Tuesday, September 15, 2026
The Brief
Leaked documents from an initiative called Project Lily reveal that human contractors are reading ChatGPT user prompts to improve OpenAI's model performance. Reviewers frequently encounter personal information and proprietary data during this process. If your team pastes sensitive code or strategy documents into a cloud-based LLM, you just received confirmation that those queries can end up on a stranger's screen.
- New AEF-1 standard. OpenAI, Anthropic, and xAI co-signed a framework for independent testing, giving enterprise buyers a standardized way to compare model reliability. Source
- Terminal agent benchmarks. We previously covered how GPT-6 Astra and Claude Fable handled custom productivity tools; the evaluation now ranks their autonomous coding performance in terminal environments. Source
- Socher launches Recursive. Natural language processing pioneer Richard Socher spun out a recursive self-improvement startup with a $5 billion valuation, introducing a heavily funded competitor to model labs. Source
- Local agents for Windows. Perplexity launched a feature allowing users to run multi-step AI agents locally on RTX GPUs, providing a privacy-focused alternative to cloud processing. Source
Do this: Verify your organization's data retention settings in ChatGPT today to ensure proprietary prompts are excluded from human review.
Hands-on
1 pick- AI Tool
How designers use AI agents for Figma and rapid prototyping
SpaceXAI designers demonstrate workflows for controlling design tools and building personal sites using AI agents. This walkthrough provides practical examples of how creative professionals can automate design tasks and bridge the gap between ideas and functional prototypes.
Lenny's NewsletterRead
The AI race
5 itemsNvidia CEO opposes AI slowdown in briefing with Trump administration
Jensen Huang has signaled strong opposition to calls from other tech leaders to pause or slow AI scaling, citing economic and competitive necessity. This highlight's the rift between hardware providers and model labs regarding the speed of industry progress.
TechCrunch AIRead the full articleOpenAI reportedly acquires camera startup Glass Imaging for $300 million
OpenAI has acquired a startup founded by former Apple engineers to integrate advanced image processing into its ecosystem. This move signals a significant push by OpenAI into hardware-level vision and mobile AI capabilities.
TechCrunch AIRead the full articleLeaked documents reveal human contractors are reviewing sensitive ChatGPT user prompts
An investigation into OpenAI's 'Project Lily' shows that human reviewers are reading user chats to improve model performance, often encountering personal information. Knowledge workers should be extremely cautious about inputting proprietary or sensitive data into LLMs.
404 MediaRead the full articleMajor AI labs back AEF-1 standard for third party model evaluations
OpenAI, Anthropic, and xAI have co-signed the AEF-1 standard to establish a framework for independent safety and performance testing. This move signals a shift toward standardized benchmarking, which will help knowledge workers trust and compare the reliability of competing enterprise models.
Latent SpaceRead the full articleRichard Socher launches Recursive to pursue recursive self-improvement
NLP pioneer Richard Socher has spun out a new startup, Recursive, focusing on Recursive Self-Improvement (RSI) with a reported $5 billion valuation. This signifies a major new player in the pursuit of AGI, backed by one of the field's most respected researchers.
Latent SpaceRead the full article
AI at work
5 itemsPerplexity launches local AI agent for Windows PCs powered by NVIDIA
Perplexity's new 'Portable Computer' allows users to run multi-step AI agents locally on Windows devices using RTX GPUs. This is a significant move for privacy-conscious professionals who want to automate tasks without sending sensitive data to the cloud.
NVIDIA BlogRead the full articleSoftware development shifts toward product engineering as coding costs collapse
Industry expert Laurie Voss argues that as AI makes writing and maintaining code nearly free, the primary value of software development shifts to product design and definition. Knowledge workers should focus on understanding user needs and precise requirements, as these human-centric tasks will soon constitute the majority of the 'work' in software creation.
Simon WillisonRead the full articleUsing ChatGPT to translate sensor data into wildfire alerts
SensoRy AI utilizes ChatGPT to convert complex environmental sensor data into actionable human language for emergency responders. This case study demonstrates how LLMs can bridge the gap between technical telemetry and critical real-world decision-making.
OpenAI (YouTube)Read the full articleOpenAI highlights how Fyxer uses specialized models for executive task automation
Fyxer leverages OpenAI models and 500,000 hours of workflow data to build an AI assistant that manages emails and meetings. The system breaks complex tasks into dozens of specialized models, resulting in 53% of drafted emails being accepted without any edits.
OpenAI (YouTube)Read the full articleApple releases iOS 27 featuring a significantly overhauled Siri assistant
The latest iOS update introduces a long-awaited Siri overhaul that improves its daily utility and conversational intelligence. Knowledge workers can benefit from a more reliable mobile assistant for managing tasks, schedules, and cross-app workflows.
TechCrunch AIRead the full article
For builders
3 itemsNew benchmarks rank agentic performance for GPT-6 Astra and Claude Fable
This evaluation goes beyond generic leaderboards to rank models like GPT-6 Astra and Claude Fable based on their actual performance in terminal-based environments. It provides crucial data for developers deciding which model to integrate into autonomous coding agents and complex engineering workflows.
IndyDevDanRead the full articleTrueForge launches as an open source platform for managed AI agents
TrueForge provides an open source alternative for deploying and managing AI agents at scale. This allows knowledge workers and teams to orchestrate complex workflows without being locked into proprietary managed service ecosystems.
Sam WitteveenRead the full articleGoogle announces DevFest 2026 focused on scaling agentic AI
Google’s global community-led developer event returns with a focus on building and securing AI agents. Knowledge workers in technical roles can explore over 800 events to stay updated on the latest Google AI developer tools.
Google The Keyword · AIRead the full article
Get tomorrow's brief, curated for you.
Subscribe →