Machines pensantes Encrage🧠, GPT-Rouge🔒, bacs à sable Perplexité 🛡️

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2026-07-16 FYI: Granola has an MCP integration. Time to put your meeting notes to work (Sponsor) Still managing a big messy folder full of meeting-transcript. txt files? You're doing it wrong 😏 With the Granola MCP integration, Claude (or ChatGPT etc. ) already has access to all your meeting notes, enriched and refined by Granola.

Automated workflows are one prompt away: → Ask Claude to review last week's client meetings and update your CRM. → Have ChatGPT extract tasks from multiple conversations and organize them in Linear. You get awesome meeting notes without bots joining your calls. Your AI gets all the context with zero copy-and-paste. Sounds like a pretty sweet deal.

And the best part? The first month is FREE with code TLDR1MO. 🚀 Headlines & Launches Thinking Machines Releases First Open Model (12 minute read) Thinking Machines has released Inkling, a 975B-parameter mixture-of-experts model with 41B active parameters, multimodal reasoning, and a one-million-token context window. The model was made available for customization through Tinker alongside a smaller preview version. Secure Sandboxes for Agents (4 minute read) Perplexity AI's SPACE is a sandbox platform ensuring functionality, efficiency, and security for AI agents handling sensitive tasks.

SPACE employs ephemeral sandboxes that are destroyed after task completion and uses layers like Control Plane and Node-level Services to manage and protect credential access. The platform delivers high-security measures including credential isolation, rolling snapshots, and encrypted storage, enabling both on-prem and offline operations. Anthropic moves closer to mega-IPO as bankers line up investor meetings (3 minute read) Anthropic plans investor meetings as it prepares for a potential IPO later this year, with Goldman Sachs, Morgan Stanley, and JPMorgan Chase leading the effort. The company recently closed a $65 billion funding round at a $965 billion valuation, surpassing OpenAI's $852 billion. This IPO could capitalize on the current AI boom and test public market interest ahead of OpenAI.

🧠 Deep Dives & Analysis The first experimental evidence of recursive self-improvement (3 minute read) Researchers used autoresearch on an autoresearch agent over eight days. The results beat the harness they hand-tuned for two years. The fully autonomous system had an inner loop that optimized code against an eval and an outer loop that optimized the inner-loop agent's harness code against the inner loop's average score across different benchmarks. It designed a novel search algorithm, reduced prompt size by 16x, and built a layered system against reward hacking. How OpenAI's Sol Finally Learned Design Taste (8 minute read) GPT-5.

6 Sol ranks first overall on Design Arena's Web Design Arena. That is a significant improvement from its predecessor, GPT-5. 5, which ranks 18 places lower. The new model appears to recognize and actively compress common AI design anti-patterns, and it combines strong templates with unusually high personalization. Examples of outputs generated by the Sol are available in the article.

Model Routing Is Simple. Until It Isn't (5 minute read) Model routing systems often face challenges beyond simple model selection, turning into complex systems optimization problems. Factors such as caching, workload interactions, and serving infrastructure significantly impact costs and latency, often more than the base model pricing or perceived task difficulty. Effective routing optimizes across multiple axes ensuring an optimal operating point for the entire system rather than just finding the best model for a task. 🧑‍💻 Engineering & Research Extending Zero Trust principles for the agentic era (Sponsor) “From Zero Trust to Agent Trust,” from Teleport, explains why Zero Trust principles must extend to accommodate agent behaviors.

Agents are neither human nor machine, with speed and scale that challenge security frameworks. Learn about Agent Trust principles, common failure modes, and what it takes to operationalize them. Read more. Grok Build Coding Agent (GitHub Repo) Grok Build is a terminal-based coding agent that can inspect codebases, edit files, run shell commands, search the web, and manage long-running tasks. It supports interactive use, headless scripting and CI, and editor integration through the Agent Client Protocol.

GPT-Red for Safety Testing (5 minute read) OpenAI trained GPT-Red to iteratively generate adversarial prompts and expose model vulnerabilities at scale. Incorporating its attacks into training reportedly reduced failures on a difficult prompt-injection benchmark by sixfold for GPT-5. 6 Sol. Open Interpreter (GitHub Repo) Runs coding agents locally and tests web or native app interfaces. 🎁 Miscellaneous NVIDIA Expanded Jetson Thor for Edge Robotics (3 minute read) NVIDIA introduced the Blackwell-powered Jetson Thor T3000 and T2000 modules for running robotics, visual AI, and agent workloads on compact edge systems.