Morning digest: GPT-6 Astra, Unreal, and AI Mode ads

Morning digest for Thursday, Oct 29. DeepSeek and Moonshot AI leaned heavily on answers from OpenAI's GPT-6 Astra to train their newest models, The Information reports. It's the latest round in the fight over "distillation" between U.S. and Chinese AI labs. Also: Epic says AI tools like Claude Code and Codex now write the bulk of its new code for Unreal Engine 6; Google is bringing ads in AI Mode to 10 more countries; Samsung says its 2027 supply of AI memory is already sold out; Cognition, the company behind the Devin coding agent, is weighing offers at up to $80 billion; U.S. computer science enrollment fell again; SenseTime released an open model that makes videos by writing code; VRChat decided against "AI-generated" labels; and Amazon launched new Echo devices built around Alexa+.

1. DeepSeek and Moonshot reportedly trained their newest models on GPT-6 Astra's answers. Distillation means training your own model on the outputs of a stronger one, so it picks up those skills faster and more cheaply. The Information reported Wednesday, citing people with direct knowledge, that Chinese labs DeepSeek and Moonshot AI "relied heavily on outputs" from OpenAI's GPT-6 Astra, according to Reuters. The models named are Moonshot's Kimi K3.1 Pro, released last week, and DeepSeek's V4.1-Flash 1017, launched Oct. 17. Those are the same two models that topped the open-model rankings on the Artificial Analysis leaderboard we covered yesterday. OpenAI, DeepSeek and Moonshot didn't immediately respond to Reuters. This isn't the first time. In a Sept. 30 blog post, OpenAI said it had linked a core part of a distillation campaign that began in July to people associated with Moonshot, Reuters notes. And earlier this month, Anthropic said new defenses cut the data harvested from Claude in these attacks by about 90% to 99%. These are reported claims from unnamed sources, and none of the companies has confirmed them. What to watch: Whether OpenAI answers with its own crackdown like Anthropic's, and whether DeepSeek or Moonshot responds. If access gets tighter, the gap between cheap open models and the frontier could widen again.

2. Epic says AI tools now write most of the new code for Unreal Engine 6. After CEO Tim Sweeney confirmed on X that Epic is "now using AI tools to develop code for Unreal Engine," Epic's executive vice president of development, Marcus Wassmer, gave This Week in Videogames the details. "Yes, we are vibe-coding things," he said. "That means our engineers are no longer writing the bulk of the code, but they are still guiding the tools to get the architectural result we want." Among the tools, he said, "probably the most obvious ones are Claude Code and Codex." He listed some wins: "almost the complete elimination of non-deterministic cooking bugs which we've been fighting for a decade" (cooking is the step that packages game assets, and those bugs bloat patches), faster shader compression, and a rewrite of a big part of the D3D12 rendering layer, the code that talks to DirectX 12 on Windows. He had estimated that job at three to six weeks for one engineer, and it took about a week. Almost all of Epic's engine work now happens in its development branches for Fortnite, UEFN (Unreal Editor for Fortnite) and Unreal Engine 6, so Unreal Engine 5.9 may be only a "relatively modest" update. "We are more concerned about the risks of moving too slowly than we are about AI skepticism," Wassmer said. The statement came out Tuesday and spread widely Wednesday, including in PC Gamer. Epic hasn't shared benchmarks or outside checks of these results. What this means for you: If you build in Unreal, Wassmer says the engine should become "significantly more modular," so you can strip out what your game doesn't need. Those are big refactors that he says were too large to do by hand before.

3. Google is bringing ads in AI Mode to 10 more countries, and adding ads to follow-up questions in the U.S. AI Mode is Google Search's chatbot-style tab. Google said Wednesday it is scaling ads there in Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore, with Brazil, Mexico and the U.K. planned later this quarter. In the U.S., it's launching "multiturn" ads, which can show up in your follow-up questions, not just the first one. Ads will be labeled "Sponsored." Google says people click sponsored and regular product listings in AI Mode at similar rates, but it didn't share the numbers. The timing matters: Search revenue grew 12% last quarter, down from 17% the quarter before, TechCrunch notes, and chief business officer Philipp Schindler told investors this takes "our monetization footprint global for the first time." When you'll feel it: Now, if you use AI Mode in one of those 10 countries. In the U.S., expect more sponsored results the longer a conversation goes.

4. Samsung says its AI memory for 2027 is already sold out. High-bandwidth memory (HBM) is the stacked, very fast memory that sits next to AI chips like Nvidia's. Samsung reported record results on Thursday (Korea time): 107.4 trillion won in quarterly operating profit, which it calls a first for any tech company. On the earnings call, memory executive Jaejune Kim said its "2027 HBM supply volume is already sold out at an agreed target price," and that key customers still want more, according to an Investing.com transcript. Revenue from HBM4, its newest version, more than tripled from the last quarter. Samsung expects to spend about 84.4 trillion won (about $63 billion, per Bloomberg) this year, mostly on chip plants for AI demand. What to watch: Samsung expects chip supply to stay tight until 2028, TokenPost reports, and plans to raise memory spending again in 2027.

5. Cognition, maker of the Devin coding agent, is weighing offers that value it at $70 billion to $80 billion. That's up from the $48 billion valuation it announced just weeks ago. Business Insider reported Wednesday, citing sources, that Cognition has received offers in that range and is "seriously considering" a round before the end of the year. The talks are early, and the price could change. What to weigh: These are offers, not a closed deal. But they show how much investors are betting that AI coding agents will take on more of what engineers do today.

6. U.S. computer science enrollment dropped again this fall. Preliminary National Student Clearinghouse data show enrollment in computer and information science programs down 12.4% at four-year colleges, 12.3% at two-year colleges and 10.4% in graduate programs. Health majors grew. Over two years, undergraduate computer science programs have lost 18.4% of their students and graduate programs 27.1%, Bloomberg reports. Bloomberg ties the drop to AI making entry-level coding jobs harder to find. What to weigh: The data is early. It covers about 46% of colleges, final numbers come in January, and it shows the trend but not the cause.

7. SenseTime released an open model that makes videos by writing code instead of pixels. SenseNova 6.8 Code2Video is a 27-billion-parameter model fine-tuned from Alibaba's Qwen3.8-27B, released under the Apache 2.0 license, which allows commercial use. Instead of generating frames directly, it writes code that draws each frame, renders it, checks the result, and fixes it. Because everything is code, you can ask for one change, like new timing or colors, without regenerating the whole clip. It's meant for motion graphics, explainers, stylized 2D and 3D scenes and procedural audio. Its GitHub page says it "does not generate photorealistic footage." SenseTime runs it inside Claude Code, with runs of "more than 10 hours and more than 2,000 turns," and admits it is "clearly behind the strongest proprietary models" on polish. There are no benchmarks, only SenseTime's own examples. What to weigh: It's free to download, but SenseTime's setup examples split it across four GPUs, and one film takes hours. For most people, the bigger takeaway is the method: video you can edit like a document.

8. VRChat won't add "AI-generated" labels, even though 95% of surveyed players wanted some kind of label. VRChat explained its decision Wednesday after 21,595 people filled out its survey. Most wanted transparency, but their answers produced 3,141 different labeling combinations, and none was picked by more than 10% of people. VRChat worries a label would invite people to "sniff out" AI content and harass creators, including ones who don't use AI at all. It also won't add filters to hide AI content. Instead, misleading world thumbnails are now against the rules, likes and dislikes for worlds are coming by the end of 2026 or early 2027, and an optional "Human-authored" mark and a "Tools used" field are being considered but not scheduled. What to watch: VRChat says it will share more on the "Human-authored" mark before the end of the first quarter of 2027.

9. Amazon launched five new Echo devices built around a smarter Alexa+. Amazon's lineup includes a new Echo Dot ($99.99, shipping now), the Echo Studio with Home Camera ($249.99), the Echo Time bedside clock ($199.99), an Alexa Photo Frame, and a redesigned Echo Show 16 (preorders from $499.99). Amazon says the latest Alexa "reasons more deeply, asks clarifying questions," and can handle multi-step requests. A new Cards view saves your conversations, so you can pick them up on another Echo, in the app or on Alexa.com. The camera on the Echo Studio lets Alexa see the room, for example to explain the rules on a board game box you hold up. Amazon devices chief Panos Panay told The Verge they "opened up Alexa Plus' brain a little bit more," including "deeper research." When you'll feel it: Some features require Alexa+, which is included with Prime. Most of the devices are available now.