Text is not the cheapest way to get text into a model.DeepSeek-OCR showed a rendered page costs roughly 10x fewer tokens, and 97% of the words still come back out.Here's what that changes about how you feed documents to an LLM:(The numbers are silly when you see them side by side).1,000 words as text: about 1,300 tokens.The same page rendered to an image: about 100 vision tokens.Under 10x compression DeepSeek-OCR decodes 97% of the text correctly. At 20x it still holds around 60%. You get to pick how much accuracy you trade for room in the window.This is not just a paper thing. It's already running in stuff you use.Gemini treats every PDF page as an image and bills it at a flat 258 tokens, no matter how many words are on the page. Glyph (ACL 2026) renders long text to images for 3-4x compression and about 4x faster prefill, which lets a 128K model take on million token workloads. Computer use agents have been reading screenshots instead of scraped text for a while now.The idea people are chasing next: downsample old conversation turns to blurrier images, so agent memory gets vaguer with age while it stays in the window.Where it breaks: anything that needs token level precision. API keys, exact quotes, code you plan to run, long IDs. Keep those as text and compress the prose around them.Try it out. Take a doc you keep pasting in, screenshot the pages, send those instead, and compare the token count against the answer quality.Would you let your agent keep a blurry version of turn 3, or does fuzzy memory break the whole thing for you?

While Generative AI has spent the last couple of years transforming IT, corporate offices, and marketing, the "Operational Technology (OT)" layer has largely remained a tough nut to crack. Traditional Gen AI is passive, it waits for a prompt. But industrial operations require autonomous action, absolute safety, and real-time reasoning. That is exactly where Agentic AI comes in. By evolving from passive chat systems to goal-driven "agents," AI is finally ready to bridge the gap between abstract digital reasoning and physical OT execution. I just published a deep-dive article breaking down the exact blueprint for making this happen.

While Generative AI has spent the last couple of years transforming IT, corporate offices, and marketing, the "Operational Technology (OT)" layer has largely remained a tough nut to crack. Traditional Gen AI is passive, it waits for a prompt. But industrial operations require autonomous action, absolute safety, and real-time reasoning. That is exactly where Agentic AI comes in. By evolving from passive chat systems to goal-driven "agents," AI is finally ready to bridge the gap between abstract digital reasoning and physical OT execution. I just published a deep-dive article breaking down the exact blueprint for making this happen.

OpenAI JUST open-sourced Codex Security CLI 🔥 It finds, validates, and fixes security vulnerabilities in your code: It's a CLI and TypeScript SDK under @𝗼𝗽𝗲𝗻𝗮𝗶/𝗰𝗼𝗱𝗲𝘅-𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆. Point it at a repo and it does four jobs: → Find: scans for security bugs, each rated by severity and confidence → Validate: separates real findings from noise, with evidence and location → Fix: suggests remediation and re-checks it on the next scan → Track: labels findings as new, persisting, reopened, or resolved across runs It runs three ways: local scans, a pre-commit hook that blocks high-severity bugs, and CI checks that fail the build past a severity threshold. It also reports its own coverage on every scan (complete, partial, or unknown) and lists what it skipped. And since it runs on a model, scans cost tokens, so there's a --𝗺𝗮𝘅-𝗰𝗼𝘀𝘁 flag to cap spend per run. Early beta. Needs Node 22+ and Python 3.10+. See what it can do: npx @openai/codex-security@latest --help Repo + npm + docs in the comments 👇 -- P.S. We’re building the best AI English tutor in the world. Try it to improve your speaking - it’s 15× cheaper than a human one → GetFluently.app

𝗞𝗶𝗺𝗶 𝗞𝟯 𝗶𝘀 𝗮 𝟮.𝟴 𝘁𝗿𝗶𝗹𝗹𝗶𝗼𝗻 𝗽𝗮𝗿𝗮𝗺𝗲𝘁𝗲𝗿 𝗼𝗽𝗲𝗻-𝘄𝗲𝗶𝗴𝗵𝘁 𝗺𝗼𝗱𝗲𝗹. 𝗧𝗵𝗲 𝗹𝗮𝗿𝗴𝗲𝘀𝘁 𝗲𝘃𝗲𝗿 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱.Moonshot AI dropped the weights last night, a day ahead of schedule.MoE, 1M-token context, native vision. Available on API since July 16 — now downloadable.━━━━━━━━━━━━━━Where it actually lands:→ #3 on Artificial Analysis Intelligence Index — behind only Claude Fable 5 and GPT-5.6 Sol Max, and cheaper than both→ #2 on the Vals AI index→ #1 in Frontend Code Arena→ Beats Claude Opus 4.8 and GPT-5.5 on coding and agentic evalsNot the best model in the world. Comfortably the best one you can download.━━━━━━━━━━━━━━The architecture is the story, not the parameter count:→ Kimi Delta Attention — hybrid linear attention→ Attention Residuals→ Stable LatentMoE — 16 of 896 experts active per token, about 1.8% of the pool→ MXFP4/MXFP8 training→ Roughly 2.5x the scaling efficiency of K2, per Moonshot2.8T total. A fraction of it firing on any given token.━━━━━━━━━━━━━━And they opened the infrastructure too:Attention kernels, the MoE communication library, and tooling for running agent environments at scale.That's the part that compounds. Weights get superseded in six months. Kernels and comm libraries get built on.━━━━━━━━━━━━━━Now the part nobody posting about this will tell you.You cannot run this.2.8T parameters at 4-bit is roughly 1.4TB of weights before you allocate a single token of KV cache. An 8×H100 node gives you 640GB. You need multi-node, high-speed interconnect, and disaggregated serving to make it work at all — Moonshot uses their own Mooncake stack for exactly this."Open weights" and "self-hostable" are not the same sentence. For most enterprises this is an API model that happens to be downloadable.If you want genuinely deployable, the 30B-class open models are still where on-prem lives.━━━━━━━━━━━━━━Two things to check before you build on it:→ Confirm the LICENSE file in the repo yourself. The previous Kimi release was Modified MIT. Terms do not automatically carry forward.→ There is an active US policy dispute around Moonshot, including public discussion of export controls. If you're deploying for regulated or government clients, factor that into procurement now, not later.━━━━━━━━━━━━━━The gap between the best closed model and the best open one is now measured in weeks, not generations.That changes the build-vs-buy maths for everyone.We help teams work out what's actually deployable inside their own infrastructure — and what only looks like it is.DM me if you're weighing open weights against an API for a regulated deployment.♻️ Repost if this is useful to your network🔔 Follow me for open model and on-prem deployment breakdowns

Most AI interview preparation goes wrong for one reason. People memorize answers. Interviewers don't care if you can define RAG, Docker, or LangGraph. They want to know whether you understand why, when, and how these technologies are used in production. After months of preparing, taking interviews, and creating AI engineering content, I realized I kept writing the same explanations over and over. So I decided to organize everything into one place. Introducing Roy's AI Lab – AI Engineer Interview Q&A 🧠 A free GitHub repository covering interview questions from fundamentals to production-level AI engineering. Inside you'll find: 📘 Python → Core concepts to advanced topics → OOP, decorators, generators, async → Interview-focused explanations ⚡ FastAPI & Backend → APIs, dependency injection, validation → Async programming → Production-ready backend concepts 🤖 AI & Machine Learning → Machine Learning → Deep Learning → Transformers → RAG → AI Agents & Agentic AI 🔗 LLM Frameworks → LangChain → LangGraph → Real-world interview scenarios → Production design questions ⚙️ Engineering & DevOps → Git & GitHub → Docker → Kubernetes → MLOps → AWS for AI 🎯 Every topic includes → 15–20 interview questions → Basic → Advanced progression → Production-level discussions → Code examples where relevant The goal isn't to help you memorize answers. It's to help you understand the reasoning behind them because that's what interviewers actually evaluate. I'll continue adding new topics and updating existing ones as I learn. ⭐ If you find it useful, consider starring the repository. 💬 Link to the repository 👉 https://lnkd.in/gFbAPTym Follow Ritesh Rai & Roy's AI Lab for more AI Engineering content. #Day154 of Documenting my Learnings & Building Meaningful Connections on LinkedIn.

This is a genuine, specific observation with self-aware irony baked in — you notice the AI-saturation, call out that the posts themselves are often AI-written, and catch yourself mid-rant realizing this post is also about AI. That's an actual, relatable insight, not a recycled LinkedIn platitude.

DeepEval is an open-source Python framework for evaluating LLM applications such as agents, RAG pipelines, and chatbots.It helps you turn quality checks into repeatable tests by combining Pytest-style test cases with ready-to-use evaluation metrics.Key features:• Ready-made metrics – evaluate answer relevancy, faithfulness, hallucination, agent task completion, and more.• Pytest-style workflow – define test cases and run them from the CLI with `deepeval test run`.• End-to-end and component evals – assess the full application or individual traced components.• Flexible evaluation models – use your chosen LLM, statistical methods, or local NLP models, depending on the metric.• Dataset and CI support – generate single- or multi-turn synthetic datasets and run evaluations in any CI/CD environment.

⚡ TikTok search keeps growing — Gen Z now discovers products through short videos more than traditional search. ⚡ Reddit ads are getting serious attention — Brands are focusing more on niche communities and authentic conversations. ⚡ LinkedIn is pushing AI content tools — Creating posts is becoming easier… but standing out is getting harder. ⚡ YouTube Shorts adds new AI features — Faster editing + faster publishing = more competition. ⚡ Canva launched powerful AI design updates — Even non-designers can now create high-quality visuals quickly. ⚡ Adobe reports massive AI traffic growth — AI-generated referrals are changing how websites get discovered. ⚡ Meta is testing AI business assistants — Customer support and repetitive tasks are becoming automated. ⚡ Shopify expands AI shopping automation — Personalized shopping experiences are becoming the new normal. The biggest lesson? Marketing is shifting from: "Who creates the most content" Which update do you think will impact marketing the most in 2026? 👇 ♻️ Save & Repost to inspire others.

I've been comparing ways to give agents web access, and what sets PinchTab apart is a security boundary. It drives real Chrome behind an HTTP API and MCP server, but defaults the browser to a local-only allowlist, so the agent can't reach the public internet until you widen it, and extracted page text is framed as untrusted content. 9.5k stars, MIT.hashtag

Shadowbroker pulls 60+ live intelligence feeds onto one map, then lets an AI agent work all of it through a single channel instead of wiring up each feed separately. The agent reads every layer and runs the lookups itself. The risky queries stay on the server, so your browser never holds the keys or reaches outside APIs directly. 10k stars, AGPL-3.0.

Fifteen years ago, Coursera and online courses changed education. It worked better than almost anyone expected, expanding access by opening up where you can learn. But how you learn remains largely the same as it has for centuries: it is still one-size-fits-all, taught the same way to each person who shows up. We now have an opportunity to change how learning happens. With advances in AI, we can now build a custom learning guide for each person. We will turn learning from one‑to‑many to one‑to‑one. I'm starting LearnVector to invent this next generation of learning. We are starting with a $100M investment from Coursera, and plan to collaborate closely with Coursera and Udemy. Good learning needs much more than just a chatbot. Research shows that chatbots without guardrails harm learning. They help complete tasks and enable students to do better on homework. But cognitive offloading to a chatbot results in them being less skilled. And, you cannot always trust what a chatbot tells you. In contrast, LearnVector will plan a path with you, adapt to how you learn, and patiently stay with you until you’ve mastered new skills. One thing has not changed in all this time. People want learning they can trust: material that is accurate, relevant, and worth the effort you put into it. Anything less wastes the most valuable thing a learner has: time. Coursera has a trusted library of materials from authoritative sources. LearnVector plans to work with Coursera to bring this trustworthy learning to everyone. I'm grateful to Greg Hart and the entire Coursera team for supporting LearnVector. I look forward to working with our talented team to change how we learn, and accelerate human development.

If semantic search compared every embedding, RAG wouldn't scale.Here's what makes retrieval fast in vector-based RAG.Your documents live as embeddings, so you can search by meaning instead of exact wording.Answering a query means finding the vectors closest to it.The obvious way: compare against every stored vector, return the closest.That is exact nearest neighbour search.↳ Cost: O(N) per query, N = number of stored vectors.Imagine searching millions of documents to answer a single query.It is too slow and does not scale.So, what do you do?𝗜𝗻𝗱𝗲𝘅𝗶𝗻𝗴Indexing is organising high-dimensional vector embeddings into a data structure that enables fast similarity search.That data structure is Approximate Nearest Neighbour (ANN).↳ Cost: O(log N) per querySince meaning is encoded into vectors, here similarity search = semantic search.For example: Hierarchical Navigable Small World (HNSW)→ It organises embeddings in a multi-layer graph.→ Each layer contains vectors as nodes.→ Every node is connected to its nearest neighbours via edges.Traversing and descending through the layers routes the query to the right neighbourhood cheaply.ANN algorithms trade a small amount of recall accuracy for massive speed gains.Which is why retrieval rarely ends at the vector search: pipelines fetch more candidates, then re-rank.𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝗮𝗹𝗹 𝗹𝗶𝘃𝗲𝘀A vector database.It is the storage system that enables fast similarity search at scale and exposes a search API on top of it.This is where indexing arranges embeddings into ANN structures.It holds:→ the embedding vectors→ the original text→ the metadata you filter on→ the ANN indexIndexing is not a one-time build. The structure is maintained on every insert, update and delete.This is what makes fast retrieval possible in vector-based RAG.💾 Save this for your next build➕ Follow for more practical GenAI insights♻ Repost if this helps someone in your networkActivate to view larger image,

Ask a coding agent where something lives and it keyword-greps, so a function sharing one word with your query outranks the one you meant. Graft seeds the lexical matches, then runs personalized PageRank over the call graph the way search ranks the web: a match wired into the code your query touches rises, an isolated namesake sinks. Deterministic, no embeddings. 154 stars, MIT.

After today, plotting in Python will never be the same!!! This github repo is pure gold!!! 100 million points. Rendered in 0.08 seconds. The exact same speed as a chart with 10 thousand points. 🤯 I had to re-read that table three times. Reflex just open-sourced xy, a Python charting library with a Rust core and WebGL2 rendering. In their live-interactivity benchmark, render time stays flat at ~0.08s from 10k all the way to 100M points. That's 34× faster than the next best at 10M, and 177× at 50M. Plotly doesn't finish at 50M. Matplotlib taps out at 100M. At 1 billion points, xy still hands you a working interactive chart in 1.24s. The one that broke my brain: a 10M-point interactive scatter exports to 258 KiB of HTML. Plotly's version is 259 MiB. And it's a drop-in swap "matplotlib.pyplot" for "xy.pyplot" and your code just runs. "pip install xy"

I built a software factory for Claude Code and rebuilt it from scratch. Here's what I learned: My first version tried to automate everything inside one giant workflow. Halfway through implementing a feature, I realized the plan needed changing. Or I'd want to review an architectural decision before continuing. Or fix one small task without waiting hours for the entire pipeline to finish. So I rebuilt Squid around three simple stages instead: 𝟭/ 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴 Better plans matter far more than better prompts. And this is where humans still matter most. I start with a rough feature idea. Then I "grill" it with an LLM to uncover: • Missing requirements• Weak assumptions• Architectural gaps• Edge cases The output is: • A polished specification• ADRs• A project glossary• A list of implementation tasks 𝟮/ 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 I stopped thinking about this as one coding agent. Instead, it's a conversation between two roles: • Software Engineer• Tester One writes the code. The other tries to break it. They iterate until the task is complete. For larger changes, I let the whole feature run autonomously. For quick fixes, I only execute a single task. The workflow adapts to what I need. 𝟯/ 𝗥𝗲𝘃𝗶𝗲𝘄 This was probably the biggest surprise... Reviews should happen against the pull request diff rather than the entire repository. Once I changed that, review quality improved dramatically. I also merged two roles into one. Instead of having a product manager and an architect, I now have a product architect. Coding agents don't suffer from the communication gaps humans do. Keeping product intent and architectural reasoning together produces much better feedback. After review comes: • CI.• Formatting.• Linting.• Integration tests.• Regression checks. If anything fails, a new implementation task is created automatically. Here's the gist: Everyone talks about agent loops. But loops are only one part of a software factory. The actual value comes from everything around them. This is the architecture I've been gravitating towards while building Squid. Check it out here: https://lnkd.in/dejXAQBc So far, it's producing some of the cleanest AI-generated code I've worked with. P.S. What's been your biggest bottleneck when building with coding agents?

Your agent's memory has no undo button. That's been true since agents got memory, one shared store, mutated in place, and a single wrong write quietly corrupts what the agent believes with nothing to isolate or roll back.Memoria just put Git on the memory itself.Snapshot the memory. Branch a risky change. Merge it only once it checks out. Roll back to any earlier state when it doesn't. The same workflow that made code safe to change, applied to what the agent knows, built on MatrixOne's copy-on-write engine, so branching is zero-copy instead of duplicating the store.It's 409 stars and early, and it leans on running MatrixOne underneath, this is the idea proven, not yet the default stack. But the framing is the part that sticks: agent memory was always treated as a database to write to. It's actually a codebase to version.Repo in the comments.

Latency Optimization Techniques 🚀 Latency Optimization: Making AI Lightning Fast Users hate waiting. If your AI takes 10 seconds to respond, they will abandon your app. Reducing latency is a top priority in production. Here's what actually happens behind the scenes 👇 ✅ Streaming: Sending text back to the user word-by-word (like ChatGPT) instead of waiting for the whole paragraph to finish ✅ Faster Hardware: Using specialized chips (TPUs/Groq) for inference ✅ Optimized Retrievers: Speeding up the Vector DB search query ✅ Parallel Processing: Running API calls and database searches at the same time, not sequentially Example: Instead of waiting 5 seconds to generate a full email draft, Streaming displays the first word in 0.5 seconds. The user reads while the AI is still typing, making it feel instantaneous. Why this matters: Dramatically improves User Experience (UX) Critical for voice-based AI applications where delays are unnatural 💡 Speed is a feature. Optimize your entire pipeline for latency.

Forget 100s of AI eng. videosI handpicked 6 videos that cover the core of it👇If you want a practical path, watch these in order:1️⃣ Python for AI & Agents - Dave Ebbelaar→ Python foundations for real AI workflowshttps://lnkd.in/dXeYSEyZ2️⃣ FastAPI Full Course - Corey Schafer→ Build APIs and ship your AI app to productionhttps://lnkd.in/d7fK3HdN3️⃣ Intro to Large Language Models - Andrej Karpathy→ Understand how LLMs train, reason, and failhttps://lnkd.in/dDTXT-UP4️⃣ RAG Crash Course - Krish Naik→ Connect LLMs with your own datahttps://lnkd.in/dt7uSG265️⃣ AI Agents Course - Marina Wyss→ Task decomposition, tools, and multi-agent flowshttps://lnkd.in/d7x3U46y6️⃣ What Exactly is LLMOps? - Myself (Shirin)→ Trace, debug, evaluate, and monitor agents in productionhttps://lnkd.in/dSgtRB-qHope this helps!

𝟭𝟳𝟴 𝗵𝗼𝘂𝗿𝘀 𝗼𝗳 𝗰𝗹𝗲𝗮𝗻 𝗠𝗮𝗹𝗮𝘆𝗮𝗹𝗮𝗺 𝘀𝗽𝗲𝗲𝗰𝗵. 𝗙𝗿𝗲𝗲. 𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝗹𝗼𝘄-𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲𝘀 𝘀𝘁𝗮𝘆 𝗹𝗼𝘄-𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲 𝘂𝗻𝘁𝗶𝗹 𝘀𝗼𝗺𝗲𝗼𝗻𝗲 𝗼𝗽𝗲𝗻𝘀 𝘁𝗵𝗲 𝗱𝗮𝘁𝗮.Malayalam has ~38 million speakers and almost no open speech corpora. That's the whole reason its voice models lag behind.So here's Malayalam-Speech-178h:🎙️ 86,799 clips, 48 kHz🧹 Fully cleaned — vocal isolation (BS-RoFormer) → de-reverb → denoise → enhancement⏱️ 178 hours of processed speech🆓 Public on Hugging FaceOne thing up front: this is unlabeled audio. No transcripts. It's built for self-supervised pretraining, voice/speaker modelling, or as raw material for your own labelling pipeline — not for supervised ASR out of the box.I'd rather tell you that now than have you download 123 GB and find out.The expensive part is already done. Source separation and denoising at this scale takes serious GPU time — that's what you're getting for free.from datasets import load_datasetds = load_dataset("psk/malayalam-speech-178h", split="train")🔗 Link in the first comment.If you're working on Malayalam speech — or any Indian language — take it and build something.And if you're sitting on processed audio nobody else has, consider putting it out there. The gap in Indian-language speech data doesn't close by itself.