Tech • IA • Crypto
GPT-5.5 Launch, OpenAI Codex Agents, Palantir Manifesto Fallout
Briefing
GPT-5.5 debuts with autonomous gains
OpenAI launched GPT-5.5 on April 23, positioning it as a shift toward sustained autonomous task execution. The model emphasizes real-world usefulness over raw benchmark hype, focusing on long-running workflows. It matches prior latency despite increased scale, signaling major infrastructure gains. The release intensifies competition with Claude Opus 4.7 and Gemini 3.1 Pro.
GPT-5.5 posts benchmark dominance
GPT-5.5 leads across multiple evaluations, including Terminal Bench 2.0 (82.7%) and GDP Val (84.9%). It also edges rivals in OSWorld Verified (78.7%), narrowly beating Claude Opus 4.7. Gains are especially strong in math and complex reasoning tasks. The results reinforce OpenAI’s focus on applied performance rather than narrow test optimization.
Codex evolves into autonomous agent system
OpenAI Codex now operates as an autonomous coding colleague with real-time desktop control. It can see screens, click interfaces, write code, and manage workflows without constant prompts. The system shifts from reactive assistance to proactive task ownership. This marks a significant step toward fully agentic software development environments.
Codex runs parallel multi-agent workflows
The latest Codex deploys multiple agents simultaneously, splitting tasks like debugging, testing, and deployment. These agents operate continuously, accelerating development cycles dramatically. Integration spans over 90 tools including Slack, Notion, Jira, and Google Drive. Persistent memory allows Codex to adapt to user habits and workflows over time.
Claude skills enable reusable automation
Claude introduces modular “skills” as reusable automation units stored in Markdown. These keyword-triggered workflows function like pre-trained employees, executing consistent tasks without retraining. A shared ecosystem, including directories like skills.sh, enables discovery of vetted community tools. This lowers the barrier to building complex AI-driven pipelines.
Claude workflows hinge on structure
Effective use of Claude code depends on mastering skills, commands, hooks, and file organization. Skills run automatically, while slash commands allow controlled execution of multi-step processes. Hooks add deeper automation by triggering scripts under defined conditions. File structure acts as Claude’s contextual memory, shaping how it retrieves and uses information.
Obsidian hype exposes RAG limitations
Popular “second brain” setups using Obsidian fall short of true RAG (Retrieval-Augmented Generation) systems. The platform relies on keyword matching rather than semantic vector search. Without embeddings or reranking, retrieval quality is limited and prone to irrelevant context. This gap undermines claims of advanced AI knowledge systems built on simple note graphs.
Palantir manifesto triggers political backlash
Palantir sparked controversy with a 22-point manifesto warning of “technofascism” and urging alignment with U.S. interests. CEO Alex Karp framed technology as a tool to defend institutions and national security. The document drew sharp criticism across France and the UK, with calls to reconsider government contracts. The episode highlights rising tensions over AI, power, and geopolitics.
Videos covered
Obsidian + Claude 4.7 = voici comment créer Le cerveau artificiel parfait !
- •Obsidian as a Knowledge Base
- •Misconceptions Around RAG and Claude 4.7
- •Challenges with Context Window and Distractors
Qui travaille vraiment pour qui avec Codex ?
- •Real-Time Desktop Control
- •Simultaneous Multi-Agent Operation
- •Multimedia Creation Beyond Code
9 Claude Skills I Can’t Live Without (steal them)
- •Introduction to Claude Skills
- •Skill #1: Dashboard Style
- •Skill #2: Find Skills Skill
Libertarianisme, Impérialisme ou Techno-fascisme : quel est le vrai visage de Palantir ?
- •Palantir’s manifesto raises international concern
- •Alex Karp, an atypical CEO at the crossroads of analysis
- •A patriotic and imperialist vision mixed with rejection of digital entertainment
3 Claude Code Concepts 99% of People Don’t Understand
- •Skills offer continuous enhancement for prompts and workflows without manual activation, ideal for embedding essential capabilities.
- •Commands are suited for specialized, user-initiated processes requiring step-by-step control.
- •Hooks automate background tasks such as file management, security protocols, or product testing, saving time and effort.
OpenAI New GPT 5.5 Is A New Kind Of Intelligence (Nothing Comes Close)
- •Terminal Bench 2.0 (complex command line tasks): 82.7%, compared to GPT-5.4’s 75.1% and Claude Opus 4.7’s 69.4%.
- •GDP Val (44 professional tasks): GPT-5.5 achieved 84.9%, edging past GPT-5.4 at 83%, Claude Opus at 80.3%, and Gemini 3.1 Pro at 67.3%.
- •OSWorld Verified (real computer environment operation): 78.7%, marginally surpassing Claude Opus 78.0% and GPT-5.4 at 75.0%.