Get tomorrow's brief in your inbox
Today: OpenAI shipped GPT-Live-1, a full-duplex voice model that can listen and speak simultaneously while delegating complex queries to GPT-5.5. Jarred Sumner rewrote Bun from Zig to Rust using coding agents in 11 days for $165,000, showing frontier models can now tackle rewrites everyone said were impossible. Zapier and Prime Intellect are helping businesses build AI agents without depending on OpenAI or Anthropic.
OpenAI launched GPT-Live-1 and GPT-Live-1 mini, new voice models that fundamentally change how we interact with AI assistants. Unlike the previous turn-based system (which waited for you to stop talking before responding), GPT-Live uses full-duplex architecture. It can listen and speak at the same time, interrupt naturally, and acknowledge you're speaking with phrases like "mhmm" or "yeah."
The bigger shift is what happens under the hood. When you ask a question requiring web search, deep reasoning, or complex work, GPT-Live delegates the task to GPT-5.5 (or whatever OpenAI's latest frontier model is) and keeps the conversation flowing while it works. You're not stuck waiting in silence for 30 seconds while it searches the web. The model can also stay quiet for long stretches, absorbing context until you need it.
OpenAI is replacing the current Advanced Voice Mode in ChatGPT with GPT-Live-1 mini for free users. Paid subscribers get access to the full GPT-Live-1 model. The company is also planning to bring these models to the API soon. If you've avoided ChatGPT Voice because the interruptions felt awkward or the long pauses killed the flow, this is worth revisiting. OpenAI is clearly positioning voice as the primary interface for complex agentic work, not just a novelty feature. (OpenAI Blog, TechCrunch, The Verge)
SpaceXAI released Grok 4.5, which Elon Musk described as "Opus-class" in capability but faster and cheaper. It costs $2 per million input tokens and $6 per million output tokens, compared to Opus 4.7's $5 and $25. Musk claims Grok 4.5 is roughly comparable to Opus 4.7 but with better speed and token efficiency. If the benchmarks hold up in real-world use, this could be the most cost-effective frontier model for tasks that don't need bleeding-edge reasoning. It's available to the public now. (TechCrunch)
Google Photos added Video Remix, an AI-powered editing tool that can transform videos in seconds. Powered by Gemini Omni, it lets you apply cinematic relighting, swap backgrounds, or add artistic styles like watercolor or oil painting. You could brighten a dark clip with a morning glow effect, change the background to a greenhouse, or paint the whole video in watercolor. It's rolling out to Google AI Plus, Pro, and Ultra subscribers in the US and 14 other countries. If you're creating social content or marketing videos and don't want to learn Premiere or DaVinci Resolve, this brings professional-looking edits to your phone. (TechCrunch)
Google's SynthID watermark caught a deepfake in a rare public win for anti-deepfake technology. A fake image of Senator Mitch McConnell in a hospital bed circulated on Reddit and X this week. Snopes debunked it by checking for SynthID's invisible watermark, which confirmed the image was AI-generated. SynthID works by embedding an invisible signature into images that survives screencapturing and cross-platform sharing. It only works when the image generator participates in the program. Gemini models have included it since 2025, and OpenAI joined in May 2026. You can check images by uploading them to OpenAI's public verification tool or asking a Gemini model. If you're dealing with misinformation risks or need to verify images, this is the most reliable detection system available today. (TechCrunch)
Jarred Sumner rewrote Bun from Zig to Rust in 11 days using coding agents, proving frontier models can now tackle projects everyone said were impossible. Bun is a JavaScript runtime competitor to Node.js, and Sumner had been worrying about memory bugs (use-after-free, double-free, forgot-to-free errors) that Zig's manual memory management couldn't prevent. Rust's compiler catches these as errors, so he decided to try a full rewrite.
The key enablers: Bun's test suite was written in TypeScript (language-independent), which meant agents could validate the port automatically. Sumner used an earlier version of what's now available as Mythos/Fable to automate the initial port. He spent most of the 11 days monitoring workflows, reading outputs, and prompting Claude to fix issues. The new Rust implementation went live in Claude Code v2.1.181 on June 17. Startup got 10% faster on Linux, and barely anyone noticed. The estimated cost at API pricing: $165,000 (5.9 billion uncached input tokens, 690 million output tokens, 72 billion cached input token reads).
The review strategy was adversarial code review combined with a million-assertion test suite. When something went wrong, he fixed the process that generated the code instead of hand-fixing the code. This is the blueprint for taking on wildly ambitious refactors with coordinated parallel agents. If you have a large codebase and a strong test suite, this approach is now viable. (Simon Willison)
n8n published an AI-powered incident response workflow for security operations centers (SOCs). When an incident payload arrives from a SIEM or ticketing tool, the workflow retrieves the closest matching playbook, similar resolved incidents from your historical record, and current threat intelligence from the web. A synthesis agent combines all three and produces a structured runbook with immediate actions, containment steps, IOCs, assumptions, and confidence levels. The goal is to capture and reuse what worked last time without making analysts read post-mortems. It's built to let teams dial AI up or down based on risk tolerance and keeps ticket data on-prem (not sent to third-party clouds). The workflow is open source on GitHub and runs on your own n8n environment. If you're in security or IT and dealing with repetitive incident handling, this is a practical starting point for automation. (n8n Blog)
Zapier vs. ChatGPT comparison clarifies when to use each (or both). ChatGPT is an AI agent harness. You open it, type requests, and get summaries, analysis, or generated text. With workspace agents (available on Business, Enterprise, Edu, and Teachers plans), it can work across your connected apps or run in Slack, handling multi-step tasks in the background.
Zapier is a governed automation layer with AI built in. You build workflows that start with a trigger (an event or a schedule) and follow with actions. You only use AI where you need it, not for every step. This saves tokens. For example, an email receipt workflow could filter for messages containing "receipt" or "invoice" (no AI), create an Asana task (no AI), and send it to your accounting tool (no AI). You'd only add AI if you wanted to extract line items or categorize the expense.
The best setup: use ChatGPT as the agent harness and Zapier as the AI automation layer that gives ChatGPT secure access and permissions across your apps. You can access Zapier directly in ChatGPT via Zapier MCP or the Zapier SDK (in open beta, not yet SOC 2 certified). If you're building automations that involve AI, this combination gives you the best of both worlds. (Zapier Blog)
Federal agencies awarded $273 billion to small businesses in FY25, with $179 billion in prime contracts (27.7% of total contracting, exceeding the 23% statutory goal). The SBA highlighted a shift toward merit-based contracting. Contracts to 8(a) firms (the socially and economically disadvantaged business program) dropped to 3.7%, the largest decline in over a decade, after the SBA audited the program and terminated nearly 800 non-compliant firms. Service-Disabled Veteran-Owned Small Businesses received $32.5 billion, exceeding the 5% federal target.
If you run a small business and haven't explored federal contracting, this is a good year to start. The focus on merit-based competition and the decline in 8(a) dominance opens doors for businesses that can demonstrate value and performance. Veteran-owned firms should actively pursue these contracts. The SBA scorecard provides detailed agency performance, which can help you identify where opportunities lie. (Small Business Trends)
Prime Intellect raised $130 million at a $1 billion valuation to help enterprises build their own AI agents without relying on OpenAI or Anthropic. The startup provides a full stack for AI agent development: compute access, a reinforcement learning framework, and evaluation tools. The platform functions like a marketplace, so customers can pick specific tools they need without being locked into an all-or-nothing system.
Prime Intellect has attracted customers like Ramp, Zapier, and Flapping Airplanes, who pay for a hosted version of its tools. The company hit $100 million in annualized revenue run rate. Ramp used Prime Intellect to build an agent that answers questions inside spreadsheets, beating frontier models on accuracy while running faster and at a fraction of the cost.
The pitch: companies are wary of providing proprietary data to OpenAI and Anthropic, and they don't want to depend on models that can be suddenly turned off (like Anthropic's Fable last month). Prime Intellect gives enterprises the infrastructure to own their enterprise intelligence and avoid those risks. If you're building on top of frontier labs and concerned about data control or model reliability, this is the alternative. (TechCrunch)
General Intuition raised $320 million at a $2.3 billion valuation on the thesis that gaming data is the key to building foundation models for physical AI. The company trained its model on millions of hours of video game data (including controller input data), which gives it a base level of reasoning about space and time. General Intuition has demonstrated that its model can both play a video game for hours and power a quadrupedal robot after fine-tuning on just eight minutes of real-world robotics data. The robot navigated an office with dynamic objects and people walking by using only a front camera and no other sensors.
General Intuition's goal is to become the foundation model of physical AI, not to build robots itself. CEO Pim de Witte argues the industry should stop collecting hundreds of thousands of hours of real-world data for specialized robot models and instead use general models that transfer intuition across many environments. If you're in robotics or building self-driving systems, this approach could make it 10 times easier to build your next product. (TechCrunch, TechCrunch)
Lovable is in talks to raise $300 million at a $13.2 billion valuation, exactly double the $6.6 billion valuation it achieved last December. Menlo Ventures is expected to lead the round. The Swedish vibe-coding startup hit $500 million in annualized revenue run rate in June. Lovable's users include founders, designers, and salespeople building websites and e-commerce storefronts. It also sells to large enterprises like Workday, Asana, and Nvidia. Vibe coding (building software by describing it) is the most popular and lucrative AI use case right now, and Lovable is one of the leaders. (TechCrunch)
Meta is reportedly standing up a cloud business, debating whether to offer access to AI models hosted on its infrastructure or to sell raw computing power. The bull case: by creating a new revenue stream, Meta hedges against overbuilding on AI infrastructure. The bear case: Meta overshot what it needs and is now trying to rent out excess capacity, which will generate lower margins (35% for cloud vs. 70% for advertising). JPMorgan estimates every gigawatt of Meta compute capacity offered in a cloud business could generate $20 billion of annual revenue. Meta stock is down 8% year-to-date, the second-worst performer among hyperscalers. If you're currently dependent on AWS, Google Cloud, or Azure and want another option, keep an eye on this. (CNBC)
OpenAI found that about 30% of SWE-Bench Pro tasks are broken, according to a detailed audit. The issues fell into four categories: overly strict tests enforcing implementation details not in the prompt, underspecified prompts omitting requirements that hidden tests enforce, low-coverage tests that let incomplete fixes pass, and misleading prompts pointing models toward wrong behavior. OpenAI created a quality assurance pipeline using investigator agents to flag likely broken tasks, followed by human review from five experienced software engineers. The findings point to the difficulty of curating hard but fair benchmarks and the growing utility of agents for scalable data quality checks. If you're using SWE-Bench Pro to evaluate models, carefully examine your results. (OpenAI Blog)
Kenton Varda declared a moratorium on AI-written change descriptions (PR and commit messages, issues, tickets) from his team. He found AI was writing descriptions that were worse than useless during code reviews: outlining details of the code visible by looking at the code, but omitting the higher-level framing needed to understand what the code is doing. If you're using AI to write commit messages, make sure they explain why the change was made, not just what changed. (Simon Willison)
Top 5 automation tools to free up your team's time: Zapier connects over 5,000 apps for multi-step workflows without coding. Airtable provides customizable automation with conditional logic for collaborative projects. Motion automates scheduling and task assignments based on priorities. Taskade generates content ideas and automates task reminders for team collaboration. Thunderbit features a drag-and-drop interface for building complex automations across CRM, social media, and email. Start by identifying tasks consuming too much time (data entry, report generation, scheduling) and implement automation to cut manual work by up to 10%. Employees report higher job satisfaction when they can focus on fulfilling tasks instead of repetitive work. (Small Business Trends)
If you need funding fast: Working capital loans provide $5,000 to $500,000 with potential same-day approvals and 3-18 month repayment terms. Business lines of credit allow access to up to $300,000, with quick approvals (sometimes within hours) and interest paid only on the amount used. Merchant cash advances offer rapid access to cash with repayments linked to daily credit card sales (though effective APRs can be high). Minimum credit score requirements are often as low as 500, making these options accessible even if traditional bank loans aren't available. (Small Business Trends)
Voice is emerging as the primary interface to AI agents. OpenAI's full-duplex architecture sets a new standard for natural conversation while complex work happens in the background. Frontier models can now tackle rewrites and refactors everyone said were impossible (Bun's 11-day Zig-to-Rust port proves it). Enterprises are moving away from dependence on OpenAI and Anthropic, building their own agents with infrastructure providers like Prime Intellect. Watch for GPT 5.6's release tomorrow and continued pressure on frontier model pricing as SpaceXAI undercuts with Grok 4.5.