- The Upload
- Posts
- GPT Just Leveled Up
GPT Just Leveled Up
GPT-4.1 brings 1M tokens and dev-friendly upgrades. Meanwhile, OpenAI is building agents that can do science.
Mark Chepelyuk
April 15, 2025
OpenAI just released its new flagship, and it’s built for more than chat. With massive context windows, faster speeds, and cheaper pricing, GPT-4.1 is the new backbone for AI builders. But it doesn’t stop there. ByteDance just stepped up in video and Google trained a model to talk to dolphins.
Today’s Upload
OpenAI launched GPT-4.1 with 1M token context and faster performance
ByteDance released Seaweed, a small but powerful AI video model
Google trained an AI to decode dolphin communication using audio and Pixel sensors
Let’s get into it. 🚀
Image source: OpenAI
🧠 GPT-4.1 Is Here
OpenAI’s new flagship model offers more context, better speed, and lower prices.
Key Details:
GPT-4.1 supports up to 1 million tokens of context
Outperforms GPT-4 Turbo on code, math, and reasoning
Costs less than GPT-4 Turbo at just 5 cents per million tokens
Now available in API, ChatGPT, and playground
Focused on dev tools, memory, and stability
Why It Matters
This is not just a better chatbot. It’s a serious platform upgrade for developers, creators, and anyone building AI-powered products. The 1M token context unlocks long documents, full video transcripts, and multi-step logic chains. GPT is no longer just a front-end tool. It’s becoming infrastructure.
Glad to share Seaweed-7B, a cost-effective foundation model for video generation. Our tech report highlights the key designs that significantly improve compute efficiency and performance given limited resources, achieving comparable quality against other industry-level models. To
— Ceyuan Yang (@CeyuanY)
3:12 AM • Apr 14, 2025
🎥 ByteDance Launches Seaweed for AI Video
A leaner, faster model outperforms bigger players on key video tasks.
Key Details:
Seaweed is a diffusion-based video model designed for low compute environments
Outperforms larger models like Sora and Kling on coherence and scene composition
Requires significantly less GPU power to run
Built for mobile and edge deployment
Released alongside ByteDance’s new research division focused on lightweight generative models
Why It Matters
While big players race to build massive models, ByteDance is optimizing for efficiency. Seaweed could be a preview of what real-time, on-device video generation looks like. For creators and developers, it opens up new workflows without massive compute needs.
Image source: Google
🐬 Google Builds AI to Talk to Dolphins
DolphinGemma is trained to understand marine mammal language using decades of audio data.
Key Details:
Built using a Gemini-based architecture
Trained on recordings of dolphin communication across multiple habitats
Combined with sensors from Pixel phones to enhance field data collection
Developed in partnership with marine biologists and oceanographic institutions
Currently analyzing patterns, pitch, and behavior-linked signals
Why It Matters
This is not a gimmick. It’s real cross-disciplinary AI being applied to the natural world. DolphinGemma pushes AI beyond screens and servers into research fields, opening doors for environmental science, conservation, and new forms of sensory intelligence.
🕐 Quick Bits
📱 Apple to train AI directly on your device
New privacy-first system keeps learning local to your iPhone or Mac.
🧬 AI outperforms doctors in skin cancer detection
UK health system rolls out model with 99 percent accuracy in trials.
🏫 Google Classroom gets Gemini-powered quiz builder
Teachers can generate skill-based assessments instantly.