
Tech • IA • Crypto
SK Hynix jumped 14% in its U.S. debut after raising $26.5 billion, the largest foreign IPO on American exchanges. Shares opened at $170 versus a $149 offer price, with demand exceeding supply by more than 7x. The listing pushed valuation to roughly $1.03 trillion, underscoring investor appetite for AI infrastructure plays. Its dominance in high-bandwidth memory (HBM) positions it at a critical bottleneck in AI system performance.
ChatGPT 5.6 introduces a Work interface designed for building real software rather than chat interactions. The system enables structured directories, local file access, and executable workflows that produce applications and automation tools. This marks a shift toward AI as a production environment rather than a conversational assistant. Token usage now directly translates into tangible outputs like codebases and systems.
AGENTS.md acts as a root-level system prompt that governs structure, memory, and workflows in ChatGPT 5.6. Without it, outputs become fragmented across files and logs, reducing reliability. Proper configuration centralizes dependencies and enforces consistent behavior across tasks. This file effectively turns the model into a programmable development environment.
Loop engineering introduces iterative cycles of observation, decision, action, and validation within AI workflows. Instead of one-shot prompts, users define conditional logic, retries, and success criteria. This allows systems to refine outputs autonomously until requirements are met. The approach significantly increases reliability for complex software and automation tasks.
ChatGPT Work is rolling out as a new desktop app, replacing the legacy ChatGPT Classic experience. The platform introduces local workspaces, persistent files, and structured project environments. However, confusion between projects and workspaces has created usability friction. Early inconsistencies suggest the product remains unfinished despite its ambitious design.
Abacus introduces an orchestration layer coordinating over 100 AI models including GPT-5.6, Fable 5, and Claude Opus 4.8. Its system dynamically routes subtasks to the most suitable model for cost and performance efficiency. A smart router assigns planning, coding, and debugging to specialized systems. This challenges the paradigm of relying on a single dominant model.
Abacus agents operate in persistent cloud environments with access to GitHub, S3, APIs, and SSH. They can install dependencies, run code, debug errors, and deploy applications continuously. This shifts AI from assistant to autonomous operator managing live infrastructure. The model enables end-to-end software lifecycle execution without human intervention.
China, Google, Amazon, Nvidia, and SpaceX are advancing plans for orbital computing infrastructure. Projections show AI data centers could reach 950 TWh annually by 2030, straining Earth-based power grids. Space offers energy and scaling advantages but introduces severe cooling and radiation challenges. A 1 MW orbital facility may require 2,500 m² of radiators, highlighting major engineering hurdles.