🤖 The Blockbrain Brief #2

High-Level Summary

• Agentic Browsers: OpenAI launched ChatGPT Atlas, a Chromium desktop browser with an optional Agent Mode that can act inside webpages; early hands-on from our team found it promising but not yet stable or fast enough for full-day workflows.
• Enterprise LLM Building Blocks: Anthropic shipped Agent Skills (folders that include instructions, scripts, and resources that Claude can load when needed) and Claude Code on the web (research preview with repo hookups, parallel tasks, and sandboxed execution), pushing Claude toward workflow-level assistants.
• Realtime Video, Open Source: Krea Realtime 14B is an open source video generation model, it’s a strong fit for real-time, steer-as-it-generates use cases where teams need control
• Long-Context Efficiency: DeepSeek OCR is an open source model that compresses document tokens by ~7–20×, aiming to cut cost/latency for structured docs with minimal accuracy loss.
• Commerce Integrations: Lovable to Shopify turns prompts into a live shop making it a one-flow handoff, signaling continued movement of agentic tools into mainstream e-commerce stacks.

AI Product Updates

OpenAI launches ChatGPT Atlas (OpenAI)

• OpenAI introduced ChatGPT Atlas, a Chromium-based desktop browser with ChatGPT built in; an optional Agent Mode (preview for Plus/Pro/Business) can perform actions inside webpages, shifting the AI assistance from pure Q&A style to in-tab task execution.
• Atlas ships with clear privacy controls: Browser Memories are opt-in, per-site visibility can be toggled from the address bar, Incognito signs you out of ChatGPT, and page content isn’t used for training unless you choose to opt in.
• In early hands-on tests by Blockbrain’s MP and Alexander, Atlas felt not yet ready for full-day productivity as it is slow and occasionally unstable”, with agent flows struggling on transactional tasks (e.g., flight booking, dinner reservations, car searches), suggesting limited practical value today despite strong potential.

Anthropic ships multiple updates (Agent Skills, Claude Code)

• Anthropic introduced Agent Skills, modular folders of instructions, scripts, and resources that Claude loads only when relevant, so you can package org-specific know-how (e.g., Excel workflows, brand rules) and reuse it across Claude apps, Claude Code, and the API; Skills are composable, portable, and can execute code within Anthropic’s secured Code Execution environment.
• How Agent Skills work: You can create Skills, which are sets of instructions or files, and toggle them on or off; when you use Claude, it automatically pulls in the relevant Skills for the task.

• Claude Code now runs in the browser (research preview), letting you connect GitHub repos, delegate multiple coding tasks in parallel, watch real-time progress, and steer changes as it works; each task runs in an isolated, network-restricted sandbox, and a mobile preview is available on iOS

Krea Realtime 14B (Krea)

• Krea launched Realtime 14B, an open sourced video-generation model that can start showing the first frames about a second after you hit “generate.” It’s designed to keep producing video smoothly (around 11 frames per second on a single high-end NVIDIA B200 GPU). It’s also much larger than most other open sourced real-time video models which struggles to model complex movements and tend to have poor high frequency details.
• Krea Realtime 14B is the better fit when you need open, real-time “steer-as-it-generates” video and tight control over deployment, while Sora 2 remains the leader for photorealism, physics awareness, audio sync, and multi-shot coherence.

DeepSeek “text-as-images” compression (South China Morning Post)

• DeepSeek released an open-source multimodal model, DeepSeek OCR, that compresses long text by turning it into images, cutting tokens by about 7× to 20× so large documents cost less to process. The system pairs a DeepEncoder vision module with a 3B MoE decoder and is available on Hugging Face and GitHub for developers to try.
• It aims to make ultra-long context cheaper and faster for LLMs without losing much accuracy on structured docs like forms and reports.

Enterprise Integrations

Lovable & Shopify integration (Lovable)

• Lovable now lets you build a storefront by chatting with AI and then push the generated app directly to Shopify, turning a prompt into a live shop in one flow.