Good morning. It’s Friday, April 4th.

On this day in tech history: In 1975, Microsoft was founded by Bill Gates and Paul Allen in Albuquerque, New Mexico.

In today’s email:

• Deepmind’s AGI Ambitions

• Midjourney V7

• OpenAI’s Research Replication And Image Craze

• 3 New AI Tools

• Latest AI Research Papers

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

Google DeepMind Charts AGI Ambitions While Shifting AI Strategy and Infrastructure

Google DeepMind predicts that AGI may outthink humans by 2030, detailed in a dense 145-page report co-authored by Shane Legg. The report examines risks including misuse, misalignment, and structural fragility, and proposes countermeasures such as MONA for interpretable decisions, AI self-assessment tools, and layers of human-led oversight.

This approach sharply contrasts with OpenAI’s automation-heavy strategy, amid critics who argue that AGI remains an elusive concept with more immediate dangers tied to AI’s propensity to reinforce false outputs.

On the infrastructure front, The Information reports that Google is finalizing a deal with CoreWeave to rent Nvidia Blackwell GPU-powered servers and negotiating to deploy its own tensor processing units within CoreWeave’s facilities, though neither side has commented on the report.

Meanwhile, DeepMind’s advanced AI agent Dreamer, mastered Minecraft’s diamond collection challenge without human intervention. Utilizing reinforcement learning and a predictive "world model," Dreamer not only cuts computational costs but also demonstrates improved generalization across domains.

In tandem with these developments, Google is shaking up its AI leadership, with Josh Woodward taking charge of Gemini while retaining his role at Google Labs. Read more.

Midjourney V7 Ups Image Quality, Introduces New Architecture, and Slashes Render Time

Midjourney has launched V7, its first new image generation model in nearly a year, featuring a redesigned architecture and enhanced prompt understanding. Available in alpha, V7 introduces default personalization: users must first rate ~200 images to train the system to their visual preferences.

The model offers two modes—Turbo, which is more resource-intensive, and Relax. A new Draft Mode renders images ten times faster at half the cost, producing lower-quality outputs that can be enhanced with a click. Image quality has reportedly improved, with sharper textures and more coherent rendering of complex features like hands and objects.

CEO David Holz noted that V7 may require different prompting styles. V7 is accessible via Midjourney’s web app and Discord. Read more.

OpenAI Elevates Research Replication While Wrestling With Rising Costs and Image Gen Fever

OpenAI is flexing its muscles on several fronts as it fine-tunes both its research and business playbooks. The firm rolled out PaperBench, a rigorous benchmark aimed at testing AI’s mettle in replicating advanced research. Built on 20 ICML 2024 papers with 8,316 tasks and graded subtasks, submissions are scored by an AI judge honed on OpenAI’s best models. Anthropic’s Claude 3.5 Sonnet took the lead by replicating 21% of results—though human PhDs still sit comfortably at 41.4%. The benchmark’s code is out on GitHub.

The company is also in the middle of a bold transformation from nonprofit to for-profit, assembling a commission to design the “world’s best-equipped nonprofit” with a board proposal due in 90 days to stave off investor clawbacks.

At the same time, OpenAI's o3 model has steeper-than-expected operational costs. Initially projected at $3,000 per ARC-AGI task using its top configuration, o3 high, the Arc Prize Foundation now estimates costs at around $30,000 per task. Although official pricing hasn't been released, o1-pro pricing is expected to be similar. Despite these high costs, enterprise customers could be charged up to $20,000 monthly for specialized AI agents. Critics note that while these models may be less expensive than human contractors, their efficiency remains in question—o3 high requires 1,024 attempts to achieve optimal performance.

Amid all this, ChatGPT’s image generator has gone viral, logging over 700 million creations in its first week, driven by 130 million users—India leading the charge. An enhanced version and a standalone API are on the horizon, signaling more disruptive innovation to come. Read more.

• Watch: China Approves World-First Autonomous Flying Taxis for Commercial Use at 81 MPH

• NotebookLM Can Now Auto-Find Sources for You

• ​​Devin 2.0 drops from $500 to $20, goes heavy on autonomy

• Augment Agent Crushes GitHub Copilot with 70% Win Rate

• Vana is letting users own a piece of the AI models trained on their data

• Perplexity.ai Revolutionizes MoE with Open-Source GPU Optimization

• OpenAI and Anthropic are fighting over college students with free AI

• Smaller AI Models Prompt Microsoft to Rethink $80B Expansion

• Spotify debuts Gen AI ads, programmatic ad buying

• Runway, best known for its video-generating AI models, raises $308M

• U.S. Copyright Office issues highly anticipated report on copyrightability of AI-generated works

• Amazon's venture arm the Alexa Fund is dialing in on AI startups because the technology 'is only going to get more relevant'

• Creating and sharing deceptive AI-generated media is now a crime in New Jersey

• Claude Models Now Ready for Federal and Defense Use via Google Cloud

• Dartmouth AI Therapy Chatbot Matches Human Therapists in Reducing Depression and Anxiety

• Pennsylvania just demolished its largest coal plant to make way for a $10 billion AI data center

• Apple leads a drop in tech stocks after Trump tariff announcement

• Lindy AI Launches Agent Swarms, Achieves #1 Ranking With 5,000+ Integrations

• Adobe launches Premiere Pro’s generative AI video extender

📄 MergeVQ: A Unified Framework for Visual Generation and Representation with Disentangled Token Merging and Quantization

📄 Improved Visual-Spatial Reasoning via R1-Zero-Like Training

📄 Open-Qwen2VL: Compute-Efficient Pre-Training of Fully-Open Multimodal LLMs on Academic Resources

📄 Envisioning Beyond the Pixels: Benchmarking Reasoning-Informed Visual Editing

📄 GPT-ImgEval: A Comprehensive Benchmark for Diagnosing GPT4o in Image Generation

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!