Raghotham Sripadraj

hot take: once you start using @modal notebook, colab feels like stone age. The attention to DX details and the swift response of the system is amazing. Kudos to @modal team. Came back to @GoogleColab after few months and feels legacy.

Just finished my talk on “Why do (Senior) Engineers struggle to build AI Agents” If you missed or want to read about, blog post below. To succeed, we have to accept: 1️⃣ Text is the new state. 2️⃣ Hand Over Control. 3️⃣ Errors Are Just Inputs. 4️⃣ Move from Unit Tests to Evals. 5️⃣ Agents Evolve, APIs Don't.

Do you understand what this means? For the first time, an Open Weight models is #1 on CyberSecutity. Sure there’s Mythos but we don’t have it :p Don’t let them trick you

@AlexPeghin Absolutely! Hence, it is even more important to have strong fundamentals today. One cannot fly blind for a long time 🙂

@raghothams Orchestrating chaos is great until you lose visibility into what tools the chaos is calling

The Test Match era of software engineering is coming to an end. The 10x developer is no longer a purist coder. They are an Agentic Engineer—an orchestrator of chaos. I just published a piece on why this is the biggest paradigm shift in tech history. (Link in comments)

@nischalhp @karpathy Absolutely!

@raghothams @karpathy And I no way I am trying to undermine the quality of knowledge that @karpathy shares with the community, there is a reason he goes viral, the content is just well thought through. That sort of original thinking should be celebrated and not just the outcome.

Everyone consumes information differently and at a varied velocities. Just building knowledge graphs because @karpathy spoke about it for your own usage might not be helpful if you don't go down the path of asking the hard question around what do you need and how do you consume?

Here is a fantastic opportunity to build, learn and grow! Nirant is one of the finest builders in India. Go grab it!

Hard agree!

Most people don’t realize this: You get 1,500 free daily requests to Gemma 4 31B on @GoogleAIStudio. That’s plenty of free inference (imo). And you can route it into @NousResearch Hermes Agent via Vercel’s AI Gateway: 1. Create an API key on Google AI Studio 2. Add it under BYOK (Google) in Vercel AI Gateway 3. Create a Vercel Gateway API key 4. In Hermes → select “Vercel AI Gateway” + your Google model Now all your Google model requests route through your free AI Studio quota. Basically: free 31B model access inside your agent stack. (Tradeoff: not as private as running locally)

Here’s what I’d recommend if you’re just getting started in AI, local or otherwise. 1. Work with the compute you have, even the dumbest LLMs can be useful if you treat them as a node in your system. Some basic problems of what could be useful to get you started - tag all your screenshots - classify your emails - recommendation algo - scanning git history for patterns 2. If you want to try larger models use prime intellect or Hotaisle (even cheaper) to rent out whatever amount of VRAM you need to run models you like. - RTX 3090 rents for cents/h - RTX 6000 rents for 1-2 dollars/h - H100 rent for 2-3 dollars/h 3. Use the right frameworks: - VLLM, and SGlang for faster inference if you have 1, 2, 4, 8, 16 GPUs - exllamav3 and llama.cpp if you have non-power of 2 GPUs - MLX if you have Mac 4. Start with problems in your life that could use intelligence for automation - shopping - research 5. Don’t expect local models to code production projects just yet

Whether you're an ML engineer building LLMs or an AI engineer building with them, understanding how to evaluate LLMs and their applications is a skill you need to have. This is the best overview of LLM benchmarks I've read. Source by @cwolferesearch: cameronrwolfe.substack.com/p/llm-bench

This is a great read for anyone interested in understanding how training is done and everything around it. huggingface.co/spaces/Hugging…

Best models to run on your hardware level I'll be doing this every week, I hope you guys enjoy. ---- 8 GB ---- Autocomplete for coding (like Cursor Tab) - huggingface.co/NexVeridian/ze… - huggingface.co/bartowski/zed-… Tool calling, assistant style - huggingface.co/nvidia/NVIDIA-… ---- 16 Gb ---- Here things get better: Multimodal - huggingface.co/Qwen/Qwen3.5-9B - huggingface.co/Tesslate/OmniC… - huggingface.co/unsloth/Qwen3.… ---- 24 GB ---- - The best model you can get (thanks Qwen) huggingface.co/Qwen/Qwen3.5-2… - Great model (strong agents) huggingface.co/nvidia/Nemotro… - Mine hehe huggingface.co/0xSero/Qwen-3.… I'm doing a weekly series

For my friends who are still using UV and might be a little weary about recent compromises to PyPi packages, stick this in your pyproject.toml. You can let all of those pip users find and report the compromises...

@deepakshenoy @atomberg_tech might build a cost efficient version for India

Can we build this here in India? Man I would pay for this.

VLMs too slow for production? Not anymore: 46ms end-to-end inference, 60+ fps on a single H100. Introducing Photon, Moondream's inference engine. Runs on everything from edge to server. moondream.ai/blog/photon-re…

Loved reading this! Thank you!

Nobody's talking about the second-order effects of GLP-1 drugs loudly enough. So let us drop some data on the effect it is having on different industries! 1. 23% of U.S. households already have someone on Ozempic or Wegovy. By 2030, that number hits 35% of all food & beverage units sold. This isn't a niche health trend anymore, it's a demand restructuring event. 2. Grocery spend drops 5.3% within 6 months of starting. For rich households, it's over 8%. Fast food and coffee shops spending falls by 8%. Savory snacks spending falls by 10%. Same story for cookies, baked goods, sweets. But here's where it gets genuinely interesting. 3. Alcohol consumption is falling but users are cutting sweets before booze. Which means this isn't just "I'm full." The brain's reward circuitry is actually being rewired. 4. Pet product spend is quietly declining among GLP-1 users. The theory? Money and attention are being redirected inward. Spending on Skincare, fragrance, cameras, wearables is spiking. Fashion shifts from bags and shoes → clothes. Body confidence replacing status signaling. 5. And the creepiest stat: stop the drug, and your basket gets worse than before you started. Candy spend spikes. Whatever the drug was doing to cravings, cold turkey reverses it hard. Scale this up. 20% obesity usage rate = 20 billion fewer calories consumed per day in the U.S. alone. Every food brand, Alcohol, restaurant chain, and FMCG company has a GLP-1 problem they haven't fully priced in yet.