In partnership with
In Today’s Issue:
🎨 OpenAI launches ChatGPT Images 2.0
🚀 Project Prometheus seeks to bring "physical AI" into the real world
🧠 OpenAI rolls out an experimental Codex upgrade
🩺 A multi-agent AI from Google and MIT autonomously discovers that "late-night doomscrolling" is a predictor of depression
✨ And more AI goodness…
Dear Readers,
OpenAI just gave its image generator a brain, and the results are hard to dismiss. ChatGPT Images 2.0 is the company's first image model that actually reasons before it renders, planning layouts, maintaining character consistency across eight images in a single prompt, and finally producing text that doesn't look like it was written by a toddler, even in Japanese, Korean, and Chinese.
But that's only the opening act today. Jeff Bezos is quietly building a $38 billion physical AI empire called Project Prometheus, poaching talent from OpenAI and DeepMind to make machines that don't just think but act in factories and on launchpads. Meanwhile, OpenAI's Codex now watches your screen to remember what you were working on, Google wants Gemini to scan your entire photo library for "personalization," and a multi-agent AI system from Google Research and MIT just discovered that your late-night doomscrolling predicts depression, naming the biomarker entirely on its own. Convenience keeps inching closer to surveillance, and the line between creative tool and autonomous researcher is getting thinner by the day.
Scroll on, this one's packed.
All the best,
🤖 Bezos Bets Big on Physical AI
Jeff Bezos’s secretive startup, Project Prometheus, is reportedly raising $10 billion at a $38 billion valuation, signaling an ambitious push into “physical AI” systems designed to interact with real-world industries like manufacturing and aerospace rather than just digital tasks. Led alongside Vik Bajaj, the company is quietly building talent from rivals like OpenAI and Google DeepMind, hinting at serious competitive intent.
Still, entering a field dominated by established players - and newer challengers - means Prometheus must prove it can translate massive funding and Bezos’ influence into real industrial breakthroughs, not just hype.
🧠 Codex Gains Contextual Memory Upgrade
OpenAI is rolling out Chronicle, an experimental feature that enhances Codex’s memory by using recent screen context, allowing it to understand vague references like “this” or “that” without repeated explanations. The system builds on-device memories from screen captures, aiming to adapt to users’ workflows over time, though it raises practical considerations around storage, rate limits, and app access to files.
Initially limited to Pro users on macOS (excluding parts of Europe), the feature reflects a cautious expansion as OpenAI evaluates how persistent, context-aware AI can integrate into real work habits.
🔒 Google Photos Privacy Line Blurs
Google’s new Personal Intelligence update lets Gemini scan a user’s Google Photos library to generate more personalized AI images, using “actual images of you and your loved ones” as context. Forbes frames this as another sharp turn in the convenience-versus-privacy tradeoff, noting Google’s promise that private photo libraries do not directly train models, while still warning that opting in gives AI access to some of people’s most intimate memories and relationships.
Demis Hassabis argues that AGI will surpass the Industrial Revolution in impact, with the main bottlenecks being compute, energy, data efficiency, and aligning increasingly powerful systems with human goals.
OpenAI's Image Model Can Think
The Takeaway
👉 OpenAI released ChatGPT Images 2.0, its first image model with built-in reasoning capabilities, allowing the system to plan layout and structure before generating visuals, rather than producing one-shot outputs.
👉 The model can now generate up to eight coherent images per prompt with character and object consistency, enabling workflows like storyboarding, multi-format ad campaigns, and sequential comic panels in a single conversation.
👉 Text rendering, long the weakest point of AI image generators, has improved significantly, including accurate non-Latin script support for Japanese, Korean, Chinese, Hindi, and Bengali.
👉 The gpt-image-2 API is live for developers at competitive pricing, with partners like Canva, Figma, and Adobe already integrating the model into production workflows.
AI image generation just stopped being a toy. OpenAI launched ChatGPT Images 2.0, its most capable image model yet, and the upgrade is substantial. The system can now render accurate text in images, handle complex layouts like magazine covers and infographics, and produce up to 2K resolution outputs across flexible aspect ratios, from ultrawide banners to vertical mobile screens.
(Yes, thats an generated image, no screenshot 🤯)
The real headline however: this is OpenAI's first image model with reasoning capabilities. When thinking mode is activated, the system searches the web, plans image structure before generating, and produces up to eight coherent images in a single prompt, maintaining character and object consistency across frames. That turns storyboarding, manga creation, and multi-asset marketing campaigns from tedious multi-step workflows into single conversations.
(could you tell if this was real of AI generated?)
Multilingual text rendering has also improved dramatically, particularly for Japanese, Korean, Chinese, Hindi, and Bengali. Previous models mangled non-Latin scripts almost comically. Images 2.0 integrates language into the design itself, producing posters and comics where the text actually reads correctly.
The model is available to all ChatGPT users, with thinking features reserved for Plus, Pro, and Business tiers. The gpt-image-2 API is live for developers at $30 per 1,000 output tokens.
Why it matters: OpenAI is repositioning image generation from a creative novelty into a production-grade visual system. By adding reasoning, web search, and multi-image coherence, Images 2.0 competes directly with design workflows, not just other AI models.
Sources:
🔗 https://openai.com/index/introducing-chatgpt-images-2-0/#textmode
Your first HR system, implemented right
Rolling out your first HR tool? Get a step-by-step guide to avoid common mistakes, drive adoption, and build a scalable HR foundation.
OpenAI’s GPT-Image-2 has taken the top spot across all Image Arena leaderboards with unprecedented margins, marking a major leap in state-of-the-art image generation and editing performance.
Google's AI Finds Depression Biomarkers
Your smartwatch data holds more secrets than you think. Google Research, Google DeepMind, and MIT just unveiled CoDaS, a multi-agent AI system that autonomously discovers health biomarkers from wearable sensor data. The system analyzed data from 9,279 participants across three clinical cohorts and found that "late-night doomscrolling" is a statistically significant predictor of depression severity. The AI named that feature entirely on its own.
CoDaS runs the full biomarker discovery lifecycle: from raw sensor data to hypothesis generation, statistical testing, adversarial validation, and even manuscript drafting. What makes it remarkable however is its built-in skepticism. When the system found a feature predicting insulin resistance with near-perfect accuracy, it rejected its own discovery after identifying it as a meaningless circular calculation. The true signal was far more modest, but scientifically real.
In a blind expert evaluation, CoDaS-generated manuscripts achieved an 86% non-rejection rate. Every competing AI system was rejected by reviewers at rates between 85% and 100%. Domain experts estimated the equivalent human research effort at roughly 37 person-days. CoDaS completed it in under eight hours.
CoDaS demonstrates that multi-agent AI can compress months of biomarker research into hours while maintaining scientific rigor that domain experts rate as publication-worthy. This shifts the bottleneck in digital health from discovery speed to validation and clinical translation.
What Defines Great PMs in the Age of AI | Live with Marty Cagan
Join Silicon Valley Product Group founder Marty Cagan for a live session on what separates great PMs: how they make decisions, balance tradeoffs, and build real products in the AI era.
