Welcome back!

Sam Altman just made a bold claim: OpenAI now knows how to build AGI — and the next step is “superintelligence”. With AI agents predicted to join the workforce by 2025, the future feels closer than ever. How will this reshape our world? Let’s unpack…

Today’s Summary:

• 🚀 OpenAI eyes superintelligence

• 🖥️ Nvidia’s desktop AI supercomputer

• 💡 Microsoft’s Phi-4 model goes open-source

• 💰 OpenAI hints at usage-based pricing

• 📚 Google AI agent whitepaper

• ✏️ NeuralSVG explores text-to-vector AI

• 🛠️ 2 new tools

TOP STORY

OpenAI sets sights beyond human-level AI

The Summary: OpenAI's upcoming o3 model scored 87.5% on the ARC-AGI test, beyond average human performance of 80%. CEO Sam Altman announced that the company now understands how to build AGI and is moving toward developing “superintelligence”. Altman predicts AI agents will begin performing complex tasks in workplaces by 2025, while emphasizing the need to balance rapid progress with safety.

Key details:

• The ARC-AGI test evaluates reasoning ability through advanced pattern recognition and problem-solving

• Altman envisions AI agents autonomously handling complex tasks in workplaces by late 2025

• OpenAI’s user base has grown from 100M to 300M in the past year

• Superintelligence aims to unlock unprecedented scientific breakthroughs and economic growth

• Joshua Achiam, OpenAI's head of mission alignment, cautions that AI will change many equilibria, from politics to human emotional dependencies

Why it matters: OpenAI’s achievements and ambitions are pushing AI into uncharted territory, raising questions about governance, energy systems, and workforce dynamics. With superintelligence on the horizon, balancing innovation with safety will be essential to managing AI’s societal impacts.

NVIDIA

Nvidia announces desktop-size AI supercomputer

The Summary: Nvidia's Project DIGITS brings AI supercomputing to personal desktops, capable of running 200B-parameter models while plugging into a standard outlet. This Mac Mini-sized device, launching in May 2025 at $3,000, packs the new GB10 Grace Blackwell Superchip, delivering 1 petaflop of AI performance. The system democratizes AI development by allowing researchers to prototype locally before scaling to cloud infrastructure.

Key details:

• Priced at $3,000, comparable to a high-end gaming PC but with supercomputer capabilities

• GB10 Superchip delivers 1 petaflop of AI performance, operating on standard wall power

• Two linked DIGITS units can handle models up to 405B parameters, matching Meta's largest Llama 3.1 model

• Includes 128GB unified memory (8x typical high-end laptops) and 4TB NVMe storage

Why it matters: Nvidia is aiming to revolutionize “personal AI computing”, similar to how PCs democratized computing in the 1980s. While the $3,000 price point isn’t cheap, it could create a new market segment between consumer laptops and data center setups, enabling a new wave of small AI startups to compete with tech giants without the need for massive cloud computing budgets.

Subscribe to Nvidia Waitlist

OPEN SOURCE

Microsoft drops Phi-4 model

The Summary: Microsoft has released Phi-4, a 14-billion-parameter language model, under an MIT license on Hugging Face. Announced in December, this compact and efficient model handles advanced tasks without requiring massive hardware. Trained on 9.8 trillion tokens, it scores over 80% on challenging math benchmarks.

Key details:

• 14B dense decoder-only Transformer, now fully open-source

• Trained on 9.8T tokens over 21 days using 1920 H100 GPUs

• Achieves over 80% on MATH and other advanced reasoning tests

• Features supervised fine-tuning and preference optimization for safety

Why it matters: Phi-4 combines advanced capabilities with efficiency, offering a smaller, cost-effective alternative to much larger models. Its open-source release allows developers and businesses to deploy state of the art tools without significant costs.

Phi-4 models download

QUICK NEWS

Quick news

• OpenAI may explore usage-based pricing for selected services

• Google releases an AI agent whitepaper

• NeuralSVG introduces text-to-vector generation

TOOLS

🥇 New tools

• TestSprite - AI agent automating entire software testing process

• AnyParser - Vision LLM for document parsing