OpenAI launches GPT-Live desktop app for agent orchestration

OpenAI launches GPT-Live desktop app for agent orchestration OpenAI launches GPT-Live desktop app for agent orchestration Jul 24, 2026 Good morning. It’s Friday, July 24th. I’ll be in NYC next weekend Aug 1-3. Hit me up for coffee if you’re in the city! -Jeff
AI Breakfast You read. We listen. Let us know what you think by replying to this email. OpenAI launches GPT-Live desktop app for agent orchestration OpenAI is executing an aggressive infrastructure grab this week. The company is dropping over $30 billion on "Project Camellia," a massive 3.2-gigawatt hyperscale facility in Georgia equipped with closed-loop cooling to anchor its compute needs through 2032. To cement its role in federal research, OpenAI also partnered with the Department of Energy’s Genesis Mission, hooking 2,000 national lab scientists and Los Alamos supercomputers directly into Codex and domain-specific GPT-Rosalind models for bioscience and superconductor discovery. OpenAI brought GPT-Live's full-duplex voice control to the ChatGPT desktop app on macOS and Windows, making hands-free workflows available across ChatGPT Work and Codex. Powered by GPT-Live, the new interface lets you have a continuous, natural back-and-forth conversation while background reasoning models simultaneously handle coding, manage emails, and run other tasks across your screen context and local files. For consumer health, they rolled out ChatGPT Health across the U.S., allowing users to link medical records, Apple Health, WeightWatchers, MyFitnessPal, and Function. Free users receive responses from GPT-5.5 Instant, while paid tiers run on GPT-5.6 Sol for higher-tier medical reasoning. OpenAI also launched Presence, an enterprise platform designed to move real-time voice and chat agents from pilot demos into controlled production. Deployed alongside Forward Deployed Engineers, Presence wraps core reasoning models in a sandboxed operational layer featuring permissions, policy simulations, API tool access, and human escalation rules. To handle drift, it monitors telemetry and uses a Codex-powered feedback loop to analyze edge cases and propose self-correcting updates. Related video: Building with ChatGPT Voice Sponsored post The Sourcing Tool That Works While You Sleep Every founder who sells a physical product eventually learns how much time sourcing can consume. Comparing suppliers, chasing quotes, negotiating terms, and following up on orders can swallow entire days, often when a small team is already stretched across product development, marketing, fulfillment, and customer service. Founders have tried plenty of workarounds. Spreadsheets can organize quotes, but they cannot chase down a supplier. Part-time sourcing support adds a cost many early-stage businesses cannot justify. Even the general AI chatbots that companies have adopted over the past few years tend to stop before the work is finished. They may draft a supplier message, but someone still has to send it, monitor the response, and follow up when the conversation stalls. The Accio Sourcing Toolkit, a plugin inside Accio Work, Alibaba.com's AI-powered business agent, is designed to carry out those steps directly. It can find suppliers, compare offers, send outreach, negotiate pricing, and support order placement without requiring a founder to manage every interaction manually. Consider a founder in Austin sourcing 500 units of a new product. They can enter the target price, quantity, specifications, and delivery requirements before signing off for the night. While the founder sleeps, the toolkit can continue conversations with suppliers in Asia, where the business day is already underway. That matters because time zones remain one of the most persistent challenges in global sourcing. A question sent at the end of the U.S. workday may not receive a response until the next morning. By then, a supplier may need clarification, creating another full-day delay. For a solo entrepreneur or small team, a negotiation that should take hours can stretch across a week. The toolkit can keep those conversations moving by asking for missing details, following up on unanswered questions, and maintaining multiple supplier discussions at once. By morning, responses can already be logged and compared, with quotes flagged when they appear high relative to prevailing market rates. Its negotiation capabilities go beyond sending a standard request for a lower price. A buyer can set a target cost, lead time, minimum order quantity, and other commercial requirements. The system can compare offers across suppliers, use the strongest available terms as a reference point, and request revised quotes. It can also negotiate around factors such as FOB and CIF pricing, sample costs, production schedules, and minimum order quantities, while escalating important decisions to the buyer for approval. For example, if one supplier offers a better unit price while another provides a shorter lead time or a lower minimum order, the toolkit can continue both conversations and surface the trade-offs. The founder can then decide whether speed, margin, or order flexibility matters most. The same applies when the scope changes. If a sample needs to move from 180 grams to 200 grams, the revised specification can be shared across the active supplier conversations at once. The founder can then review the resulting changes in price, lead time, or minimum order quantity rather than repeating the same message and rebuilding the comparison manually. For sellers running lean operations through Amazon, Shopify, or their own storefronts, the practical benefit is time. Fewer hours are lost to monitoring inboxes and maintaining sourcing trackers, leaving more capacity for product decisions, customer acquisition, and the other work that determines whether a young business grows. Businesses can access the Accio Sourcing Toolkit through Accio Work and begin a sourcing task right away by entering their product requirements, target terms, and delivery needs. xAI introduces 1,024 parallel agents and skeptic verification to Grok Build xAI just added Workflows to Grok Build, making it easy to run complex coding tasks across hundreds of background agents at once. Instead of handling everything in a single chat, Grok automatically writes an orchestration script, splits up the job for parallel sub-agents, and runs skeptic checks to verify findings before handing over a final report. Concurrently, xAI is expanding Grok Build into an app-publishing platform, enabling users to deploy, host, and publicly share apps using custom domains. Mobile credit top-ups are also in development. On the foundation model front, Musk expects xAI's 2-trillion-parameter Grok 4.6 to wrap up initial training ahead of an August launch. The model promises to beat Grok 4.5 and Moonshot's Kimi K3 while matching their speed and token efficiency. Looking further ahead, Musk also publicly backed claims of an AI extinction event for math and physics.

Tesla is also preparing to train its Optimus humanoid robot using camera-equipped worker backpacks at its Grünheide Gigafactory, capturing real-world tool manipulation and assembly data. This effort complements component development at its Reutlingen robotics hub. Musk highlighted that scaling Optimus remains Tesla's toughest manufacturing hurdle. Model & Product Releases

Black Forest Labs expands FLUX 3 into factory robotics with Audi test (Black Forest Labs)

Google's expanding access to Gemini Spark (The Verge)

ElevenLabs adds copyright filter to block artist reference tracks in AI music generator (ElevenLabs)

Sakana AI rolls out Fugu-Ultra v1.1 for agentic workloads while keeping prices fixed (X)

New Ling-3.0-flash model uses 1/64 expert activation to cut multi-agent compute costs (RuntimeWire)

New Ling-3.0-flash model uses 1/64 expert activation to cut multi-agent compute costs (RuntimeWire) Model Routing & Orchestration

Runway launches media router to send creative prompts to rival AI models (TechCrunch)

Cursor uses live developer feedback to train new enterprise model router (Cursor)

AMD-backed Lemonade 11.5 lets small AI models act as routers for larger workloads (Phoronix)

AMD-backed Lemonade 11.5 lets small AI models act as routers for larger workloads (Phoronix) AI Agents & Developer Tools

Amazon adopts MCP to streamline third-party Alexa+ agent tools (Amazon)

1 in 4 software developers now design APIs for AI agents instead of humans (Fortune)

The Numbers.com rebuilds infrastructure after agentic AI scraping overloads legacy site (Stephen Follows)

The Numbers.com rebuilds infrastructure after agentic AI scraping overloads legacy site (Stephen Follows) Security & Safety

Google adds selfie video verification to defeat deepfake account recovery scams (Google)

Flaw in OpenAI Workspace Agents allowed URL links to silently deploy rogue AI agents (The Decoder)

AISI researcher proposes semi-verified operating systems to contain rogue AI models (X)

AMD's Lisa Su defends open-source AI following Hugging Face security breach (Fortune)

AMD's Lisa Su defends open-source AI following Hugging Face security breach (Fortune) Chips, Energy & Infrastructure

AMD takes on Nvidia with its Helios AI rack-scale system (TechCrunch)

Nvidia is sending GPUs to the moon (TechCrunch)

Quantinuum and SoftBank lay framework for hybrid quantum AI data centers (Quantinuum)

Deutsche Bank report connects El Niño energy shocks to data center grid stability (Deutsche Bank Research)

Deutsche Bank report connects El Niño energy shocks to data center grid stability (Deutsche Bank Research) Business & Market

Google's Gemini nears billion-user milestone (TechCrunch)

Microsoft claims new in-house AI models undercut OpenAI costs by up to 89% (VentureBeat)

Microsoft claims new in-house AI models undercut OpenAI costs by up to 89% (VentureBeat) China AI Debate

Experts say exploiting Anthropic's Fable isn't how Kimi K3 got so good (TechCrunch)

Nvidia's Jensen Huang defends Chinese AI amid Kimi panic (Axios)

Arcee CTO says Chinese AI models offer reviewable code and useful research data (TechCrunch)

Arcee CTO says Chinese AI models offer reviewable code and useful research data (TechCrunch) Policy & Regulation

Delaware weighs legal personhood for autonomous AI companies in new regulatory sandbox (Bloomberg Law)

US House passes military ban on Chinese humanoid robots (SCMP)

Little Tech Association lobbies against federal restrictions on Chinese AI models (Politico)

Little Tech Association lobbies against federal restrictions on Chinese AI models (Politico) Research & Benchmarks

Columbia researcher uses GPT-5.6 workflow to crack six open Erdős math problems in five days (X)

Synthetic LLM benchmarks hide critical memory issues found in production (LeadDev)

Prompt training helped custom AI tool clear 1,848 extra cases per year in court study (The Decoder)

Prompt training helped custom AI tool clear 1,848 extra cases per year in court study (The Decoder) Labor & Economy

New labor data shows AI automates early-career tasks without targeting specific demographics (Fortune)

Amazon cuts jobs in its AGI group (Reuters)

AI will not trigger employment collapse, staffing company Adecco Group says (Reuters)

AI will not trigger employment collapse, staffing company Adecco Group says (Reuters) Media & Culture

Meta sets its upbeat AI optimism commercial to an apocalyptic David Bowie anthem (TechCrunch)

Award-winning author calls out publishing industry's AI disclosure double standard (Fortune)

Award-winning author calls out publishing industry's AI disclosure double standard (Fortune) Devices & Automotive

Samsung and Google target all-day wear with 9-hour battery life on Android XR glasses (9to5Google)

Ford feeds Apple Maps road data to its Latitude AI hands-free driving team (9to5Mac)

Ford feeds Apple Maps road data to its Latitude AI hands-free driving team (9to5Mac)
Teable 3.0 turns business data into custom apps and AI workflows while connecting systems and automating data migration. Humalike x Hermes gives Hermes agents social intelligence to analyze tone, remember participants, and speak in group chats. Manifest converts web pages into JSON action maps, allowing AI agents to navigate buttons, forms, and required inputs. PromptQL serves as a multiplayer AI agent that connects workspace data and builds shared team context through human corrections. runevr combines media review, project management, and AI agents into a single collaborative workspace for creative teams. Thank you for reading today’s edition. Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter. Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X! Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months. Keep Reading AI Breakfast Curated weekly analysis of the latest AI projects, products, and news

正在读取已有翻译

这里只读取已保存内容，不会调用 AI 或消耗 Token。

正在读取已有改写

这里只读取已保存内容，不会调用 AI 或消耗 Token。