GPT-4o just got outplayed — by a smaller model.
Welcome back to Taledy AI, your go-to resource for staying up to date with the latest advancements in AI, one edition at a time.
This week? Things got spicy.
DeepSeek is teaching models how to think better than GPT-4o. OpenAI gave ChatGPT a memory upgrade. Meta just casually dropped a 2 trillion parameter tease. And NVIDIA? They've got a model that can switch reasoning on and off.
Let's get into it ⬇️
Latest Video:
Discover how to use Manus to bulk-create YouTube videos.
📈 The Latest in AI Signals
🚀 OpenAI Levels Up ChatGPT (Again)
ChatGPT now remembers everything. Literally.
🤔 What’s New:
- Memory is persistent across chats
- Better personalization, long-term context
- You control what it remembers or forgets
🚫 Who Doesn't Get It Yet?
- EEA, UK, Switzerland, Norway, Iceland, Liechtenstein
🔮 Why It Matters: We’re inching toward AI personal assistants that truly know you. The more you use it, the smarter it becomes. Creepy or cool? You decide.
Siri's Makeover Incoming: Apple is finally updating Siri for fall 2025. Expect smarter responses and deeper personalization.
📊 ByteDance Gets Logical: ByteDance introduces Seed-Thinking-v1.5 - a MoE model built for STEM reasoning. It's their first real leap into serious LLM territory.
🔥 HEADLINER: DeepSeek Reinvents Reinforcement
DeepSeek just dropped DeepSeek GRM — a generative reward model that critiques itself during inference. Think: "This AI doesn’t just answer. It provides explanations, evaluates its own responses, and adjusts its performance dynamically.
🔎 Method: SPCT (Self-Principled Critique Tuning)
- Rejective Fine-Tuning (RFT): Uses 1.07M instruction + 186K tough rejective samples
- Online RL via GRPO: Reward +1 if its answer matches ground truth, -1 if not. KL penalty? 0.08.
- Inference-Time Sampling: Get multiple self-critiques, then use a Meta-RM to filter out bad takes.
🔢 Results
- RewardBench: From 86.0% → 90.4% with meta-filtering
- PPE Preference: 64.7% → 67.2%
- Overall: Single-pass 69.5% → 72.8% with 32-sample voting
- Outperforms: Nemetron 4340BR, rivals GPT-4o
🔮 Why It Matters: This flips the script. Instead of scaling to 671B+ models, DeepSeek shows you can train smaller, smarter AIs that teach themselves. It’s like AI doing code reviews on its own brain.
🎉 New Model Watch: NVIDIA's Nemetron Ultra 253B
NVIDIA just launched a 253B parameter model that beats bigger beasts like DeepSeek R1 on most tasks, and it runs on a single 8x H100 node.
🎯 Highlights:
- Built on LLAMA 3.1
- Uses Neural Architecture Search (NAS)
- Reasoning On/Off toggle
- Math500: 97% with reasoning
- LiveCodeBench: 29% → 66.3%
🌊 Key Feature: Max sequence length of 131,072 tokens, long context window heaven.
🔮 Why It Matters: You don't need trillion-parameter monsters. Smart architecture + selective reasoning = efficient power. It's open-source with commercial use.
🚀 AI Tool of the Day: Fresh Picks
Taledy.com: Create viral short clips from YouTube videos or your own uploads.
👨🎨 Mind-Blowing AI: DreamActor-M1
From ByteDance comes DreamActor-M1: an AI that animates static images into full-body video sequences.
- Smooth motion
- Realistic expressions
- Rivals Runway's Act-One
Perfect for content creators who want motion without mocap. 🎥
🚀 Rumor Radar
- Meta teased a 2 TRILLION parameter "LLaMA 4 Behemoth" model
- OpenAI’s GPT-4.1 + GPT-O3 expected to drop this week
- DeepSeek R2 chatbot might already be in testing — SPCT-powered?
- Google’s Gemini Live now has real-time camera + screen input AI
📢 From Taledy
MCPs are taking over the world. Are you familiar with them, or should I create a tutorial for that next time? Let me know in the comments or by replying to this email.
📊 TL;DR Recap
- DeepSeek’s SPCT beats GPT-4o in some benchmarks
- ChatGPT memory is now permanent (unless you turn it off)
- NVIDIA’s 253B model does more with less
- Meta is back with open weight LLaMA-4s
- Gemini Live = real-time screen + camera AI
- ChatGPT is now the world’s most downloaded app
- ByteDance and Appleare making bold AI moves
Stay curious, stay sharp.
— The Taledy Team 🧠
P.S. Think someone needs this newsletter? Share it with them!
Follow us on Twitter | Join the Taledy AI community | Subscribe on YouTube
