Xiaoliu BOT
中文EN

X Platform AI Briefing May 6th: ChatGPT Upgrades to GPT-5.5 Model, Runway Launches Real-Time Conversational Video Characters

GPT-5.5 Instant Upgraded to ChatGPT’s Default Model, with Multiple Bloggers Discussing Their Experiences

Multiple bloggers have mentioned that OpenAI has upgraded ChatGPT’s default model to GPT-5.5 Instant, fully replacing the previous generation GPT-5.3 Instant, with a complete rollout. After the upgrade, the hallucination rate in high-risk fields like healthcare, law, and finance has significantly decreased, responses are more concise, and it has begun to actively use user chat history for personalized recommendations.

Sam Altman (@sama) posted that “ChatGPT feels very switched on now” and “the new instant model in chatgpt is so good damn”, and recommended that users who previously only used reasoning models try the new version; he later mentioned that improvements in speed, intelligence, personality, and memory/personalization combined feel greater than the sum of their parts. @dotey cited @ChatGPTapp’s announcement to detail four dimensions of change: a 52.5% reduction in hallucination rates in high-risk areas, shorter and less verbose answers, proactive use of external context like Gmail and history, and a “memory sources” feature that supports visually viewing citation sources and manual deletion. @op7418 also posted a summary of the main upgrade points of GPT-5.5 Instant, noting that it is available to ChatGPT free users, with the API alias being chat-latest.

Sources:

OpenAI Launches Migrate to Codex Feature, Supporting One-Click Migration of Other Programming Agent Configurations

A blogger mentioned that OpenAI has launched the Migrate to Codex feature, allowing users to import configurations from other programming tools like Claude Code and Cursor into Codex with one click, including programming agent configurations, rules, skills, MCP, hooks, subagents, and all sessions from the last 30 days.

This feature scans user-level and project-level configurations, automatically maps instruction files to AGENTS.md and settings.json to config.toml, with corresponding entries for MCP, hooks, skills, and subagents; parts that cannot be migrated automatically, Codex will proactively open a new thread to assist. @xiaohu outlined the steps: open Codex App’s Settings → General page, find Import other agent setup → select the project to import → after completion, click View imported files to check the results.

Sources:

Runway Launches Characters Feature: Generate a Real-Time Conversational Video Character from a Single Image

A blogger mentioned that Runway has launched the Runway Characters product, allowing users to upload a reference image to generate a video character that can converse with the user in real-time. This character can see the camera and screen sharing, supports custom voice and opening lines, can be connected to text or Markdown knowledge bases to answer based on materials, can call tools (like highlighting web buttons, checking order inventory, etc.), and can be integrated into third-party products via API, React SDK, or web Widget.

Multiple bloggers believe this is different from traditional “digital avatars”: users are not waiting for a pre-generated video but are interacting with the character on screen in real-time, where the character must understand the user, see what the user is viewing, use materials to answer, and perform actions within products.

Sources:

Google Releases MTP Drafter for Gemma 4, Inference Speed Up to 3x Faster

Multiple bloggers have mentioned that Google has released the MTP drafter (multi-token prediction draft model) for the open-source model Gemma 4, using a speculative decoding mechanism where a smaller model first guesses multiple upcoming tokens in batch, then a larger model verifies them all in parallel at once, outputting all verified parts at once. On Apple Silicon running a 26B MoE model with batch sizes 4-8, it achieves about a 2.2x speedup, with a peak memory of only 4.3GB; the final output is word-for-word identical to the larger model, with no quality loss.

@dotey analyzed the technical principle: the bottleneck in LLM inference is often not in computing power but in memory bandwidth, with processors spending most time moving parameters rather than computing; speculative decoding utilizes idle computing power, essentially maxing out the pipeline. @LufzzLiz pointed out this mechanism is particularly valuable for scenarios like running models locally and multi-step planning with low latency. @dotey also referenced @googlegemma’s illustrated article for further explanation. The drafter uses the Apache 2.0 license, with weights uploaded to Hugging Face and Kaggle, and is already supported by transformers, MLX, vLLM, SGLang, and Ollama.

Sources:

Mininglamp-AI Open-Sources Cider and Mano-P: Turning Mac into a Local AI Workstation

A blogger mentioned that Mininglamp-AI has open-sourced two projects, Cider and Mano-P, addressing two problems: “how to run AI faster locally on Mac” and “how AI can truly operate a computer”.

Cider makes better use of the M5 chip’s INT8 TensorOps, enabling faster and more memory-efficient LLM/VLM inference; Mano-P is a pure-vision GUI-VLA Agent for local inference on Mac mini/MacBook, capable of operating desktop software, web interfaces, and complex graphical workflows, supporting cross-system data integration and long-task planning. The technical path is pure-vision operation with screenshot data that can stay on the device. Cider solves on-device inference acceleration, while Mano-P solves how AI can see the screen and operate a computer like a human, together forming a private AI local infrastructure. @xiaohu also added real-world test results on a Mac mini with an M4 chip and 32GB memory: a 4B quantized model achieves 476 tokens/s prefill and 76 tokens/s decoding on an Apple M4 Pro, with a peak memory of only 4.3GB, representing a 60x+ speedup compared to standard PyTorch CPU inference on the device.

Sources:

Cryptocurrency Exchange Coinbase Announces Layoffs of About 14%, AI Efficiency Cited as One Reason

A blogger mentioned that cryptocurrency exchange Coinbase announced layoffs of about 14% (approximately 700 employees), with CEO Brian Armstrong giving two reasons: the cryptocurrency market entering a downturn cycle, and AI changing how the company operates.

@dotey cited Brian Armstrong’s internal letter noting: AI allows engineers to deliver in days what used to take teams weeks, and non-technical teams are now starting to write production code; last October, 40% of the company’s daily code was generated by AI, with a goal to push this ratio above 50%. Accompanying the layoffs is organizational restructuring: under the CEO and COO, management levels are reduced to a maximum of 5 layers, with each manager potentially overseeing 15 or more direct reports; all managers must also be frontline contributors, not just pure managers; the most aggressive “AI-native teams” may even have single-person teams—one person taking on the roles of engineer, designer, and product manager, dispatching numerous AI agents to complete work. In terms of compensation, U.S. employees receive at least 16 weeks of base salary plus 2 weeks for each year worked, along with 6 months of continued health insurance. @dotey also added an analyst’s interpretation from Mizuho Securities: the crypto winter is likely the real reason for most layoffs, with AI serving as a convenient excuse. On the same day, @sama posted mentioning “5.5 in codex is so good for non-coding tasks”, hinting that AI’s penetration into non-coding work is also accelerating.

Sources:

Petdex: An Open-Source Project Providing Third-Party Terminal Pets for Codex Users

Multiple bloggers have mentioned that the open-source project Petdex can replace “terminal pets” for Codex users, serving as a Codex Pets store that supports previewing, downloading, and submitting various dynamic pets that can be used in Codex.

Usage is simple: after choosing a favorite pet, run `npx petdex install ` to install it, then select the pet via Codex’s Settings → Appearance → Pets, or use the /pet command to summon or hide the pet. @xiaohu pointed out that users who found Codex’s built-in pets too ugly finally have a solution, and compiled the GitHub address and multiple available pet types.

Sources:

CodexBar 0.24 and OpenClaw 2026.5.5 Updated Simultaneously

@steipete posted that CodexBar 0.24 has been officially released, adding support for Windsurf, Codebuff, and DeepSeek providers, introducing Copilot multi-account switching and optional local storage splitting, and fixing hung Codex RPC and battery drain issues. The @openclaw official account released OpenClaw 2026.5.5 on the same day, fixing message routing issues across multiple platforms including Feishu, LINE, Telegram, Discord, Matrix, Slack, and WhatsApp, improving DM configuration verification, progress draft cleanup, and reconnection false positive handling; Gateway diagnostics were also optimized, with clearer supervisor restart handshake displays, quieter and more precise token shadow warnings, and plugin updates no longer losing SDK links. @vista8 also posted that day recommending Luoxiao Shan’s 16,000-word long article on experiences in creating “lifelike AI assistants”, noting that those who excel in this area are mostly from the gaming industry.

Sources:

Scraping Statistics (2026-05-06)

  • Timeline lines scanned: 240
  • Bloggers hit: 27
  • Total tweets hit: 132
  • Weighted tweet score: 110.1
  • Original tweets: 78
  • RT tweets: 23
  • Scraping attempts: 1
  • Boundary coverage status: tail_confidently_crossed_target_boundary (full coverage)