The Summary AI
Posts
📅 ChatGPT Goes Autonomous

📅 ChatGPT Goes Autonomous

PLUS: Google Titans: Transformers 2.0?

The Summary AI
January 17, 2025

Welcome back!

OpenAI just made ChatGPT more autonomous with the launch of its Tasks feature, letting you schedule reminders and updates — even offline. It's an early glimpse into OpenAI's ambitions for autonomous AI. Let’s unpack…

Today’s Summary:

📅 ChatGPT Tasks launches with reminders
🎥 Luma's Ray2 revolutionizes AI videos
🧠 Google Titans mimics human memory
🎨 Stability unveils SPAR3D for 3D generation and editing
📱 xAI releases Grok free app
🧪 Microsoft’s MatterGen creates new materials
🛠️ 2 new tools

TOP STORY

ChatGPT Tasks introduces recurring reminders

The Summary: OpenAI launched ChatGPT Tasks, a beta feature that allows ChatGPT to run scheduled tasks - even when offline. Premium users can now set ChatGPT to deliver reminders, weather updates, and more through push notifications. The feature marks OpenAI's first step into autonomous AI systems.

Key details:

Tasks run on GPT-4o with a cap of 10 active tasks per user
Users can edit existing tasks in the new Tasks menu
Beta rollout begins with Plus, Pro, and Teams users and excludes voice chats, file uploads, and custom GPTs
Early user reports show some glitches, with some tasks failing without explanation, or executing at incorrect times
Tasks works across platforms (Web, iOS, Android, MacOS) with Windows support planned for Q1 2025, but task management is web-only

Why it matters: This move hints at OpenAI's broader plan for autonomous AI systems, with rumors pointing to an upcoming "Operator" agent to move beyond simple reminders toward more complex tasks. The feature's early bugs and limitations suggest OpenAI rushed the launch to stake its claim in the AI agent space.

VIDEO AI

Luma Ray2 model brings life to AI videos

The Summary: Luma's new Ray2 model, trained with 10x more compute power, generates videos with natural motion and physics that rival real footage. The model creates 5-10 second clips from text prompts in seconds. Early demos show high quality visuals with fluid movements. Available now to paid subscribers through Dream Machine, Ray2 marks a breakthrough in making AI video generation production-ready.

Introducing Ray2, a new frontier in video generative models. Scaled to 10x compute, #Ray2 creates realistic videos with natural and coherent motion, unlocking new freedoms of creative expression and visual storytelling. Available now. Learn more lumalabs.ai/ray.
— Luma AI (@LumaLabsAI)
6:13 PM • Jan 15, 2025

Key details:

Model uses novel multi-modal architecture, trained with 10x compute
Generates fluid, fast-motion scenes in seconds, breaking free from the slow-motion limitations common in other AI video tools
Early testers report a high success rate for usable generations, though complex prompts may still produce glitches
Launch includes $8,000 in creator prizes through Ray2 Awards, with $5,000 for most-viewed content by January 22nd

Why it matters: Ray2's ability to create videos with natural motion and realistic physics brings AI video generation closer to matching real-world footage expectations. Its high success rate for usable outputs could further transform video production workflows, especially for teams with limited VFX budgets.

GOOGLE DEEPMIND

Google Titans introduces human-like memory to AI models

The Summary: Google researchers have built a new AI architecture called Titans that mimics human memory systems. By separating processing into short-term attention and long-term neural memory, Titans can handle sequences over 2 million tokens, far beyond current limits. The system learns what to remember based on how surprising information is, outperforming GPT-4 on long-context tasks while using fewer parameters.

Attention has been the key component for most advances in LLMs, but it can’t scale to long context. Does this mean we need to find an alternative?
Presenting Titans: a new architecture with attention and a meta in-context memory that learns how to memorize at test time. Titans… x.com/i/web/status/1…
— Ali Behrouz (@behrouz_ali)
5:38 PM • Jan 13, 2025

Key details:

Titans architecture uses "surprise" metrics to decide what information to store long-term
Achieved 95% accuracy on 16,000-token needle-in-haystack tests with only 760M parameters
Beats GPT-4, Llama3-70B, and RAG systems on the BABILong benchmark for long document comprehension
Three variants (MAC, MAG, MAL) offer different memory integration approaches for specific use cases

Why it matters: Titans’ design improves efficiency for tasks requiring deep comprehension of very long documents, from legal analysis to scientific research. Its ability to achieve high performance with fewer parameters suggest a path toward more accessible and cost-effective AI systems.

Research paper