Good morning. It’s Monday, January 27th.

Did you know: Apple introduced the first iPad on this day in 2010? They have since sold 670 million tablets.

In today’s email:

• “Proxy” Challenges OpenAI’s Operator - and it’s free today

• Deepseek Crashed the AI Markets Yesterday

• OpenAI’s Canvas Gets Major Upgrade

• 3 New AI Tools

• Latest AI Research Papers

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

Convergent AI Releases “Proxy” To Challenge OpenAI’s Operator Agent - And It’s Free Today

London-based startup Convergence AI has launched Proxy, an innovative AI assistant designed to transform how AI interfaces with an internet browser. Founded in April 2024 by machine learning engineers Marvin Purtorab and Andy Toulis, Proxy leverages Large Meta Learning Models (LMLMs), enabling it to learn, adapt, and remember over time.

Proxy is aimed at challenging OpenAI’s Operator, the AI Agent capable of using an internet browser on you behalf currently available for OpenAI’s $200/mo Pro users.

Proxy was announced overnight, and you can try it here for free today - going forward it will be $20/mo to use the service.

DeepSeek’s AI Efficiency Sinks Nasdaq Futures Over 200 Points as Tech Sector Reels

Chinese AI startup DeepSeek is upending the global tech hierarchy with its groundbreaking, cost-efficient AI models, triggering a seismic response from industry titans and financial markets. DeepSeek’s open-source large reasoning model, R1, rivals OpenAI’s top models in performance while operating at a fraction of the cost, challenging the prevailing strategy of relying on massive compute power. Marc Andreessen hailed R1 as a “profound gift to the world,” while Meta’s Yann LeCun emphasized its validation of open-source AI over proprietary models.

Meta CEO Mark Zuckerberg countered with ambitious plans for Llama 4, a 2GW data center, and a $60-65 billion investment to maintain U.S. AI leadership. Meta has also mobilized four specialized task forces to dissect DeepSeek’s methodologies, aiming to replicate its low-cost training techniques and operational efficiencies.

The market reaction was swift and severe. Nvidia’s shares dropped over 5%, while Nasdaq 100 futures tumbled nearly 2.5%. Japanese chipmakers Advantest and Disco fell 8.2% and 2.9%, respectively, as investor confidence wavered. Simultaneously, DeepSeek’s app has dethroned ChatGPT to claim the top spot on Apple’s App Store, further fueling anxiety over U.S. tech dominance.

The sell-off followed a week of bullish momentum driven by the Project Stargate announcement by OpenAI and Oracle. However, whispers from Scale AI CEO Alexandr Wang about DeepSeek’s access to 50,000 NVIDIA H100s—despite U.S. export controls—added to the unease. Microsoft CEO Satya Nadella, meanwhile, invoked the Jevons Paradox, suggesting that AI’s growing efficiency could paradoxically drive even higher demand. Read more.

OpenAI's ChatGPT gets a major Canvas upgrade with HTML and React code rendering

OpenAI has enhanced ChatGPT’s Canvas feature with significant upgrades, including HTML and React code rendering, making it more versatile for users. The new functionality allows users to preview code directly within the interface, eliminating the need to download and run it separately.

While this feature is available to all users, access to the advanced o1 model remains limited to Pro, Plus, and Team subscribers. Canvas, introduced in October 2023, also includes a built-in Python emulator for immediate code execution and visual outputs, streamlining AI-generated content workflows.

The update appears to be a response to Anthropic’s Claude.ai, which offers similar capabilities. Mac users benefit from full Canvas integration in the ChatGPT desktop app, with Enterprise and Education users set to gain access soon. Read more.

• China’s $137 Billion AI Push Escalates Tech Race with U.S.

• AI can now replicate itself — a milestone that has experts terrified

• Meta's Llama stack makes it easier to develop typical LLM use cases

• Retro Biosciences, backed by Sam Altman, is raising $1 billion to extend human lifespan

• Brazil bans Sam Altman’s World from paying Brazilians for their iris scans

• Frontier models fail hard at "Humanity's Last Exam" but experts question if it matters

• Humans Can Oversee 100+ Robots, DARPA Research Reveals

• OpenAI has upped its lobbying efforts nearly sevenfold

• Trump Signs Order Calling for AI Development ‘Free From Ideological Bias’

• Yale’s New AI Tool Identifies Risk of Future Heart Failure

• HeyGen Launches Motion Control Feature for Dynamic AI Video Avatars

• Chinese OpenAI o1 challenger Kimi k1.5 now available as free web version 

• Perplexity AI revises Tiktok merger proposal that could give the U.S. government a 50% stake

• Billionaire Larry Ellison says a vast AI-fueled surveillance system can ensure 'citizens will be on their best behavior'

📄 RealCritic: Towards Effectiveness-Driven Evaluation of Language Model Critiques

📄 Humanity's Last Exam ​​

📄 Chain-of-Retrieval Augmented Generation

📄 SRMT: Shared Memory for Multi-agent Lifelong Pathfinding  

📄 Improving Video Generation with Human Feedback

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!