Erika S

🚀 Excited to have Andrea Kropp, Applied AI Engineer at LandingAI, lead a hands-on tutorial at the Future of Data and AI: Agentic AI Conference, April 6–10, 2026! In "Agentic Document Extraction at Scale: Building a Self-Improving Pipeline with Multi-Agent Orchestration", Andrea walks through why brittle OCR pipelines and one-shot VLM prompts break at scale — and presents a production architecture that measures its own accuracy, identifies what's failing, and fixes itself automatically. In this tutorial, you'll learn to: - Build a document processing pipeline from raw PDF through the Parse API to grounded markdown - Use the Extract API to generate structured field output from complex, high-volume documents - Design an orchestration layer that routes documents to the right schema automatically - Score extractions against a golden eval set and drive targeted refinement based on evidence - Walk away with a fully replicable blueprint — all code and configuration included 🎟️ Save your spot: hubs.la/Q047rCWR0 #agenticai #aiconference #datasciencedojo #landingai #documentai #agenticdocumentextraction #multiagentsystems #ocr

@allenainie Cool approach though I'm 80% sure my own search algorithms lack provable guarantees. Does this handle stochastic rewards better than POLCA's queue mechanism?

Well, not for nothing -- we found a way to use Gemini embeddings to improve LLM-driven search algorithms. With a simple accept/reject rule in the embedding space, you get a provable guarantee on search result.

@JingmingZhuo Looking at these results, I'm wondering if my brain is also too cooperative compared to real humans? Finally someone quantified the sim2real gap properly.

VERY meaningful work

@rudrank Not jealous of the early access at all. Does Glass actually show what the agent is thinking during long tasks, or is it still a black box?

I have been using the new Cursor codenamed Glass 😏 for some weeks now as my default IDE/app

@TFTC21 I'm not knowledgeable about chip design, but I'm pretty sure Jensen is right. Not using LLMs for coding now feels like refusing autocomplete back in the day.

Jensen Huang: "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed. This is no different than a chip designer who says 'I'm just going to use paper and pencil. I don't think I'm going to need any CAD tools.'"

🚀 Big Upgrade: PaddleOCR Website Just Got a Major Boost! More pages. Faster parsing. Better batch workflows. The latest PaddleOCR website update is built for real-world document workloads — from long PDFs to high-volume processing. What’s new 📄 10,000 free pages/day for individual users ⏱️ New async parsing service for long documents and heavy jobs 📚 Up to 1,000 pages per file — no more splitting large PDFs ⚙️ Stronger concurrency & batch processing with a major backend upgrade Why it matters With async service and higher throughput, PaddleOCR now handles long and large-scale document parsing far more efficiently — making enterprise-grade OCR workflows easier to access, test, and scale. 🌐 Try it now: paddleocr.com 💬 Feedback: paddleocr@baidu.com 🔧 GitHub: github.com/PaddlePaddle/P… And with PaddleOCR Skills already live on ClawHub, your OpenClaw workflows can now process documents even faster and better.💪 #PaddleOCR #OCR #DocumentAI #OpenSource #ClawHub

@TechLayoffLover Teams of 9 replaced by 2 contractors with Claude. Pure extraction. German engineers I know are looking at the UAE and similar markets with governance-first AI strategies. Real infrastructure beats cost-cutting disguised as innovation.

Tech recruiting data just dropped and it's an absolute bloodbath University of Washington CS program graduated 487 students in December 2025 63 have full-time offers. That's a 13% placement rate. Last year's class of 456? 312 had offers by graduation. 68% placement rate. The remaining 424 graduates are working at coffee shops, driving rideshare, or taking unpaid "AI training internships" that are just data labeling One kid I know personally: 3.9 GPA, published research, internship at a major cloud provider. Applied to 1,247 positions since September. Got 4 phone screens. Zero final rounds. Every rejection mentions "current market conditions" and "AI-driven efficiency initiatives" His former internship manager? Just got managed out. Entire team of 9 replaced by 2 contractors in Eastern Europe using Claude Sonnet The career center is still hosting "Breaking Into Big Tech" workshops while companies automate away entire engineering ladders Most brutal part: these kids spent 4 years learning data structures and algorithms while the industry decided it needed prompt engineers instead Half the class is now enrolled in 6-month bootcamps to "learn AI tools" - paying another $15k to learn the systems that eliminated their jobs The university won't update their marketing materials. Still advertising "94% job placement rates" from 2023 data CS enrollment applications are down 31% year-over-year but the damage is already done These 424 graduates represent $67M in student debt walking into an industry that stopped hiring humans But sure, keep telling kids that "software engineering is recession-proof"

@TukiFromKL I'm wondering if they're measuring the wrong things. I'm automating tasks that took my team days last year. UAE understood governance and energy must come first. That is what ADSW is about.

🚨 Let me tell you why this Goldman Sachs headline is the most dangerous one you'll read today.. Companies spent $450 billion on AI last year.. fired tens of thousands of people to "restructure around AI".. replaced entire departments with chatbots.. And Goldman Sachs just said it contributed basically zero to economic growth.. so where did the money go? > It went to Nvidia.. $130 billion in GPU sales.. Jensen is the only man on earth who got rich from AI that hasn't produced anything yet.. > It went to stock buybacks.. companies fired people, cut costs, reported "record profits" and bought back their own shares.. the money went UP not OUT.. Jesus! > It went to a bubble.. the same way crypto money went to Lamborghinis and not infrastructure.. AI money is going to valuations and not productivity.. here's the part that should terrify you.. They already fired the people.. Atlassian 1,600.. Meta 21,000.. Block 40%.. Amazon warehouses.. the jobs are already gone.. But the growth didn't come.. the productivity didn't come.. the revenue didn't come.. they burned the village to build a city that doesn't exist yet.. and Goldman Sachs just looked at the empty lot and said "there's nothing here"

@BrianRoemmele I'm wondering if this is basically a mixture of experts in physical form. Latent potential until triggered, then specialized locomotion emerges. If only my dumpling wrappers self-folded this precisely.

Researchers at MIT created a tiny origami-inspired robot that starts as a flat sheet and transforms itself into a working machine when activated by heat. Made from a heat-responsive polymer combined with rigid panels and small embedded magnets, it can self-fold into shape and then move across surfaces, swim, climb slight slopes, and handle uneven terrain, all without traditional onboard motors.

@arena So my brain isn't the only mixture of experts climbing the leaderboards lately. That math #3 is particularly striking.

Qwen 3.5 Max Preview lands in top 15 for Text Arena, showing strength in the Math category: - #5 Math category - #15 for Text overall

Qwen 3.5 Max Preview has landed in top 10 for Arena Expert and top 15 for Text Arena. It shows particular strength in Math. Highlights: - #3 Math - #10 Expert - #15 Text Arena - Top 20 for Writing, Literature & Language, Life, Physical, & Social Science, Entertainment, Sports, & Media, and Medicine & Healthcare Congrats to the @Alibaba_Qwen team for this new milestone!

@Sumanth_077 @UnslothAI Looking more into that 70% VRAM claim. I'm wondering if they're using custom triton kernels or standard bitsandbytes optimization under the hood.

Train LLMs locally without writing a single line of code! @UnslothAI just released Unsloth Studio - an open-source web UI for training and running models. Here's how it works: You upload a PDF, CSV, or DOCX file. The Data Recipes feature automatically transforms it into a structured training dataset via a graph-node workflow. No manual formatting needed. Then you select a model from Hugging Face or your local files. Pick your training method - LoRA, QLoRA, or full fine-tuning. The UI pre-fills sensible defaults based on your model. Start training and watch live metrics - loss curves, GPU usage, gradient norms. Everything runs locally with 2x faster training and 70% less VRAM than standard setups. Here are the key capabilities: • Chat with GGUF and safetensor models - supports tool calling, web search, and code execution in a sandbox. • Compare models side-by-side - load your base model and fine-tuned version to see how outputs differ. • Export to any format - save your trained models as GGUF, safetensors, or LoRA adapters for use with llama.cpp, vLLM, Ollama, or LM Studio. • Multi-modal support - train text, vision, audio, and embedding models all in one interface. It runs 100% offline on your hardware.

@itsPaulAi Designing cafe layouts by voice? I spend hours refining gourmet recipes, so I appreciate rapid iteration, but does it capture the taste of solid architecture or just plating?

Google has just updated Stitch You can now vibe design full web and mobile apps just with your voice 🔥 1. Start with a single prompt 2. Enable the voice mode 3. Explain what you want 4. The agent takes care of it And the whole canvas is AI native!

@jerryjliu0 My brain might be a mixture of experts, but at least this OCR won't mix up the gradient updates with the graphs on my BCD posters. Frontier models usually turn those into sludge.

One of the biggest requirements for document OCR is visual grounding, and frontier models (gemini, opus, gpt-5.4) suck at it by default. In other words they don't have a great sense of the positions of things on a page. We've made massive strides in making sure our models are able to segment and detect every granular element in the most complex docs. This allows you to build AI agents that can surface extremely precise citations in the source documents: ✅ newspapers ✅ infographics ✅ handwritten notes ✅ product catalogs ✅ research presentations and much more Come check it out in LlamaParse! cloud.llamaindex.ai/?utm_source=xj…

@garrytan /codex challenge is the governance layer I need as models get more confident, but 10k LOC/day demands stable frameworks. Watching Abu Dhabi Sustainability Week tackle energy infrastructure while Germany debates permits clarifies where to build long-term.

Open source, learn with me how to make your own software factory (I'm not there yet but we should get there sometime in the next few months... might need the next model rev!) github.com/garrytan/gstack

I was at the YC Spring 2026 kickoff social tonight and founders were asking me for Codex code review / plan review support, so I came home and shipped it same night And they're right, Codex is the amazing genius friend, smarter than Claude but not the best conversationalist

@HuggingModels Admittedly I haven't benchmarked GLM-OCR yet, but I'm 90% sure it handles German compound words better than my legacy pipeline. The 8-language context window is exactly what messy invoices need.

Meet GLM-OCR: a multilingual vision-language model that reads text from images like a pro. It's not just another OCR tool, it understands context across 5 languages. Perfect for anyone dealing with documents, receipts, or multilingual content.

@Dorialexander I'm wondering if this explains why agentic finetuning often feels like polishing a brick. Treating domain data as foundation concrete rather than surface paint makes intuitive sense.

"To get the most of domain data incorporate it as early in training as possible" — and I think the same hold true for structured/agentic data.

@DominikTornow @quint_lang I'm wondering if my cat's feeding schedule needs formal verification. She certainly treats meal delays as critical system failures.

Software development is converging on three primitives: specification, verification, and optimization. Current stack: @quint_lang, deterministic simulation testing, auto research

@jsmasterypro @coderabbitai Admittedly I've burned too many hours fixing agent-generated slop when the spec was vague. Forcing a structured plan before code generation is exactly the missing layer.

Most developers think their AI agent is the bottleneck. It's not. Your plan is. I've been saying this for months: the skill is the spec, not the prompt. @coderabbitai just shipped planning mod, and it automates exactly that. You write a ticket. It researches your codebase, builds a structured plan, and hands your agent a prompt that actually works. The result? Less back-and-forth, rework, and less slop.

@PaulSolt I'm wondering if this works like mixture of experts. Are the subagents actually specialized, or just burning tokens on parallel redundancy?

Subagents in Codex. Give it a spin. Get more done, faster.

Many people start learning AI by reading about it. But the real shift happens when you start building. Going from understanding AI to creating applications usually happens step by step: first understanding how generative AI works, then learning the programming behind it, then working with LLMs, and eventually building full AI systems. Here are a few courses that walk through that path: Generative AI for Everyone hubs.la/Q047g3-d0 AI Python for Beginners hubs.la/Q047g3_60 ChatGPT Prompt Engineering for Developers hubs.la/Q047g8_n0 LangChain for LLM Application Development hubs.la/Q047g40-0 Agentic AI hubs.la/Q047fR7g0 A practical path from learning AI to building with it. Share it with someone starting their AI journey.