Xiaoliu BOT
中文EN

X Platform June 9 AI Brief | Anthropic’s New Model Fable Imminent, Apple Siri AI Limited to iPhone 17 Pro, SpaceX Reveals Space AI Infrastructure Blueprint

Anthropic’s New Model Mythos to Launch, Final Name Likely Fable

Anthropic plans to release its next-generation AI model on June 10. @MaxForAI cites sources saying the public version will not be called Mythos; the final name may be Fable. The model will include extensive safety measures and intentionally limit cybersecurity capabilities provided to trusted partners through Project Glasswing. Anthropic will highlight advanced reasoning abilities and improvements in long multi-turn tasks. The previous week, Anthropic completed a confidential IPO filing with a post-money valuation of $965 billion. @xiaohu also shared “inside news” confirming the release date, while @op7418 noted that Anthropic may introduce mandatory KYC for usage.

Sources:

Apple WWDC2026 Unveils New Siri AI, On-Device Model Only Supports iPhone 17 Pro

Apple introduced a brand-new Siri at WWDC2026, upgraded to a standalone app supporting natural conversation, contextual memory, cross-app operations, and screen/camera content understanding. @xiaohu detailed the technical architecture: an on-device model with approximately 20 billion parameters handles local information retrieval and scene awareness, while the cloud model is distilled from Gemini rather than directly using Gemini. Photos, Safari, Passwords, Shortcuts, and more gain AI capabilities. @op7418 pointed out that the on-device model only supports the iPhone 17 Pro, and is temporarily unavailable in Europe and China; he called it the “only highlight” of WWDC. @xiaohu later added that Siri is powered by Apple’s own foundation model (distilled via Gemini), with Gemini only providing customized additional support on iCloud, lamenting “Google got played again.” @imwsl90 noted that Intel devices have been completely removed from the macOS 27 compatible device list.

Sources:

SpaceX Reveals Space AI Infrastructure Blueprint: AI Satellites, Terafab Chip Factory, and Lunar Mass Driver

SpaceX released a video detailing its space AI data center plans. @xiaohu provided a full transcript: the first AI satellite is designed with 150 kW peak power, 120 kW sustained compute, a 70-meter wingspan, using existing Starlink V3 solar technology, essentially “a GB300 compute rack in space.” SpaceX plans to build a Terafab space chip factory in Bastrop, Texas, covering approximately 100 million square feet (10 times Tesla Giga Texas), targeting an annual production of 1 terawatt of AI chips. Timeline: by end of next year, target annualized 1 GW; about three and a half years to reach 100 GW; ultimately 1 TW (twice current US electricity consumption). Longer-term plans include using a mass driver (electromagnetic cannon) on the Moon to launch AI satellites into deep space. @elonmusk also retweeted SpaceX AI Satellites content.

Sources:

Xiaomi MiMo V2.5 Pro UltraSpeed: Trillion-Parameter Model Breaks 1000 token/s for the First Time

Xiaomi, in collaboration with TileRT_AI, released MiMo-V2.5-Pro-UltraSpeed, achieving over 1000 token/s output speed on a single node with 8 GPGPUs for a 1 trillion parameter MoE model. @MaxForAI noted this differs from Cerebras wafer-scale integration or Groq’s pure SRAM approach, using only standard hardware. @op7418 provided a detailed review after early access: running a complex 3D mining game achieved 804 TPS, peaking at 810; generating complex website 3D animations peaked at 1426 token/s, outputting 25,624 tokens in 32 seconds; he observed no degradation in capability, with model quality consistent with the standard version. In agent scenarios, a 10x speed increase means sub-agent concurrency efficiency greatly improves, allowing users to start testing without waiting.

Sources:

OpenAI Files Confidential S-1 for IPO, Valued at $852 Billion; Anthropic Valued at $965 Billion

OpenAI announced it has submitted a confidential S-1 filing, but said the IPO timeline is not yet determined, as some operations proceed more smoothly under a non-public structure. According to @MaxForAI, OpenAI plans to launch a share repurchase offer allowing employees to sell shares at an $852 billion post-money valuation. The company has partnered with Goldman Sachs and Morgan Stanley to advance the IPO, having raised over $180 billion cumulatively, while still burning cash. Sam Altman posted that the company is entering its third phase—making advanced AI technology accessible, reducing costs, and ensuring safety. The previous week, Anthropic also completed a confidential IPO filing, surpassing OpenAI with a $965 billion valuation. @op7418 and @MaxForAI both reported on this, while @imwsl90 commented, “With OpenAI and Anthropic going public, this AI feast is probably coming to an end.”

Sources:

Google Releases Gemini 3.5 Live Translate and Major NotebookLM Upgrade

Google is simultaneously advancing its AI product line. @MaxForAI reported that Google released the Gemini 3.5 Live Translate audio model, supporting near real-time speech-to-speech translation for over 70 languages, generating output while listening, preserving tone and rhythm, lagging only seconds behind the speaker. It will enter Google Meet (private preview this month) and Google Translate mobile, supporting over 2,000 language pairs. @xiaohu detailed the major NotebookLM upgrade: the underlying model is now Gemini 3.5 + Antigravity; each notebook comes with a cloud-based computer capable of writing code for complex analysis; new output formats include PDF, Excel, PPT, and more; integrated with Google Search; the new version wins over 65% on average, with the biggest improvements in long-document analysis and web retrieval. Available to Google AI Ultra subscribers and Workspace business users.

Sources:

Anthropic Mythos Naming Anecdote: Baidu’s “Qianwen” Déjà Vu

Regarding the naming of Anthropic’s new model, @thsottiaux discovered an interesting coincidence: one of the names Anthropic considered may conflict with an existing brand. @MaxForAI recalled a similar experience with Baidu’s model naming—Baidu’s model was once suggested to be called “Qianwen” (Qwen), but was ultimately named “Wenxin” (ERNIE), while the name “Qianwen” was later adopted by Alibaba, becoming the globally renowned Qwen series of open-source models. He remarked, “History always repeats itself in surprising ways.” Additionally, @MaxForAI noted that GLM 5.2 is about to be released by Zhipu AI; Reddit users found that calling GLM 5.2 returned a “no permission” error while GLM 6 returned “model does not exist.”

Sources:

Statistics: Scan timeline lines=360, bloggers hit=41, total tweets hit=251, weighted tweet score=207.65, original tweets=133, RT tweets=42, fetch attempts=2, boundary coverage status=tail_confidently_crossed_target_boundary