Oct 7 · AI Daily

Cursor has rolled out remote control for local agents in its iOS app, letting users view and reply to agents running on their computers. The feature is enabled by default for all users except enterprise organizations. Once logged in, the user's computer appears automatically, and pairing can be approved on the desktop app. The agent still runs locally, so the computer must stay on and connected to the internet. An option to prevent sleep can be turned on in settings, and enterprise administrators can enable this feature in Org settings.

OpenAI has released Openai/math. The author believes this suggests that mathematical problems, or any problem that can be verified, have been solved. The author also notes that OpenAI exhausted all Pro tokens, strayed from the main task, and launched countless subagents to run meaningless tests, ultimately concluding that everything was done "just to verify."

Anthropic is expanding its Cyber Verification Program and integrating Project Glasswing. Vetted security professionals can now access Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 across three access tiers.

Melius, founded by former Ramp engineers, has raised a total of $25 million, including a $20 million Series A led by CRV and a $5 million seed round led by General Catalyst. The company originally built AI performance marketing tools, but after six months scrapped its entire codebase to pivot to a "creative work agent lab" platform that generates ad creative assets and marketing campaigns. Within two months of exiting stealth in July, its annualized revenue surpassed $1 million.

OpenAI has announced a new Meetings plugin for ChatGPT that automatically records meeting notes and saves summaries to a ChatGPT Space. The plugin is currently in beta on the macOS ChatGPT desktop app for Pro and Business users, with the Enterprise version coming soon. Additional details noted that users can add notes or ask questions during meetings, tag documents, files, and attendees from connected apps, and send action items to ChatGPT as prompts.

PyTorch has released the FBTriton kernel design for Table Batched Embedding (TBE) forward and backward calculations, outperforming original CUDA kernels on recommendation system workloads.

The BestBlogs morning report covers three pieces: Amin Vahdat, head of AI infrastructure at Google, discusses the physical and economic constraints of frontier AI, arguing for Goodput to measure system useful output. Google doubles its useful token serving capacity roughly every six months rather than its chip count.

Every robot starts with a board you buy. The good ones end up on a board you build.

It is that time again when every AI account pretends to understand number theory and opines on the importance of the latest mathematical breakthrough.

The US Department of Justice instructed internal employees on Tuesday to use "superintelligence" and the abbreviation SI to replace AI in external communications, policy documents, official records, and appropriate court filings "to the maximum extent permitted by law." The directive was signed by Principal Deputy Associate Attorney General Trent McCotter, implementing an executive order signed by Trump the previous week. Reports indicate that Trump is disinclined to issue new regulatory rules, but stated that the Department of Justice will apply constraints if necessary.

According to the Financial Times, SpaceX plans to raise $40 billion to purchase NVIDIA chips, led by Apollo Global Management. The funding will be raised through about $10 billion in bank loans and $30 billion in investment-grade bonds, with the deal expected to close in 2027. Elon Musk previously stated on an earnings call that construction will be based entirely on the NVIDIA Vera Rubin architecture. MarketAxess data shows that the yield on SpaceX bonds maturing in 2056 has risen to about 2.27 percentage points above US Treasurys.

OpenAI has published a batch of documents containing 722 manuscripts and 372 result families, demonstrating solutions to hundreds of unsolved math problems by an unreleased frontier model. The findings were released in a GitHub repository. The advisory panel AGMAI called for publishing through academic channels and disclosing the model name, prompts, and compute costs, noting that the compute required for an ordinary result is roughly equivalent to ChatGPT Pro thinking continuously for three hours.

The October 7 BestBlogs morning report summarizes 10 AI items: Google introduced EmbeddingGemma 2, a multimodal embedding model with fewer than 1 billion parameters that supports text, code, images, video, and audio, utilizing a modular architecture and Matryoshka representation.

The Document Foundation stated in a blog post that LibreOffice will not feature built-in AI capabilities for the foreseeable future. This is a deliberate design choice intended to ensure documents are not uploaded to servers for processing and that software execution requires no network connectivity at any stage. Users with needs can install extensions to connect to local AI models. The organization stated that no integration option currently meets conditions for keeping data local while avoiding single-vendor dependencies, and it will continue to track technological developments.

Jump Trading used GPT-6 Astra to delegate longer, more ambiguous quantitative research tasks to AI agents. Lucas Baker, head of LLM R&D, said the model unlocked a new level of autonomy for long-cycle tasks that previously required frequent manual intervention.

OpenAI announced the release of a batch of new mathematical results produced by its internal frontier model, with materials made open on GitHub. The release followed the recommendations and public suggestions of the Institute for Advanced Study's Independent Advisory Group on Math and AI.

The author analyzes Claude Code's feature of pre-filling suggested next messages after tasks conclude, suggesting it may serve as a low-cost channel to collect human preference data rather than merely saving input time. Sending the user prompt as-is forms a positive label, while edits create preference pairs with diffs from real, in-distribution codebases that can serve as RLHF material. Predicting the next round of user messages is also a valuable training target. The author notes this is speculation, and whether it is used this way depends on user privacy settings.

OpenAI has released a batch of new mathematical results produced by its internal frontier model via a GitHub repository, accompanied by paper revisions and citation protocols. Many of the proofs have been formalized using Lean for computer verification.

Google and MIT released a report introducing Coco, an agent platform for TPU architects to use in hardware-model co-design. Coco logs every simulation scan into a normalized relational database, meaning every number reported by agents comes from an SQL query, allowing them to autonomously combine typed tools and perform comparative analysis. The near-term goal is to cut the time architects spend building simulations and gaining insights in half.

Now this is really taking it to another level.

"China already has spies inside OpenAI and Anthropic. China has the capability to launch cyberattacks at any time. Even if we lead, it doesn't give us much advantage; they can step in at any time and steal our progress."

— Jacob Coxon, former OpenAI and Anthropic researcher

Full video on the "PBD Podcast" YouTube channel (link in comments).