Aravind Srinivas

Nice to have 'Top New Release' status on Amazon for the latest book (my 25th) that explores an alternative paradigm for quantum computation. Illustrations (111) by Dr. Narendra Madhav Joshi. My take is that we owe it to ourselves at this early stage of the industry's lifecycle to explore alternative approaches. Here is a brief overview: "This book attempts to reveal something of the vastly different and fundamentally creative quality of computing that must accompany any computation involving the quantum-levels. To linearly project digital computing laurels manifest as increasing speeds and the ability to process vaster amounts of information, as the inevitable trajectory of quantum computing is perhaps, in the aphorism attributed to the Buddha, to look only at the finger and to miss the moon and the sky that it is pointing to. The cover figure highlights all that is being missed, summarized as an egg-like structure synonymous with the term ‘Hiranyagarbha’ in Sanskrit. This depicts the womb of creation abundant with many layers and patterns reflecting luminescence, due to the constant superposition and entanglement inherent to the quantum computation that derives from a light-centered interpretation of quantum dynamics." The book is available here: https://lnkd.in/gXu4H9uW #quantumcomputation #quantumtech #technologyleadership #cosmology #systemsthinking

16

5 Comments

I believe a solution to this is currently being developed.

5

if you know, you know Amjad Masad of Replit x's. & paul graham of y combinator demonstrate precisely what (ai) OpenAI nor Google DeepMind nor Perplexity nor any other frontier model, can achieve....personal...experiential...unique...at times communal...context & meaning. Amjad Masad discovers the McDonald's in germany. his perspective living in the u.s. is this https://lnkd.in/eT-h86tJ. this is (his) experience as anyone in the u.s. experiences it. right? then a trip. crossing a border. but it's more than that. it's crossing a mindset...a culture...and another way of advertising the same 'ting somewhere else. germany. Amjad Masad discovers that in asking for a quarter pounder there is a blank stare. vaaaat. do. you. vant. sir. a quarter pounder. we don't have that here. what do you mean. this is McDonald's. sorry sir. we don't have that here. wait. wait a minute. there it is. that picture over there. i recognize it. that's what i want. (here image supersedes audio ;) ) ah. royal. wtf says Amjad Masad. royal. wtf. in chimes in paul graham. he replies with a movie reference. clearly one that masad has not seen. here https://lnkd.in/eqs-Zg2U pulp fiction. https://lnkd.in/erGURkRx at the time quintin tarantino https://lnkd.in/enQ6M2H8 became a phenomenon. a former video store employee who had a library of movies. unlimited. he watched all of them. also had a love of video games. and also a love of foreign famous japanese film directors, cinematographers. a retrieval machine can and does have the ability to retrieve these 'tokens'. what it does not in this moment offer is a time machine a time machine where the context and meaning is transferred, transposed over time from one spot to another seemlessly. now that would be magical wouldn't it. and so in this X we have paul graham attempting to take what we experienced at the time. pulp fiction. was a runaway hit. indeed a palme d'or at cannes. the olympics of cinema. quintin. a lover of film. respectfully, artfully, pirately crafting into a new perspective. we all felt it. we all watched it. many times over. yet here we are back at the beginning. which is to say that removing yourself from one cultural bubble you notice the differences. you notice the little 'tings. and with that offers a new experience. Amjad Masad has crossed the cultural chasm. and with that is opening his mind to new euro/german/berlin ways of communicating the same 'ting. ahhhhhh. who will (b)e the one to cross this context, meaning window will rule the world truly. as the greatest communicators are indeed the greatest leaders. clear in their perspective. culturally sensitive. aware of themselves. and when they ask for a quarter pounder in the future in euro land it will immediately jump to. hey. what's your order? royal. yeah. a royal. pg, in euro. a royal. with cheese. as he remains a pulp fiction, quitin fan. ;) < . >

4

6 Comments

It took us < 30 mins to integrate Llama3 to our health AI stack. It takes our customers < 3 mins. Within a week Llama3 went live with all the 101 GenAI healthtech customers. We’ve curated and integrated 15+ models over last 7+ months to enhance healthcare workflows. Don’t spend weeks or months integrating with just one model; instead have a “team of models” power your “network of copilots”. Great post by Prashanth Aditya Susarla about how to build a high performance AI product with team of LLMs.

16

Hiring full-time NLP engineer for fast-growing and funded generative AI startup! At Arrowhead, we’re becoming the AI partner for enterprises’ customer support and sales teams. We are working with enterprises globally and are targeting close to $1 mn in ARR by end of the year. We’re building proprietary in-house models that will create 100x value compared to what we can do today. At Arrowhead, you’ll experience a culture where:  🚀 You will have ownership and empowerment to take decisions, even as an intern 🚀 You will see what you’re building directly in the hands of users, and be able to iterate on a weekly basis based on user feedback 🚀 You will be challenged and pushed to do some of the best work you’ve ever done! If you are interested, please fill out the google form in the comments below 👇 #hiring #NLP #startuphiring

136

11 Comments

Great work here Rohan Sidankar . Thanks to Rohan and Justin, hushh is known for its work in optimizing neural network search through dimensionality reduction, specifically with ViT-B-32 CLIP for various search tasks as we build out our search index. This allows for maintaining recall while reducing vector size, enhancing search efficiency and effectiveness against large datasets like the LAION Benchmark Repository. Dimensionality reduction impacts neural network search by enabling the maintenance of recall while reducing vector size. This optimization, as seen in Rohan Sidankar's work at Hushh with ViT-B-32 CLIP, enhances search efficiency and effectiveness against large datasets like the LAION Benchmark Repository. ViT-B-32 CLIP can enhance search tasks by allowing for dimensionality reduction in neural network search, as demonstrated by Rohan Sidankar's work at Hushh. This approach enables the maintenance of recall while reducing the vector size, which is crucial for efficient and effective search tasks against large datasets like the LAION Benchmark Repository.

11

Wondering what top research got published this last week? 📑🤖🤔 So essentially, 1. "AI Model StyleFeatureEditor gives highly detailed image editing results!" 2. "We have a Universal Quantum Chemistry dataset to find new drugs now!" 3. "Stylebreeder clusters diverse art styles from 6.8M images on Artbreeder, enhancing personalized AI-driven artistic expression." 4. "LongRAG addresses the limitations of traditional RAG by using long retrieval units" 5. "Language model tuned to avoid toxic language in English will also avoid it in other languages!" 6. "YOUDREAM Auto-Generates Perfect 3D Creatures from Text!" From cutting-edge research on Artificial Intelligence to advances in Machine Learning, Computer Vision, and Neuroscience, we distill the latest technical papers into bite-sized insights you can apply today! For more subscribe at https://lnkd.in/gJ_wJxAC

21

1 Comment

Really enjoyed the SWE-Agent technical paper. My overall takeaway is the importance of context management for building complex agentic workflows. There are a lot of key components necessary for LLM-based agents to autonomously complete software engineering tasks. This paper goes into a few areas like dealing with error propagations, how to efficiently manage context, simplifying actions, and the importance of navigation/search. If you are curious about my detailed thoughts on this paper, I just published a summary on my YouTube channel. You can watch at 2X here: https://lnkd.in/eUmpCqZ3

78

1 Comment

𝗪𝗶𝗹𝗹 𝗔𝗜 𝗛𝗶𝘁 𝗮 𝗗𝗮𝘁𝗮 𝗪𝗮𝗹𝗹 𝗶𝗻 𝟮𝟬𝟯𝟬? 🆘 Very interesting analysis from epoch.ai about data corpus available for training large language models and whether we will hit the ceiling soon. https://lnkd.in/gD78f9DW The authors estimate that all available data will be used by ~2030, but they're also the first to admit that by then we will have ways to address "the shortage" (such as synthetic data, data efficiency improvements, use of other modalities). What I found more interesting though is why the latest predictions differ from their 2022 paper, where they speculated that all the data will be used by 2024: ☑ Web data proves more valuable than expected (on par with scientific papers) ☑ Models can "reuse" datasets multiple times 2030 is right around the corner, so we will see a lot of innovation on this front and creative ideas on how to 'recycle' and make best use of the data we already have. These estimates also don't take into account proprietary data owned by individual companies. In this No priors episode (https://lnkd.in/g2BkTJBz) there is a mention of JPMorgan owning 150PB of proprietary data, which is at least 150x of what was used to train GPT-4. JPMorgan is monetizing this data today (https://lnkd.in/gXjE92qx), but it gets exponentially more valuable in a world where data is a finite (and extremely sought after) resource.

12

If you’ve hung around long enough on Linkedin you notice a lot of AI counter programming and anti-bodies from loud voices. Maybe because they are often not business types, they don’t realize that tech working 75% or more is plasterer over by endless bandaids and tweaks until it gives you an edge in business. We cobble together dozens of SaaS products, fill in the holes with humans and processes. Everyone knows it is a huge hack that only sort of works. Maybe the mistake was to call this new thing intelligent that caused the anti bodies to appear. The silent majority just want to get something done and AI works for them. #ai #aiadoption

3

1 Comment

RLHF was thought to be the secret behind aligning ChatGPT to give 'smart' responses. Implementing RLHF using PPO optimization is a complex engg task but initially there seemed to be no way around it. 🛠️ The road from the original complex RLHF implementation to the more recent simpler model fine-tuning/alignment strategies is long and unwinding. 👉 The initial RLHF setup used 3 (or 4?) copies of the pretrained model. First is the next-token pretrained model (also called SFT model), one for external reward model and one for the new model being fine-tuned. PPO algorithm was used for learning. Led to many asking do we really need RLHF/PPO ? 👉 DPO got rid of the external reward model by reformulating the loss fn and reduced the copies to two - the original SFT model and the new model being trained. 👉 SPO finally reduced the copies to a single one. They construct a new intrinsic loss function which trains the model while learning from (good, bad) samples. The average log probability of a sequence is used as the implicit reward, same as the usual generation likelihood. Very clean formulation and writing. 🤔 In hindsight, SPO is obvious but why so long? (ORPO was close enough). Two key ideas seem to make it work: length normalization of loss and an extra margin parameter. 🤨 While both DPO and SPO claim full victory on compute and memory efficiency, the jury is still out on whether they yield “better” models than PPO (used in RLHF) setting, on say downstream reasoning tasks like GSM8k. 🎇 I really wish that implementing alignment training becomes as easy as next-token training. Hope that SPO holds its ground and whatever remains of PPO can be brought back to SPO++. --- Stay informed about latest LLM updates with our newsletter: offnote.substack.com

39

3 Comments

Hallucinations in LLMs are inevitable, no seriously! Josh Fritzsche and I have worked hard at Voze to find a way to tune, prompt and even plead 🙏 with LLMs to prevent them from the occasional 🍄 hallucination. It seemed like the better we have done the harder the edge cases have been to solve and eventually we realized that they tend to frequently happen when LLMs predict or make anticipations on "nothing". I remember at one point nearly going insane with frustration because when LLMs had a hallucination, they were EPIC! 😱 Then we pushed through it and benefited for the effort! How did we solve this common problem? We didn't (Not completely), we got better at identifying the issues as they happened. Once we identified in the data we were able to start looking for patterns. The same catalysts that cause it to happen once if left unchanged caused it over and over again. Our solution is astoundingly simple. Once you know that it's likely to happen..... let it! Accept it will 🍄, but also be preemptive in not focusing on it and use alternative data sources, generators or providers. Change the conditions that cause it but continue to monitor it as it happens. As Dan Caffee is fond of saying to me on a regular bases. "Dusty we need the system to be adaptive and provide mechanisms to learn from it's past." Well he is right! All we needed was data and the people that contribute to it's quality. Huge props 👏 to strong support from Kathy and Janelles audit teams. We were armed with a knowledge and a desire to learn from past mistakes. Armed with confidence that we could correct these problems we could focus on the data to prevent their impact going forward. When doing millions of LLM (Generative AI) calls regularly Josh and I have learned that hallucinations will happen. We also know exactly how we will handle and prevent it going forward! Failure is not doing anything and accepting defeat.

4

2 Comments

I'm thrilled to share an exciting breakthrough in the world of large language models (LLMs) - a game-changer that eliminates the need for matrix multiplication (MatMul) operations, without compromising on performance! In the recent paper "Scalable MatMul-Free Language Modeling," researchers have introduced a novel approach that utilizes ternary weights and element-wise operations. The results are nothing short of amazing: 🔹 Performance: MatMul-free models deliver performance on par with traditional Transformers, even scaling up to 2.7 billion parameters! 🔹 Efficiency: By removing MatMul operations, the researchers have seen reduced memory usage by up to 61% during training. 🔹 Speed: Inference speed has increased by a staggering 4.57 times, making real-time applications more viable than ever before. See the results in the slides below. This innovation doesn't just promise better performance; it represents a significant leap towards more efficient and scalable AI solutions. Whether you're working with #gpu or #fpga, this approach can drastically cut down on computational costs and energy consumption. Read the full paper here - https://lnkd.in/dBaeNGmm As someone deeply invested in the future of AI and machine learning, I can't wait to see how this technology evolves and gets adopted across various industries. The potential applications are vast, from natural language processing to real-time data analysis. #AI #MachineLearning #Innovation #LanguageModels #TechTrends #FutureOfAI

20

3 Comments

In the first innings of GenAI, much of the attention and $s have flowed into foundation models and GPUs. But that could change significantly. The foundation model layer is moving so fast that it may soon commoditize itself and move the action to higher layers of the tech stack. New, increasingly capable foundation models are coming out every day. Open source models are catching up rapidly with premier closed models. AI apps are being designed to select models for each request based on cost and task complexity. Model pricing is coming down rapidly even as quality goes up. For most pure-play, horizontal foundation models, marginal pricing may approach marginal cost, even as usage rises exponentially. In previous platform shifts such as PC, Internet and Mobile, a large portion of value eventually accrued to applications (e.g. Google, Amazon, Facebook), software platforms (e.g. Windows, Cloud platforms, Salesforce), and full solutions (e.g. Apple, Tesla). Some leading pure-play innovators in the hardware and enabling layers did create huge successes (e.g. Intel, Cisco, TSMC), but wider value accretion over time happened in the layers above. I did a rough, directional exercise looking through the top 50 public tech companies today by market cap, and classified them by the primary category they operate in (highly subjective, as many large companies eventually straddle multiple categories). Over half of the ~$20T combined market value of these top 50 tech companies comes from those I classified as Applications and Platforms. This included Microsoft, Alphabet, Amazon, Oracle & many others. A quarter came from those classified as Full Solutions, including Apple & Tesla. And only about a fifth is from those classified as Hardware or Enabling Tech, including Nvidia, TSMC, Intel et al. This is in spite of the rapid recent ascent in market caps in this category from the promise of AI. In contrast, the list of top 10 tech companies by market cap in the early Internet era frequently included pioneers in Hardware and Networking (key Enabling Tech for the Internet), including Cisco, Intel, Sun, Dell, Compaq, Lucent, 3Com, AT&T and Silicon Graphics. A large % of the premier tech talent in those days wanted to work in Hardware and Networking. Investors were disproportionately focused on those areas too. This led to those layers becoming so good that much of the new value accretion moved to the layers above. Similarly, apps and platforms for the enterprise AI era are just getting started, and that’s going to be a multi-decade, multi-trillion journey. Some leading foundation model players will expand to offer wider platforms or full solutions they can monetize better. Others leading models may come from those with a highly profitable, data-generating business at the layers above. But there will be many, many opportunities for pure-play app and platform companies that smartly build upon the underlying layers that seem to be getting ready for prime time.

57

4 Comments

“Artificial intelligence, deep learning, machine learning  are everywhere — whatever you’re doing it's going to affect you, if you don’t understand it — learn it. Because otherwise, you’re going to be a dinosaur within three years.” Mark Cuban, Entrepreneur and Investor AI has taken the world by storm and has already changed how we work. Now is the time to get ahead of the curve and learn about these emerging technologies. Launching Minor in AI from IIT Ropar. The course is designed to provide a comprehensive AI education to students and early-career professionals. It focuses on equipping participants with advanced AI tools and techniques through a hybrid setup. Course Overview > 15 credits > Course Designed by Prof. Sudarshan Iyengar Head CSE Dept, IIT Ropar > 12 months, part-time program Click here for more details https://lnkd.in/gTwY8ejE

42

1 Comment

India needs lot of electricity in the future for AI? No, it just needs 1-bit LLMs. Bhavish Aggarwal made an interesting comparison between training AI models and running an Ola S1 scooter. “1 H100 NVIDIA GPU consumes 30x electricity in a year as an Ola S1 scooter....Need a lot of electricity in the future!” On the other hand, Pranav Mistry disagreed with Aggarwal completely. “No, you need optimised AI models like SUTRA Light and innovations like 1-bit LLM.” Though this might simply be a promotion of TWO’s new AI model, the part about the 1-bit LLM does make sense. Mistry further said that Aggarwal probably overlooked the fact that most of the energy consumption by H100 is during training the model, which is a one-time process. “ Inference is the main energy-hungry task,” he said. For inference, H100s are not the most-efficient GPUs. The conversation around 1-bit LLM started around February, when Microsoft released its paper titled The Era of 1-bit LLMs. “One-bit LLMs open new doors for designing custom hardware and systems specifically optimised for 1-bit LLMs,” said Furu Wei, one of the researchers of the 1-bit LLM paper. Wei explained that these quantised models have multiple advantages as they fit on smaller chips, require less memory, and have faster processing. This substantial reduction in bit usage is the cornerstone of the proposed model. It performs as well as the traditional ones with the same size and training data in terms of end-task performance. Most importantly, it is more cost-effective in terms of factors like latency, memory usage, throughput, and energy consumption. This is why it is necessary for Indian AI research in the future. Though the performance of these 1-bit LLMs in Indic language is yet to be measured and evaluated, discussions on Hacker News point to the fact that the existing models can also be converted into 1-bit LLMs in the future using the changed hardware. India is mostly experimenting with AI’s inference part. One-bit LLMs are the way forward for India’s AI models, not just increasing the electricity requirement since that may not be the ideal way forward. https://lnkd.in/gqUhHg6K

63

4 Comments

Looks like fine tuning out performs RAG and Large context LLMs for data accuracy with longer contexts Paper: https://lnkd.in/gj9-3jsK Summary of "BABILong: Testing the Limits of LLMs with Long Context Reasoning-in-a-Haystack" - Introduction to BABILong Benchmark: - Designed to test LLMs' reasoning abilities over extremely long documents. - Comprises 20 diverse reasoning tasks, such as fact chaining, induction, deduction, counting, and handling lists/sets. - Uses natural text from the PG19 corpus, extendable to document lengths up to 1 million tokens. - Current LLM Performance: - Popular LLMs utilize only 10-20% of the context effectively. - Performance declines with increased reasoning complexity. - Retrieval-Augmented Generation (RAG) methods achieve 60% accuracy on single-fact QA, independent of context length. - Top Performing Models: - Recurrent Memory Transformers (RMT) show the best performance, processing up to 11 million tokens. - Small models like Mamba (130M) and fine-tuned RMT (137M) achieve high accuracy on long-context tasks. - Evaluation Findings: - Current benchmarks (e.g., Longbench, L-Eval) are insufficient for LLMs with capabilities beyond 40,000 tokens. - BABILong tests models' efficiency in distinguishing relevant facts from irrelevant text over long contexts. - Popular LLMs fail to maintain high performance as context length and complexity increase. - Fine-Tuning Insights: - Fine-tuning improves model performance significantly. - Mamba and RMT demonstrate successful QA with long contexts, outperforming RAG models. - RMT's memory mechanism allows efficient processing of long sequences, with marginal quality degradation up to 11 million tokens. - Benchmark Significance: - BABILong provides a rigorous evaluation framework for LLMs' long-context reasoning abilities. - Highlights the limitations of current LLMs and the need for improved context processing mechanisms. - Key Contributions: - Introduction of BABILong, a scalable generative multi-task benchmark. - Evaluation of over 20 recent LLMs on various context lengths and tasks. - Demonstration of the limitations of current models in long-context utilization. - Setting a new record for sequence size processed by a single model with RMT.

4

5 Comments

Two AI Releases SUTRA: A Multilingual AI Model Improving Language Processing in Over 30 Languages for South Asian Markets SUTRA’s architecture comprises two mixture-of-experts transformers: a concept model and an encoder-decoder for translation. The concept model is trained to predict the next token, leveraging publicly available datasets primarily in languages with abundant data like English. Concurrently, the translation model learned from 100 million human- and machine-translated conversations across multiple languages, allowing it to map concepts to similar embeddings in all languages it supports. The innovative integration of these models involves the translation model’s encoder generating an initial embedding from the input text, which the concept model processes and feeds into the translation model’s decoder to produce the final output. This approach ensures that SUTRA can effectively handle a diverse range of languages, making it a robust tool for multilingual communication. SUTRA is available in three versions: Pro, Light, and Online. SUTRA-Pro and SUTRA-Online offer high performance and internet connectivity at $1 per 1 million tokens, while SUTRA-Light provides a low-latency option at $0.75 per 1 million tokens. This pricing structure makes SUTRA an attractive option for users and businesses in cost-sensitive markets. Read our take on this: https://lnkd.in/gSN-Rk85 Paper: https://lnkd.in/gpDbGDCe? Model: https://www.two.ai/sutra? Chatbot: https://chat.two.ai/ TWO.AI

35

2 Comments