Meta Faces Largest Fine in Tech History, Four States Seek $1.4 Trillion
On August 12, Meta will face a critical trial in the U.S. Oakland Federal Court. California, Colorado, Kentucky, and New Jersey jointly accuse Meta of addicting minors through the product design of Facebook and Instagram, while misleading the public about platform safety. The states calculated a maximum fine of $1.4 trillion based on the number of affected users multiplied by per-violation penalties under state consumer protection laws, equivalent to 92% of Meta’s market value. Meta calls this calculation “unprecedented in the history of consumer protection enforcement.” Previously, New Mexico won a similar lawsuit, with cumulative fines reaching $942 million. The core issue of this case is: when does making user addiction part of product design translate into legally recognized harm.
Source:
Kimi’s Valuation Surges Over 10x in a Year, FOMO Grips Primary Market
Moonshot AI’s Kimi valuation has soared from approximately $4.3 billion to $50 billion in less than a year. According to primary market sources, a $18 billion valuation Kimi secondary share offering was snapped up instantly upon release, with funds arriving for some deals even before a term sheet was secured. Kimi’s fundraising pace this year has been intense: three rounds from January-February raised valuation from $10B to $18B; it surpassed $20B in May; after the K3 release, the Series F round was oversubscribed multiple times and closed early, with a post-money valuation of $35B; the Pre-IPO round now values the company at $50B pre-money. Fundraising materials show Kimi’s ARR grew from about $100M in March to ~$300M in June, with API revenue exceeding 70% and overseas revenue up 400% year-over-year.
Source:
Meta Open-Sources Muse Glimmer 30B, Optimized for Local Agent Scenarios
Meta founder Mark Zuckerberg announced the open-sourcing of Muse Glimmer 30B, a dense architecture model with ~29.6B parameters, supporting image+text input and 131K context, distilled from Muse Spark, with focused training on Agent, multi-step reasoning, Tool Use, and Coding. After 4-bit quantization, it’s under 20GB and can run on 24GB VRAM. An RTX 5090 with DFlash speculative decoding achieves 233 tok/s, while an M4 Max reaches ~37.8 tok/s. Weights are directly released under Apache 2.0 license. Benchmarks: SWE-Bench Verified 76.0, TerminalBench 2.1 at 51.7, MCP Atlas 75.5. Muse Spark 1.2 weights will be released later.
Source:
- @MaxForAI: https://x.com/MaxForAI/status/2086774638793928771
- @vista8: https://x.com/vista8/status/2086822225601241588
OpenAI Splits Top-Tier Model Product Lines, Acquires HTML PPT Company Nextslide
According to multiple bloggers, OpenAI has divided its top-tier models into two product lines: a deep reasoning and complex task backend model codenamed Doug, focusing on logic chains and tool use; and a real-time multimodal assistant codenamed Astra, capable of seeing, hearing, and speaking, emphasizing low-latency interaction. Simultaneously, OpenAI acquired Nextslide, an HTML PPT company whose functionality is considered mediocre, even inferior to open-source alternatives, with speculation that the acquisition is for the team. Claude Code’s Auto mode is now enabled by default, allowing the model to autonomously decide when to ask the user during coding tasks, with added defenses against malicious instructions.
Source:
- @vista8: https://x.com/vista8/status/2086822225601241588
- @vista8: https://x.com/vista8/status/2086823315780636774
Doubao Enters Hotel Booking Commissions, AI Commercialization Reaches Transaction Flow
According to media reports, starting August 10, Douyin Life Services implemented a new software service fee policy: for hotel bookings where users search and inquire via Doubao and jump to Douyin’s merchant platform to complete the order, a 11.4% software service fee plus a 0.6% payment processing fee will be charged to the hotel, resulting in a total rate of approximately 12%. This indicates ByteDance is experimenting with moving its AI assistant from a chat interface into real consumer transaction flows—the AI handles demand understanding and purchase decisions, while Douyin facilitates the transaction and takes a commission from the GMV.
Source:
Open Design Launches Team Workspace, Codex Enters Multi-User Collaboration Mode
Open Design officially released v0.18.0, introducing the world’s first collaborative Design Agent. This update includes over 200,000 lines of code, supports collaboration between dozens of agents and humans, and is compatible with products like Claude Code, Codex, OpenCode, and Kimi. Codex users can now create, review, and iterate with teammates around the same real-time canvas. A plugin system was also released, supporting 10+ design outputs including prototypes, PPTs, and motion videos.
Source:
- @tuturetom: https://x.com/tuturetom/status/2086829438961119668
- @OpenDesignHQ: https://x.com/OpenDesignHQ/status/2086749655354786159
Genspark Intensively Releases Products, Open-Sources GenOffice and Launches Recording Hardware
Multiple bloggers noted Genspark’s extremely fast product cadence—having just open-sourced the office suite GenOffice a few days ago, it today launched recording hardware with supporting software. Comments suggest its strategy is to address any demand and compete on volume. Meanwhile, preparations are underway for the second GEO conference in Shenzhen.
Source:
- @vista8: https://x.com/vista8/status/2086834401669677364
- @vista8: https://x.com/vista8/status/2086721028810846556
Stats: Timeline scans=360 Bloggers matched=33 Total tweets matched=204 Weighted tweet score=164.1 Original tweets=96 RT tweets=39 Crawl attempts=2 Boundary coverage status=tail_confidently_crossed_target_boundary