I create content about deep learning | The DevRel and Deep Learning Dude

Hacking on LLMs, curating multimodal datasets, and writing a book about RAG.

Harpreet Sahota 🥑

i've literally been waiting for years to have an OPEN SOURCE model like qwen3-vl-embedding

i tested its multimodal retrieval capabilities on a corpus of 412 short video clips and the results literally blew my mind

here are the queries i tested:

1. a cartoon guy drinks merlot wine

i like this query because we see how it can retrieve based on semantics (a cartoon), text (the label merlot), and temporal action in the video (the cartoon guy drinks the wine mid-way through the video)

2. garfield runs out of the door

semantics are awesome here, the model knows i'm talking about the cartoon character...

3. a woman in a blue dress sleeping on a red bench

the skeptic in you might think that it's retrieving just based on the red color block in the 2/3 of the video...but notice the specific part of the query "a woman in a blue dress..." and this is only shown in 3s out of the full 10s

this is such a huge release 

and it's gonna open up SO much more for multimodal video retrieval this year

on my wishlist is natural language search of pcd datasets, who gonna ship that?

you can hack around with the model using the resources below

check out the docs: here: https://lnkd.in/gh7yfBNP

and the quickstart nbhere:  https://lnkd.in/gbinsx7p

🤖 Hacker-in-Residence @ Voxel51| 👨🏽‍💻 AI/ML Engineer | 👷🏽‍♀️ Technical Developer Advocate | Learn. Do. Write. Teach. Repeat.

for those who missed it, or were there in-person and know the trouble i was facing 😆, my ODSC webinar is now live on youtube!

in this session i discuss visual ai techniques for automotive damage detection. 

i used the cardd dataset, which is a public dataset built specifically for vehicle damage analysis and demonstrated how fiftyone helps you explore, visualize, and build models to detect dents, scratches, cracks, glass shatter, tire flats, and broken lamps.

i feel like its a super practical workshop with some insightful workflows for visual AI in automotive settings, but as the creator i might be biased.

only one way to find out for yourself, watch the webinar!

https://lnkd.in/ewfz5vZD

Open video: Mastering Visual AI with Vision-Language Models & Advanced Evaluation Techniques by Harpreet Sahota on youtube.com

Mastering Visual AI with Vision-Language Models & Advanced Evaluation Techniques by Harpreet Sahota

sometimes i look back at the shit i get done and i genuinely impress myself

how teams really build Visual AI in the wild? what's the he messy reality?  

i don't know...and maybe its because nobody’s asking practitioners and compiling their answers so we can find out the landscape actually looks like. 

so, we at Voxel51 are kickin off the...

wait for it...

 State of Visual AI survey...vey...vey

for those out ther building visual AI systems i want your hot takes

what’s the hardest part about making visual ai a reality? 

let us know in this survey

it'll take you 30 minutes and is a much better use of time than baby sitting your runs and staring at your loss curves

i can't offer much beyond my heartiest thanks, but you will get early access to the full report with benchmarks, common pain points, and where you stand within the landscape. 

here's a link: https://lnkd.in/gnsADACy

MedGemma 1.5 supports detection, but for best results, you'll need to fine-tune. 

there's a kaggle competition using the model, i created a starter notebook to give you a jump start on how to fine-tune it for detection

checkout the starter notebook and additonal resources 👇🏼

cc Omar Sanseviero Megan Risdal Logan Kilpatrick

in the asu area and wanna get started with open source contributions? then don't miss this hack

On March 21, 2026, we’re hosting Agents World: Visual AI Hackathon @Arizona State University, together with Pavan Turaga 🚀 

This is a hands-on day: meet people, build with AI, eat pizza, and leave with something you created after a short workshop and a full day of hacking. 
You’ll work with the FiftyOne ecosystem and real models used in practice, like Qwen, Gemini Vision, Sharp, and more. 

If you’re an ASU student who wants to build, learn by doing, and work with real Visual AI tools, this is for you.

Registration is open:
https://lnkd.in/gBCHW5Fg

Hope to see you at ASU. I’ll be there with the amazing Voxel51 team:  Sid Mehta, Ritchie Martori, Jimmy Guerrero ! 🙌

looking forward to the next meetup in boston!

Join our in-person meetup at the Microsoft Research Lab at the Massachusetts Institute of Technology in Boston, Massachusetts on Feb 18 to hear talks from experts on cutting-edge topics across AI, ML, and computer vision. Register here: https://lnkd.in/gi4kpSMB

Talks will include:

• SNAP: Towards Segmenting Anything in Any Point Cloud - Hanhui Wang at Northeastern University

• Culturally Adaptive AI - Anku R. at Massachusetts Institute of Technology

• Data Foundations for Vision-Language-Action Models - Harpreet Sahota 🥑 at Voxel51

• Better Together: Leveraging Unpaired Multimodal Data for Stronger Unimodal Models - Sharut G. at Massachusetts Institute of Technology

• Neural Radiance Fields for Image Verification - Sage Simhon at Pasteur Labs & ISI

#computervision #ai #artificialintelligence #machinevision #machinelearning #datascience

Syed Abdul Sami Rizvi, Matthew Ebisu and 118 other attendees

View event: Feb 18 - Boston AI, ML and Computer Vision Meetup

Feb 18 - Boston AI, ML and Computer Vision Meetup

Inspired by the amazing Hugging Face’s 2025 AI Timeline, we created one for the Fiftyone community. The FiftyOne Community Development Timeline highlights more than 131 contributions from 2025, including datasets, models, plugins, and integrations.

You can filter by type or modality and jump straight to each integration.
https://lnkd.in/ev2kpCGN

Big thanks to the entire Voxel51 community for building all of this together.
Harpreet Sahota 🥑, Daniel Gural, Paula Ramos, PhD, Jimmy Guerrero, Derek Van de Streek, Nicholas Lotz 🚀

Gonna be a great series of talks, lots of technical depth!

Chief Community Officer @ Voxel51 | Startup Advisor

Join us on Feb 18 for the Boston AI, ML and Computer Vision Meetup! Pre-registration is mandatory to clear building security - https://lnkd.in/gHtBuH49

Talks will include:

* SNAP: Towards Segmenting Anything in Any Point Cloud - Hanhui Wang at Northeastern University

* Culturally Adaptive AI - Anku R. at MIT

* Data Foundations for Vision-Language-Action Models - Harpreet Sahota 🥑 at Voxel51

* Better Together: Leveraging Unpaired Multimodal Data for Stronger Unimodal Models - Sharut G. at MIT

* Neural Radiance Fields for Image Verification - Sage Simhon at Pasteur Labs & ISI

***********

Want to build better computer vision models? FiftyOne is an open source toolkit from Voxel51 (our Meetup sponsor) that helps you curate datasets, evaluate model performance, visualize embeddings, catch annotation errors, and eliminate duplicate images—all in one place.

“pip install fiftyone” is all it takes to get started - https://docs.voxel51.com/

#computervision #ai #artificialintelligence #machinevision
#machinelearning #datascience #physicalai

AI product leader at Google, Author, and community builder making complex AI concepts simple for technical and business audiences. Published author of two books on GPT-3 and Neural Search. Lead and manage an active community of 200K+ AI developers across my social media channels and Unwind AI platform. Creator of Awesome LLM Apps, a #1 top-trending GitHub repository with 90k+ stars featuring practical AI Agents and RAG tutorials.

Building a community of 1M+ AI Developers at Unwind AI, an open AI education platform where 200k+ AI developers today learn to build with AI in just 3 minutes a day:
→ Daily AI news that actually matters
→ Code-first tutorials on AI Agents, RAG and LLMs
→ Open-source AI Agent and RAG apps you can use today with full code implementation

Author of books on: 
→ "GPT-3: The Ultimate Guide To Building NLP Products With OpenAI API"
→ "Neural Search - From Prototype to Production with Jina"

Previously:
→ Built Tenstorrent's developer community from scratch & launched first generation of AI hardware products from AI Accelerator cards to AI Desktop Workstations.
→ Built AI systems for Wall Street hedge funds @Arcesium (DE Shaw)
→ Led data initiatives at 256 Network
→ Tech Leaders Fellow @Plaksha

I share daily tips and tutorials on:
→ Building production-ready LLM apps
→ Implementing RAG systems
→ AI Agents and opensource agentic frameworks

🤝 Let's connect if you're:
→ Building with AI and need practical guidance 
→ Looking for ready-to-use LLM implementations
→ Want to stay ahead of AI developments and grow your career in AI

Follow me for daily AI insights and tutorials on AI Agents, RAG and LLMs.

Shubham Saboo

"Clawdbot is an overnight success."

No. I just looked at the creator's GitHub profile.

50+ projects. Years of building. Before anyone noticed.

Here's what Peter shipped before Clawdbot blew up:
🧑‍💻 Claude Code MCP - An agent inside your agent
🚇 VibeTunnel - Turn any browser into your terminal for remote agent control
🎚️ CodexBar - Track your agent token limits in real time
📸 Peekaboo - macOS screenshots and GUI automation for MCP
🧲 mcporter - Call MCPs from TypeScript or package them as CLI
🍭 sweetlink - Playwright vibes in your browser to close the agent loop
📖 llm codes - Transform docs for AI agents
🧮 tokentally - LLM token and cost math
🧿 oracle - Token management for AI workflows
📜 agent-rules - Shared rules and knowledge for coding with agents
🛑 Terminator MCP - Terminal output for agents
🪢 Conduit MCP - File ops and web prowling for agents

Every single one solving a real problem in agentic engineering.

Then Clawdbot happens and people ask "how did he build this so fast?"

He's been building the pieces for years.

The MCP servers. The token trackers. The browser automation. The CLI tools.

Clawdbot wasn't a lucky first try.

It was project number 50+.

Everyone sees the viral launch. Nobody sees the 49 projects before it.

Ship beats perfect. Consistently. For years.

That's the whole secret.

Link to Peter's GitHub in comments.

AI Product Manager @ Google | Open Source Awesome LLM Apps Repo (#1 GitHub with 90k+ stars) | 3x AI Author | 100k+ on X | Views are my Own

No alternative text description for this image

Big News: Claude Code agent can now run locally for free 🔥

Ollama just added official support for Claude Code.

Run it with open source models. No API costs. 100% local.

Here's what just happened:

Claude Code is Anthropic's agentic coding tool. It reads, modifies, and executes code in your working directory.

Until now, you needed Anthropic API credits.

Not anymore.

In just three commands.

You get Claude Code agent harness. Running locally. Free forever.

Models that work great:

Local:
→ qwen3-coder (built for coding)
→ gpt-oss:20b (strong general purpose)
→ gpt-oss:120b (complex tasks)

One requirement: You need at least 32K context window. Ollama handles this.

Why this matters:
→ No API bills
→ No rate limits
→ No data leaving your machine
→ Full Claude Code experience with open models

Cursor costs $20/month.
GitHub Copilot costs $10/month.
Claude Code + Ollama costs $0/month.

The agentic coding era just got accessible for all.

AI Product Manager @ Google | Open Source Awesome LLM Apps Repo (#1 GitHub with 91k+ stars) | 3x AI Author | 100k+ on X | Views are my Own

Open source Claude Cowork just dropped 🔥

It runs locally and that too for free with Ollama local models.
You bring your own API keys. 
Your files never leave your machine.

Introducing Openwork.

Think Claude Cowork, but:
• 100% Open Source
• Runs entirely on your machine
• Works with any model (OpenAI, Anthropic, Google, xAI, Ollama)
• No subscription. No cloud. No trust required.

Most AI desktop agents are cloud services. Your files get uploaded. Your data gets processed somewhere else. You pay a subscription. You hope they don't change the terms.

Here's what it actually does:

📁 File Management
Sort, rename, and move files based on content or rules you define.

✍️ Document Creation
Draft, summarize, rewrite docs and reports.

🌐 Browser Automation
Research workflows. Form entry. Repetitive web tasks.

🔗 Tool Connections
Works with Notion, Google Drive, Dropbox through local APIs.

Setup takes 2 minutes:
1. Download the app
2. Add your API key (or use Ollama)
3. Choose which folders it can access
4. Start working

Link to the GitHub Repo in the comments.

AI Product Manager @ Google | Open Source Awesome LLM Apps Repo (#1 GitHub with 88k+ stars) | 3x AI Author | 100k+ on X | Views are my Own

Google quietly dropped a 71-page prompting guide 🔥

And it's a masterclass in getting better outputs from AI.

Here's what most people miss:

The 4-part prompt framework that changes everything:

• Persona: Tell AI who you are
• Task: Be specific about what you need
• Context: Give background that matters
• Format: Describe your desired output

Simple? Yes.
But 90% of people skip at least 2 of these.

The guide breaks down prompts by role:
• Executives
• Marketers
• Sales teams
• HR professionals
• Customer service reps

Each with copy-paste examples you can steal today.

My 3 favorite takeaways:
1. "Make it a conversation" – iterate, don't one-shot
2. Use @file references to ground outputs in YOUR data
3. Ask AI: "What questions do you have for me?"

That last one? Game-changer for complex projects.

The best part: Google says prompting is a skill anyone can learn.

No engineering degree required.

Prompting Guide 101

Claude Code can now create videos 🤯

Remotion just dropped Agent Skills for Claude Code.

One prompt. 2 minutes. Full video. Done.

This is the ChatGPT moment for video creation.

🌟 Opensource Awesome LLM Apps Repo: https://lnkd.in/dW6b_dEn

Here's what just happened:

I typed a single prompt. Claude Code generated a complete animated video for my open source repo.

No After Effects. No Premiere. No learning curve.

Just words → video.

But wait. It gets crazier.

I generated a rap song with Suno. Added it to the video.

Sound on for fun 🎵

The entire thing took minutes. Video animation + custom AI rap track.

Install it with one command:

`npx skills add remotion-dev/skills`

Then prompt Claude Code like:

"Create a 30-second animated explainer for Opensource Awesome LLM apps repo (get the content from: repo link) "

And watch it build the entire thing.

Why this hits different:
→ Videos that used to take hours? Minutes.
→ Animations that needed a designer? Just describe them.
→ Custom soundtrack? Suno. 30 seconds.
→ Explainers, demos, social content? One-shot.

I've been playing with this for a few hours.

It's way too addictive.

The barrier between "I have an idea" and "here's a full video with music" just disappeared.

Developers making videos. AI writing the soundtrack.

We are so cooked.

AI Product Manager @ Google | Open Source Awesome LLM Apps Repo (#1 GitHub with 89k+ stars) | 3x AI Author | 100k+ on X | Views are my Own

Google Colab GPUs now run directly inside VS Code 🔥

VS Code now has an official extension for Colab.

Free T4 GPU. Inside your editor. Takes just 2 minutes to setup.

You used to choose between them. Local development in VS Code. Notebook execution in Colab's browser.

Not anymore.

What you get:
• Your VS Code setup (themes, keybindings, extensions)
• Colab's compute (including Pro-tier GPUs and TPUs)
• Local notebooks powered by cloud runtimes
• No workarounds. No hacks. Official extension.

Get started in 3 steps:
1. Open Extensions in VS Code
2. Search "Google Colab"
3. Install. Connect. Done.

Open any ipynb file. Select Colab as your kernel. Sign in with Google.

Your local notebook now runs on Colab hardware.

This is huge for:
• ML engineers who want VS Code features with Colab compute
• Researchers working on notebooks inside larger Git repos
• Anyone tired of switching between editor and browser
• Students who need GPU access without the setup pain

The community has been asking for this for years. Blog posts. Forum threads. GitHub repos with workarounds.

Google finally shipped the official solution.

NEW RESEARCH: What if your LLM could call itself recursively to handle infinite context 🤯

Researchers from MIT just dropped Recursive Language Models (RLMs).

And the results are wild.

GPT-5-mini with RLMs outperforms GPT-5 on long-context benchmarks.

By more than double. At lower cost.

Here's why this matters:
2025 gave us plenty of evidence that "context rot" is real.

Long-running Claude Code sessions get dumber. 
ChatGPT forgets instructions after 50 messages. 
Your agent loses its personality halfway through a workflow.

Critical instructions disappear into the noise.

Everyone knows this happens. Nobody could benchmark it. Until now.

Here's how RLMs work:

Instead of stuffing everything into one massive context window:
🔄 The model recursively calls itself (or other LLMs) over chunks of context
🐍 It operates in a Python REPL environment with your context as a variable
🔍 It can peek, grep, partition, and summarize before answering
📊 The root LLM never sees the full context, so it never rots

The model decides how to break down the problem. Not you.

The results speak for themselves:
• RLM(GPT-5-mini) beats GPT-5 by 114% on the long-context benchmarks
• 49% better performance at lower cost per query
• Zero degradation even with 10M+ tokens
• Model autonomously decides how to decompose tasks at inference time

Agents decompose problems based on human intuition.

RLMs let the model decide how to decompose problems for itself.

This could be the next milestone after CoT reasoning and ReAct agents.

The era of "just make the context window bigger" is ending.

The era of recursive reasoning is beginning.

AI Product Manager @ Google | Open Source Awesome LLM Apps Repo (#1 GitHub with 87k+ stars) | 3x AI Author | 100k+ on X | Views are my Own

Just SHIPPED: A visual guide to AI agent protocols 🔥

Here’s why I wrote this:
People keep comparing protocols like they’re rivals.

MCP vs A2A. AG-UI vs A2UI.

This framing is completely wrong.

These aren't competing standards. They're layers in a stack.

New protocols are dropping every month. MCP. A2A. AG-UI. A2UI. AP2. UCP.

Most developers are confused about when to use what.

Here's the mental model that makes it click:
Think of it like the networking stack. TCP/IP handles packets. HTTP handles requests. HTML handles content. Each layer does one thing well.

Agent protocols work the same way.

🔧 MCP: Agent → Tools 
USB-C for AI. One standard connector for all external resources. Databases, APIs, services. Build once, works everywhere.

🤝 A2A: Agent → Agent 
How agents discover and coordinate with each other. Agent Cards advertise capabilities. Tasks get delegated. Results flow back.

🖥️ AG-UI: Agent → Frontend 
Why your agent demo feels janky while Cursor feels like magic. Streaming, tool calls, human approval, state sync. All standardized.

🎨 A2UI: Agent → Interfaces 
Why should agents only return text? Generate forms, charts, dialogs, comparison cards. The agent outputs structure, the frontend renders it.

💳 AP2: Agent → Payments 
Cryptographic proof of user intent. Mandates that prove authorization. Built with Mastercard, PayPal, Stripe, Coinbase.

🛒 UCP: Agent → Commerce 
Standardized shopping across every retailer. Discovery, inventory, checkout, returns. One integration for all merchants.

Here's how they work together:
You ask your agent for running shoes under $150.
→ MCP pulls your size and purchase history 
→ A2A coordinates with price-comparison agents 
→ UCP queries merchants for real-time inventory 
→ AG-UI streams results to your phone 
→ A2UI renders a comparison card 
→ You tap buy 
→ AP2 signs the payment cryptographically

Six protocols. One flow. Each layer did exactly one job.

What to actually use:
Simple agent with tools? MCP only. 
Multiple agents? Add A2A. 
User-facing UI? Add AG-UI.
Rich interfaces? Add A2UI. 
Handling money? Add AP2. 
Commerce features? Add UCP.

Start simple. Add layers as complexity grows.

Agent Protocol 101

Stop describing UI bugs to your AI agent. Just point at them. 🔥

Agentation just dropped.

Here's the problem we all have:

You see a bug in your app. You open Claude Code. You type:

"The blue button in the sidebar isn't working"

Agent guesses wrong. Fixes the wrong button. You try again.

"No, the other blue button. The one below the nav."

Back and forth. Wasted time. Frustration.

Agentation fixes this.

Click on any element in your running app. Add a note. Copy. Paste into your agent.

The output includes:
→ Exact class names
→ CSS selectors
→ Element positions
→ Your feedback

It greps your codebase. Finds the exact file. Makes the fix. Done.

How it works:
1. Click the icon in the bottom corner
2. Hover over elements to see their names
3. Click any element to annotate
4. Write your feedback
5. Copy the markdown output
6. Paste into Claude Code, Cursor, Windsurf, whatever

Works with any AI coding tool. Zero dependencies beyond React.

The key insight:

Agents find and fix code much faster when they know exactly which element you mean.

Stop playing guessing games with your agent.

Start pointing at problems.

Link to the tool in the comments.

Airtable just dropped Superagent and it spins up a team of AI agents on the fly.

One prompt. Multiple specialized agents. Working simultaneously.

I asked it to do a competitive analysis of AI web agent infrastructure companies. Browserbase, Browser Use, Tinyfish, and others from a VC point-of-view.

Then I sat back and watched.

One agent pulling company backgrounds.
One agent digging through funding rounds.
One agent mapping technical moats.
One agent analyzing GTM strategies.
One agent synthesizing the investment thesis.

All at once. Like watching a research team speedrun due diligence.

But here's where it got crazy.
The output wasn't a text dump I'd spend 3 hours reformatting.

It was a Super Report.

• Interactive competitor matrices I could click through
• Scrollable visualizations
• Expandable sections with full citations
• Shareable URL ready to send to anyone

This would normally take me a weekend.

Superagent did it in just ~10 mins.

The part that surprised me: it didn't flatten a hard question into generic fluff.

It actually broke down the problem. Worked each dimension. Gave me something I'd actually present to a partner.

If you do competitive analysis, market research, investment memos, strategy decks - anything that needs depth + polish.

This is for YOU.

Try it at: https://lnkd.in/gejqXS3a

#SuperagentPartner

Google just dropped a free course on Gemini CLI with DeepLearning AI🔥

An open-source AI agent that lives in your terminal.

Here's why this matters:

Most AI coding tools work in isolation. They help you write code. That's it.

Gemini CLI is different. It has access to your local files, your dev tools, AND cloud services. All from the command line.

You give it high-level instructions. It plans. It executes. Multiple steps. Autonomously.

What you'll actually build in this course:
• A conference website session catalog
• An interactive data dashboard combining local CSVs with cloud data
• Social media content generated from video recordings
• Automated PR review workflows with GitHub Actions

The real power? Tool orchestration.

Connect MCP servers. Use Gemini CLI extensions. Coordinate across:
• Local filesystem
• GitHub
• Google Workspace
• Canva

All from one terminal command.

The course also covers:

🧠 How Gemini CLI reasons through tasks and calls tools
📁 Context management with gemini MD files
🔌 Connecting MCP servers and extensions
🔍 Web search grounding for accurate, cited responses

Who should take this:
• Developers who want to automate workflows
• Anyone prototyping apps faster
• People learning agentic AI patterns

Basic command line and Python helps. That's it.

Free. Hands-on. Available now.

Taught by our rockstar DevRel Jack Wotherspoon

Link to the free course in the comments.

taplio

The Battle

Harpreet Sahota 🥑

Shubham Saboo

Followers

Posts

Engagements

Likes

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance

Performance ScoreShubham Saboo's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Harpreet Sahota 🥑

Shubham Saboo

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Harpreet Sahota 🥑's LinkedIn Performance

Performance ScoreShubham Saboo's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance