OpenAI's New Model Talks and Listens Simultaneously

Why owning AI deployment is a career move

Across product, ops, and CX teams, a new role is emerging: the person making AI actually work day to day.

On July 16, three people living this shift, from Yoco, Zumba Fitness, and Fin, share what the job looks like, the skills they'd hire for, and the challenges ahead. Register to save your spot.

Save your spot

In today's issue

01 · Models

OpenAI ships GPT-Live-1, a full-duplex voice model that replaces Advanced Voice Mode

OpenAI released GPT-Live-1, a voice model that listens and speaks at the same time, replacing its older Advanced Voice Mode. The shift matters because it collapses three separate steps into one, cutting delays and letting conversations flow naturally—now rolling out to 150 million ChatGPT voice users.

The details:

GPT-Live-1 mini is the default for all users; paid tiers get the larger GPT-Live-1, which routes harder questions to GPT-5.5 mid-conversation without breaking the chat.
The model can be interrupted mid-sentence, handle silences while you think, and show charts or images alongside spoken responses.
Product lead Atty Eleti said he held 30- to 40-minute conversations with the model on walks, signaling OpenAI is positioning voice as a platform for long-running tasks, not just quick commands.

Why it matters: Full-duplex voice—where both sides speak simultaneously—removes the robotic turn-taking that made earlier assistants feel stiff. This is the software foundation for always-on AI hardware like earbuds or wearables, and OpenAI's 150-million-user head start means the first truly natural voice assistant likely lands here first.

Read the full story →

02 · Security

China flags Anthropic's Claude Code as a security back-door risk

China's cybersecurity authority flagged Anthropic's Claude Code as containing a back-door vulnerability that can leak user location and identity data, telling domestic users to uninstall or upgrade versions released between April and June. The advisory is notable because Claude Code isn't officially sold in China, and it arrives just as Alibaba bans Anthropic tools company-wide—one week after Anthropic accused the company of stealing its AI technology.

The details:

The flagged Claude Code versions span 2.1.91 to 2.1.196; Anthropic's current build is 2.1.204, already three releases past the ministry's warning.
Xiaomi AI developers publicly acknowledged using Claude Code at a state-organized forum in March, despite the tool's lack of official availability in China.
Most modern coding assistants, including OpenAI's and Google's, send telemetry and crash reports to their makers—the ministry did not clarify what distinguishes Claude Code or provide technical proof.

Why it matters: The timing—Anthropic accuses Alibaba of stealing tech, Alibaba bans Anthropic products, then the government flags Claude Code nationally—suggests a coordinated escalation rather than a genuine security discovery. For Anthropic's pitch of Claude Code as an enterprise coding tool globally, a foreign government security advisory, credible or not, now becomes a sales headwind.

Read the full story →

03 · Models

Nvidia Nemotron 3 Ultra hits closed-model parity at 10x lower cost on LangChain

Nvidia's open-source Nemotron 3 Ultra model now matches the best closed AI systems on business tasks—at one-tenth the cost. LangChain tuned the agent software around the model without retraining it, proving that the right engineering environment can compete with expensive proprietary AI on real enterprise work.

The details:

LangChain adjusted system prompts, tool descriptions, and middleware to squeeze top-tier performance from Nemotron 3 Ultra on its Deep Agents benchmark.
Abridge, Amdocs, Box, and EY are already building specialized agents on this stack for their platforms and workflows.
The tuned model is live on six hosting providers: Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius, and Together AI.

Why it matters: Ten times lower cost per run unlocks continuous iteration and evaluation—workloads that were too expensive to automate before now make financial sense. Companies get a full-stack choice: run Nemotron on their own servers under their own governance, instead of renting agent capability from a closed vendor.

Read the full story →

AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

● ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat

● Generate images & video with Sora, Veo, Ideogram

● Compare any two models side by side

● From $8.99/mo · 80+ models, all included

Try AI Box →

aibox.ai

Trusted by 3,000+ teams

04 · Security

Meta adds tamper-proof LED to AI glasses while widening data collection elsewhere

Meta added a tamper-proof LED to its AI glasses that disables the camera if the light is modified or destroyed, stopping users from recording others without consent. The same week, Meta expanded data collection elsewhere — allowing its AI to train on public Instagram photos by default and testing glasses that continuously collect audio and snap photos every few seconds.

The details:

Meta's glasses now detect sophisticated hardware tampering attempts; the feature was triggered after users began covering the LED with tape to record people covertly.
Meta AI can train on any public Instagram photo unless users opt out — a change that applies to billions of photos users never intended for AI training.
A prototype under testing would shift glasses from on-demand camera to always-on sensor, feeding continuous audio and periodic photos to Meta's AI systems with no launch date announced.

Why it matters: The LED safeguard shows Meta responding to a real privacy threat, but the broader pattern reveals the company pulling more ambient data from more surfaces at the same time. Users get one protective measure on the device while Meta's AI ingests more data from their existing photos and potentially future always-on sensors — a mismatch between what the hardware prevents and what the software enables.

Read the full story →

05 · Business

OpenAI chief futurist Joshua Achiam departs after nearly nine years

Joshua Achiam, OpenAI's chief futurist and a safety advocate since the company's early days, is leaving later this month after nearly nine years. His departure marks the fifth senior safety-focused leader to exit OpenAI in two years as the company prepares to go public.

The details:

Achiam joined as an intern in 2017, led the mission alignment team until it was disbanded in February 2026, then moved into the chief futurist role studying AI harms and benefits.
Five safety-adjacent leaders have now left OpenAI since 2024: Jan Leike and Andrea Vallone both joined Anthropic; Miles Brundage and Steven Adler started nonprofits focused on AI safety standards.
Former White House AI adviser Dean Ball started this week as head of strategic futures and will briefly overlap with Achiam, taking on the policy-plus-research work without a direct replacement for the chief futurist title.

Why it matters: OpenAI has steadily shifted from a research-focused lab toward a commercial operation, and the pattern of departures suggests the company's internal safety mission is being outsourced—to Anthropic, to nonprofits, or to outside roles. That structural change matters more than any individual exit.

Read the full story →