has room left to run. Wait for an independent benchmark before you buy the "faster, cheaper" pitch. X post
• 🧬 AWS built a GraphRAG stack (Neptune Analytics, Bedrock's Claude 3.5 Sonnet) that cut a pharma team's target-screening cycle from months to weeks and lifted their hit rate, per AWS's own case study. Steal the architecture even if you've never touched a knowledge graph. AWS blog

🔬 RESEARCH HIGHLIGHTS

• NVIDIA + LangChain's Deep Agents. Nemotron 3 Ultra, tuned through LangChain's Deep Agents harness, now matches the top closed models on LangChain's own agentic benchmark at roughly a tenth of the inference cost. Nobody retrained the model. The gains came entirely from rebuilding the harness around it: system prompts, tool descriptions, middleware. If that holds outside this one benchmark, it says more about how much performance is sitting unclaimed in the orchestration layer than about Nemotron itself. Blog

• OpenAI re-audits SWE-Bench Pro. OpenAI ran its own audit of the coding benchmark half the industry cites and found a sizable chunk of its tasks are broken: hidden requirements, contradictory instructions, tests too strict for even a correct fix to pass. It's retracting its own recommendation of the eval. The audit process looks solid, but clock who's doing the debunking: a lab that sits near the top of a lot of other coding leaderboards. Take the finding, discount the timing. Post

• Mistral's Robostral Navigate. An 8B model that takes one RGB camera feed and a plain-language instruction and gets a robot where you told it to go. No LIDAR, no depth sensor, state of the art on the R2R-CE navigation benchmark. Small and single-camera by design. That's the actual news: navigation stacks that used to need a full sensor suite might not for much longer. Thread

🏛️ POLICY & REGULATION

• AWS shipped the Claude apps gateway: a self-hosted control plane that routes every Claude Code and Claude Desktop request through one choke point, letting IT set access rules and spend caps instead of leaving that to whoever holds an API key. Introducing Claude apps gateway for AWS
• AWS and Jamf paired Amazon Bedrock with Jamf's AI Governance so admins can push the same access and usage rules across a whole Mac fleet, the same lever they already pull for VPN profiles and app whitelisting. Neither of these is a regulator moving. It's enterprise IT catching up to tools employees already installed. Manage AI applications on Mac with Jamf's AI Governance and Amazon Bedrock

🛠️ TRY THIS

Skip the boilerplate in today's 8-K, and every one after it

Today's Quick Links carries six of these: Nocera, AXT, LM Funding, Netcapital, Inuvo, Two Hands. Most 8-Ks are three pages of disclosure filed to satisfy a checkbox, one paragraph of substance. You don't need to read all three pages to find the one that matters.

1. Feed the raw EDGAR link straight to your coding agent or Claude. Skip writing your own summary first, let it work from the primary source.

2. Ask for triggering items only (1.01, 2.02, 5.02, whichever box is checked), dollar figures, dates, named parties. Nothing else.

3. Have it open with BOILERPLATE or MATERIAL on the first line. Stop reading the moment it says boilerplate.

4. Running this daily across a watchlist? Install Caveman (below) so terse is the agent's default output, not a prompt you retype every morning.

Prompt: Read this 8-K: [URL]. First line: BOILERPLATE or MATERIAL. If material, list only: triggering item number, dollar amounts, dates, named parties. No summary, no context, no filler.

Worth a look

• Caveman Claude Code skill that rewrites agent output in terse fragments instead of full explanations. Claims sizable token savings across a batch of tasks with no drop in technical accuracy, the kind of number worth running against your own workload before you trust it. No telemetry, runs local. github.com/JuliusBrussee/caveman
• MemPalace Local-first memory for agents. Stores conversations verbatim, retrieves by semantic search scoped to a project instead of one flat log. Claims strong recall on LongMemEval with zero API calls, which is the kind of number worth checking yourself before you trust it. github.com/MemPalace/mempalace
• Hermes Desktop A community-built GUI for Nous Research's Hermes Agent. Chat, scheduled tasks, persona and skill management, wired into 16 messaging platforms. Not Nous's own build, so treat it as third-party until they say otherwise. github.com/fathah/hermes-desktop

🔗 QUICK LINKS

• Nocera signed a binding deal for an equity stake in INERGX, an AI-datacenter power and battery-storage platform, its latest step from solar panels toward a full AI-infrastructure holding company. 8-K
• AXT locked in a three-year, $22.3M prepaid supply deal with Coherent for indium phosphide wafers that feed AI-datacenter optics. Stock jumped nearly 20% on the news. 8-K
• LM Funding America mined 8.7 BTC in June, down on summer heat, and still calls itself an HPC and AI infrastructure play while the stock sits at $0.15. 8-K
• Netcapital filed an 8-K pitching Project Orion, a "private AI operating environment," plus talk of data center acquisitions. Long on vision, short on numbers. 8-K
• Inuvo put out a new investor deck built around IntentKey, its in-house LLM for ad targeting, leaning on a claimed 24-hour lead reading consumer intent. 8-K
• Two Hands Corp is delisting from the Canadian Securities Exchange to cut costs, freeing up cash for its quantum computing and AI push. 8-K

That's the one worth watching today.

Pradeep Perugu