Oct 2 · AI Daily

Musk announced on X that Tesla will reduce the memory allocation of the AI5 and AI6 chips to allocate enough memory for the mass production of Optimus robots. The RAM of AI5 will drop from the originally planned 144GB to 72GB LP5 (a 50% reduction), and AI6 will drop from 216GB to 144GB LP6 (a reduction of about 33%). He stated that the memory bandwidth remains unchanged, and the impact on Optimus performance is expected to be negligible.

The Claude Code team introduced Mods, which allows customizing Claude's appearance and working methods through prompts, and can be shared as plugins.

Grok Build launched a new Agent Dashboard, allowing users to view all agents on the same screen via /dashboard. Author Andrew Milich added that the view supports managing multiple agents and worktrees simultaneously within a single /dashboard interface.

Now live on Grok Build: the brand-new Agent Dashboard. All agents on a single screen.

Try it with /dashboard

a16z analyzed a new shopping workflow: users directly tell the AI their budget, arrival time, and dietary restrictions, and the AI selects a restaurant and places the order, with Toast handling order payment and DoorDash responsible for delivery. Users no longer open food delivery apps, nor do they look at recommendation slots and ads; the orders remain, but the customer relationship is taken away by the AI. a16z believes that legacy platforms will continue to provide merchant networks, payment, fulfillment, and other infrastructure, while the traffic and profit entry points that decide what users buy and from whom may shift to AI.

This article compiles the first-hand experiences of several web development education providers, pointing out that generative AI is destroying their livelihoods and industry ecosystem. Baldur Bjarnason said training programs are disappearing one after another and peers are shutting down their businesses.

https://suno.com/s/nLnF9DZvqyIFxi2u Thanks, we really tried our best

Chamath bluntly stated at Stanford AI Club that long-horizon tasks "simply don't work," complex problems are still not well solved, and refused to use benchmarks or 48-hour scripts as counter-arguments. He warned that if AI fails to enter highly complex environments, the industry will fall into a "trough of disillusionment," while hundreds of billions or even trillions of dollars have already been invested. The direction he offered is: a symbolic space is needed to guide the embedding space.

Google's in-orbit computing prototype satellite was launched aboard a SpaceX rocket on the 1st local time to verify whether Tensor processors can operate properly in space, marking the first time Google has sent advanced chips into space. The satellite was manufactured by Planet Labs, and after commissioning, the processor runs for 15 minutes at a time. Google envisions forming an in-orbit data center consisting of a constellation of 81 satellites in the future.

OpenAI launched two new shopping features globally in ChatGPT on the 1st local time: virtual try-on for clothing and accessories, and a save feature to keep items alongside try-on images. The new features are based on the ChatGPT Images 2.5 model, which reportedly features more natural lighting, richer textures, more stable editing instructions, and faster generation. Users can upload selfies or full-body photos to see try-on effects, describe a clothing style for ChatGPT to pick individual pieces, or upload celebrity outfit photos to find purchasable matching items.

Black Forest Labs announced the release of its new image model, FLUX 3 Image, focusing on precise editing and supporting 4K. Users can already try the new features in Tools Playground, with open weights planned for release in the coming weeks.

According to the Wall Street Journal, OpenAI announced on the 1st local time that it terminated cooperation with three researchers because they bypassed established procedures and improperly handled sensitive information. The individuals involved are Jasmine Wang, Tomek Korbak, and Mikita Bolesny, with Korbak belonging to the safety team and the other two working on model alignment. The three were accused of providing company confidential information to a third-party AI safety organization. OpenAI also stated that it supports independent evaluations of safety arguments and provides evaluators with deep access to training, evaluation, and deployment stages.

Microsoft released its first real-time streaming speech transcription model, MAI-Transcribe-2-Streaming, on October 1. It supports 60 languages and automatic language detection, and can continuously output text while speech is underway.

Artificial Analysis locally deployed and evaluated Alibaba's Qwen-Image-2.1, released with open weights on September 20. The model ranked 18th on both AA-Image-T2I v2.0 and AA-Image-Editing v2.0, making it the top-ranking open-weight model on both leaderboards, surpassing Ideogram 4.0 (Quality) and HunyuanImage 3.0 Instruct.

Ben Affleck (Hollywood star and Artists Equity CEO) introduced his method for fine-tuning open video models: unfreezing weights and training only the final cinematic layers to help film crews achieve true production standards.

GLM 5.3 and GLM 5.3 Flash are now available on Cursor!

GLM 5.3 Max is the highest-scoring open-weight model on CursorBench 4.0.

MIT CSAIL, in collaboration with Google and Northeastern University, launched InstructMesh, combining Microsoft TRELLIS's 3D generation capabilities with GPT-4's reasoning abilities. Users can use natural language to annotate and modify AI-generated 3D designs before printing.

Suno announced that the Speech feature has entered Beta, which officials claim is the first model capable of generating speech accompanied by matching background music. Users need to update the app to get the latest version; details can be found at https://suno.com/blog/introducing-speech-beta.

"Today, we are releasing two decision models trained by Cloudflare, Clef and Clef-flash" https://blog.cloudflare.com/clef-decision-models/

Never seen an idea spread so fast.

Epoch AI, in partnership with YouGov, launched the ChatGPT usage explorer, publishing chat metadata from a sample of 5,000 US YouGov users. It covers approximately 660,000 conversations and 8.3 million messages, with some records tracing back to the weeks following its release in November 2022.