Strong Compute

Thrilled to share that our team — Ramprasadh Kumar, Bernett Orlando, and I — won the DeepSeek AI Fine-tuning Hackathon, hosted by Strong Compute. We developed a novel approach to fine-tune DeepSeek for a Math AI Tutor use case, enabling the transformation of natural language prompts into conceptual visualizations through Python-generated animations. We played around with a bunch of GRPO reward modeling strategies — super fun and a great mix of technical challenge and creative thinking. 😄

🎉 Outthinking the ARC! I'm thrilled to share that our team won the Strong Compute 24 hour hackathon, the ARC agi2 prize track, with $2.5k Open Source Research Compute Grant! What we built: • We built a synthetic data generation pipeline for the ARC-AGI-2 challenge benchmark, demonstrating superior general pattern recognition capabilities in our vision-LLM model. • By structuring the model responses using custom token blocks, which improved interpretability and enabled systematic reasoning on abstract visual patterns. • This resulted in achieving 75+% resolution rate on the training set by optimizing the architecture for LIMO-based reasoning, significantly enhancing the model's ability to generalize across unseen pattern recognition tasks. I had a super incredible team: Aman Priyanshu and Mudit Sinha. A big shoutout and thanks to the organisers from Strong Compute for the amazing hackathon event and delicious food!!

Happy to wrap up another hackathon with Strong Compute! We proposed a novel way with https://lnkd.in/g4zsjAZG to reinforce fine-tune DeepSeek-R1-Distill-Qwen2.5-1.5B on GSM8K. Our results shown strong potential in enhancing LLM agent's ability with RFT to: 1. Perform better in multi-hop tool use settings - the ability break down a complicated problem into many small sub problems in a divide-and-conquer manner; 2. Plan ahead with CoT; 3. Better tool retrieval and better slot filling (experiment results appended below). To teach small model to do better with high school level math, we introduce the LLM agent with multiple tools like python_sandbox_tool, wolfram_alpha_tool, etc. With the help of GRPO/PPO and reward functions designed to reward correct answer + correct tool use format, the model quickly learns how to get the correct answer and call the right tool with right arguments instead of generating gibberish. This shows incredible potential for LLM agent post-training with RFT! Huge shout out to my third-time veteran hackathon teammate Hanyi (Laurie) Luo and Ms. upcoming Ph.D Maxine Liu. It was a fun all-nighter hacking at the Strong Compute hacker house. Also I want to thank Ben Sand, Miles Whiticker, Tim Smoothy and Rebecca Pham for granting us lots of top-notch GPU hours and great hosting you can ever expected at a hackathon. It's a privilege to have worked with y'all and can't wait what it can do in the future. Our slides: https://lnkd.in/gHWQNC4a cc: Daniel Pang, Tingyu Zhang, Jaymari Chua

We’ve been running GPU hackathons in San Francisco and Sydney to see what happens when you give smart people full access to compute. 100+ engineers across SF and Sydney got access to powerful GPUs for 24 hours. Some fine-tuned LLMs. Others trained chess bots on hundreds of thousands of games. A few built things we definitely didn’t see coming. We loved it. No infra setup. Just ideas → code → running jobs → results. Read the full recap and see what people built at our last event: 👉 https://lnkd.in/gcfJBQwi Join our next hackathon to work on ARC Prize or Fine-Tune Deep Seek April 18–19. Apply now — https://lnkd.in/g58tAs2H #gpu #ai #hackathon #infra #developerexperience #machinelearning #strongcompute #deepseek #arcprize #finetuning

🚀 Tech Events Roundup: Top Tech Events You Can't Miss 🚀 This week is packed with exciting technological events that are set to spark innovation and shape the future! 🔹GTC Hack Night - SF Bay Area Builders, Adam Chan, Chia Hwu & Dr Punnet Anand 🔗https://lnkd.in/g-RFPY7w 🔹Built on Bedrock Demo Nights - AWS GenAI Loft Events & Arte Merritt 🔗https://lu.ma/c65g5wju 🔹Networking in AI & Tech, hosted by Microsoft & NVIDIA - Microsoft Events & Insurgi Media Group 🔗 https://lu.ma/70uceufh 🔹Unstructured Data Workshop with Milvus and OpenAI - Milvus & Stefan Webb 🔗https://lu.ma/ncr07fnd 🔹Engineers Only: GPU Hackathon – Fine-Tune DeepSeek + Master Chess 🇺🇸 - Strong Compute, Rebecca Pham & Ben Sand 🔗https://lu.ma/7nbbmk0w #TechEvents #Innovation #TechnologyTrends #DigitalTransformation #Learning #Networking #BusinessNetworking #AIConference

Congrats to the Winners of Our Latest AI Chess + Deep Seek Hackathon! ♟️ 🐋 🎉 Last month, we wrapped up two epic hackathons—our Chess AI Hackathon and the first-ever Deep Seek Hackathon! It was incredible to see the energy, innovation, and cutting-edge research projects that emerged over just 48 hours. A huge shoutout to the winning teams and everyone who pushed the boundaries of AI, from mastering strategic gameplay to fine-tuning Deep Seek for creative use cases. 💡 Want to join the next one? 🔜 SF: March 21-22 https://lnkd.in/gS_bTH4a 🔜 Sydney: March 22-23 https://lnkd.in/grbwSzMB Engineers only. Spots fill up fast—see you there! 🔥 #AI #Hackathon #ChessAI #DeepSeek

Announcing Our 7th Hackathon: Fine-Tune DeepSeek + Chess AI ♟️🐋 Join us for an intense 2 day hackathon where you’ll have the option to fine-tune DeepSeek or build AI-powered chess bots—all while getting exclusive early access to Strong Compute’s newest features. 📅 Dates: 📍 San Francisco: Feb 21-22 📍 Sydney: Feb 22-23 🔗 Register now: SF: https://lnkd.in/g6kZFShe Sydney: https://lnkd.in/gYSsJPGU ⚡ Engineers only. Limited spots. Don’t miss out!

This week, China’s AI lab DeepSeek rattled global tech — particularly Nvidia, which suffered the largest single-day corporate value loss in US history on Monday. DeepSeek announced it had built a competitive AI model for just US$6 million using older AI chips, raising questions about the future demand for high-end chips. However, some Australian investors and founders aren’t sold on this narrative. They believe that this kind of innovation will ultimately drive AI chip sales up, not down. Capital Brief’s Daniel Van Boom spoke to Munro Partners’ Nick Griffin and Alex Pollak, CIO at Loftus Peak as well as Ben Sand founder of Strong Compute, a Y Combinator, Blackbird, and Peak XV Partners backed startup that helps companies develop AI more efficiently. Read the full story 👉 https://lnkd.in/g4aYXjci

Last chance to join our final chess hackathon of the year this Friday & Saturday in SF! Calling all engineers/AI researchers experienced with PyTorch. Help us refine the ultimate platform for building and training models while having a blast. 📅 Event details & registration: https://lnkd.in/g94c8DG6 ♟️ Check out the latest chess winners’ write-up: https://lnkd.in/grdXRBAg #chess #hackathon #ai #ml #pytorch

Scaling AI Research from 4 to 60+ GPUs: How Strong Compute Enabled InSite's AI for Construction Monitoring. Insite Project Solutions Pty Ltd is developing AI monitoring for Construction Projects. The challenge was getting large scale compute and managing 160 people across 26 research teams. With Strong Compute, Insite scaled from 4 to 60+ GPUs, enabling seamless collaboration across teams and accelerating AI development—all without the need for hands-on infrastructure management. Curious how Insite tackled these challenges and achieved: ✅ Real-time construction monitoring using high-resolution AI models ✅ Improved performance and cost efficiency ✅ A streamlined experience for his entire research team Check out the full case study here: https://lnkd.in/ghHVFDt6 Cian Byrne MACS CT, MAICD Consilium Solutions #AI #ConstructionTech #ResearchAtScale #StrongCompute #AIInfra #ClusterOps #FinOps #ML