OpenAI Releases 722 AI Math Manuscripts in One Go, Math Community Begins Review

Beating AI Express News: AI startup Interfaze has open-sourced its multi-model system interfaze-1-lite, which integrates capabilities including OCR, speech recognition, image segmentation, classification, and prediction into a single interface. Qwen3.8-27B serves as the scheduling core, responsible for understanding user requests and invoking specialized models to perform tasks: document recognition uses Chandra OCR and PaddleOCR, speech-to-text leverages Whisper, image segmentation relies on SAM 2.1, and time series prediction is handled by TimesFM. Developers still need to download the model weights for local deployment, but Interfaze eliminates the work of assembling the system from scratch. The startup has pre-implemented model scheduling, data conversion, and unified output, exposing only one interface to upper-layer applications. Tasks like OCR also return confidence scores and text positions, helping subsequent programs determine if manual review is required. Interfaze refers to this design as a Mixture of Architectures, and the entire system can run on a single 80GB H100 GPU.

Stacks founder Muneeb Ali has unveiled the five key priorities from his first 30 days back as CEO of Stacks Labs, including expanding institutional business development, simplifying the management structure, releasing a two-year technical roadmap, strengthening engagement with large STX holders, and further clarifying STX’s value capture mechanism. He noted that institutional demand for Bitcoin Bonds is growing, and the future roadmap will also cover institutional-grade Bitcoin (BTC) privacy and quantum-resistant technologies. Ali said Stacks Labs plans to transfer all its intellectual property to a shareholder-free Endowment.

Beating AI (Dynamic Insights) Flash News: Web special effects tool Shaders has announced open-sourcing. Designed to create GPU-powered dynamic effects including gradients, fluids, glass, metal, and distortions, it enables developers to add complex visual effects to websites without writing low-level graphics code from scratch. Shaders has open-sourced over 200 WebGPU effect components, its core rendering engine, and adapters for multiple frontend frameworks under the MIT license. Developers can use the tool for free in both personal and commercial projects. They can also fine-tune effects in its web editor before exporting production-ready code for frameworks like React, Vue, and Svelte at no cost. This suite is also highly practical for AI coding: agents such as Claude Code, Cursor, and Codex can directly search for and call these components via MCP. Previously, when agents needed to build complex web effects, they often had to generate low-level WGSL code on the fly, leading to inconsistent results and performance. Now, they can modify existing components directly, eliminating massive work of writing code and tuning performance from scratch. However, Shaders has not open-sourced its full product. The visual editor remains operated by the official team, while some advanced presets, complete web blocks, and image/video rendering capabilities are still exclusive to its paid Pro version.

Beating AI News Flash: Customer service agent firm Sierra has partnered with Meta to launch the Personal Agent Protocol (PAP), with companies including Walmart, Shopify, Stripe, Genesys, and Instinct co-developing the standard. PAP is designed for personal AI agents, enabling enterprises to verify an agent’s identity, the user it represents, and the actions the user has authorized the agent to take. Currently, many personal agents still simulate human web interactions—opening sites, clicking buttons, filling forms—making them vulnerable to blocks from captchas and anti-bot systems. For instance, Amazon previously banned Meta Muse from accessing its shopping platform; even Walmart, a Muse partner, deploys human verification to block agents. PAP seeks to standardize this identity and authorization framework: when a user asks an agent to check an order, only read access is granted, while separate authorization is required for modifying or placing orders. Enterprises can also define the operations agents are allowed to perform, as well as integration methods (web, API, or their own agent connections). While MCP addresses how agents call external tools and data, PAP focuses on agent identity, user authorization, and enterprise permissions. PAP remains in preview, with its first v0.1 specification set to be released later this month. On the same day, Sierra’s competitor Decagon open-sourced a similar PACT protocol and announced its entry into the PAP working group. PACT also handles agent identity and user authorization, using A2A to connect personal and enterprise agents, then completing user authorization via OAuth.

Former Messari analyst Mason Nystrom announced on X that he has officially joined Multicoin Capital as an investment partner. Prior to this, Nystrom held roles at Pantera Capital and Variant Fund. He noted that leveraging his experience at ConsenSys and Messari, as well as his in-depth research on stablecoins, tokenization, and consumer cryptocurrencies, he will closely focus on venture investments in stablecoins, consumer applications, and financial infrastructure.

Beating AI News (from Dongcha) – Figure AI founder Brett Adcock’s personal AI assistant Hark Pro is now fully available, launching on web, iOS, and Android, with free access open to all. The free version of Hark Pro has no feature restrictions, only usage limits. The $20/month Pro2 plan offers double the usage, while the $100/month Pro3 plan provides 10x the usage. All three plans have identical functionality; upgrades are solely for higher usage quotas. Hark can retain user information long-term and uses its own cloud computer to operate websites on behalf of users. It supports running up to six browsers simultaneously, handling tasks like food ordering, shopping, emails, and bill payments, and integrates with Google Workspace and Microsoft 365. Hark also includes long-term project conversations, on-demand small app generation, and action cards that proactively identify to-dos and process them with one click. For instance, after detecting bills or unused subscriptions, users can directly ask Hark to make payments or cancel them. Since Hark stores website passwords and credit card details, the company has built a dedicated encrypted vault named "Secured by Hark". The official notes that passwords and card information remain encrypted, and even Hark itself cannot access them directly.