🧠 DeepSeek Builds Software for Huawei Chips

The companies open-sourced tools for Huawei Ascend, now the main Chinese AI accelerators used by local startups. The software gap is the hard part: the industry has spent 10+ years writing for Nvidia CUDA.

The release includes DeepGEMM for matrix multiplication and DeepEP for chip-to-chip communication. Their APIs match Nvidia versions, so existing code needs almost no rewriting.

The stack runs on TileLang. Its compiler now handles scheduling and data synchronization instead of requiring manual code. DeepSeek says its training kernels already run on Ascend. This covers kernel writing, not the full CUDA stack.

📊@tech

🍏 Apple May Break Its September Cycle

According to Bloomberg, new CEO John Ternus wants Apple to release products more often than at its major autumn and spring presentations. He says faster launches and more experiments are needed to stay competitive in the AI era.

Ternus is also considering fewer management layers between engineers and top executives. Over the past 2 weeks, the hardware development team has shed some engineering program managers. Bloomberg reported earlier cuts in the Siri and Vision Pro teams in August.

The rumored pipeline includes a new HomePod mini and iPad mini. Apple is also reportedly working on a new Apple TV 4K, camera-equipped AirPods, touchscreen MacBooks, plus a special iPhone for its 20th anniversary.

📊@tech

🤖 Anthropic Adds Robots to the AI Risk Map

Anthropic estimates that robots can already perform 74% of physical work tasks in the US under at least some conditions. With LLMs added, the figure reaches 81% of all work.

The most exposed roles are mostly vehicle operators. Waste sorting is another example. Waymo robotaxis and autonomous tractors show how the shift works. Sorting robots target recycling work.

With LLMs alone, transport and moving work had under 15% exposure. Robots raise it to about 90%. But exposure means technology can perform part of a job. Robots are currently cost-competitive with humans in only 0.3% of tasks.

📊@tech

📹 HeyGen launches its first video model

The model generates clips from text, animates an image, or uses another video as a reference for motion and visual style. Audio is created with the video, including speech, background sound, and effects.

It runs on MiniMax H3, with additional training by HeyGen. Clips run from 5 to 15 seconds in 480p or 768p, with aspect ratios from 21:9 to 9:16.

HeyGen’s blind comparison put the model first against 4 others, based on 4,800 votes. 768p generation costs $0.03 per second, with a 50% discount through the end of October. It is available through API, OpenRouter, Runware, and ComfyUI. The web version will come later.

📊@tech

🍟 McDonald’s AI Suggests Local Big Mac Prices

In Fresno, a Big Mac costs $5.69 at one McDonald’s and $6.89 at another 2 miles away. That is a 21% gap, linked by Reuters to the chain’s AI pricing system.

The system seeks an “optimal price” for each US restaurant and some overseas markets, based on local customers’ willingness to pay. Franchisees see prompts about price sensitivity. McDonald’s tracks deviations and sends suggestions at least 3 times a year.

Owners formally set the price, but some told Reuters that deviations can trigger calls from headquarters. New franchise standards require them to “constructively engage” with approved pricing tools.

📊@tech

🤖 ChatGPT Gets Virtual Try-On

OpenAI added virtual try-on to ChatGPT. When a clothing product card appears, tap “Try on” and upload a selfie plus a full-body photo. ChatGPT Images 2.5 generates the result.

The feature can also use clothing screenshots, not just items suggested by the bot. Favorite pieces can be saved to Library. Google added the same feature to Search in 2025.

The catch is privacy. By default, images uploaded from a personal account may be used to train models unless you opt out. That includes full-body photos.

📊@tech

🍏 Apple’s Camera Will Not Record Video

Apple is reportedly developing J450, a small cylindrical home security camera with a very low frame rate. AI analyzes what its sensor sees and produces only text descriptions, such as someone entering a room. Face recognition is also expected.

J450 is designed as a companion to Apple’s J490 smart home display. It is part of a broader smart home ecosystem that may include Apple’s own hub, plus security devices from Apple and third parties. Its technology is reportedly similar to what Apple is developing for camera-equipped AirPods.

There will be no video archive to watch later. No video leaves less to leak, while also leaving nothing to review when something happens at home.

📊@tech

🤖 Karpathy Wants AI to Stop Writing Text

Karpathy suggests asking AI to explain complex topics in ASD-STE100 simplified English, with short sentences and little ambiguity. For harder subjects, a diagram or a custom HTML page can work better than plain text.

Modern models can already build one-off interfaces for a specific question. Karpathy is especially interested in personal teaching videos, with animation, a 3Blue1Brown-style explanation, and voice from ElevenLabs or local TTS.

As code and AI work become cheaper, temporary interactive artifacts become practical. The next AI interface may be an answer assembled for each task, rather than an endless chat with walls of text.

📊@tech

🔗 AI Progress Could Compress Years

A Cambridge paper by 22 authors, including Geoffrey Hinton, Yoshua Bengio, OpenAI’s Jakub Pachocki, and Anthropic’s Jack Clark, examines compressing years of AI progress into months.

At Anthropic, model-written code rose from single digits to over 80% since early 2025. Autonomous R&D tasks rose from 1% to 26% in 6 months. Full automation could make progress 10x faster in about 1.5 years.

Compute or data could limit the scenario. Risks include lost control and a winner gaining too much advantage. The authors propose reporting R&D automation, external audits before deployment, and ways to pause specific data-center workloads.

📊@tech

🙂‍↔️ DeepSeek Brings Harness to Desktop

Harness is available for macOS and Windows as an open-source alternative to Claude Code and Codex.

DeepSeek models act as autonomous agents. They write and test code, work with files and the terminal, then handle long tasks. Plugins allow customization.

📊@tech

🤖 GPT-6.1 Astra May Arrive Next Week

Thibault and other OpenAI developers are hinting at the release.

GPT-6.1 Astra, or another strong model, was supposed to launch at DevDay but was postponed “for safety reasons.”

📊@tech