Oct 5 · AI Daily

SCM (Screen Memories) has been released, allowing local AI search frame by frame through photos and videos in any folder on macOS, with no account, no cloud, and no uploads.

You can build custom vertical AI applications within Computer, such as guessing the spatial location of an image using 3D and satellite views.

Over the next 28 days, every single day we will either ship a feature that is clearly improved and relevant for the majority of Codex/working users, or do a hard reset. Let the improvements begin.

TechCrunch's Equity podcast discussed Trump signing an executive order requiring official use of the term super intelligence, alongside multiple AI executives meeting him this week and signing the Joint Commitment on Frontier Responsibilities.

When you're building your own company, the last thing you need is one more thing to worry about. Put your energy where it's needed most: your creativity. Leave the rest to us.

@SRNA_LI empowers herself and her business with Cohere.

Yes

(Note: The primary post is simply "Yes," serving as a pure exclamation/response with extremely low information density. It quotes @VulcanBench stating "Grok 4.7 writes great code," meaning Grok 4.7 has excellent coding capability. The main post is an affirmative response to that assessment.)

A practical guide on hiring your first @Grok @Bot employee.

It really is like hiring an amazing helper that learns what you need and gets smarter almost every day!

The U.S. Department of Justice has charged Greg Lui, 38, CEO of Daidi Intelligent Manufacturing Computer Company, with smuggling servers equipped with export-controlled Nvidia chips into China, with a cargo value exceeding $300 million. The indictment alleges that he transshipped A100, H100, and GeForce RTX 4090 and RTX 5090 chips through freight forwarders in Malaysia, Singapore, and other locations, using falsified documents and others' identities to conceal the transactions, running from October 2023 to August 2026.

Semiconductor analysis firm Kurnal Insights published microphotographs of the Huawei Kirin 9050 Pro die, showing for the first time at the physical level that it is vertically stacked using two 11.13 × 10.84 mm dies, with a single-die area of approximately 120.65 mm², slightly smaller than the previous-generation Kirin 9030S at 122 mm².

Pavan Davuluri, head of Windows and devices at Microsoft, stated on the Inside Windows podcast that due to rising memory costs, lowering Windows 11 memory usage has become a major focus for the remainder of 2026. The team is improving memory utilization across multiple layers, including the memory manager, memory compression mechanisms, WinUI 3, Chromium, and WebView2 frameworks.

Recently, Alibaba T-Head announced the Zhenwu V900 at the 2026 Computing Conference. The Zhenwu V900 features 216 GB of memory capacity and 1,200 GB/s of interconnect bandwidth. The performance of the V900 is expected to reach "3x" that of the Zhenwu M890 and will begin shipping in Q1 2027. (1/3)🧵

Trump announced on Truth Social the formation of the "Super Intelligence Force" (SIF), responsible for coordinating federal government engagement with consumers, public interest groups, religious organizations, critical infrastructure operators, and AI companies, reporting directly to the President and Chief of Staff Susie Wiles.

Anthropic has begun inviting more users through its Anthropic Interviewer to participate in a 15-minute voice interview.

"Excited, worried, or somewhere in between? We want to hear how AI is fitting into your life and what you hope for the future."

Afterwards, users can decide whether to make the interview content public.

This seems like an interesting way to ensure people have a chance to express their views on AI. It looks like it will expand to a broader audience soon.

MarkTechPost compiled and compared four frontier models released within 30 days: Claude Fable 5.1 (Sept 1), GPT-6 Astra (Sept 3), GPT-6.1 Sol (Sept 29), and Gemini 4 Argon (Sept 30).

The open-source tool RemoveMacAI can disable Apple Intelligence on macOS 27 and delete its downloaded models. macOS 27 no longer provides a single toggle switch, leaving models on disk after the feature is turned off; the tool applies Apple's restriction keys via configuration profiles, deletes models through Apple's asset service, and redirects model downloads to a closed local port to block re-downloads. All changes can be reverted using removemacai revert.

Google paused its Open Source Software Vulnerability Rewards Program (OSS VRP) starting October 1, citing a significant increase in automated AI submissions, the vast majority of which are invalid. The company promised an update in the first quarter of 2027, during which participants are advised to use its other bug bounty programs. According to Tom's Hardware, engineers and open-source maintainers have been swamped with invalid or model-hallucinated reports.

Major release from the Manus team.

Thrilled to see H3 Max—our post-trained video model—integrated into Manus 2.0.

Elvis Saravia responded to Karpathy's post on understanding LLM outputs, stating he has practiced similar methods for over a year and sharing the universal human-AI collaboration interface he is building. His current approach connects advanced agents to Notion-like flexible pages that support embedded artifacts, text, video, and visual walkthroughs. Agents output these kinds of artifacts per task instead of default text replies, and he collaborates directly via comments and suggestions, covering use cases such as code review, note-taking, writing, research, and prototyping.

Surpassed 300,000 followers today!

Never imagined this day would come when I started tweeting in 2024. (I thought you needed Andrej Karpathy's level of fame to grow this much)

Looking back, the journey has been fascinating:

• Shared side projects like running nanoGPT & llm.c, and Muon
• Brought many users to my previous startup
• Met many amazing people offline here

I basically try to consistently post about 1 tweet a day, writing whatever comes to mind.

X is the best place for AI and tech.

Will keep chatting!

Product designer Mete wrote an analysis on why Meta's Muse topped the US App Store, arguing that its core components have existed in products like Codex and Claude Code for over a year. Meta's success lies in using a single primary chat interface as the mental model and abstracting away complex options like model selectors.