X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products & Apps¶
humans& releases &Tutor: a tutoring model trained against a simulated student¶
What happened: humans& released Ampersand Tutor (&Tutor) on October 30, available via API. The approach trains a simulated-user model called Persimmon to play the student, then has &Tutor teach it, with reward based on how much the student learned rather than on rewarding any particular teaching behavior. The model card shows &Tutor is post-trained on GLM 5.3 (MoE, NVFP4 weights), with the API identifier ampersand-tutor-v0.1. On Ai2's TutorMoments benchmark (key moments taken from real math tutoring transcripts), &Tutor scores 0.754 with a plain prompt, above the GLM-5.3 base at 0.500, Claude Sonnet 5.5 at 0.667 and GPT-6 Astra at 0.634; with a prompt that explicitly states the benchmark requirements it scores 0.919, slightly below GPT-6 Astra's 0.926.
Why it matters: this is the first concrete product on the "user model + reinforcement learning" path, with the goal shifted from "answering correctly" to "making the learner understand". The official post was one of the more widely shared new product launches in this window.
Key posts: @humansand original post (10-30 23:11, 58K views, 448 likes, 239 bookmarks); team member's explanation: @alexisjross (6,666 views, 123 likes).
Verification: the training method, base model and scores were verified, see humans& official blog, &Tutor model card, Ai2 TutorMoments announcement. "Best performance" only holds under the plain prompt; the evaluation is done by simulated students and an LLM judge, and the model card itself notes the results may not generalize to real students.
Assessment: the method is worth watching, but so far there is only automated evaluation, with no data on real-student learning outcomes.
Pika releases its first open-weight model, Pika H3 Flex: built on MiniMax H3, video with sound in 6 steps¶
What happened: Pika released Pika H3 Flex in the early hours of October 30 (01:37 Shanghai time) and then published the weights on Hugging Face. It consists of two LoRA adapters post-trained on MiniMax H3, one for text and first/last-frame generation and one for multi-reference input with up to 12 reference assets, generating video and stereo audio together in 6 sampling steps. The local open-weight version supports native 480p/768p; 1080p through 4K and separated audio tracks are only available through Pika's paid API, which the company says starts at $0.01 per second. DesignArena says it ranks 7th on its video leaderboard with 1293 Elo, the highest open-weight video model on the board.
Why it matters: Pika has stuck to a closed product strategy, and this is the first time it has released weights, post-trained on top of an open model from Chinese vendor MiniMax. The official post had relatively high reach among AI launches in this window; a Japanese user ran a comparison test on an RTX 4070 the same day.
Key posts: @pika_labs launch post (830K views, 1,575 likes, 1,019 bookmarks); technical notes: @pika_labs technical thread; notable shares: @DesignArena leaderboard, @HuggingApps (9,499 views, 146 likes), @sep_is_heim RTX 4070 test (5,423 views, 78 likes, personal test).
Verification: the structure, resolutions and license were verified, see the Hugging Face model page and the ComfyUI tutorial. The weights are distributed under the MiniMax H3 Community License, with regional and commercial conditions; running locally requires about 210GB of H3 base components. The leaderboard ranking comes from DesignArena's own voting board.
Assessment: "open weights" covers only the adapters and local low-resolution capability; high resolution still requires the paid API, so it reads as an "open-source funnel, paid API" model.
Other¶
- Edge0's open-weight speech model Audio8-TTS-Realistic (0.6B) spread widely on X: @SamuelZengML introduction post (10-30 03:00, 74K views, 1,475 likes, 1,995 bookmarks), which says it can clone a voice from a short audio clip, supports 23 languages, and ranks first on the Hugging Face TTS Arena. The Hugging Face model page shows about 601 million parameters and an Apache-2.0 license; but the repository was created on October 14, so it may not have been released in this window. The Arena ranking was not verified directly this time.
- Rumor, unverified: @dummerspast39 (10-31 03:59, 149K views, 1,089 likes) says OpenAI's GPT-6.2 "Bel" (internal codename betelgeuse) ships this week, and cites an anonymous "insider" for benchmark scores, parameter counts and pricing. OpenAI has said nothing officially, so none of these numbers should be treated as fact; the post only shows the rumor is hot on X.
2. AI Compute & Industry¶
Google's space TPU completes AI inference in orbit for the first time: the Suncatcher prototype satellite sends back a haiku¶
What happened: Google announced on October 30 that the Project Suncatcher prototype satellite launched on October 1, after finishing commissioning, powered up a data-center-class TPU in orbit for the first time, had the on-board open Gemma 4 model write a haiku about space on command, and transmitted it to the ground as the satellite passed over Svalbard, Norway. The satellite was built with Planet and carries 4 TPUs; TechCrunch reports that power constraints allow only 1 to run at a time, for about 20 minutes each. The next step is testing how the TPUs handle thermal stress and radiation in orbit, and the following milestone is two interconnected satellites in 2027.
Why it matters: this is the first real measurement point for the "put AI compute in space" path. Sundar Pichai's post was the most widely shared AI-related post in this window.
Key posts: @sundarpichai original post (10-31 01:05, 1.528M views, 16,743 likes, 1,619 reposts); @Google official post (95K views, 895 likes).
Verification: the launch date, Gemma 4, the haiku content and the follow-up plans were verified, see Google official blog; "only one TPU at a time, about 20 minutes" comes from TechCrunch, while another outlet wrote about 15 minutes, so the accounts differ.
Assessment: so far this only proves a chip can boot and run inference in orbit; economic viability for a "space data center" is still far off, and the truly useful part will be the coming radiation and thermal data.
Report: Nvidia in talks to provide multi-billion-dollar credit support for Nscale's Anthropic data center¶
What happened: according to The Information, relayed by Reuters, Nvidia is in talks to provide several billion dollars of credit support to UK cloud provider Nscale for building a 460 MW data center for Anthropic at the Monarch campus in West Virginia. The support mainly covers power infrastructure and construction financing rather than GPU purchases; the compute contract between Anthropic and Nscale is worth about $45 billion. No agreement has been signed and terms may change.
Why it matters: Nvidia's role is expanding from selling chips to using its own credit to help customers finance data center construction. On X this was mostly relayed by finance accounts, with low engagement per post.
Key posts: @allday_stocks news flash (10-30 06:52, 756 views); fuller relay: @Dragon21Soul.
Verification: the report's key points were verified, see Reuters copy republished by Business Standard. The "more than $5 billion" figure in the X posts is not confirmed by that article and is a relay.
Assessment: this is a reported in-talks story, not a signed agreement.
Other¶
- According to relays on X, Sam Altman said on the Cheeky Pint podcast that OpenAI's robotics project will be used to build data centers, calling robot-built, self-replicating data centers "within reach"; he also said that amid a global DRAM shortage, everyone keeping memory in their own computers is "terrible" because a PC only uses it about 5% of the time, and pooling it in data centers makes more sense. Relays on X: @UAPWatchers (4,645 views), @Chain_Duck (476 views). These are relays; the podcast itself was not checked this time.
3. Chinese AI Ecosystem¶
LatePost: Volcano Engine MaaS revenue 4.8 billion yuan in H1, gross margin above 55%¶
What happened: LatePost reported exclusively on October 30 that the MaaS business (selling model usage by volume) of ByteDance's Volcano Engine had 4.8 billion yuan of revenue in the first half of 2026 with a gross margin above 55%, making it ByteDance's largest AI business by revenue; Volcano Engine has made commercialization revenue its top metric and raised its original 15 billion yuan annual revenue target.
Key posts: @latepostnews original post (10-30 21:53, 3,014 views, 14 bookmarks).
Verification: the report exists and the key points match, see Sina Finance republication and NetEase republication. The numbers come from LatePost's sources; ByteDance has not confirmed them publicly.
Assessment: the business of selling model calls in China is starting to show decent gross margin, which differs from the "losing money in a price war" impression, and is worth reading alongside the IDC share data below.
IDC: China LLM MaaS market 10.63 billion yuan in H1, Alibaba Cloud at 46.6%¶
What happened: IDC's "China AI Software Market Semi-Annual Tracker, 2026 H1" shows the China large language model MaaS market at 10.63 billion yuan in H1, up 889.4% year on year; by revenue, Alibaba Cloud holds 46.6%, Zhipu 10.6%, Volcano Engine 10.2%, Baidu 6.7% and iFlytek 3.5%. The data was published on October 29 and relayed in Chinese circles on October 30 and 31.
Key posts: @NFT_Chen (10-31 02:45, 3,325 views), @AutonomyAdam share chart.
Verification: the market size and top-five shares were verified, see Yicai and NBD. The X posts map each company's share directly onto specific models like "Qwen 3.8" and "GLM 5.3", which is the posters' own interpretation; IDC counts by vendor.
Assessment: Volcano Engine is third by share, yet by LatePost's figures it alone booked 4.8 billion yuan in H1; the two sets use different scopes (IDC counts "large language model MaaS" revenue), so they should not be directly subtracted or compared.