In partnership with

Good morning. It’s Monday, March 24th.

On this day in Tech History: 1984: Apple introduced the Macintosh, a groundbreaking personal computer featuring a graphical user interface, which set new standards for user-friendly computing.

In today’s email:

• OpenAI At Strategic Crossroads

• Grok’s Deep(er) Search and Image Editing

• Claude’s Sharpened Complex Reasoning

• 3 New AI Tools

• Latest AI Research Papers

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

​​OpenAI Faces Intensifying Competition and Strategic Crossroads

OpenAI is navigating mounting challenges as it seeks to expand in India and defend its subscription model. The company is in discussions with Reliance Industries to distribute ChatGPT via Jio and offer its AI models to Indian enterprises through an API—potentially requiring Microsoft’s involvement due to its resale rights.

To drive adoption, OpenAI is also considering reducing ChatGPT’s subscription fee in India by 75-85%, lowering the cost from $20 to as little as $3-$5 per month. Reliance, meanwhile, has proposed hosting OpenAI and Meta models at its planned three-gigawatt Jamnagar data centre, aligning with India’s data sovereignty push.

At the same time, competition from China is escalating. Kai-Fu Lee’s 01.AI has fully pivoted to Deepseek’s open-source models, abandoning proprietary development to focus on enterprise customization. Lee argues Deepseek’s free availability is undercutting OpenAI’s paid model, leading some users to cancel ChatGPT subscriptions. He estimates Deepseek’s operating costs at just 2% of OpenAI’s $7 billion expenditure in 2024, raising concerns about the sustainability of OpenAI’s high-cost approach.

While OpenAI and Anthropic have urged the U.S. government to regulate Deepseek, calling it state-controlled, Lee dismisses these claims as paranoia, suggesting that CEO Sam Altman is “probably not sleeping well” as Deepseek’s momentum accelerates. Read more.

Grok 3 Expands with Deeper AI Search and Text-Based Image Editing

xAI is upgrading its Grok chatbot with two new features: DeeperSearch, an advanced research tool, and text-based image editing. DeeperSearch refines queries by taking longer to process fewer, higher-quality sources, contrasting with the existing DeepSearch, which retrieves results more quickly but includes lower-credibility sites. Initial tests suggest DeeperSearch prioritizes mainstream sources like The Guardian and BBC but may overlook key developments.

xAI is also quietly introducing image editing capabilities, allowing users to modify uploaded images with simple text instructions—similar to Google’s Gemini models. The feature is rolling out gradually, with an “Edit Image” button appearing in both X and Grok’s iOS app. Additionally, xAI has announced that its Aurora image model, known for photorealistic rendering and fewer content restrictions, is now accessible via API. Read more.

AI That Knows When to Hit Pause: Claude 3.7 Sharpens Complex Reasoning

Anthropic just gave Claude 3.7 a cognitive edge with the "think" tool—an embedded pause button for AI reasoning. Unlike extended thinking, which front-loads analysis before response generation, the "think" tool lets Claude stop mid-process, reassess, and refine its approach in real-time. This upgrade is a game-changer for high-stakes, tool-heavy workflows where missing a step isn’t an option.

τ-Bench testing proves it’s not just theoretical. In the airline domain—where rigid policies and layered decisions make AI stumble—the tool with optimized prompting pushed Claude’s pass¹ metric from 0.370 to 0.570, a 54% leap. In retail, where rules are looser, it still hit 0.812 with no extra guidance. This means Claude now knows when to slow down and think. Anthropic advises developers to implement domain-specific prompts for maximum effectiveness. Read more.

• Tencent launches T1 reasoning model amid growing AI competition in China

• Microsoft’s KBLaM is a new way to feed knowledge into LLMs without the retrieval bottleneck

• Google’s Bolt3D Turns Photos into 3D Worlds in Just 6 Seconds

• Apple explores adding AI cameras to Apple Watch and AirPods

• AI users can match two-person team performance, study suggests

• Cambridge researchers unveil faster and more accurate AI weather system that rivals supercomputers

• Texas private school’s use of new ‘AI tutor’ rockets student test scores to top 2% in the country

• Meta spotted testing AI-generated comments on Instagram

• Perplexity confirms proposal to acquire U.S. ops of TikTok with detailed plans

• New bipartisan bill threatens to sunset key internet platform protection law

📄 One-Step Residual Shifting Diffusion for Image Super-Resolution via Distillation

📄 JARVIS-VLA: Post-Training Large-Scale Vision Language Models to Play Visual Games with Keyboards and Mouse ​​

📄 MARS: A Multi-Agent Framework Incorporating Socratic Guidance for Automated Prompt Optimization

📄 Reinforcement Learning for Reasoning in Small LLMs: What Works and What Doesn't 

📄 RoboFactory: Exploring Embodied Agent Collaboration with Compositional Constraints

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!