In partnership with

Good morning. It’s Friday, February 28th.

On this day in tech history: in 1990, ARPANET, the precursor to the modern internet, was formally decommissioned. Developed by the Advanced Research Projects Agency (ARPA), ARPANET was one of the first networks to implement the TCP/IP protocol suite, laying the groundwork for today’s internet.

In today’s email:

• GPT-4.5

• Meta Scouts $200B AI Data Site

• Meta’s Aria Gen 2 Glasses

• 4 New AI Tools

• Latest AI Research Papers

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

GPT-4.5 launches as OpenAI’s largest model with subtle upgrades and steep costs

OpenAI has unveiled GPT-4.5, internally known as ‘Orion', marking the final iteration under its traditional scaling paradigm. CEO Sam Altman described it as the first model that “feels like talking to a thoughtful person.” Unlike OpenAI’s reasoning models, it adheres to a traditional LLM architecture. It supports text-based multimodal inputs but lacks voice, video, and screen-sharing features.

Originally intended as GPT-5, Orion underperformed against expectations, prompting a rebranding. The model improves emotional intelligence and reduces hallucinations but lacks chain-of-thought reasoning, leaving it behind competitors like Anthropic’s Claude 3.7.

With API pricing at $75/$180 per million tokens, the model has sparked debate over its cost-to-value ratio. Its computational demands render it prohibitively expensive—up to 30 times costlier than GPT-4o. Access also remains limited to ChatGPT Pro subscribers ($200/month) and API users due to GPU shortages, delaying its rollout to lower-tier plans.

While early reactions are mixed, OpenAI positions GPT-4.5 as a stepping stone toward more sophisticated AI systems. Read more.

Meta Quietly Scouts Sites for $200 Billion AI Powerhouse

Meta is reportedly plotting a $200 billion AI data center project, eyeing Louisiana, Wyoming, and Texas for the buildout, according to The Information. Executives have met with developers and toured potential sites, though Meta dismissed the claims, calling them speculative and reaffirming that its disclosed capital expenditures remain unchanged.

The alleged scale dwarfs even the most aggressive AI infrastructure bets. Microsoft is pouring $80 billion into data centers next year, while Amazon is set to exceed $75 billion. Meta itself has already allocated $65 billion for AI expansion in 2024. Read more.

Meta’s Aria Gen 2: AR Meets AI

Meta has introduced Aria Gen 2, a research-focused headset designed to advance AI, robotics, and machine perception. Without a display, it prioritizes sensor-driven spatial intelligence, featuring upgraded RGB and eye-tracking cameras, spatial microphones, and Meta’s custom silicon for real-time AI processing. New additions include a photoplethysmogram (PPG) sensor for heart rate monitoring and a contact microphone for precise voice isolation.

Designed for on-device AI tasks, Aria Gen 2 runs hand and eye tracking, speech recognition, and SLAM-based navigation, making it a proving ground for future AR intelligence. Read more.

Related Story: Meta to launch standalone app with paid subscription model in testing

• Pika Labs upgrades AI video with 10-second clips, 1080p, and key frame control

• Hugging Face launches FastRTC to simplify real-time AI voice and video apps

• Microsoft releases new Phi models optimized for multimodal processing, efficiency

• Amazon debuts LLM-powered Alexa+ with expanded automation features

• IBM debuts new Granite 3.2 family of models that include reasoning when you want it

• Tencent’s latest AI model blends intuition with structured reasoning

• You.com unveils AI agent that automates market research with 400+ sources and instant verification

• ElevenLabs' Scribe model can transcribe the world's fastest speaker without breaking a sweat

• Nvidia’s AI Empire Grows as Revenue Hits $39.3B

• In just 4 months, AI coding assistant Cursor raised another $100M at a $2.6B valuation led by Thrive, sources say

• Hume launches new text-to-speech model Octave that generates custom AI voices with adjustable emotions

• TikTok sunsets its creator marketplace for TikTok One, a broader solution with AI tools

• OpenAI brings Sora video generator to UK and EU

• Figure AI accelerates home robotics timeline, moves up testing by 2 years

• Microsoft Copilot is Now Available for Mac Computers

• Ideogram Launches 2a Model: Faster, Cheaper Text-to-Image Generation

• OctoTools: Stanford's open-source framework optimizes LLM reasoning through modular tool orchestration

• AI in movies: ‘Emilia Pérez’ and ‘The Brutalist’ controversies, explained

📄 SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution

📄 ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation ​​

📄 KV-Edit: Training-Free Image Editing for Precise Background Preservation

📄 GHOST 2.0: generative high-fidelity one shot transfer of heads

📄 R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts

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!