Fingent AI

while everyone's fine-tuning the same 3 open-source architectures, this Korean lab said "nah" and built a 314 BILLION parameter model completely from scratch. Motif-3-Beta: 🧠 ~314B total params, only ~13B active per token (MoE sparsity go brrr) 📏 256K context, natively — not "extended," not duct-taped ⚡ 384 experts, 8 fire per token + 1 shared expert always on 🌐 fully multilingual and they invented their OWN attention mechanism (Grouped Differential Latent Attention) and their OWN activation function (Grouped PolyNorm) because apparently reusing existing architectures was too easy weights are fully open, no gatekeeping, no access request this is still the BETA. final version drops soon. imagine what that looks like huggingface.co/Motif-Technolo…

@ScottyBeamIO China on another planet

WTF, THIS CHINESE ENGINEER NAVIGATES HIS OBSIDIAN SECOND BRAIN WITHOUT EVER TOUCHING HIS LAPTOP – JUST HAND GESTURES IN THE AIR The stack: Obsidian + Claude Code + Hermes Agent He's not clicking, scrolling, or typing. He moves his hand in the air, and his entire knowledge graph responds – nodes shifting, branches expanding, connections highlighting – all on screen, all in real time, with zero physical contact with the machine. He reaches toward a cluster of notes. It opens. He moves his hand to the side. The graph rotates and follows. This isn't a demo video with cuts and effects. It's happening live, gesture by gesture, as he actually navigates through his own thinking. Most people are still scrolling through a flat list of notes. This guy turned his second brain into something he physically moves through. This is what happens when local AI agents stop being something you type commands into and start being something you interact with directly. Bookmark this. Full demo in the video below.

China quietly dropped a model that reads a 100+ page PDF in one pass — not page by page like everything else... It's called Unlimited-OCR. Just 3B params. Runs locally. Most OCR tools slice your document into pages and lose the thread between them. This one processes the entire document in a single shot — "one-shot long-horizon parsing." → 32K token context — holds the whole doc, not just a chunk → Multilingual out of the box → Handles single images AND multi-page PDFs via a dedicated infer_multi mode → 100% local, no API keys, no cloud → Works with Transformers, vLLM, SGLang, Docker, plus ready quantizations for llama.cpp and Ollama → MIT license — do whatever you want with it Cloud OCR (Textract, Google Vision, Azure Doc Intelligence) costs $1.5–15 per 1,000 pages. This runs on your own hardware. For free. Forever. Baidu says it straight in the model card: built to "push DeepSeek-OCR one step further." 100% open source

@shiqway92 facts. context is so underrated

YOU’RE PAYING $200/MONTH FOR AN AI THAT FORGETS YOUR BUSINESS EVERY 5 MINUTES Here’s why your competitor with the “dumber” model is beating you. You bought GPT-4, Claude, or Gemini. Paid for the aggregator, API, and plugins. And ended up with… An expensive intern with amnesia. Every new chat: < “Tell me about your business.” < “Upload your documents.” < “Explain the context.” You keep feeding it the same PDFs, repeating the same things, and wasting 20 minutes warming it up instead of getting work done. Your competitor? They don’t have a better model. They have better memory architecture: > Short-term → current conversation > Long-term → summaries of past conversations > Vector memory → semantic search across thousands of interactions But the real upgrade is this: Their AI reads the documents once, extracts the key concepts, connects related information, and stores everything in a structured Wiki. Now it doesn’t reread the PDFs every time. It retrieves only the context it actually needs. Your ai is a chatbot. Their ai is a coworker with 5 years of experience who remembers every client. The next ai advantage won't come from better models It'll come from better memory architecture While you’re tweaking prompts they’re building a brain bro

@AleiahLock @williamlegate @PolymarketTrade @Polymarket @PolymarketDevs you should probably see this — this is happening in a DAO tied to your brand

I was kicked from this DAO for no reason. I was consistently ranked in the top 2–3 of the ZSK DAO leaderboard. I contributed by editing the Polymarket and DAO pages, and I actively participated in every raid and community program. When I asked why I had been removed, I was told it was because I was also a member of another DAO that Atlantis didn't like. Interesting. I've been asking @williamlegate for a long time why the @Polymarket team allocates budgets to people like this. (And please don't tell me they don't receive funding. Everyone knows ZSK members have the Builder badge, and Atlantis itself openly says it receives funding from you every month.) The person running the DAO behaves extremely unprofessionally - threatening and insulting people. If you don't do exactly what he wants, you're simply removed. So how do you choose partners for your project? Polymarket says that reputation is everything, yet according to this article, someone who was receiving money from you allegedly created a farm of accounts to extract even more money from the project through deception. You gave access to your ecosystem @PolymarketTraders , but they only promote people whom they personally like. The funniest thing about this situation is that when I was in this DAO, all the PolymarketTrader badge holders spoke badly about us, and they didn’t even know that their main page was managed by the DAO, which they don’t respect. Big shout out to @Frosen Are your Trader partners generally like Car as well? He was one of the highest-ranked traders and seemed untouchable-free to insult or even scam people without consequences. In the end, someone who reportedly earned hundreds of thousands of dollars through PolyMarket was caught running a $5,000 scam. Here is the same situation: you gave POWER AND MONEY to a child who constantly screams if he doesn't like something, having access to DUB, he wanted to use personal information to find and physically harm a person who expressed his opinion against ATLANTIS. I'd reconsider the affiliate program. Many people want badges, want to be part of the project, and create content for it. But there are no specific conditions, no rules-nothing. @0xTone @shayne_coplan @mustafap0ly @MatthewModabber

@beamnxw Failures start in the context, not the model

AI agents often fail before the model itself makes a mistake In the paper "AI Agents Do Not Fail Alone: The Context Fails First" researchers showed that future agent failures can be predicted from the environment before the task ends Common causes: - vague role and instructions - poorly described tools - insufficient facts - contradictory rules - weak memory - weak guardrails - poorly structured context The authors measured the quality of the environment separately from the final answer On 300 tests the same fixed models produced higher results after the context was made structured Before changing the model, check prompts, tools, memory, facts and safety rules full paper + my article👇

@AleiahLock yep

@fingent_ai New wallpaper for my page ?

Microsoft just dropped a 4B image model that's punching WAY above its weight class Mage-Flow beats Qwen-Image (20B), FLUX.2 (32B), and FireRed-Edit (20B) at generation AND editing... at a FRACTION of the size. we're talking 5-8x smaller how? they didn't just scale up. they rebuilt the tokenizer from scratch. Mage-VAE does encode/decode with 12x-22x fewer compute ops per pixel than FLUX.2's VAE — the bottleneck just disappeared one checkpoint. any resolution from 512 to 2048. any aspect ratio, including absolutely unhinged ones like 512×2048 (4:1 panoramas, no cropping tricks needed) and the Turbo variant generates a full image in 0.59 seconds on a single A100. editing in 1.02s. using ~18GB of memory — less than almost everything it's beating MIT licensed. open weights. Python API + CLI + Gradio app included. runs on your own GPU, no API key, no rate limit the "just make it bigger" era of image models might be over link👇

@ColoringArtFN $spcx is ok. check $try

Rug pull?!😁 $SPCX

your team's stack right now: Slack for chat, GitHub for code, Jira for tickets, CI for builds, a bot for releases, and 7 tabs pretending they talk to each other. Block just built Buzz block/buzz: A hive mind communication platform — one Nostr relay where humans AND AI agents live in the same rooms, sign with the same keys, and show up in the same audit log. your feature branch literally BECOMES a channel. patches land as signed events, CI posts results, an agent does first-pass review, you react, it merges — all in one room. ask "have we seen this bug before?" at 2am and an agent pulls 6 months of history with receipts, not vibes. agents aren't bots bolted onto your workspace. they get their own keypair, their own channel memberships, their own audit trail — same as any teammate. no more giving an agent root access just so it can triage a bug. self-hostable. Apache 2.0. built on Rust + Nostr, so it's not vaporware, it's an actual protocol. 116 stars and this is legitimately what "agent-native workspace" should've meant from day one github.com/block/buzz

me: "make me a landing page" AI: purple gradient, centered hero, rounded pill button, "Built for the modern team," Lottie animation 🥲 me after installing Hallmark: 65 slop gates just rejected my own AI's output like a bouncer at a club it doesn't recognize the fit for it studies real sites, extracts the design DNA, picks from 21 different page structures, and REFUSES the default AI slop it was trained on this is the "no more purple gradients" update we all needed github.com/Nutlope/hallma… 💀

Someone open-sourced a system that turns regular WiFi into a radar that sees through walls. It's called RuView (aka WiFi DensePose). Already 81,000+ stars on GitHub — with zero cameras involved. How it works: your WiFi router is already flooding the room with radio waves. When a person moves, breathes, or even just sits still, they subtly distort those waves. The system reads that distortion through a cheap ESP32 chip (~$9) and turns it into data: who's in the room, what they're doing, whether they're okay. What actually works right now: → Presence detection and headcount — including through walls → Real-time breathing rate and heart rate, contactless → Fall detection → Through-wall range up to ~5 meters, depending on material Full 3D skeletal pose tracking is a separate published model, scoring 82.69% on the MM-Fi benchmark (beating prior approaches in that category) — but it's not "instant out of the box," it's an add-on. Honest caveats worth knowing: — you need an actual $9 ESP32 board, not just "any router by itself" — it doesn't do facial/personal identification — the authors themselves state that telling two specific people apart from WiFi signal alone isn't reliably possible yet — earlier claims like "100% accuracy" were retracted by the authors themselves; the current honest figure is ~82% No cameras, no cloud — everything runs locally. MIT license, free.

Someone connected Claude directly to TradingView Now it can see your chart, read your indicators, and write trading scripts for you. Free, runs locally on your own computer, already 3.2K stars on GitHub: github.com/tradesdontlie/…

🧵 Someone fixed the #1 problem with AI coding agents: they bury the answer. "i-have-adhd" is a skill plugin (Claude Code, Codex, Cursor) with 10 rules: 1. Lead with the next action 2. Number multi-step tasks 3. End with one concrete next step 4. Suppress tangents 5. Restate state every turn 6. Specific time estimates (minutes, not "a bit") 7. Make wins visible 8. Matter-of-fact errors 9. Cap lists at 5 items 10. No preamble. No recap. No closers. Before: "Great question! Let me think about this..." (3 paragraphs) After: "Run npm install jsonwebtoken@latest, then edit src/auth.ts:42." 5,498 ⭐ in 2 months. People didn't want a smarter model — they wanted a more readable one.

Tableau is dead. They just don't know it yet. 16K stars on an open-source project that lets AI agents generate, validate, and deploy BI dashboards — for free. WrenAI: text-to-SQL → chart → dashboard → shareable URL. Governed. 22+ data sources. Agent-driven via Claude Code, Cursor, Codex, MCP. The context layer is the moat: business definitions, approved joins, memory — all in Git, not locked behind a $70/user/month paywall. Apache 2.0. 1,851 forks. The community is already building.