OpenAI's LifeVal grades ChatGPT on everyday errands

Overview

8 stories in this issue. The first 3 are today's priorities.

Hot Model Updates

1. Top · OpenAI unveils LifeVal; GPT-6 Astra reportedly scores 80%
2. Top · Grok Bot quotas reportedly halved as Musk calls costs "unsustainably high"
3. Top · Anthropic confirms Claude Code prompt change to avoid cloning others' games

Global AI News

4. SynthID Detector goes global, covering OpenAI, NVIDIA, Kakao and Apple content
5. Goldman Sachs pulls Claude Code and GitHub Copilot from traders and bankers
6. Uber builds an MCP gateway for 800+ servers and 5,000+ tools

Regional and Early Signals

7. Qwen Office sells 100,000 monthly passes in 30 minutes on a livestream (Chinese-language source)
8. MiniMax H3 open weights pass 24 million downloads and 300+ derivatives in a month (Chinese-language source)

Jiufeng graphic based on the sources cited in this issue.

Hot Model Updates

OpenAI is extending model evaluation from work tasks to personal errands like finding childcare, budgeting and dealing with government services.

OpenAI announced LifeVal on November 2. It is a benchmark that grades AI models on everyday personal tasks, including finding childcare, creating a budget, ordering groceries and accessing government services. OpenAI said it has used LifeVal internally over the past year to track and improve model performance on those tasks. RuntimeWire, citing CNN's reporting, says GPT-6 Astra scored 80%. RuntimeWire frames LifeVal as extending the GDPVal approach for work tasks into household life, giving OpenAI a training target for personal agents that browse, use computers and act through connectors.

Limitations: LifeVal is an internal OpenAI benchmark, and the 80% figure comes secondhand through media reporting. The report gives no scores for other models and does not say whether the task set or grading method is public.

Source: Original source OpenAI · RuntimeWire

Less than two weeks after Musk called Grok Bot the fastest-growing product he had ever seen, its usage limits were tightened.

The Decoder, citing the Telegraph, reports that several users saw their Grok Bot quotas cut in half over the weekend. TechCrunch reports that over Halloween weekend some users' weekly usage jumped from as low as 50% to 100% overnight. On Sunday, Musk wrote on X that "the cost of running Grok Bot was unsustainably high relative to revenue," which TechCrunch reads as confirmation that the change was not a glitch. Grok Bot browses websites, answers emails and logs into accounts, and it shares its quota with Grok and the coding platform Cursor. In early October, SpaceX president Gwynne Shotwell cited 800,000 paying users per week, more than double the previous month.

Limitations: Musk did not explicitly confirm a cut. The reports give no new quota numbers and no breakdown by subscription tier. One user reported waiting days for a fresh quota.

Source: Original source Elon Musk on X · TechCrunch · The Decoder

Original games, genre games and games built with your own assets should still work, but Claude Code is steered away from re-creating someone else's specific title.

The Verge, citing Tom's Hardware, reports that Anthropic updated Claude Code's system prompt to stop the tool from being used to re-create existing video games. Last week, a user on X posted a screenshot allegedly showing two new sentences, beginning "Do build original games, games in a genre, and games…".

Anthropic's Addy Osmani replied, confirming that with Opus 5.5 the company added two sentences to Claude Code's prompt to avoid cloning other people's specific games, because IP infringement violates its usage policy. He said original games, genre games and games made with your own assets should keep working normally. The Verge notes that retro gaming is currently overrun with AI-made decompilations, recompilations and ports.

Limitations: Osmani confirmed the prompt update but did not confirm the screenshot's exact wording. The report does not explain when the rule applies or how strictly it is enforced.

Source: Original source Screenshot post on X · The Verge

Global AI News

One upload now checks a file against every SynthID partner's watermark, but Google says it is not a general AI detector.

Google DeepMind has made SynthID Detector available worldwide. Users upload an image, audio or video file at SynthID.com and get a yes-or-no result.

• Coverage: Per Google's blog, updated October 29, it detects content from Google and partners including OpenAI, NVIDIA, Kakao and Apple.
• Before: Users could only ask Gemini, which recognized Google-generated content only.
• Technique: Introduced in 2023, SynthID uses a key to shift probability patterns in pixels, tokens or audio frequencies, embedding an invisible watermark.

Limitations: Google says a "no clear signal" result does not prove content is human-made. SynthID cannot detect Meta or Microsoft AI output, which use different watermarking, and heavy edits weaken the signal and can cause misjudgments. Reddit users criticized the false-positive rate and said light denoising removes the watermark.

Image source: Google.

Source: Original source Google Blog · Google DeepMind · InfoQ

The bank's non-engineer "vibe coding" experiment is over, and AI coding tools are now limited to engineers.

The Decoder, citing Bloomberg, reports that under internal guidelines from October, Goldman Sachs cut most employees off from AI coding tools such as Claude Code and GitHub Copilot, leaving access to engineers only. Traders and bankers had used the tools for months to monitor trading flows and create content. In July, CIO Marco Argenti publicly praised vibe coding by non-technical staff. Spokesperson Nick Carcaterra cited "a variety of factors including productivity and risk management" and said the experimentation phase is over.

Limitations: The report gives no count of affected staff and does not specify the risks. The rollback is reportedly causing frustration inside the firm.

Source: Single report The Decoder

Uber now routes how its internal AI agents find and call tools through a single gateway.

InfoQ reports that Uber built an MCP Gateway to standardize how AI agents discover and invoke internal services. It covers more than 800 MCP servers and 5,000 tools, and it separates tool registration from runtime invocation.

Limitations: Only a summary is available. It gives no latency figures, no detail on the permission model and no word on whether the gateway is open source.

Source: Single report InfoQ

Regional and Early Signals

A token-based AI office subscription was sold as a Singles' Day hit in a top livestream.

Leiphone reports that on the evening of November 2, Qwen Office personal monthly passes went on sale in Li Jiaqi's livestream and sold 100,000 units in half an hour, totaling about 2 trillion tokens. Qwen Office CEO Chen Yusen joined the stream to demo a few uses: organizing children's homework and tests, turning stream replays and comments into a business report, and building a shareable multi-user fishing map through conversation.

Limitations: Chinese-language source. The sales and token figures come from the report and have not been independently verified. The source gives no price and no per-pass token allowance.

Source: Single report Leiphone

After its weights were opened, H3 became a base model for video companies and NVIDIA researchers.

InfoQ China reports that MiniMax opened the weights of its multimodal model H3 in August. Within a month it passed 24 million downloads and more than 300 public derivative models. fal trained H3 Max on top of it, focusing on faster inference. Besides Open-H3-Regenerate-2K, NVIDIA researchers built the StreamRender-H3 streaming renderer and a fast video decoder, and they improved the generation pipeline without changing H3's weights, prompts or generation settings.

Limitations: Chinese-language source. The report does not explain how downloads were counted or on which platforms, and gives no quantified comparison of the derivative projects.

Source: Single report InfoQ China