CJ Zafir's banner

CJ Zafir

@cjzafir • 63,112 subscribers

I fine-tune small language models (SLMs) and beat large language models (LLMs)

Shorts

Bad news... Deepseek still has the purple gradient curse by default.

Bad news... Deepseek still has the purple gradient curse by default.

76,958 просмотров

Truth: AI coding models are good for code execution, but not good at planning. If you let AI plan next step while coding, it will mess up the codebase. You can solve this by attaching a detailed "Implementation Plan" Here's how I create my plan: ↓

Truth: AI coding models are good for code execution, but not good at planning. If you let AI plan next step while coding, it will mess up the codebase. You can solve this by attaching a detailed "Implementation Plan" Here's how I create my plan: ↓

64,900 просмотров

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

Here's a teaser of our Mac-1 model. > 6.6B model > runs locally (on any Mac) > requires 7GB RAM (12GB ideal) > can use 487 MacOS native tools > perform multi-tool chained tasks > reasoning: ON > output: ~65 tok/s We built a robust application layer around the model to make UI/UX MacOS native. The "model-focused" SaaS era is here. Stay tuned for more.

Here's a teaser of our Mac-1 model. > 6.6B model > runs locally (on any Mac) > requires 7GB RAM (12GB ideal) > can use 487 MacOS native tools > perform multi-tool chained tasks > reasoning: ON > output: ~65 tok/s We built a robust application layer around the model to make UI/UX MacOS native. The "model-focused" SaaS era is here. Stay tuned for more.

1,068,678 просмотров • 1 месяц назад

I built 36 projects in last 14 months, with AI code. Today, I'm sharing my Coding Docs with you. I use these 7-docs in Cursor/Windsurf to build a knowledge base for AI models. Bookmark this post so you can copy the templates. ↓ 1. Project Requirements Document (PRD)

I built 36 projects in last 14 months, with AI code. Today, I'm sharing my Coding Docs with you. I use these 7-docs in Cursor/Windsurf to build a knowledge base for AI models. Bookmark this post so you can copy the templates. ↓ 1. Project Requirements Document (PRD)

2,054,550 просмотров • 1 год назад

vibe coding:

vibe coding:

708,859 просмотров • 8 месяцев назад

Claude Fable 5 took my 4 months of fine-tuning work and made an end-to-end 7-stage pipeline that i can sell. It took 3 hours of /goal. > Built TUI, > html dashboard, > html dataset viewer, > crafted 39 special skills, > wrote 8700 lines of code, > ran 235 tests. And the pipeline in 98% ready in one-shot. Shipping this soon.

Claude Fable 5 took my 4 months of fine-tuning work and made an end-to-end 7-stage pipeline that i can sell. It took 3 hours of /goal. > Built TUI, > html dashboard, > html dataset viewer, > crafted 39 special skills, > wrote 8700 lines of code, > ran 235 tests. And the pipeline in 98% ready in one-shot. Shipping this soon.

77,154 просмотров • 1 месяц назад

Mac-1 6.6B searching files on my Macbook with 10/10 accuracy. Mac's Finder and Spotlight are broken and painful to use. I have 650GB of data on my Mac, and it's all scattered. I can't find files manually. Now I ask Mac-1 to find any file (PDF, JSON, PNG, DOC, etc.), and it finds it in 2–4 seconds. Just amazing. To get early access to Mac-1, join the waitlist:

Mac-1 6.6B searching files on my Macbook with 10/10 accuracy. Mac's Finder and Spotlight are broken and painful to use. I have 650GB of data on my Mac, and it's all scattered. I can't find files manually. Now I ask Mac-1 to find any file (PDF, JSON, PNG, DOC, etc.), and it finds it in 2–4 seconds. Just amazing. To get early access to Mac-1, join the waitlist:

67,358 просмотров • 1 месяц назад

Before Claude Fable 5 got banned, I turned all my fine-tuning research and experiments into a product: (A CLI for data generation to fine-tune AI models). Finetuner uses Codex 5.5 or Opus 4.8 as the orchestrator and Chinese models (DeepSeek v4 Pro, Kimi K2.7, MiMo v2.5, etc.) to generate dataset rows. This system is unique because these are not Python-script-paraphrased datasets. Each and every row is handcrafted by Chinese models based on the orchestrator model's instructions. Now anyone can generate datasets to fine-tune small language models (1B to 30B models). I achieved 10x lower costs and 5x better dataset quality. Releasing the product in a few days. (I'll open-source the skills.)

Before Claude Fable 5 got banned, I turned all my fine-tuning research and experiments into a product: (A CLI for data generation to fine-tune AI models). Finetuner uses Codex 5.5 or Opus 4.8 as the orchestrator and Chinese models (DeepSeek v4 Pro, Kimi K2.7, MiMo v2.5, etc.) to generate dataset rows. This system is unique because these are not Python-script-paraphrased datasets. Each and every row is handcrafted by Chinese models based on the orchestrator model's instructions. Now anyone can generate datasets to fine-tune small language models (1B to 30B models). I achieved 10x lower costs and 5x better dataset quality. Releasing the product in a few days. (I'll open-source the skills.)

39,716 просмотров • 1 месяц назад

Turn any AI Model into Reasoning Model with Deepseek r1 <thinking> Architecture. Models like GPT4o and Sonnet 3.5 are Implementation Models But a new breakthrough with Deepseek can make them a Reasoning model. Here's a step-by-step Explanation: 🧵

Turn any AI Model into Reasoning Model with Deepseek r1 <thinking> Architecture. Models like GPT4o and Sonnet 3.5 are Implementation Models But a new breakthrough with Deepseek can make them a Reasoning model. Here's a step-by-step Explanation: 🧵

192,667 просмотров • 1 год назад

GitHub + Cursor is just INSANE GitHub has 28 million open-source projects that anyone can fork and build a new product So I built a new product: Repo to Docs The Agent will scan entire GitHub repo, analyze all files and create a knowledge base (docs) for AI coding tools: ↓

GitHub + Cursor is just INSANE GitHub has 28 million open-source projects that anyone can fork and build a new product So I built a new product: Repo to Docs The Agent will scan entire GitHub repo, analyze all files and create a knowledge base (docs) for AI coding tools: ↓

123,772 просмотров • 1 год назад

Introducing CodeGuide You have an Idea to turn into Software? Use CodeGuide to turn that idea into proper developments documentation for AI coding models. AI Coding = Detailed Context + Clear Prompting. Build web apps, mobile apps, landing pages, extentions...

Introducing CodeGuide You have an Idea to turn into Software? Use CodeGuide to turn that idea into proper developments documentation for AI coding models. AI Coding = Detailed Context + Clear Prompting. Build web apps, mobile apps, landing pages, extentions...

153,116 просмотров • 1 год назад

Introducing CodeGuide 2.0 A complete planning and prototyping tool for AI coding tools. Generate PRDs, wireframes, skills, prompts, tasks for your new project or existing codebase. (I'll explain 16 new updates. 1/day.) Enjoy Spec it, vibe it, ship it

Introducing CodeGuide 2.0 A complete planning and prototyping tool for AI coding tools. Generate PRDs, wireframes, skills, prompts, tasks for your new project or existing codebase. (I'll explain 16 new updates. 1/day.) Enjoy Spec it, vibe it, ship it

56,698 просмотров • 5 месяцев назад

I built "Clade Tasks" to make Claude Code 10x better and 2x cheaper. I integrated 3 engines in 1 flow: 1. Claude Code (best agentic coding interface) 2. CodeGuide (to provide deep context to AI) 3. Taskmaster (to create detailed implementation plan) Here's how it works: ↓

I built "Clade Tasks" to make Claude Code 10x better and 2x cheaper. I integrated 3 engines in 1 flow: 1. Claude Code (best agentic coding interface) 2. CodeGuide (to provide deep context to AI) 3. Taskmaster (to create detailed implementation plan) Here's how it works: ↓

107,918 просмотров • 1 год назад

If you use bolt.new let me save you alot of tokens. Here's how: Step 1: Install CodeGuide NextJS Starter Kit with pre-installed packages - NextJS 14 (ideal for AI tools) - Typescript - TailwindCSS - Shadcn UI - Supabase - ClerkDev

If you use bolt.new let me save you alot of tokens. Here's how: Step 1: Install CodeGuide NextJS Starter Kit with pre-installed packages - NextJS 14 (ideal for AI tools) - Typescript - TailwindCSS - Shadcn UI - Supabase - ClerkDev

119,918 просмотров • 1 год назад

I tested Gemini Pro 2.5 as my main coding model for 40+ hours. Here're 2 documents that are working brilliantly well with Gemini. "App flow document + App flowchart." This made my Cursor workflow 10x better. Here's why it is working: ↓

I tested Gemini Pro 2.5 as my main coding model for 40+ hours. Here're 2 documents that are working brilliantly well with Gemini. "App flow document + App flowchart." This made my Cursor workflow 10x better. Here's why it is working: ↓

85,260 просмотров • 1 год назад

Bolt 9/10 Lovable 8.5/10 Replit 8/10 I built a 1-doc system for Bolt/Lovable/Replit to one-shot applications. A hybrid of PRD x Implementation plan to provide detailed context in just 1 document. I call it "Unified Project Documentation." Simple steps: 1/ Open CodeGuide 2/ New Project "Write your app idea" 3/ Select AI tools (Replit, Lovable, Bolt etc) 4/ Copy unified project doc. by clicking "copy prompt" 5/ Paste the prompt in any selected tool. 6/ Start building your app. Here're the demos of all 3 tools: 1. bolt.new

Bolt 9/10 Lovable 8.5/10 Replit 8/10 I built a 1-doc system for Bolt/Lovable/Replit to one-shot applications. A hybrid of PRD x Implementation plan to provide detailed context in just 1 document. I call it "Unified Project Documentation." Simple steps: 1/ Open CodeGuide 2/ New Project "Write your app idea" 3/ Select AI tools (Replit, Lovable, Bolt etc) 4/ Copy unified project doc. by clicking "copy prompt" 5/ Paste the prompt in any selected tool. 6/ Start building your app. Here're the demos of all 3 tools: 1. bolt.new

60,033 просмотров • 1 год назад

Introducing Codespace A virtual computer that runs custom Claude Code on server. It is using 2 models: Kimi K2 + Claude Sonnet 4 in a sync. I achieved same code quality with 2x better reasoning and 53% less model cost. Access it today CodeGuide (500 credits for everyone)

Introducing Codespace A virtual computer that runs custom Claude Code on server. It is using 2 models: Kimi K2 + Claude Sonnet 4 in a sync. I achieved same code quality with 2x better reasoning and 53% less model cost. Access it today CodeGuide (500 credits for everyone)

31,070 просмотров • 11 месяцев назад

My Researcher Agent - {Xinstein} was a HIT. It is performing very well. But now I want to take Xinstein to the next level. I am implementing the "Chain of Density" Summarization Technique. To gather the best context for the Agent to explain the concept even better. Here are a few ideas to build your next AI Agent in ChatGPT: 1. Add Writer Agent prompt structure after the summarization step. And you'll have your Blog Writer Agent. 2. Add "Doc Maker A+" Plugin in the end to save all your Research in Google Doc or PDF format. 3. Wait till DALL-E 3 and you'll be able to design infographics for your research, posts, blogs. Xinstein was 1 of 100 weird experiments I did with ChatGPT. Results: I think I have found a technique that erases hallucinations & assumptions from AI Agents. I am experimenting with this technique. Once I'll format it correctly, I might publish a paper on that. I call it the "Agent Benchmark Technique." This week was crazy! ✦ I have cracked the best offer for AI Automation Agency. ✦ I got solid breakthrough with HypeGenius. ✦ I got invited to speak in one of the biggest LLM Firm. ✦ I got another viral post after 25 days of gap. ✦ 100 days left in New Year: I have an exciting ANNOUNCEMENT for AI Hustlers, I will share that in Monday's Newsletter Edition. If you haven't tried Xinstein Agent in ChatGPT, try it. It'll sort your research with just ONE prompt. Here's a link: --- That's a wrap for today. Get Ready for Monday's Newsletter Edition. *Something coming to push you forward and make you take Action! Follow me CJ for daily AI breakdowns and Crazy AI experiments. Peace.

My Researcher Agent - {Xinstein} was a HIT. It is performing very well. But now I want to take Xinstein to the next level. I am implementing the "Chain of Density" Summarization Technique. To gather the best context for the Agent to explain the concept even better. Here are a few ideas to build your next AI Agent in ChatGPT: 1. Add Writer Agent prompt structure after the summarization step. And you'll have your Blog Writer Agent. 2. Add "Doc Maker A+" Plugin in the end to save all your Research in Google Doc or PDF format. 3. Wait till DALL-E 3 and you'll be able to design infographics for your research, posts, blogs. Xinstein was 1 of 100 weird experiments I did with ChatGPT. Results: I think I have found a technique that erases hallucinations & assumptions from AI Agents. I am experimenting with this technique. Once I'll format it correctly, I might publish a paper on that. I call it the "Agent Benchmark Technique." This week was crazy! ✦ I have cracked the best offer for AI Automation Agency. ✦ I got solid breakthrough with HypeGenius. ✦ I got invited to speak in one of the biggest LLM Firm. ✦ I got another viral post after 25 days of gap. ✦ 100 days left in New Year: I have an exciting ANNOUNCEMENT for AI Hustlers, I will share that in Monday's Newsletter Edition. If you haven't tried Xinstein Agent in ChatGPT, try it. It'll sort your research with just ONE prompt. Here's a link: --- That's a wrap for today. Get Ready for Monday's Newsletter Edition. *Something coming to push you forward and make you take Action! Follow me CJ for daily AI breakdowns and Crazy AI experiments. Peace.

44,626 просмотров • 2 лет назад

Day 1/6 of CodeGuide 2.0 🧡 Introducing Codie - AI Agent that can help you Code on any AI Coding tool Just Screen Share and Ask for Realtime help: - Setting up Project Files - Debugging Errors - Deployment etc In this Demo I used CodeGuide x StackBlitz to show you how it works You can also use Codie with: - Cursor AI - Windsurf - Lovable - Replit - v0 Enjoy.

Day 1/6 of CodeGuide 2.0 🧡 Introducing Codie - AI Agent that can help you Code on any AI Coding tool Just Screen Share and Ask for Realtime help: - Setting up Project Files - Debugging Errors - Deployment etc In this Demo I used CodeGuide x StackBlitz to show you how it works You can also use Codie with: - Cursor AI - Windsurf - Lovable - Replit - v0 Enjoy.

23,084 просмотров • 1 год назад

Every AI startup is focused on "Build me an ........... app" (New projects) But what about millions of existing codebases on GitHub with billions of lines of code? To focus on existing codebases, I built a coding environment: Codespace. (Explained below)

Every AI startup is focused on "Build me an ........... app" (New projects) But what about millions of existing codebases on GitHub with billions of lines of code? To focus on existing codebases, I built a coding environment: Codespace. (Explained below)

13,832 просмотров • 11 месяцев назад

If this post get 400+ reposts. I'll build it for you guys. The best way to validate the idea is to share it with right audience. Check out the app flow below. It can create: ✅ System prompts ✅ PRDs ✅ App flow Docs ✅ Schema Design ✅ v0 System Prompts and so on/

If this post get 400+ reposts. I'll build it for you guys. The best way to validate the idea is to share it with right audience. Check out the app flow below. It can create: ✅ System prompts ✅ PRDs ✅ App flow Docs ✅ Schema Design ✅ v0 System Prompts and so on/

17,687 просмотров • 1 год назад