Ankit Tripathi @AnkitBuildsAI

Introducing Unsloth Desktop 🦥 The first desktop app to run and train models locally. • Open-source. Runs on Mac, Windows and Linux • Supports MLX, diffusion image/video, audio, GGUF • Connect Claude Code and Codex to local LLMs • 50% more accurate, self-healing tool calls + sandboxed code exec • Works for CPU + multiGPU setups - NVIDIA, AMD, Intel, Mac • Train models 2× faster with 70% less VRAM • Private web search, deep research, RAG, MCP and exports (NVFP4, GGUF) • Use Unsloth’s OpenAI-compatible API and cloud models • Securely deploy LLMs remotely and access anywhere Unsloth Desktop is now available on unsloth.ai and GitHub. GitHub: github.com/unslothai/unsl… Blog and Guide: unsloth.ai/docs/desktop

@Powercommitment @aryanXmahajan what does AI specialist skill stack look like

Business Idea : The "SOP-to-Skill" Agency (High-Ticket B2B Consulting) The Problem: Mid-sized companies (10-100 employees) have all their "how-to" knowledge trapped in messy Notion docs, Slack messages, or the brains of their senior employees. When a senior employee quits, that knowledge leaves with them. The Workflow (Your Service): 1.)You interview their top 3 employees to understand their daily workflows (e.g., how they process vendor invoices). 2.) You build a private, version-controlled GitHub repository of SKILL.md files for their company. 3.)You create the invoice-processor Skill: It contains the exact rules for approving invoices, a Python script to extract totals from PDFs, and instructions on how to use their accounting software via MCP. Why They Pay: You aren't selling them "AI." You are selling them Organizational Memory and Time. You charge a $5,000 setup fee + $1,000/month to maintain and update their skill repository as their business changes.

If you are entering AI space know some jargons and their meanings ! 🧠LLM (Large Language Model): The "Brain" (e.g., GPT-4, Claude). It’s very smart but trapped in a text box until you give it tools. API (Application Programming Interface): The "Digital Handshake." It’s the plug that lets the AI talk to other software, like connecting the AI to your Gmail, Slack, or Shopify store. Function Calling / Tool Use: Giving the AI "Hands." Instead of the AI just writing, "Here is an email draft for you to send," Function Calling allows the AI to actually log into your email client and hit "Send" itself. Vector Database: The AI's "Long-Term Memory Filing Cabinet." Normal AI forgets everything when you close the tab. A Vector Database allows the AI to read 10,000 company documents, remember them forever, and find the exact right page in seconds. RAG (Retrieval-Augmented Generation): Giving the AI an "Open-Book Test." Instead of letting the AI guess (and potentially lie/hallucinate), RAG forces the AI to look up the facts in your private company data before it answers. Orchestration (e.g., CrewAI, LangChain): The "Project Manager." If you have one AI acting as a Researcher, one as a Writer, and one as a Coder, orchestration is the framework that forces them to talk to each other, divide the work, and act as a team.

True. Software → LLM fine-tuning → SLM training

Naval @naval

a day ago

People who are serious about software train their own models.

565 580 10K 792K 1K

what jobs AI will be creating AI era? I know only froward deplloyed engineer and AI engineer

What are "Agent Skills" (The Foundation) Before we dive into the deep research and the roadmap, let's establish exactly what we are talking about. Most people think of AI as a Chatbot (like standard ChatGPT). A chatbot is like a smart consultant sitting in a locked room. You slip a question under the door, and they slip an answer back. They are great at writing text, but they can't do anything. An AI Agent is that same consultant, but you’ve given them a computer, internet access, a phone, and a checklist. They can figure out how to get things done, use software, and do the work autonomously. Therefore, Agent Skills are the technical and business abilities required to build, deploy, and sell these autonomous "worker bots." It is no longer just about writing good prompts; it is about building systems that can take action.

Installing the ChatGPT browser plugin will save you hours of watching YouTube videos and having conversations with AI on other people's intelligence. For example : If you need someones prompt in youtube video you can simply ask in chat, and it will give the long prompt after analyzing the video.

Market Pain Problem and its solution The front desk staff is drowning. They spend 4 hours a day answering repetitive questions on Instagram DMs, SMS, and website chats ("How much is lip filler?","Do you have openings next Tuesday?", "What is the downtime for CoolSculpting?"). Solution >They do not want "AI Agents." They want a 24/7 Digital Concierge that instantly answers questions, sends before-and-after photo galleries, qualifies the patient's budget, and books a paid consultation directly into the doctor's calendar.

@cjzafir @TheAhmadOsman @codyschneider @EXM7777 I want feedback on below thing

Bindu Reddy @bindureddy

2 days ago

TBH, you just have to spend 10x more time making an open-source model work in production - K3 is super slow - Deepseek flash will spin on more complex problems - GLM and Flash don't have visual capabilities I am close to totally giving up and going back to closed frontier

97 11 327 54K 58

@CodeWithTamara Really?

Should I build this Here is the breakdown of exactly what we are building What We Are Building We are building an **Agent Operating System (AOS)**. Think of it as a smart, secure control center that sits between a company and giant AI models (like Anthropic’s Fable 5). Instead of just letting an AI loose to do work, our platform gives that AI a permanent memory, a strict manager to double-check its work, and a financial controller to keep it from spending too much money. --- Why We Are Building It Right now, companies want to use autonomous AI agents to handle big jobs (like managing customer support pipelines or running real estate analysis) over days or weeks. But they run into three massive roadblocks: * **They forget:** AI agents have short-term memory loss. After a few hours of work, they forget what they did on Day 1. * **They are wildly expensive:** Advanced models cost a fortune. Running them continuously in loops can accidentally run up thousands of dollars in bills overnight. * **They make mistakes:** AI is notorious for grading its own homework and missing its own errors. --- How It Works We use a modern, fast tech stack (**Lovable, Supabase, and Cloudflare**) to solve these three problems automatically behind the scenes: A Shared Brain (Memory):** We automatically save the AI’s progress, successful skills, and mistakes into a permanent database. If the AI gets interrupted, it can pick up right where it left off without losing its mind. * **The Smart Router (Saving Money):** Our system acts like a traffic cop. When a massive task comes in, it gives the hard thinking to the smartest, most expensive AI model, but automatically passes the easy, boring tasks (like formatting text) to cheaper, faster models. * **The Supervisor (Quality Control):** Before the AI pushes a button to change a company’s data, our system automatically pauses it, spins up a completely separate "reviewer" AI to check the work, and only gives the green light if it passes inspection. --- Why People Will Care Enterprises don't want to buy another basic AI chatbot wrapper. They want reliable digital employees. People will care—and buy this—because **we turn erratic, expensive AI into safe, predictable corporate infrastructure.** We allow businesses to deploy autonomous agents that actually work for days at a time, slash their AI operating costs by up to 80%, and give management total peace of mind that the AI won't go rogue.

The amount of papers I read in last 90 days, I can confess that I am obsessed with post-training LLM models. I read research paper, then brainstorm with Fable/Sol. We spin up a server and test its logic, evaluate, tweak and write final report. Chinese models are built different for a reason. Just read their research papers and you'll understand how fast they are innovating the LLM training space. The future is private. Small language models (SLMs) running on your laptop, phone, car making a private ecosystem of personal intelligence. These models will use internet, connect with other models, use custom skills, tools and databases but in a way that 0.1% data gets leaked at max. Current fine-tuning techniques will be redundant in next 2 years. Fast, tap and tune techniques will emerge (they're already showing great signs in research testing). I am stoked about the open source AI future. I am thinking about starting a lab to publish all my work.

Forward deploy engineer= AI product manager+ Coder😇

@AlishaJain01 And lunges will make you dead

@CodeWithTamara Looks like a mermaid scene