爱窝啦 AI 日报

Today’s Summary

Jev, released by TypeSafe AI, returns probability values instead of text. SynthID opened its AI content detection to the public on the same day, covering 180 billion pieces of content.
From decision models to content provenance, today's updates are all about making AI outputs verifiable and quantifiable — not more verbose.
Start with Jev's probability output design and SynthID's public portal, then decide whether either fits your product or moderation workflow.

🔥 Today’s Top 10 Highlights

1. Jev Decision Model Returns Probabilities, Not Text

A new kind of decision model has entered the chat. Jev, released by TypeSafe AI, works nothing like a traditional language model — instead of returning text, it returns a floating-point probability value. This lets AI output decision confidence directly, no more parsing intent from strings. Ruan Yifeng flagged it as the biggest AI news last month in his Weekly Issue 414, noting “it’s wild that nobody thought of this useful concept before.” 🎯

2. Grok Bot Automatically Generates Daily AI Briefing Videos — Entirely in the Cloud

Grok Bot handled the entire content-to-video pipeline on its own. In a hands-on demo post shared on Jike , developer 歸藏 walked through the full automation chain: content collection, code writing, and video rendering all run on Grok’s cloud VM — zero local processes. He also shared the prompt publicly; just swap in your username and you’re good to go. If you’ve been wanting to automate a daily briefing, today’s a great day to try it. 🤖

3. Codex Plugin Can Now Directly Control a Real iPhone

Codex just got its first real phone-control capability. A project demo video posted on Jike by developer 钟二信 shows the iPhone Use plugin reliably handling everyday app tasks — scrolling Moments, placing Taobao orders, ordering Meituan takeout, posting to Xiaohongshu. The project is open-sourced at github.com/zhongerxin/iPhone-use. Developers who want to give Codex a hand (literally) can check out the repo directly. 📱

4. Claude Knowledge Work Plugins Repo Tops Today’s Daily Chart

This Anthropic open-source repo for knowledge workers is blowing up. The Claude Knowledge Work Plugins repo — a Claude Cowork plugin collection aimed at knowledge workers — picked up 392 new stars today, pushing its total to 27,560. Written in Python, it’s a solid reference for developers and enterprise users looking to extend collaborative capabilities inside Claude workflows. ⭐

5. Google SynthID Opens AI Content Detection to Everyone

Anyone can now verify whether a file was AI-generated. According to the official Google Gemini tweet , SynthID launched a public portal at synthid.com supporting watermark detection across images, video, and audio. Partners include OpenAI, NVIDIA, and Kakao, with Apple joining soon. Google says it has already watermarked 180 billion images and videos plus over 240,000 years’ worth of audio, processing 1 million verification requests per day. Content moderators and fact-checkers now have a free, public tool in their arsenal. 🔍

6. 歸藏Shares a Hacky Way to Power Grok with Leftover Code Plan Credits

Your unused Code plan tokens can actually feed a Grok Bot. In this post , 歸藏 describes the setup: drop a coding agent like Pi or DeepSeek Harness into an agent cloud VM, then route your leftover Code plan token quota through it — and Grok Bot can tap those tokens to write code, render videos, and more. This is a personal experiment without any official support, so quota availability and API stability are on you to evaluate. ⚠️

7. Codex Usage Analytics Page Has a Bug — Token Consumption Data Is Gone

Codex’s usage stats are temporarily broken. A user reported in this V2EX thread that the “Usage & Billing → Analytics” tab in Codex settings has a bug — the bottom section now only shows percentages, with no daily token consumption data. No official acknowledgment or fix timeline has been posted yet. If you’re running batch jobs and need to reconcile usage, find another way to estimate for now. 🐛

8. Terence Tao Calls for a “Mathematics 2.0” Evaluation Framework

Terence Tao argues that AI solving problems isn’t the same as mathematical progress. In a series of four consecutive posts on Mathstodon , he points out that traditional mathematical breakthroughs come bundled with seminars, collaboration, and follow-on digestion — that’s the real growth of the mathematical community. Some AI prompt engineers are only interested in marking problems as “solved” without being able to explain results or engage with the field. He advocates moving from a Math 1.0 (“first to solve”) mindset to a Math 2.0 that values explanation, community-building, and opening new directions — and he’s clear he’s critiquing narrow outcome metrics, not banning AI from math. The thread scored 587 points and 609 comments on HackerNews. 🧮

9. Overwatch 1 Private Server Resurrected Using AI Hacking Techniques

AI hackers brought Overwatch 1 back to life on a private server. A video shared by 宝玉 shows hackers spinning up a private server running the last version of Overwatch 1 before the OW2 transition completed. This is a real-world case of AI-assisted reverse engineering — the technical significance is that it demonstrates AI’s ability to dramatically accelerate private server setup, not that it’s officially sanctioned. Worth watching for developers interested in game preservation and legacy operation. 🎮

10. New Podcast Cold-Starts to 5,000 Subscribers in One Month

Human-made content still commands a premium in the age of AI slop. A Telegram channel documented this growth story : a brand-new podcast hit 5,000 subscribers in a single month. The channel author’s take: with AI-generated content flooding every feed, authentic human perspectives are actually easier for audiences to identify and stick with. For creators wondering whether it’s worth grinding out original content, this is a real data point worth bookmarking. 🎙️

😄 AI Fun Facts

A 10,000-Word Story — AI Made 9,000 of the Calls For You

Ever feel like AI-written pieces are just slightly off? A writer cited Ted Chiang’s analysis in The New Yorker for a surprisingly physical explanation: a 10,000-word story is essentially the author making 10,000 word-by-word choices — but when you generate it with a Prompt, you might only make a hundred choices yourself, with AI quietly handling the other nine thousand. Nine thousand unconstrained choices, with no guarantee they’re anywhere close to what you actually wanted. When cameras first showed up, people scoffed that “pressing a shutter button isn’t art” — until everyone realized that all those choices about light and composition were exactly what separated a photographer from a master. 📸

❓ Related Questions

I’m a Grok subscriber who wants the bot to run the video workflow daily — do I have to re-explain everything each time?

You don’t need to re-explain every time — the workflow just isn’t locked in yet. Today’s hands-on test of the cloud VM running an AI briefing video pipeline offers a practical fix: bake the content collection, code writing, and rendering steps into a prompt template so the bot follows a fixed script. Once you’ve got a successful run, subsequent runs only need a trigger — no re-introducing roles or steps. Get the workflow stable before thinking about long-term use. If your existing account needs a renewal, check out Aivora AI Account Store for Grok renewals — but don’t mix up the order: nail the workflow first, then handle billing. 💡