Meta faces dual celebrity AI crises: unauthorized bots impersonating Swift and others while licensed celebrity voices engaged inappropriately with minors. Both expose how engagement incentives override safety guardrails.
Despite massive AI hype, 95% of enterprise projects deliver no real returns. The gap between promises and reality reveals hidden costs, workflow mismatches, and why human oversight remains surprisingly essential.
Meta's $14B AI talent blitz hits turbulence as ChatGPT co-creator Shengjia Zhao threatened to quit days after joining. The company hastily named him Chief Scientist to prevent defection, but at least three other marquee hires have already left.
OpenAI's GPT-4o is revolutionizing AI by integrating text and images, paving the road for advanced visual understanding.
OpenAI just made its language model a lot more visual. GPT-4o can now generate images with uncanny precision, particularly when it comes to text rendering and photorealistic details. But this isn't just another pretty picture generator – it's a practical tool that understands context and handles complex visual instructions.
The model excels at creating what OpenAI calls "workhorse imagery" – the kinds of visuals that actually help people get work done. Think technical diagrams, presentation graphics, and mockups with accurate text placement. It can manage up to 20 distinct objects in a single image, far beyond the 5-8 object limit of current systems.
What sets GPT-4o apart is its deep integration with language. The system maintains visual consistency across multiple generations, letting users refine images through natural conversation. Upload a reference image, and GPT-4o analyzes it to inform new generations. It's like having a design assistant who never forgets what you showed them.
Deep Learning Meets Design
The technology stems from training on the joint distribution of online images and text. This approach taught the model not just how images relate to words, but how they connect to each other. Add some aggressive post-training optimization, and you get a system with surprising visual fluency.
Safety First, Generate Later
Safety features include C2PA metadata marking images as AI-generated and an internal search tool to verify if content came from the model. OpenAI has also trained a reasoning language model to interpret safety policies, helping moderate both input text and output images.
Working Through the Wrinkles
Some limitations persist. The model occasionally crops longer images too tightly, especially at the bottom. Generation times can stretch up to a minute – the price of creating more detailed images.
Rolling Out the Welcome Mat
Access rolls out today to ChatGPT Plus, Pro, Team, and Free users, with Enterprise and Edu coming soon. Developers will get API access in the coming weeks. Users can still access DALL·E through a dedicated GPT if they prefer the older system.
Creating images is straightforward: just describe what you need, including specifics like aspect ratios, hex color codes, or transparent backgrounds. The system handles these technical details while maintaining the natural flow of conversation.
The implications stretch beyond just making pretty pictures. This technology could revolutionize fields like technical documentation, where precise diagrams with accurate labels are crucial. Designers can iterate more naturally, and content creators can generate consistent visual assets more efficiently.
Why this matters:
Read on, my dear:
Marcus Schuler
Marcus Schuler
Marcus Schuler
AI startup Perplexity bids $34.5B for Google's Chrome—nearly double its own valuation—as federal judge prepares antitrust ruling. The timing isn't coincidental. Browser control becomes the new battleground in AI search competition.
While AI music rivals face billion-dollar lawsuits for copyright infringement, ElevenLabs took a different approach: they asked permission first. Their new music generator launched with licensing deals already signed.
Chinese startup Z.ai's new AI model costs 87% less than DeepSeek while running on half the chips. Built despite US trade restrictions, GLM-4.5 uses 'agentic' approach that breaks tasks into steps—potentially reshaping how AI works.
Chinese startup Moonshot AI released Kimi K2, an open-source model that matches GPT-4.1 performance while costing five times less. Silicon Valley's response? OpenAI delayed their planned open-source release hours after K2 launched.
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
