Claude Code Desktop Goes Multi-Session with Full Redesign

🧠 LAUNCH

Claude Code Desktop Goes Multi-Session with Full Redesign

The Claude Code desktop app just got a ground-up rebuild — you can now run multiple Claude sessions side by side from a single window, with a new sidebar to manage them all. For anyone juggling agent tasks across repos, branches, or projects, this is the workflow shift you've been waiting for. Parallel sessions mean you can have one agent running tests while another writes docs while a third debugs a production issue — no more alt-tabbing between terminals. (20,761 likes | 1,411 RTs) Read more →

Felix Rieseberg, Claude Code engineering lead, confirms: the redesign is the team's own daily driver now. The speed improvements are real — this isn't a skin change, it's a rearchitecture built around how power users actually work with coding agents. If you haven't updated yet, do it today. (6,623 likes | 303 RTs) Read more →

For a full rundown of the keyboard shortcuts that make multi-session workflows fly, see our Claude Code keyboard shortcuts guide.

Gemini Robotics-ER 1.6 Brings Enhanced Spatial Reasoning to Physical Tasks

DeepMind upgrades its robotics foundation model with enhanced embodied reasoning, better spatial understanding, and multi-view scene comprehension. This isn't a demo reel — ER 1.6 is designed for real-world manipulation tasks where the robot needs to reason about 3D space, object relationships, and physical constraints. A concrete step toward general-purpose robot foundation models, and a reminder that the frontier isn't just text and code. (1,577 likes | 257 RTs) Read more →

OpenAI Launches GPT-5.4-Cyber — First Frontier Model Fine-Tuned for Defense

OpenAI just created a new category: domain-specific frontier models. GPT-5.4-Cyber is fine-tuned specifically for cybersecurity defenders and available through an expanded Trusted Access tier. This is the first time a frontier lab has shipped a model explicitly optimized for a sensitive domain with gated access — the tiered access model here could become the template for healthcare, finance, and other high-stakes verticals. (1,836 likes | 193 RTs) Read more →

🔧 TOOL

Claude Code Routines Turn Agents into Programmable Automation

Claude Code Routines let you trigger templated agents via GitHub events, API calls, or cron schedules. This shifts Claude Code from an interactive coding partner to a programmable automation platform — think "if PR opened, run security review agent" or "every morning, audit dependency updates." The implications for CI/CD workflows are significant: your agent can now respond to your repo's lifecycle without you being in the loop. (2,252 likes | 173 RTs) Read more →

Chrome Skills packages AI prompts as one-click browser tools. Google is turning prompt engineering into a shareable, remixable format — anyone can bundle an AI workflow into a one-click tool inside Chrome. Given the browser's billions of users, this could be how non-developers actually start using AI beyond chatbots. Read more →

HuggingFace Kernels: publish GPU code like you push models. You can now version, share, and distribute GPU kernels through the Hub with the same workflow you use for model weights. This dramatically lowers the barrier for distributing high-performance inference code — no more "clone this repo and pray the CUDA versions match." (1,165 likes | 159 RTs) Read more →

Kontext CLI brokers credentials for AI coding agents. A Go-based credential broker that handles secrets without exposing them in the agent's context window — solving one of the gnarliest practical problems with agentic coding. Your agent gets authenticated access; your API keys never touch the prompt. (64 likes | 26 RTs) Read more →

🔬 RESEARCH

Anthropic tests whether Opus 4.6 can accelerate alignment research itself. The Anthropic Fellows program developed an Automated Alignment Researcher using Claude Opus 4.6 — the recursive bet is straightforward: if AI can meaningfully contribute to its own safety research, alignment scales with capability. If it can't, we're building increasingly powerful systems faster than we can understand them. (1,247 likes | 142 RTs) Read more →

A few malicious documents can poison any model, regardless of size. Joint research from Anthropic, UK AISI, and the Turing Institute demonstrates that data-poisoning attacks are far more practical than assumed — injecting a small number of crafted documents into training data can introduce vulnerabilities regardless of model scale. If you're fine-tuning on scraped or user-submitted data, this is your wake-up call to audit sources. (1,544 likes | 244 RTs) Read more →

Introspective Diffusion merges diffusion generation with language models. A novel architecture that lets LLMs use diffusion-based processes during generation — enabling the model to iteratively revise and refine outputs rather than committing token-by-token. Early results suggest this could unlock better handling of uncertainty and self-correction. (220 likes | 42 RTs) Read more →

📝 TECHNIQUE

Pydantic creator's 15-minute MCP masterclass: you're doing it wrong. Samuel Colvin — the person who built Pydantic — walks through how most MCP implementations misuse the protocol. At 1.7K likes, the community clearly needed this authoritative correction. If you've set up an MCP server in the last three months, watch this before you ship anything else. (1,729 likes | 220 RTs) Read more →

💡 INSIGHT

Mistral Compute is Europe's most concrete AI sovereignty play yet. Mistral's infrastructure initiative ensures European nation states and enterprises can train frontier models without US cloud dependency. While others talk about AI sovereignty in white papers, Mistral is actually building the compute layer — and that makes this the most serious non-US infrastructure commitment to date. (2,387 likes | 330 RTs) Read more →

Mollick: if you've used agentic tools, Mythos cyber claims are entirely plausible. Ethan Mollick pushes back on the skeptics dismissing Mythos-scale cyber threats as marketing hype. His point is pragmatic: anyone who has spent real time with current agentic coding tools would find large-scale autonomous cyber operations believable, not fantastical. The gap between "demo" and "weapon" is narrower than comfortable. (647 likes | 42 RTs) Read more →

Simon Willison: cybersecurity now looks like proof of work. As AI makes attacks cheaper to generate, Willison argues defense increasingly requires proving computational effort was spent — reshaping security economics from "can we detect this?" to "can we prove we did enough work to stop it?" A framework-level shift in how we think about defense. Read more →

🏗️ BUILD

27-tool security MCP server turns any AI agent into a security analyst. CVE lookup, EPSS scoring, CISA KEV, MITRE ATT&CK, Shodan, VirusTotal — 21 APIs accessible through one MCP server. Install it, point Claude at it, and you've got a security research assistant that can cross-reference vulnerabilities, check exploit probability, and map attack techniques in natural language. Production-grade and open source. (205 likes | 37 RTs) Read more →

HuggingFace OCR'd 27,000 ArXiv papers to Markdown with an open 5B model. A fully reproducible pipeline using 16 parallel jobs — fork this for your own corpus of PDFs. The fact that a 5B parameter model handles this at scale says a lot about where open-source document understanding has gotten. (818 likes | 94 RTs) Read more →

🎓 MODEL LITERACY

Data Poisoning Attacks: Today's Anthropic/AISI research highlights a threat that scales with the AI ecosystem itself: data poisoning. The attack is deceptively simple — an adversary injects a small number of crafted documents into a model's training data, embedding vulnerabilities or biases that survive the training process. The critical finding is that model size doesn't protect you — a 70B model is just as vulnerable as a 7B one because poisoning exploits the learning process, not the model's capacity. As more teams fine-tune on scraped web data, user-submitted datasets, or community-curated corpora, the attack surface grows with every new data source you trust. The defense starts with provenance: know exactly where your training data comes from, and treat any external source as potentially adversarial.

⚡ QUICK LINKS

• Latent Space's definitive local models list: Comprehensive April 2026 roundup of what actually runs on your hardware. Link
• LtxVideo 2.3 LoRA: High-quality image-to-video generation without training your own model. (105 likes | 2.7K downloads) Link
• AMD GAIA: Official framework for building local AI agents on AMD hardware. (74 likes | 16 RTs) Link
• Anthropic appoints Novartis CEO Vas Narasimhan to board: Pharma leadership signals AI-for-science as a strategic priority. (1,126 likes | 88 RTs) Link

🎯 PICK OF THE DAY

Anthropic using Opus 4.6 to accelerate alignment research is the highest-stakes recursive bet in AI. The Anthropic Fellows program isn't just another research project — it's a direct test of the thesis that AI safety research can scale with AI capability. If Opus 4.6 can meaningfully contribute to alignment work, you get a virtuous cycle: better models produce better safety insights that make the next generation safer. But the failure mode is stark. If the model can't reliably advance alignment — if it hallucinates safety guarantees or optimizes for the appearance of aligned behavior rather than genuine understanding — then we're building increasingly powerful tools we fundamentally cannot steer, and every capability jump narrows the window to course-correct. What makes this moment different from previous AI-for-AI-safety experiments is the capability level: Opus 4.6 is genuinely good enough at research synthesis and technical reasoning that the experiment is no longer a toy. The results will tell us whether the recursive bet is viable or whether alignment will always require human researchers who can't scale at the speed of compute. Either answer reshapes the industry. Read more →

Until next time ✌️