Sid Rajaram

we really liked this Bierstadt inspired picture we made with gpt4o, it has absolutely no connection to our product.

Personal update! Been working on @clodoai.bsky.social - an AI assistant for real estate agents. Happy to announce we're now backed by YC! Check us out at clodo.ai If anyone is in real estate or know someone who is, would super appreciate a DM! 🙂 youtu.be/WDXtM2T6NwA #ycombinator #ai #realestate

The timeline after I get back from doing my thing at Edwards Air Force Base, 1990. 🕷️⏳ (credit: Adam Burch/Hangar-B Productions)

A way to help this hero's family: fundthefirst.com/campaign/sup...

gpt4o image gen still insane (also, if you know anyone in real estate, please hmu! 🙂) at @disclodocs.bsky.social (yc s25) we're building agents for real estate agents!

yeah man let's just go ahead and turn these apps into lava lamps

Was this the worst trade in NBA history?

A terribly sad situation coming out of Berkeley. But we can help, even if just a little. Consider donating and/or sharing! gofund.me/62cb0098

Insane aura maxing

In light of recent pure HTTP support by MCP, if you're still interested in doing something SSE-based, check it out! github.com/sidharthraja... #mcp #llm

The duality of man

🥵

🎉 mcp server for interacting with object store! 🎉 blog.min.io/mcp-server-f... #mcp #s3

Fully stateless MCP servers are now possible 🙂 github.com/modelcontext...

A little piece (along with Deepseek-style GRPO code) about on-prem RL on a small LLM against object store. blog.min.io/deepseek-rl-... ~Based on the awesome GRPO notebook by Will Brown.~ #grpo #rl #qwen #deepseek

thats it for #GTC. Area under the curve, compute pareto frontier, and free Denny’s —we will be there. Scaling laws have worked out quite nicely with object store and will continue to do so. Great meeting everyone! 🙂

Give your agents the gift of time 🙃⌚ github.com/sidharthraja... #mcp #agents #gtc

In honor of #GTC Quantum Day, we’re announcing quantum object store along with a new API. Your objects may be stored…or not.

inference availability/throughput has become the new stated goal of compute scaling — as opposed to what it used to be: emergent behavior from training bigger models Feels like an implicit confession that LLM scaling (even “post training” scaling) is NOT going to lead to ASI.

extremely clutch move by nvidia #gtc staff making the lanyards green. (auto pinch armor on this st. patricks day)

anybody else going to be at GTC this week? the minio team and I will be there preaching the gospel of exascale object store. See you there! #gtc

“I believe now is the right time to start preparing for AGI” The same warnings are now appearing with increasing frequency from smart outside observers of the AI industry who do not gain from hype, like Kevin Roose (below) & Ezra Klein I think ignoring the possibility they are right is a mistake

Gemma 3 27B in the same ballpark as R1 and way smaller GPU memory footprint. Model download: huggingface.co/google/gemma...

a great snapshot of how ppl are building RAG on top of object store Video: www.youtube.com/watch?v=47np...

we are in a (multi-head) attention economy

crazy how the words "Attention is All You Need" encapsulate both the technology and societal meta of the 2020s.

it's interesting that the gemini 2.0 series are the only top-tier LLMs trained and served entirely on non-nvidia (and specifically, in-house) hardware. Not to mention some of the longest context windows too.

just discovered Albert Bierstadt. Wallpapers and zoom backgrounds have accordingly been updated.

I thought this was kind of funny (one of the calculations from github.com/sidharthraja...)

Why are so many GPUs *actually* needed? Where does the model go? How is it actually getting trained on the cluster? What are FLOPdays? Some notes I took in trying to find the answer to these questions. Fact-checking welcomed!

Only fools think Elon is incompetent

Harlequin, by Pablo Picasso, 1923

Posting my first ever paper on BSKY as we inch towards the 7-month anniversary of its publication in Materials Advances.

Never have my biases been confirmed so thoroughly. Great read by @glennklockwood.com (for anyone who broadly wants to see how LLM training in the data center works)

🤖🛫

a team from berkeley has put out a reproduction of DeepSeek R1 Zero in countdown and multiplication tasks (built upon the qwen2.5 series base model) "Through RL, the 3B base LM develops self-verification and search abilities all on its own"

nvidia staying a top 5 market cap company in the long term would be like if the store that sold the shovels during the Gold Rush was still one of the most valuable companies in SF. We’ll know AI’s utility and bounty is actually diffusing throughout society when that turns.

en.wikipedia.org/wiki/Mikoyan...

For anyone experimenting with Anthropic’s MCP who wants the server to be a separate process from the client, I put together an example client/server for this using SSE: github.com/sidharthraja...

a single image from a film that immediately makes you hear the song from the soundtrack playing in your head

You can run Qwen2-VL-2B on your Mac with Llama.cpp (instructions in the README). The Qwen series has been one of the leading multimodal models for a couple months now. Fun to run it on the Mac. Get the GGUF files and instructions here: huggingface.co/sidrajaram/Q...

bring back over-the-top, artistic ads for computing

Really cool new work out of Deep Mind for video game world generation using latent diffusion! Soon you'll be able to speed run a game just by tricking a model to morph you from one location to another. deepmind.google/discover/blo...