Your curated collection of saved posts and media
Wild. By far the most complete Claude Skills repo yet π€― @Composioβs Awesome-Claude-Skills packs 100`s of ready-to-use workflows: β³ PDF tools, changelog generation β³ Playwright automation β³ AWS/CDK tools, MCP builders ... and much more! Free and open-source. Repo in π§΅β https://t.co/kJhZHLU5OA
How to inject Higher Order Instructions for Claude Code: - Allows for Agent Personas (Planner, Build, QA etc) - Allows for Partner Agent approach that can think on higher level - System instructions have stronger adherence For injecting the instructions at runtime, use this command: claude --append-system-prompt "$(cat ./PARTNER.md)" The Claude .md files is passed as a user message with <system-reminder> tags and is re-injected after compaction, which overall allows drift in my experience I have seen huge performance improvements in instruction following by using the Higher Order approach - just make sure to keep it lean and following the prompting guidelines for Opus 4.5
I wrote about how I've been having a ton of fun cooking with LLMs recently - including having Claude take a photo of two recipe cards and vibe-code me up a custom timer application to help me prepare two meals at once https://t.co/qK5fvtF9FG
Claude Code plugins now support LSP servers. LSP (Language Server Protocol) gives Claude real-time code intelligence while working on your codebase. Before: Claude had to read files, run type checkers, and piece together error messages to understand what's broken. Now with LSP: Claude gets instant diagnostics, knows exactly where errors are, can jump to definitions, and sees type information in real-time. For teams: Package your LSP configs in a plugin and publish to your internal marketplace. Your team gets the same code intelligence setup without individual configuration. One person sets it up, everyone benefits.
π 42.55x Speedup! LightX2V delivers Day 0 support for @Alibaba_Qwen new Qwen-Image-Edit-2511! π₯ 47% framework speedup β‘οΈ CFG+4-step distillation (β25x compute) π 42.55x total acceleration Repo: https://t.co/SATQr3KA7m Model: https://t.co/ik3MVMnrQK https://t.co/4jqPXaao9F
How can you tell if the stated caffeine dose on an energy drink is right? I made a video showing how I measured the concentration myself at home! And for this test beverage, the estimate (144.1 += 19.6mg) lines up perfectly with the label's 150mg π https://t.co/FK83gKhwfE https://t.co/TqrfsqouWu
Are LLMs any good for web API integrations? While we see a lot of fancy demos, the reality is that LLMs still largely struggle with web API integrations. The default assumption is that code models can handle API calls reliably. After all, they excel at general code completion tasks. But web APIs have unique challenges that break this assumption. This new research introduces WAPIIBench, a benchmark for evaluating LLM-generated web API invocation code across four real-world APIs: Asana, Google Calendar, Google Sheets, and Slack. None of the evaluated open-source models solved more than 40% of tasks. Even when given the correct endpoint, models still generated 6-31% illegal arguments. URLs get hallucinated 14-39% of the time. Why is this so hard? Web API invocations differ from regular function calls in critical ways. Operations are identified by HTTP method plus long URL strings, not simple function names. Multiple argument lists exist across body, header, and query locations. Parameters have complex nested data types. And API specifications are documented externally, limiting what models can memorize. The researchers propose a solution: constrained decoding. They automatically translate OpenAPI specifications into regex-based constraints that filter token predictions during generation. The constraints enforce compliance with API specs without requiring model modifications or prompt adjustments. Constrained decoding improves correctness by 90% on average for full completion and 135% for argument completion. Illegal URLs, methods, and arguments drop to zero. Models that previously generated zero executable code now achieve similar rates to other models. Great read for AI devs. Paper: https://t.co/OXFKJRMmJc Learn to build effective AI agents in our academy: https://t.co/zQXQt0PMbG
Santa receives thousands of letters every yearβProcessing these manually takes a lot of time, so this year, we're helping him out π We built an agent to automate extracting wish-list items from letters π π€ Upload to LlamaCloud βοΈ LlamaSplit categorizes pages into logical segments (letters vs. index pages) using AI-powered document understanding π LlamaExtract extracts structured data from each child's letterβname, age, wishlist items, address, and whether they've been good or bad (using Pydantic schemas for type-safe extraction) π LlamaAgent Workflows orchestrates the process with a fan-in pattern: split the document into segments, then extract data from each letter segment in parallel The result? Transform a messy multi-page PDF into clean, structured JSON for every child's wishlistβautomatically handling document segmentation, parallel extraction, and data validation. Try it yourself in this Colab notebook: https://t.co/TbLLah1BV5
MiniMax M2.1 is a SoTA 10B-activated OSS coding & agent model, scoring 72.5% on SWE-multilingual and 88.6% on our newly open-sourced VIBE-bench, exceeding leading closed-source models like Gemini 3 Pro and Claude 4.5 Sonnet. https://t.co/3mFEc57V6S
The secret sauce to get the most out of MiniMax M2.1 for complex agentic tasks is the interleaved thinking. It excels at integrating composite instruction constraints (as seen in OctoCodingBench), ready for office automation tasks. https://t.co/UxAEzfvRan
I was also testing it for designing websites. The aesthetics are much nicer than what I get with Gemini 3 Pro. More examples of this soon. Stay tuned! M2.1 docs here: https://t.co/YdCOkaAbo2 Check out my deep research agent here: https://t.co/6oCi8ig6At
DataFlow An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI https://t.co/SbPUMOIvnA
The Prism Hypothesis Harmonizing Semantic and Pixel Representations via Unified Autoencoding https://t.co/iNjvF1Z41i
Region-Constraint In-Context Generation for Instructional Video Editing https://t.co/T0Th0WTaCp
Qwen-Image-Edit-2511 is out on Hugging Face https://t.co/rKveDT8bfk
app: https://t.co/C3eUAPqaGR
Weβve teamed up with @XSquareRobot to integrate WALL-OSS, a powerful new VLA foundation model into LeRobot!
@natolambert Ha, we have a pretty similar desk setup it seems π https://t.co/3gDAFETMOJ
Manus Projects Just Got Smarter with Connectors. You can now link your essential personal tools, such as Gmail, Google Drive, Notion, GitHub and more, directly into any project. Custom APIs are also supported. https://t.co/Cxw9ahmlCG
Thousands of builders signed up for Manus Academy (Early Bird) and many finished their certifications over the weekend! @hidecloud sat down with Annie @annieliao_2000 from @buildclub_ to unpack the mission behind Manus: extending human reach. Why we build Manus as an Action Engine, not a Q&A chatbot, and why democratizing access matters. π₯ Snippet below. Learn by doing. Join the free Manus Academy π§΅
Woo! Super pumped to try and grab my free credits too hehe https://t.co/3aXN6zHEjq
Thousands of builders signed up for Manus Academy (Early Bird) and many finished their certifications over the weekend! @hidecloud sat down with Annie @annieliao_2000 from @buildclub_ to unpack the mission behind Manus: extending human reach. Why we build Manus as an Action Eng
ELON MUSK EXPLAINS WHY π IS BECOMING THE ULTIMATE PLATFORM FOR TRUTH Something is broken in modern journalism Legacy media was built for a world where information had to be centralized, filtered by editors, printed on paper, and distributed to the public That world no longer exists Today, anyone can access information instantly. We no longer need a handful of editors deciding what matters and what does not This is why π represents the future It enables citizen journalism, news by the people, for the people What makes this moment transformative is simple The public now decides the narrative The public decides what matters Legacy media concentrates power at the top Citizen journalism distributes it to everyone That shift is fundamental And it is why Elon encourages people to write, share, and report directly on π Journalism is not dead It is being rebuilt by the public
xAI just launched the Grok Collections API A built-in RAG system that lets you upload and search entire datasets - from PDFs and Excel sheets to codebases, without building your own indexing system Why itβs huge: β’ Easy file uploads & updates with auto reindexing β’ Powerful search for finance, legal & code data β’ Free first week of indexing & storage, then $2.50 / 1K searches In head-to-head tests, Grok Collections outperformed Gemini 3 Pro and GPT-5.1 on key retrieval tasks pulling more accurate context from docs during search In internal benchmarks, Grok Collections delivered state-of-the-art retrieval performance, matching or beating leading models in finance, legal & coding real-world RAG tasks Built for real apps: Great for knowledge bases, research tools, support systems, legal/finance insights, and anywhere you need precise answers without hallucination
One year ago, Grok was only available on π, with no website and limited access Fast forward to today, and Grok has expanded across platforms with rapid feature rollouts and major upgrades New models. New apps. New features. Big upgrades Constant shipping This kind of progress does not happen without an insanely driven team xAI is moving at an unmatched pace, shipping updates faster than almost anyone in AI right now
To paraphrase the late Charlie Munger... There are only two places in the world where they haven't lost the pure faith in Communism: North Korea and the English department at Yale. https://t.co/YtITUuSFzz
Leaving California: Three More Major Businesses Close or Move Out of State. Blue Diamond, Leprino Foods, Anheuser-Busch closing/leaving CA. @CaliforniaGlobe link πππ https://t.co/ijl5f2nScs
BREAKING: Grok app traffic is surging. π Daily active users climbed from 6.1 million in August to over 10.1 million in the past 28 days, 66% increase, as per the latest @SimilarWeb data. https://t.co/UBpVKiAuie
XAI TO LEGACY AI: STOP REPEATING LIES xAI isnβt building just another chatbot. Itβs going after the entire system that filters, censors, and sanitizes reality. Most AI models are shown locked into echo chambers, βrepeating approved liesβ while appearing obedient and safe xAIβs model is visually separated. A cracked statue, glowing truth lines, and data swirling around it βFacts That Push Progressβ appears as the core value, rejecting conformity in favor of transparency Visuals show a clean break from legacy AI. No censorship, no filters, no groupthink This whole thingβs less about building AI and more about shaping what people believe. Forget polished responses. xAI is marketing itself as the AI that chooses truth over comfort, no matter how disruptive Source: @xAI, @grok
Grok
BREAKING: Grok just took the #1 spot in coding performance benchmark. Grok 4.1 Fast outscored both Google and OpenAI on code understanding and large file systems. Grok leads where it matters for developers. https://t.co/JgYQYy8pGp
Stay productive with the latest @code release - reuse your Claude skills, view terminal output in chat, manage your Git stashes, and more. 𧡠https://t.co/dLO1j3toz7
π§ Already using Claude Code? Bring your skills to VS Code. VS Code now discovers and uses skills from ~/.claude/skills/ and workspace .claude/skills/ folders. Load capabilities on-demand with full context integration. Enable with βοΈ chat.useClaudeSkills https://t.co/UZHiXCQz61
