Agent memory has a hidden bill: the tokens and calls spent building and updating it, often more than answering the question. LightMem compresses each conversation before storing it and moves consolidation to an offline pass, so the costly work never sits in the query path. On LongMemEval it reports large token and call cuts against Mem0 and A-MEM, on its own harness. 970 stars, MIT.

the dataset exists, is genuinely comprehensive, and the tooling (SQL generators, API scaffolding, HTML viewers) is legitimate and useful.

 parameter count, CPU-only operation, ~200ms latency, 6-language support, open-source/pip-installable nature are all independently verified.

Someone open-sourced a full AI operating system that runs 100% on your machine. It's called Aiden. 1500+ skills. 89+ tools. 14+ LLM providers. It controls your screen, browses the web, runs code, sends emails, and holds full conversations, all offline via Ollama. No cloud. No telemetry. No API bill. 100% Open Source. https://lnkd.in/d83EfvND

14x faster and 90% cheaper LLM inference. (100% open-source, KV cache management) LMCache is an open-source KV cache management layer that plugs into vLLM, SGLang, and TensorRT-LLM. Here's how it works: LLMs recompute their understanding of the same content on every request. The same system prompts, the same documents, processed from scratch every time, and a single GPU throws away roughly 15 TB of this reusable cache per day. LMCache stores that cache and serves it back on repeat requests, running as a separate process completely outside the inference engine. The engine just asks for the cache blocks it needs. LMCache handles all the heavy data movement across GPU, CPU, disk, and remote storage in parallel, so cache work never steals compute from inference. It also reuses cache beyond exact prefixes. Their CacheBlend technique (EuroSys 2025 Best Paper) keeps RAG documents cached no matter what order they appear in. On H200s with a 235B model, that adds up to 14x faster time-to-first-token and 4x faster decoding. And since reuse skips the compute entirely (the same reason providers discount cached tokens by 90%), the cost savings follow directly. The repo has the full architecture breakdown, benchmarks, and a Kubernetes operator for production use. Link in the first comment. ____ Share this with your network if you found this insightful ♻️ Follow me (Akshay Pachaar) for more insights and tutorials on AI and Machine Learning!

🫠 Kind of a hot take, but any seasoned AI engineer would agree. You don't need to cram machine learning and deep learning to become an AI engineer today. I've seen too many people get obsessed with the wrong things, spend months learning concepts they rarely apply, and then never actually move toward the applied side of AI. In today's world, AI engineering at most companies isn't about training models from scratch. Unless your company is hosting and training its own models, which I'd usually call ML engineering, most AI engineers are working with hosted models, whether open source or closed source. So a lot of AI engineering today is about learning how to: - Understand how Large Language Models work under the hood (Briefly, you don't have to learn how to design attention and build transformers from scratch) - Choose the right AI model for different real world problems - Build AI agents using tools, knowledge through RAG, and memory - Evaluate AI systems before they fail in production - Know when to use a single agent versus a multi agent architecture - Deploy AI applications that improve through AI Ops And honestly, the biggest thing is mastering system design, which is essentially all of the above, but contextualized to a real problem with clear thinking around tradeoffs. I published a simple beginner friendly video to help people understand what to learn, and how to think about learning it, if they want to build AI systems. It's not another recycled AI engineer roadmap. It's a mind map for people who want to build but are looking for a starting point, check it out! https://lnkd.in/dJCnxguA

Wait what...there's an open-source alternative to Claude Science 🤯 And it runs the entire scientific research LOOP right in your browser. 100% Open Source. Meet OpenScience. It works with any model you already use. Claude, OpenAI GPT, Gemini, GLM, Kimi, DeepSeek, or your own local fine-tune. Give it a research goal and it reads the papers, forms a hypothesis, writes and runs the code, runs the experiment on real compute, and writes up what it found. Most research tooling locks you into one vendor's model, one vendor's stack, and your data sitting on someone else's servers. OpenScience runs on your machine with your own keys. Requests go straight to the provider, nothing leaves your box. You give it a goal and it works the loop the way a good collaborator would: • A research agent by default, plus biology, physics, and ML specialists, with critique and literature-review sub-agents • 290+ editable skills across training, evals, LaTeX, and cloud compute • 30+ scientific databases as native tools: UniProt, PDB, ChEMBL, arXiv • A browser workspace with a file tree, editor, terminal, and inline rendering for molecules, genomes, structures, and plots Bring-your-own-key usage is free and never throttled. The workspace opens in your browser. No account required. Every skill is readable and editable, so you can open any research playbook, change how it works, and add your own. The tools we discover with shouldn't be owned by one company. Link to the GitHub repo in the comments. ____ 100+ AI Agents, Agent Skills, and Agent teams (free and open-source with 118,000+ stars): https://lnkd.in/dW6b_dEn ♻️ Repost to share it with your network. Follow me at Shubham Saboo for more such AI Agent, Loop Engineering and Agent Skill workflows.

Agent Skills just hit 71,000 stars! 24 engineering skills for AI coding agents inspired by industry best practices. npx skills add addyosmani/agent-skills ⭐ Repo: https://lnkd.in/gqFGTYUK AI coding agents default to the shortest path. That usually means skipping the spec, the tests, and the review - the exact things that separate production code from prototype code. Agent Skills exists to stop that. It encodes the workflows, quality gates, and judgment senior engineers bring to real software, packaged so agents follow them consistently across every phase: define → plan → build → verify → review → ship. This week we shipped 0.6.3, and the theme is verification. If review is the bottleneck on agent autonomy, the review skill has to be sharp: → Deeper code review. The code-review-and-quality skill gains: structural-depth lenses - reduce-don't-relocate, a named Structural Remedies catalog, a file-size gate, and finding prioritization - with severity labels that line up with its own taxonomy. → Two new reference checklists. An observability checklist (on-call questions, RED/USE metrics, symptom-based alerting) and a project-wide Definition of Done. → A comparison doc. A side-by-side of Agent Skills vs. Superpowers and Matt Pocock's skills - how each is shaped differently and when to reach for each, including a link to a controlled head-to-head. → Several stability and polish improvements. Plus a wave of community fixes and polish from 13 contributors this release. Thank you all. 🙏 24 skills, 8 slash commands, 4 specialist personas - and it runs across Claude Code, Codex, Cursor, Antigravity, Windsurf, OpenCode, Copilot, Kiro and anything that reads Markdown. Grateful to everyone building with it. #ai #programming #softwareengineering

 the technical workflow, multi-server support, and Unsloth's actual published guide are all verifiable and accurately described.

𝗜 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲𝗱 𝗺𝘆 𝗣𝗶 𝗰𝗼𝗱𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 𝘀𝗲𝘁𝘂𝗽 𝗮𝗻𝗱 𝗶𝘁 𝗰𝗿𝗼𝘀𝘀𝗲𝗱 𝟭𝟬𝟬+ 𝗚𝗶𝘁𝗛𝘂𝗯 𝘀𝘁𝗮𝗿𝘀 𝗶𝗻 𝗷𝘂𝘀𝘁 𝗮 𝗱𝗮𝘆 🚀I’ve been using this as my main coding harness for over a month with local/OSS models. For my workflow, it feels almost as good as and in some ways better than other coding-agent harnesses, mainly because I can tweak it exactly to my needs and make local models fit my workflow better.𝗖𝘂𝗿𝗿𝗲𝗻𝘁 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄:→ Pi as the coding agent→ Qwen3.6-27B as the local daily-driver model→ Advisor extension backed by a stronger frontier model (GPT-5.5 / Opus 4.8 / DeepSeek-V4-Pro)→ Local-first coding for most tasksThe setup includes local model onboarding, token/cost/speed tracking, themes, extensions, context breakdown, configurable permissions, custom skills and a sync/backup script to make it easy to backup and recreate the setup anywhere.It’s wild how capable local coding workflows have become. They may not replace frontier models for everything but they absolutely have a place and can be incredibly valuable.

the curriculum structure is technically sound and comprehensive, the terminology is used correctly throughout, and the tiered progression (mental model → mechanisms → systems) reflects a genuinely sensible way to teach this material.

the tool exists, does what's described, comes from LangChain, and the technical distinction (agent-context docs vs. human-facing READMEs) is a legitimate and well-articulated point.

The core mechanism, package structure, and design philosophy are all confirmed accurate against the actual open-source repository, and the technical approach (LLM as rule-compiler, not inference-time dependency) is clearly and specifically explained.

I just built a 4-agent software team.(100% local and runs on open-source stack)Everything runs from Telegram and is managed on a Kanban board.- A project manager who plans the work- A backend developer- A frontend developer- And a testerThe PM reads a goal, breaks it into linked tasks, and assigns each to the right agent.The thing that makes them a team instead of four strangers is a shared kanban board. Every task is a row, and when an agent finishes, it writes a summary of what it built and what the next agent needs to know.The next agent reads that summary before it starts. So the frontend developer never has to guess the API shape, and the tester knows exactly what to verify.The hardest part I faced when setting this up was building an agent that could actually act like a backend engineer.A backend engineer has to stand up a database, wire auth, manage storage, deploy functions, and keep all of it consistent while the rest of the team builds on top.An agent doing this from scratch failed almost every time in my run.It burned its context window, remembering which tables existed and which endpoint it created. Due to this, the work context exhausted quickly.I solved this by adding InsForge as the backend context engineering layer. It is an open-source, agent-native backend, and I added it to our backend developer agent as a skill.With InsForge installed, the agent stopped improvising infrastructure and followed a reliable path to create the project, define the database, set up auth, and deploy functions.To test the whole team, I had them build a working Google Docs clone, AI features included.The backend agent spun up the full service on its own, like database tables, user auth, document handling, and edge functions running real TypeScript, all in one dashboard.The frontend agent read that summary and built the UI on top of it, and the tester closed the loop.

The benchmark figures, Wikipedia scale, author attribution, model architecture, and token-cost reduction are all precisely sourced and match the original paper and its coverage closely.

Agent memory has a hidden bill: the tokens and calls spent building and updating it, often more than answering the question. LightMem compresses each conversation before storing it and moves consolidation to an offline pass, so the costly work never sits in the query path. On LongMemEval it reports large token and call cuts against Mem0 and A-MEM, on its own harness. 970 stars, MIT.

: Every number is precise, verifiable, and matches the tool's own published benchmarks and third-party coverage almost exactly — the token conversion ratio, dollar figures, and the honest caveat about byte-exact strings are all accurate.

HOLY SHIT. SOMEONE JUST KILLED EVERY SKETCHY VIDEO DOWNLOADER SITE ON THE INTERNET.It's called ReClip. Self-hosted. Open source. Free.Paste a link from YouTube, TikTok, Instagram, X, or 1000+ other sites. Pick MP4 or MP3. Done.No ads. No popups. No trackers. No rate limits.Here's how it works: 👇What ReClip actually does:→ 1000+ supported sites (powered by yt-dlp)→ MP4 video or MP3 audio→ Resolution picker before you download→ Batch downloads. Paste multiple URLs at once.→ Automatic URL deduplication→ Clean web UI in any browser→ Docker supportHere's the part developers will love:Python + Flask backend in ~150 lines.Frontend is vanilla HTML, CSS, JS. No framework. No build step.Two dependencies total: Flask and yt-dlp.That's it. The whole thing is auditable in one sitting.Setup takes one command if you have brew:brew install yt-dlp ffmpeggit clone https://lnkd.in/eYMuXxxjcd reclip && ./reclip.shOpen localhost:8899. You're done.Prefer Docker? One line:docker build -t reclip . && docker run -p 8899:8899 reclip100% Open Source. MIT License. Free forever.

Andrej Karpathy: "There's room for an incredible new product in the AI second brain space". This might be it.Everyone is suddenly building a second brain. Karpathy's LLM wiki pattern went viral, and half of X is now hand-wiring Obsidian to Claude Code so an agent maintains their notes for them.The idea is beautiful: stop making your AI re-read raw notes on every question. Let it build a wiki that compounds. As Karpathy put it, "LLMs don't get bored, they don't forget to update a cross-reference (backlinks), and can touch 15 files in one pass."But if you start doing it manually, it becomes a project in itself. You wire up the vault, the agents, the schedules, the integrations, and then you babysit all of it.So I sat down with Arjun Maheswaran, who actually built the open source version of this, and we broke down what it looks like when the whole thing already works out of the box.It just crossed 15K stars on GitHub.Think Claude's desktop app, open source, with two things layered on top:→ A work brain: background agents index your emails, meetings, and notes into a living knowledge graph that updates itself as you work.→ Work surfaces: chat is not the best interface for real work, so you get an email client, a meeting note taker, a browser, and a code mode where you and the AI actually collaborate.The part that got me: a customer email comes in asking for a product change, a background agent triages it, spins up Claude Code in its own worktree, and the feature is written before you are back at your desk.Bring your existing Obsidian vault, connect Slack, X, and Fireflies, and let it run your day.Here's what we covered in this session:- what is Rowboat (an open source ai co-worker)- the second brain (a knowledge graph of your work)- bringing your existing Obsidian vault in- meetings and automatic note taking- connecting Slack, X and other sources- background agents that run your day- code mode (Claude Code and Codex)- demo: from an email to written code- guardrails: approvals and agent workspaces- scheduling agents on a cron- the browser work surface (browser use)- wrapping up: automating your whole day

The five pieces of advice are specific and actionable — "scope the task, define done, calibrate trust, verify the result" and "own what you ship" are concrete behavioral guidance, not vague platitudes, and reflect a coherent, well-reasoned position on where human judgment adds value.