Sign in Subscribe
2 min read

OpenAI Teases GPT-5 & Google Deprecates Gemini—New AI Arms Race Heats Up

New AI Models & Announcements

OpenAI Teases Upcoming Models, Possibly GPT-5 or Specialized Variants
Sam Altman hinted at new OpenAI models via tweet, sparking speculation about a general-purpose successor (e.g., GPT-5) or a specialized coding model (e.g., GPT-5 Codex). The community debates potential features, with some referencing Brian Fioca’s upcoming discussion on The Roo Cast (Nov 5) about Codex and coding agents.

Google Deprecates Gemini Models, Hinting at Gemini 3 Launch
Google’s changelog reveals several Gemini models will be deprecated by November 18th, fueling speculation about the imminent release of Gemini 3. Users discuss potential upgrades and the strategic timing of the deprecation.

Research & Technical Breakthroughs

LLMs Achieve Direct Semantic Communication Without Text (Cache-to-Cache)
A new paper introduces Cache-to-Cache (C2C), enabling LLMs to communicate via KV-cache fusion instead of text, improving accuracy and reducing latency. The method uses a neural network to project and merge caches between source and target models.

Scaling Coding-Agent RL to 32x H100s Yields 160% Performance Boost
A developer scaled Orca-Agent-v0.1 (14B orchestrator model) on 32x H100 GPUs, achieving a 160% improvement on Stanford’s TerminalBench. Key insights include the superiority of simple unit tests over complex reward functions. All code is open-sourced.

Tools & Platforms

llama.cpp Launches Official WebUI
The llama.cpp team released a new official WebUI, improving usability with enhanced features for interacting with local LLMs. The update includes streamlined workflows and community-driven feedback.

GPT-OSS Reimplemented from Scratch in Pure Python (No PyTorch/GPU)
A developer built a pure-Python GPT-OSS implementation without PyTorch or GPU dependencies, covering Grouped Query Attention (GQA), Mixture of Experts (MoE), and Rotary Position Embeddings (RoPE). The project includes a detailed blog post and open-source code.

Benchmarking & Evaluation

France Unveils Comparia: An LLM Leaderboard for European Languages & Energy Efficiency
The French government launched Comparia, a leaderboard akin to LMarena but focused on European language support and energy efficiency. The platform aims to reduce biases in LLM evaluations and promote sustainable AI.

Model Comparisons & Coding Agents

Cursor Composer-1 vs. Windsurf SWE-1.5: Real-World Coding Task Showdown
A user benchmarked Cursor’s Composer-1 and Windsurf’s SWE-1.5 by building a typing game and 3D solar system simulator. SWE-1.5 proved faster and more reliable, while Composer-1 showed stronger reasoning but occasional freezing.

OpenAI Faces Billion-Dollar Lawsuit Over Alleged Book Piracy for Training Data
OpenAI is accused of pirating books to train models, with evidence suggesting employees deleted the dataset and discussed it internally. The lawsuit could impose $150,000 per book in damages, totaling billions.