In partnership with

Good morning. It’s Monday, September 1st.

On this day in tech history: In 2000, the release of OpenCV delivered a widely used, open-source library of computer-vision and AI algorithms (k-NN, SVMs, decision trees, etc.), laying groundwork for accessible machine learning tools in academia and industry.

In today’s email:

• ChatGPT’s “Juice 200”

• Microsoft To Distance Themselves from OpenAI?

• $10k/mo UBI Feasible With AI Growth?

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

ChatGPT experiments with a max-thinking mode called Juice 200 

OpenAI is experimenting with a new “Thinking Effort” selector in the ChatGPT web app, letting users adjust the AI’s cognitive intensity. Options range from Light thinking (5) to Max thinking (200), with intermediate tiers like Standard (18) and Extended (48). Max thinking, or “Juice 200,” is currently limited to Pro and Enterprise users due to the heavy compute it demands.

The update also includes other experiments: showing the selected model in the composer, a fully collapsed tool menu, and model visibility in the Plus menu. The company wants to give users smarter AI without forcing them to guess which model or power level to pick. Early tests suggest this could let ChatGPT scale from casual chat to heavy reasoning tasks while keeping the experience smooth. Read more.

With new in-house models, Microsoft lays the groundwork for independence from OpenAI

Microsoft is no longer content being just OpenAI’s biggest customer. Last week, it rolled out two homegrown AI models designed to show it can build world-class systems in-house while keeping costs under control. MAI-Voice-1 is a speech-generation model tuned for speed and efficiency: it cranks out a full minute of expressive, multi-speaker audio in under a second on a single GPU. It’s already powering Copilot features like Daily and Podcasts, a clear play for voice to become a mainstream interface.

The second model, MAI-1-preview, is a large language model trained on around 15,000 Nvidia H100 GPUs, far fewer than rivals like xAI’s Grok, but engineered to run inference on a single GPU. That balance of scale and efficiency is no accident; AI chief Mustafa Suleyman calls it “punching above its weight,” crediting data curation and training discipline over brute-force compute. The model is in public testing now, with a wider Copilot rollout coming. These systems give Microsoft a hedge against overreliance on OpenAI. Read more.

Ex-OpenAI researcher says $10K UBI payments 'feasible' with AI-growth

Miles Brundage, who once led AGI readiness at OpenAI, argues that current UBI pilots offering $500–$1,500 a month are relics of a pre-AI economy. With trillion-dollar data center buildouts and Nvidia’s Blackwell Ultra GPUs already reshaping global output, he says $10,000 monthly stipends could be both economically and politically viable in the near term. The bottleneck isn’t money, it’s whether governments can rewrite policy fast enough to manage a post-work transition without stagnation.

Nvidia CEO Jensen Huang sees the same forces driving not redistribution but reconfiguration: AI as the long-awaited solution to the “productivity paradox.” By automating drudge work, he argues, companies can finally execute on shelved ideas, making a four-day workweek not just possible but efficient. Pilot programs back him up - 24% productivity gains, burnout cut in half, turnover down.

Brundage and Huang sketch two diverging but linked paths: direct redistribution of AI’s economic surplus, or structural redesign of work itself. Read more.

• Lovable’s CEO isn’t too worried about the vibe-coding competition

• Cracks are forming in Meta’s partnership with Scale AI

• Cohere launches open weights AI model Aya 23 with support for nearly two dozen languages

• Forget data labeling: Tencent’s R-Zero shows how LLMs can train themselves

• Andrej Karpathy says reinforcement learning isn’t enough, LLMs need experience

• Nvidia says two mystery customers accounted for 39% of Q2 revenue

• Ethan Mollick on the Dawn of Mass Intelligence

• Accenture CEO weighs in on why so many AI projects have failed with 3 red flags to watch out for

• DeepConf can greatly reduce computational effort in language model reasoning tasks

• The White House apparently ordered federal workers to roll out Grok ‘ASAP’

• Alibaba develops a new AI chip for a wide range of inference tasks 

• Musk's xAI sues engineer for allegedly taking secrets to OpenAI

• From drones to robots, tourists flock to China to glimpse a ‘cyberpunk’ future

• Alibaba is firing back as the US-China AI war heats up

• A firewall for science: AI tool identifies 1,000 'questionable' journals

• New method enables AI models to forget private and copyrighted data

• AI spots ‘covert’ consciousness before doctors

• Xcode 26 beta 7 adds GPT-5 & Claude account support

• ChatGPT-powered dolls are becoming caregivers in South Korea

📄 TiKMiX: Take Data Influence into Dynamic Mixture for Language Model Pre-training

📄 AWorld: Orchestrating the Training Recipe for Agentic AI ​​

📄 EmbodiedOneVision: Interleaved Vision-Text-Action Pretraining for General Robot Control

📄 Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning

📄 USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning

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!