Aktivitäten
3040 Follower:innen
-
Punjal Agarwal hat dies repostetPunjal Agarwal hat dies repostetAnimals and humans get very smart very quickly with vastly smaller amounts of training data than current AI systems. Current LLMs are trained on text data that would take 20,000 years for a human to read. And still, they haven't learned that if A is the same as B, then B is the same as A. Humans get a lot smarter than that with comparatively little training data. Even corvids, parrots, dogs, and octopuses get smarter than that very, very quickly, with only 2 billion neurons and a few trillion "parameters." My money is on new architectures that would learn as efficiently as animals and humans. Using more text data (synthetic or not) is a temporary stopgap made necessary by the limitations of our current approaches. The salvation is in using sensory data, e.g. video, which has higher bandwidth and more internal structure. The total amount of visual data seen by a 2 year-old is larger than the amount of data used to train LLMs, but still pretty reasonable. 2 years = 2x365x12x3600 or roughly 32 million seconds. We have 2 million optical nerve fibers, carrying roughly ten bytes per second each. That's a total of 6E14 bytes. The volume of data for LLM training is typically 1E13 tokens, which is about 2E13 bytes. It's a factor of 30. Importantly, there is more to learn from video than from text because it is more redundant. It tells you a lot about the structure of the world.
-
Punjal Agarwal hat dies geteiltPunjal Agarwal hat dies geteiltMeet ‘Instant Avatars’. Photorealistic avatars are likely to become the most powerful remote connection technology that has ever existed. The feeling of actually being together with other people despite physical distance. The opportunity to unlock remote social presence in a way video conferencing can only dream of. … but the current technology requires *a lot* of computing power. Now, imagine if you could instead easily create a photorealistic avatar simply by scanning your face with your phone. Technology like this will never replace meeting friends and family in person… but why not have truly awesome alternatives when being together in the same location isn’t an option. I’m incredibly excited about this work. #technology #innovation #virtualreality #meta
-
Punjal Agarwal hat dies geteiltPunjal Agarwal hat dies geteiltHow many times have we walked the wrong direction following #GoogleMaps? #Google is now piloting Augmented Reality walking directions into maps - what do you think of it? You can also follow me on twitter (https://lnkd.in/fjddMYP) and YouTube (https://lnkd.in/fcGsrfj) #innovation #technology #augmentedreality #ar #computervision #machinelearning #gps
-
Punjal Agarwal hat dies geteiltPunjal Agarwal hat dies geteiltThis wearable tech gives you a stylish way to track your lifestyle via http://bit.ly/2Nybucz
-
Punjal Agarwal hat dies geteiltPunjal Agarwal hat dies geteiltAugmented Reality Sneakers! This is just cool! I want to be able to try my new shoes like this. This is a great example of how Virtual and Augmented Reality for Retail and eCommerce (vCommerce) is going to take the world by storm. As traditional retail struggles with how to get people to come to stores, online retailers are doing the opposite, trying to find ways to serve their customers from the device in their hand most of the day. My prediction is that vCommerce will be the largest growing segment of retail marketing in the next 3-5 years. The only question is... Would you use this AR app to find your next pair of Jeezy's? #WannaKicks #AR #vCommerce #Sneakers #Marketing #Retail #Nike #Adidas #MetaVRse
Punjal Agarwal hat auf einen Kommentar geantwortet
1 Jahr
Aviral Bhatnagar How can one work with you at ajvc?
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt das🚨 University of California, Berkeley JUST OPEN-SOURCED FREETOKEN, AND THE RESULTS ARE WILD It is a new local inference engine running 2-4x faster than @ollama by exploiting Mixture-of-Experts architectures. The initial benchmarks point to a massive shift for local capabilities: → Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s → DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s → GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s A 35B model normally requires 70GB for weights. FreeToken serves it on an 8GB GPU because compute is no longer the bottleneck. Instead of loading the full model: > it targets the MoE router. > it measures your machine's bandwidth to dynamically split memory misses between the CPU and PCIe. This architecture is also a massive win for agents. Coding agents constantly rewrite history, forcing thousands of tokens through prefill. FreeToken saves checkpoints exactly at agent framework boundaries, dropping first-token latency from 232 seconds (llama.cpp) to under 44 seconds. Apache 2.0. OpenAI and Anthropic API compatible. 100% free and open-source. Repo: https://lnkd.in/ec88AU3E Paper: https://lnkd.in/ey9PURfd Shoutout to University of California, Berkeley for building this and making it open-source for the community 🤗 Don't forget to drop a ★!
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt dasA 10 million document RAG corpus takes 31 GB of RAM in float32. With pyturboquant, it fits in 4 GB. When Google released Gemma 4 this month, the community quickly adopted Google Research's TurboQuant algorithm to compress the KV cache during inference. TurboQuant is a data-oblivious vector quantizer that matches the Shannon lower bound on distortion — with zero training and zero data passes. That same property makes it extraordinary for the other side of the LLM stack: retrieval. I open-sourced pyturboquant — a PyTorch implementation of TurboQuant aimed at embeddings and RAG. What that unlocks for on-premise deployments: No codebook training. Add a document, it's indexed. True streaming ingestion. No rebuild when the corpus grows. Pure local. Pair it with an open-source embedding model and nothing leaves your machine. Drop-in LangChain VectorStore today; LlamaIndex and Haystack next. If you're building RAG where privacy or memory budget actually matters, this is worth a look. GitHub: https://lnkd.in/dt8xQ7he Paper: https://lnkd.in/dckD5VgV #RAG #LLM #OpenSource #Gemma4
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt dasThere is no "Training" and "Inference." There is only the update. For decades, ML has been stuck in a rigid cycle: 1. Train (massive compute, weights change). 2. Freeze. 3. Inference (weights are static, only context changes). What if that distinction is the reason our models can't learn continuously? I'm reading a fascinating new paper for this week's newsletter that proposes a paradigm shift called Nested Learning. It argues that the "Depth" of a network matters less than the "Frequency" of its updates. It proposes a model that never stops training, using a fractal-like structure of optimization loops nested inside each other. It’s one of the most theoretical yet practical papers I've seen this year. Make sure to sub to ML@Scale to make sure to get it!!
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt dasThis might be what a 10x engineer looks like in 2026. 😅 This picture of 🦄 Peter Steinberger, creator of OpenClaw, went viral on X last week. He shows off his setup of creating one of the most popular open-source projects ever on Github. He wrote his learnings down in a blog post called "Shipping at Inference-Speed". Some learnings: 1. He uses Codex by OpenAI for coding everything, he doesn't rely on Claude as it would otherwise "be too buggy". 2. He "doesn't read much code anymore", he only comes up with the high-level architecture, and then lets the LLM implement the pieces. 3. He runs many Codex instances in parallel using the Codex CLI, which is led by Thibault Sottiaux, a fellow Belgian 🇧🇪 in the Bay Area. 4. As Codex can feel a bit slow, he uses multiple screens to run various terminal windows in parallel. However, OpenAI is pushing hard in this regard, with a 40% speed improvement recently, and a partnership with Cerebras which will run the model even faster. 5. He pushes directly to main. 6. He heavily relies on the filesystem and Bash to make his agent do work, using CLIs, Skills, a "docs" folder, and an AGENTS.md file. Link: https://lnkd.in/ejNaERxy
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt dasMy closest friend, who is now a postdoctoral fellow after his PhD at Berkeley, would usually be found with his headphones on before our exams at IIT Bombay Most of us strugglers had no clue about what was taught in class (because we never went to any), so we would pretend to study till he finished. When I asked him why he would keep his headphones on he would say “They help me focus till I am able to break down every problem and solve it” He would not stop till he had it done, and would keep persevering. When we would ask him how on Earth he solved an intricate equation, he would genuinely laugh “I just got lucky this time!” His humility made him pretty easy to interact with and most of our (small) department would get taught by him. Despite him having his own exam, he used to take the responsibility of having all of us pass by teaching us. He ended up as department rank 1, probably more because he studied harder for us than himself! I realised that if you take others along with you, you will do well for yourself. College is a lot about learning a way of life, and as impressionable students we should focus on building a great way of life. As you can see, three skills that you should develop are perseverance, being humble and taking people along. These skills, if learned very early, will help you an incredible lot as you progress in your career
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt dasIndia has just 5 firms with a market cap of greater than $100B, with the youngest being Airtel at 31 years old US has 89 firms, the youngest being AppLovin at 13 years old China has 17 with the youngest being PinDuoDuo at 10 years India needs more $100B cos who are also young
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt das🥇 The paper we’ve been waiting for is finally here. Amazon Web Services (AWS) researchers just published a wild paper on arXiv, demonstrating that a 350M fine-tuned SLM outperformed several fine-tuned LLMs on tool use and agentic tasks. How did the experiment play out? First, they fine-tuned a SLM (facebook/opt-350m) on the ToolBench dataset for three specific tasks generally performed by LLMs: 1. Document summarization 2. Query answering 3. Structured data interpretation Then, they compared the performance of their fine-tuned SLM to proprietary CoT LLMs (ChatGPT, Claude) and open LLMs fine-tuned for tool use on the ToolBench dataset. The results were wild. The fine-tuned SLM outperformed ALL LLMs on all aspects of tool use on the ToolBench evaluation framework. They also report that: • 350M is the sweet spot for SLM size for tool use • The SLM learned to suppress irrelevant behaviors and focus better on tool-use only Why is this interesting? For a few reasons: • The SLM is very, very small — well under 1B parameters. The fact that it could outperform all LLMs in the experiment is very impressive. • The SLM is a decoder-only model that performs generative tasks. Many modern SLMs are fine-tuned to perform non-generative tasks like classification and NER. It’s important to remember that a SLM like this isn’t intended to replace LLMs across multiple applications and domains. A fine-tuned SLM can replace a LLM for a specific domain and application. But still, this is wild. 📄 Paper: https://lnkd.in/gTqmzfTg 🔗 ToolBench Repo: https://lnkd.in/gCCfyNyp
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt dasi was tired of stupid people on road so i hacked my helmet into a traffic police device 🚨 while i ride, ai agent runs in near real time, flags violations, and proof with location & no plate goes straight to police. blr people - so now ride safe… or regret it.
-
Punjal Agarwal gefällt dasPunjal Agarwal gefällt dasHabits that changed my life: 1. Sleeping before 11 PM 2. Drinking more water 3. Running everyday 4. Gym 5 days/week 5. Reading everyday 6. Writing every week 7. Zero sugar 8. Heavy breakfast, light dinner 9. Phone < 90 mins / day 10. Being grateful
Berufserfahrung und Ausbildung
Veröffentlichungen
-
A robust security framework for cloud-based logistics services
2018 IEEE International Conference on Applied System Invention (ICASI)
-
Epileptic Seizure Prediction over EEG Data using Hybrid CNN-SVM Model with Edge Computing Services
22nd International Conference on Circuits, Systems, Communications and Computers (CSCC 2018)
-
Effective Colour Reduction Using Grey Wolf Optimisation
European Congress on Computational Methods in Applied Sciences and Engineering ECCOMAS (2017): VipIMAGE (2017)
-
VGR-Net: A View Invariant Gait Recognition Network
2018 IEEE 4th International Conference on Identity, Security, and Behavior Analysis (ISBA)
-
A novel image detail preserving impulse noise removal algorithm
2016 5th International Conference on Multimedia Computing and Systems (ICMCS) (IEEE)
-
Natural Image Interpolation Using Extreme Learning Machine
International Conference on Soft Computing and Pattern Recognition SoCPaR 2016