OpenAI Declares "Chat Is Dead" and Set to Overhaul ChatGPT Into an Agent App; Gemma 4 MTP Merged into llama.cpp

OpenAI Declares "Chat Is Dead" — Set to Fully Rebuild ChatGPT as an Agent App

OpenAI is planning the biggest overhaul of ChatGPT since its launch. According to a report by The Decoder, the company has adopted an internal mantra of "Chat is dead," and intends to transform ChatGPT from a chatbot into a "super app" that brings together coding tools, AI agents, and partner apps such as Canva and Booking.com.

Until now, ChatGPT has centered on a chat format where users enter a prompt and get a response. Under OpenAI's vision, however, autonomous task-executing agents will take the lead. Writing and running code, arranging bookings, creating designs — OpenAI envisions a future where AI, acting on user instructions, works across multiple apps and services.

This move aligns with the broader industry-wide shift from "chat" to "agents." That said, specific release timing and feature details remain unknown at this point.

Gemma 4 MTP Support Merged into llama.cpp

Multi-Token Prediction (MTP) support for Google's Gemma 4 has been merged into llama.cpp, the de facto standard engine for local LLM inference (PR #23398). As a result, running Gemma 4 models on llama.cpp can now benefit from the inference speedup that MTP provides.

On Reddit's LocalLLaMA community, users have welcomed the change, saying that their local Gemma setups will get faster than ever. Coming on the heels of the recent report of a 31% speedup from QAT optimization, running Gemma 4 locally looks set to become even smoother.

KV Cache Quantization Benchmarks for Qwen 3.6 27B Published Across 75 Configurations

A community member has published benchmark results covering 75 pairs of configurations for KV cache quantization on the Qwen 3.6 27B model. The exhaustive tests span four quantization levels — q8/q6/q5/q4 — combined with implementations such as KVarN and TurboQuant/TCQ.

The benchmarks were run with BeeLlama.cpp, a fork of llama.cpp. KV cache quantization is a key technique that can dramatically reduce memory usage during long-context inference, and it has a major impact on practicality in local environments where VRAM capacity is constrained.

Detailed analyses are available in two posts: one on the KV cache quantization benchmarks and another on the KVarN implementation.

Nemotron 3.5 ASR Dockerized — 4.5x Real-Time Speed on CPU

NVIDIA's speech recognition model Nemotron 3.5 ASR is now easy to run as a Docker container. The developer cites multilingual support covering 40+ languages and a native streaming architecture as reasons for switching from the Parakeet model.

Particularly noteworthy is that it achieves 4.5x real-time speed on CPU using onnxruntime-genai as the backend. Since practical speech recognition speeds are attainable without a GPU, this holds promise for edge devices and other low-cost environments.

The repository includes API call examples for both streaming and file upload. CUDA support is unverified at this point, but it appears achievable with some configuration adjustments.

Other Topics

**AI boom fuels anti-tech extremism** — According to The Guardian, anti-tech extremism is on the rise against the backdrop of the rapidly accelerating AI boom. Hostile moves toward technology companies have also sparked discussion on Hacker News.

**Paul Krugman on the pitfalls of AI and innovation** — Economist Paul Krugman published a Substack post examining the limits and pitfalls of the innovation that AI brings.

**CognatoAI announces a version control tool for AI agents** — A project building a version control tool for AI agents, motivated by the concern that "Git/GitHub hasn't evolved for the agent era," was announced on Show HN.

**AI-generated short story wins the Commonwealth Short Story Prize** — A short story written with AI has won the Commonwealth Short Story Prize, making headlines. It is likely to reignite debate over AI's role in literary creation.

**antirez discusses AI-powered automated QA and testing** — antirez, the creator of Redis, published a blog post on leveraging AI for automated QA and testing. It is also generating active discussion on Hacker News.