Happy Valentine’s Day, she wants flowers even if she says she doesn’t.

Today in Tech History: On February 14, 2005, YouTube was established by Chad Hurley, Steve Chen, and Jawed Karim, all of whom were early employees of PayPal.

In today’s email:

• Roadmap to GPT-4.5, GPT-5

• Superfast Perplexity Model

• Grok-3 Release Imminent

• 4 New AI Tools

• Latest AI Research Papers

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

OpenAI phases out legacy models as GPT-4.5 leads into GPT-5

OpenAI is preparing to release GPT-4.5, internally codenamed "Orion," as its final incremental language model before transitioning to a new AI architecture. Initial reports indicated that Orion delivered only marginal improvements over GPT-4, reinforcing the idea that scaling traditional language models is approaching its practical limits. Rather than deploying o3 as an independent release, OpenAI is integrating it into GPT-5—a system designed to autonomously determine when to engage in complex reasoning, leverage external tools, and adapt dynamically across tasks.

GPT-5 represents a strategic shift away from OpenAI’s fragmented AI offerings, consolidating multiple ChatGPT versions, API models, and feature sets into a unified intelligence framework. The model will integrate capabilities such as voice interaction, search, and advanced research into a single system.

OpenAI is also introducing a tiered access model: free users will interact with the standard version, Plus subscribers will access an enhanced reasoning layer, and Pro users will unlock the highest level of performance. This transition aligns with OpenAI’s investment in Large Reasoning Models (LRMs), which have already propelled its o3 model into the top 50 global coding benchmarks.

By combining the general linguistic capabilities of LLMs with the structured precision of reinforcement-trained reasoning models, OpenAI is positioning GPT-5 as a more adaptive and functionally robust AI system. Read more.

Perplexity AI launches new ultra-fast AI search model Sonar

Perplexity AI has launched an upgraded version of its Sonar search model, now powered by Meta’s Llama 3.3 70B and optimized with Cerebras Systems’ specialized hardware. In internal tests, Sonar outperforms models like GPT-4o mini and Claude 3.5 Haiku in user satisfaction, and competes closely with premium models like GPT-4o and Claude 3.5 Sonnet in search accuracy.

The key to Sonar's speed lies in Cerebras' Wafer Scale Engines, massive chips capable of processing 1,200 tokens per second. This allows Sonar to generate responses nearly instantly, setting a new standard for AI-powered search. Currently, Sonar is available exclusively to Pro users, though Perplexity plans to expand access soon. Details on the financial terms of their partnership with Cerebras remain under wraps. Read more.

Musk says Grok 3 is almost here and already outthinking the competition

Elon Musk announced that Grok 3, xAI’s chatbot and competitor to OpenAI’s ChatGPT, is in its final development phase and set to launch within two weeks. Speaking via video at the World Government Summit in Dubai, he claimed Grok 3 exhibits superior reasoning capabilities, outperforming all existing chatbots.

During the same address, Musk warned that human intelligence will soon be dwarfed, stating that digital intelligence will comprise "more than 99% of all intelligence in the future." He also referenced futurist Ray Kurzweil’s predictions that humans may eventually merge with machines, accessing cloud computing directly with their brains.

Musk has long positioned brain-computer interfaces, such as those developed by his company Neuralink, as a way to keep pace with AI. Read more.

• Elon Musk ready to pull $97.4B OpenAI bid if Board upholds nonprofit mission

• Anthropic takes aim at OpenAI’s lead with upcoming hybrid reasoning model

• OpenAI adjusts Model Spec and removes certain content restrictions in ChatGPT

• Sam Altman shares plans to bring o3 Deep Research agent to free and ChatGPT Plus users

• Meta in talks to acquire AI chip firm FuriosaAI, according to report

• Adobe's Firefly AI video generator is now in public beta, turning text and images into short clips

• Apple announces February 19 launch for upcoming budget iPhone with AI and FaceID

• PIN AI launches app for custom, on-device intelligence

• YouTube brings AI video creation to Shorts with Google's Veo 2

• Humanoid Robots Master Fall Recovery with Advanced RL Framework

• Artificial Analysis launches Intelligence Index V2, highlighting reasoning models

• Epoch AI research uncovers rapid AI compute scaling and predicts future bottlenecks

• Watch: Google DeepMindCEO: We Want To Build A Virtual Cell

• Stable Diffusion 3.5 Large is Now Available on Microsoft Azure AI Foundry

• Gemini Advanced now remembers past chats for personalized responses

• Report says companies ‘playing with fire’ as AI chatbots fail when trying to summarize news

• Snowflake expands AI tools with Anthropic partnership — what it means for businesses

• Scarlett Johansson calls for deepfake ban after AI video goes viral

• U.K. drops AI safety focus as it signs up Anthropic to help transform public services

📄 Light-A-Video: Training-free Video Relighting via Progressive Light Fusion

📄 TextAtlas5M: A Large-scale Dataset for Dense Text Image Generation ​​

📄 Exploring the Limit of Outcome Reward for Learning Mathematical Reasoning

📄 LM2: Large Memory Models

📄 Fino1: On the Transferability of Reasoning Enhanced LLMs to Finance

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!