Bhiman Kumar Baghel

🚨“How DeepSeek-R1 beat OpenAI at reasoning — a deep dive” 🧵 1/ I’ve launched a 3-part breakdown. In Part 1: – LLaMA 3 vs DeepSeek-R1-zero – MoE vs Dense – RLHF Let’s dive in 🧵 Full Video: 🔗 youtu.be/iLBdjINKR0U

@IterIntellectus Hi mate, there are a few people here claiming this is fake news. Rose and I were on Aussie national television this morning: youtube.com/watch?v=COYSRb…

this story is absolutely insane 🤯 > tech guy with zero biology background. > his dog got terminal cancer. > vets said 1 - 6 months left. > bro said nah not on my watch. > asked ChatGPT for a treatment plan. > sequenced tumor DNA for $3k. > used AlphaFold AI to model mutated proteins. > designed world’s first personalized mRNA vaccine for a dog. > partnered with universities to synthesize it. > ethics approval took 3 months. > vaccine design took 2 months. > first injection December 2025. > tumors shrank 75% within weeks. > dog happy. > universities confirmed it worked. > now designing version 2 for remaining tumor. AI + a guy determined to save his dog just outperformed the pharma industry 💀 the cure for cancer will be open source.

PittNLP @emnlpmeeting @pitt_sci #emnlp2025 !! @xiang_lorraine Dr. Litman @bhinubkb @JoeyHou5 @janetlauyeung César GS @YuyaAsano47 Multi-Lingual Word Grouping Games ✅ LLM Creativity ✅ Neighbor-Assisted Model Editing ✅ Free Response Math Problems ✅ Human Label Variation✅

🧵 9/ 🎓 I’m Bhiman, a CS PhD student @ Pitt, studying model editing, LLM reasoning, and efficient fine-tuning. As I go through my PhD journey, I share what I learn. Let’s learn together. 👉 Follow me: x.com/bhinubkb #phdlife

🧵 8/ Watch Part 1 with more details: 🔗 youtu.be/iLBdjINKR0U Follow for the next parts 👇 #DeepSeekR1 #GPT4 #LLM #RLHF #MoE #AI #MachineLearning

🧵 5/ LLaMA 3: → 405B parameters (all active during inference) DeepSeek-V3-base: → 671B total → Only 37B active at inference via MoE routing ⚡ That’s massive savings in compute during training and inference.

🧵 4/ Starting with the architectural difference … 💡 Instead of a dense decoder-only transformer (like LLaMA 3), DeepSeek-R1-zero uses a Mixture of Experts model. 🧠 Only a subset of experts are active per input → faster inference + modular specialization.

🚨“How DeepSeek-R1 beat OpenAI at reasoning — a deep dive” 🧵 1/ I’ve launched a 3-part breakdown. In Part 1: – LLaMA 3 vs DeepSeek-R1-zero – MoE vs Dense – RLHF Let’s dive in 🧵 Full Video: 🔗 youtu.be/iLBdjINKR0U

🧵 3/ 🔧 Let’s start with the training pipeline. Conventional LLMs typically follow: 1. Pre-training (language learning) 2. Supervised fine-tuning 3. RLHF (reinforcement learning with human feedback) DeepSeek-R1 uses a twist on this…

🧵 2/ 📌 Why is DeepSeek-R1 a big deal? It beat OpenAI’s O1 (their reasoning-tuned GPT-4) on math & code — traditionally the hardest LLM tasks. And it did it using mostly reinforcement learning.

Do these expenditures seem like a good use of your taxes?

“That's one small step for man, one giant leap for mankind” - Neil Armstrong

Updating all my NeurIPS papers.

@likhithnj @outofofficedaku Not required in transit

@outofofficedaku London stopover, transit visa would be required right?

Delhi ✈️ New York Return Jan - Feb - Mar 2025 Rs. 57K 🎒💼 Fly with British Airways Share | RT #DakuFlightDeals

@sondesix Innovation 💡

We've seen green or pink lines before, but have you seen white lines on a phone's screen before? 🤔

@Akshat_World Holistic planning is crucial.

Burj Khalifa was built in 2004. In 20 years, it helped transform Dubai's economy. Lakhs of Tourists come to see Burj Khalifa. Most of them international tourists, which adds more money to Dubai's economy. Now, Indian politicians look at this case study: and say "let's build giant structures, tourists will come, jobs will be added. And, money will flow like river" In fact: many politicians in India have tried emulating this model. For example: In Noida, Mayawati spent 2600 Crore of Public Money on building numerous statues. Did it transform Noida's economy? Absolutely not. What people miss is: a structure is the centerpiece. Planned infrastructure needs to be built around that (that's the hard part). Near Burj Khalifa, for example:- - Females can go out at the middle of the night (safe) - Easy to get a Cab/travel (good transport) - If you have a baby, you can use a baby stroller. And walk for kilometres. A developer called EMAAR not only built Burj Khalifa, but did the masterplan development for entire Downtown. Their vision was to give tourists a great experience. And, they succeeded. Unfortunately, in India: the vision is to build a structure. Make reels and vlogs with it (take media credit) Therefore, we have reached a point where even our newly constructed Parliament is leaking, despite spending several crores.

@TalkingTree__ @punekarnews Dekh chunotiyon se nahi chutiyon se pareshan hai

🚨 Disturbing Video Alert 🚨 #Pune: A tragic incident occurred in Colony No. 2, Ganeshnagar, near Bopkhel, Pimpri Chinchwad, Pune, where a minor girl lost her life as a gate fell on her on Wednesday (31st July 2024). The heartbreaking event was captured on CCTV. Parents keep a watch on your children and please ensure the safety of your surroundings and regularly check for any potential hazards to prevent such unfortunate accidents. #Pune #TragicIncident #Ganeshnagar #PimpriChinchwad #Heartbreaking #StaySafe