Snehal Talati

I have been following arXiv publications, and over the past couple of years, it has become evident that significant advancements in AI research are predominantly led by companies such as Meta, OpenAI, Microsoft, and Baidu, as well as various startups involved in the Generative AI (LLM/LMM) development ecosystem. In contrast, Academic research institutes seem to be generating less attention. This brings up following questions: 1. Has the center of progress in AI research shifted more towards industry, because it has become increasingly dependent on substantial financial resources? 2. Is academic research becoming less impactful as many prominent researchers join these leading organizations? Much of the current academic research appears to focus on enhancing industry efforts, which is a notable shift from a few years ago. P.S.: I acknowledge that my perspective might be completely incorrect, however insights into these questions could be valuable for aspiring researchers in making informed career decisions. #AIResearch #TechIndustry #AcademicResearch #CareerAdvice #ArtificialIntelligence #ResearchTrends

24

1 Comment

🚀 Beyond ChatGPT: Exploring the Diverse World of LLMs Curious about how different AI language models stack up? So was I! After experimenting with ChatGPT-4o, Claude 3.5 Sonnet, and Perplexity, I've put together a comparison of these leading LLMs. Check out the visualized table in this post for a detailed breakdown of my findings, including research from Perplexity. But here's a quick summary of what I've discovered through hands-on experience: 📊 Each LLM has its own strengths: - ChatGPT-4o: Excels in creative writing - Claude 3.5 Sonnet: Shines in factual accuracy - Perplexity: Stands out with real-time info and source citations My personal workflow: 🔍 Research: I lean on Perplexity. Its real-time information and solid citation game are game-changers! (Side note: I even used it to plan my weekend getaway - but that's a story for another post! 😉) 🎨 Content Creation: - For creative projects, ChatGPT-4o is my go-to. - When accuracy is key, I trust Claude 3.5 Sonnet. - Pro tip: I always have Claude apply my writing style guide to ensure a cohesive tone and voice across my content. Thanks Kieran Flanagan for coming up with this ingenious workflow! The takeaway? These LLMs aren't competing – they're complementing! It's about choosing the right tool for the type of task you’re undertaking. Remember, this is just scratching the surface. The AI world is evolving rapidly, and these insights might change tomorrow! What's your experience with different LLMs? Share your thoughts below – I'm always eager to learn more! #AIComparison #LLMs #ContentCreation

14

4 Comments

Discover "Jury Learning: Integrating Dissenting Voices into Machine Learning Models," an innovative study by researchers from Stanford University and Apple Inc. This groundbreaking approach challenges the traditional majority vote system in supervised machine learning by incorporating diverse perspectives explicitly, akin to a jury's function in a courtroom. Instead of overriding minority views, Jury Learning models each annotator individually, allowing for dynamic jury compositions that better reflect varied stakeholder perspectives. For instance, in tasks such as online toxicity detection, the model might feature jurors who are often targets of online harassment, like women and Black individuals. The study introduces a deep learning architecture that samples from annotators’ models to populate the jury and run inference to classify data. This approach enables the creation of juries that can adapt their composition dynamically, explore counterfactuals, and visualize dissent among jurors. A field evaluation revealed that practitioners using Jury Learning constructed more diverse juries, significantly altering 14% of classification outcomes compared to standard methods. This shift highlights how Jury Learning captures individual jurors’ views more effectively than conventional aggregated models. Check out the study: https://lnkd.in/g33bZ3vA #MachineLearning #AI #Inclusion #TechInnovation #MLResearch #Jury

5

I've been following the AI debate closely, and there's a critical aspect that's often overlooked: the risks and harms of delaying AI implementation. Why are we more concerned with a single accident caused by an autonomous vehicle than the millions of accidents that could be prevented by reducing driving human error through AI? As a physician, I've witnessed the devastating impact of poor adherence to heart failure treatment guidelines. Why focus solely on potential errors in diagnostic and treatment related AI tools when millions of lives could be improved or saved by accelerating their use? The ethical considerations are indeed complex, and I don't have all the answers. However, it's imperative to discuss the profound harm caused by fear and bureaucracy in delaying AI deployment, especially when it has the potential to save lives and alleviate suffering on a massive scale. Let's push for a balanced perspective and a more profound conversation on this critical issue. #AI #Healthcare #Innovation #Ethics #FutureTech #SaveLives #ReduceSuffering

20

2 Comments

The subject of "Mind Control" is the remit of science fiction writers and the immense fecundity of their imagination. Elon Musk's Neuralink is in the process making those works of fiction prophetic. With Brain Computer Interfaces (BCI), Neuralink is enabling the control of a computer or mobile device through the human mind. Yes, you read that right! With a fully implantable, cosmetically invisible chip in the human brain, neural signals are wirelessly transmitted to a smartphone app that decodes the signals into actions. This has the potential of aiding people who are physically challenged, restoring their independence and transforming their lives for the better. We are truly living in a time of wondrous innovations. You can read more details of this fascinating device at https://neuralink.com/

9

The competition between "open-source" and commercial models is becoming indirect. - The former (open-source) are becoming focused (e.g. visual, domain-specific, SLMs for mobile etc) ... Thanks to Hugging Face for accelerating this change. - Whereas the latter (commercial) are for-the-most staying general use. The developers are investing in training methodologies + wrappers to ensure effectiveness (reasoning, reproducibility etc) when deployed in end-customer software products. The above strategies are still work-in-progress, but emerging model developers need to do their SWOT analyses. 🫣 Kalyan KS, Philipp Schmid, Elvis S., Aishwarya Naresh Reganti, Anthony Alcaraz, Pavan Belagatti, Eduardo Ordax Am I missing something? Please punch holes in my observations above. 20/15 Visioneers HitchhikersAI

2

👉 The Foundations of AI Agents Architectures 🤖 Single Agent System: This is where it all begins. A single agent, often an LLM, calls on one or multiple tools. For example, An LLM that calls on Google to search the web. It's the simplest form of agent system and serves as a starting point for more complex architectures. 🤖🤖 Network of Agents: Imagine multiple agents, each with their own tools, communicating by deciding who goes next. However, this architecture can be too loose in communication patterns, leading to reliability issues. 👮🏻‍♂️ Supervisor Agent Approach: Here, a single agent acts as a supervisor, routing tasks to other agents. This makes the system more manageable as subagents focus solely on their tasks without deciding the next steps. 👮🏻‍♂️🔫 Supervisor with Tools Architecture: A variation where the supervisor agent uses subagents as tools. The communication is simplified as the LLM generates tool call parameters, which are passed to subagents. 🦁 Hierarchical Approach: This involves layering supervisor agents, where each supervisor can call another subagent, which itself might be a supervisor. This architecture is beneficial for organizing subagents in specialized ways. 💍 Custom Cognitive Architectures: Highly customized systems tailored to specific domains. They borrow aspects from other architectures but are designed to meet the unique needs of the domain they serve. To learn more about these ai agents architectures please join our AI Masterclass. To get the invite type in the comments: #aimasterclass #artificialintelligence #competitiveintelligence #pharmaceuticals #biotechnology

39

14 Comments

These smaller, faster, more accurate models can be used in RAG-based chat applications, enabling smaller organizations to demonstrate and build more on GenAI. #GenAI #OpenAI

15

Here is an idea for a new AI startup that can have significant upside potential with rapid scale: Tackling enterprise code merge and Git management problems. This has been a major challenge and I still see this issue persist. We need a fully AI managed CI/CD. Current tools are still old school. I will take 2% equity for the idea and to mentor the team/startup into a winning proposition (funding, product, market access) 😇 My back of the envelope math says this is a $5billion TAM. Yup TAM. There will be 45million human developers and probably an equal number of AI agents coding and merging code by 2030. This issue requires a ground-up rethinking and redesign. As development teams grow, codebases scale, and branching strategies become complex, the challenge of managing code merges becomes increasingly gnarly. Large teams often encounter difficult merge conflicts, review bottlenecks, and inconsistent processes. This is compounded by the need to integrate with CI/CD pipelines, automated tests, and quality checks, which can slow down development and cause significant disruptions. Also, picture an AI that learns the structure of your codebase, tracks developer patterns, and autonomously predicts where conflicts might arise, offering instant solutions before they become roadblocks as well. It could intelligently handle branching strategies, propose code changes to prevent conflicts, and streamline pull request reviews with context-aware recommendations. An AI-powered solution could optimize and automate much of this process, detecting potential conflicts early, suggesting resolutions, and streamlining reviews. With the right approach, this startup could significantly improve developer efficiency, reduce downtime, and accelerate the release cycles for large-scale enterprise teams. #startup #idea #AI

7

💥💥💥 $100K or 100 Days: Trade-offs when Pre-Training with Academic Resources Apoorv Khandelwal, Tian Yun, Nihal V. Nayak, Jack Merullo, Stephen H. Bach, Chen Sun, Ellie Pavlick Abstract Pre-training is notoriously compute-intensive and academic researchers are notoriously under-resourced. It is, therefore, commonly assumed that academics can't pre-train models. In this paper, we seek to clarify this assumption. We first survey academic researchers to learn about their available compute and then empirically measure the time to replicate models on such resources. We introduce a benchmark to measure the time to pre-train models on given GPUs and also identify ideal settings for maximizing training speed. We run our benchmark on a range of models and academic GPUs, spending 2,000 GPU-hours on our experiments. Our results reveal a brighter picture for academic pre-training: for example, although Pythia-1B was originally trained on 64 GPUs for 3 days, we find it is also possible to replicate this model (with the same hyper-parameters) in 3x fewer GPU-days: i.e. on 4 GPUs in 18 days. We conclude with a cost-benefit analysis to help clarify the trade-offs between price and pre-training time. We believe our benchmark will help academic researchers conduct experiments that require training larger models on more data. 👉 https://lnkd.in/d5Vh3ycg #machinelearning

1

Recently Google has pulled back on some of the AI capabilities it rolled out in support of search.    While I’m confident that the skilled engineers at Google will take this practical user testing and turn it into an even better product, it does remind us of how critical robust training data – and a well-defined use case – is in AI applications.    We’ve been talking a lot lately about AI for ag, and the ways in which it can enhance agronomic knowledge when done right.    Crop Science CIO Amanda McClerren and I will be recording another Inside Innovation conversation together in a month, this time focused on AI applications in ag.    What questions do you have for how we’re using artificial intelligence and large language models to help farmers? Drop them below and we’ll try to address them in the recording! #AI #LLM #DigitalFarming

102

3 Comments

Clever framework proposed by these researchers for evaluating safety of MLLMs. Huge improvements over the current approach in terms of having a systematic approach to generating and using a diverse and comprehensive set of harmful testing queries. Creative way of using LLMs to rethink and deliberate to adjust their evaluation. Great step forward in improving AI safety. Highly encourage reading this paper. #ResponsibleAI #MLLM #LLM #AI https://lnkd.in/gv3BQSs

5

$100M is what it took Meta to develop #Llama3 (and 24k GPUs ... and even if you have the 💰 you can't get them as they are manufactured with #Unobtainium 😂 ). As the post says below: there is no point in trying to create your own model: you should spend your time on clarifying the use-cases (what do you want to use it for ... in other words, it should not be a solution in search of a problem), and how to integrate it within your existing workflows.

3

🔍 Enhancing AI Reliability Through Effective Problem Framing and Prompt Engineering Unlocking the full potential of generative AI goes beyond simple queries—it involves precise problem framing and prompt crafting. Here’s how you can achieve more dependable AI outcomes: 1. Set Clear Objectives: Define specific, well-articulated goals for your AI project. 2. Provide Context: Include comprehensive background information and relevant constraints. 3. Be Precise: Use straightforward and unambiguous language in your prompts. 4. Use Examples: Offer examples to guide AI, establishing clear quality and style benchmarks. 5. Apply Constraints: Limit the scope to ensure the AI remains focused on pertinent aspects. 6. Refine Iteratively: Begin with a broad prompt and refine it based on initial feedback. 7. Request Explanations: Ask the AI to explain its reasoning to uncover any logical gaps or biases. 8. Feedback Loops: Continuously provide feedback to enhance future AI responses. Implementing these strategies helps frame problems effectively and craft prompts that result in reliable, high-quality AI outputs. Let’s use AI thoughtfully and efficiently! 💡 #AI #PromptEngineering #ReliableAI #GenerativeAI #MachineLearning #ArtificialIntelligence #TechInnovation #ProblemSolving #DataScience

2

💥Google's new Med-Gemini for healthcare outperforming GPT-4 Med-Gemini is a family of multimodal AI models built upon their Gemini models, specifically designed for the healthcare industry. From what I've read, these models are built on Google's existing Gemini AI architecture, but they've been specifically designed and trained for medical and clinical applications. It can process long-form medical records and research papers, and it can output text, images and videos to help explain things. Google is claiming that Med-Gemini has even outperformed OpenAI's powerful GPT-4 model on certain medical reasoning tasks, which is pretty impressive. On standard medical benchmarks like the MedQA exam, the Med-Gemini-L 1.0 model has scored really high, beating out even Google's own previous Med-PaLM 2 model. One cool feature is that Med-Gemini integrates web search, so it can pull in reliable information from the internet to provide more nuanced and factual responses. The models aren't publicly available yet - Google is still working on improving them and making sure they adhere to responsible AI principles around privacy and fairness before they release them. But from what I've seen, Med-Gemini is shaping up to be a really powerful set of medical AI tools.

1

🚀 Can #AI Unlock the Biotech Revolution? The AI biotechnology market is experiencing rapid growth, poised to revolutionize precision and efficiency in research and industry applications. Our latest data visualization underscores this transformative potential, highlighting key trends and breakthroughs. 📈 Market Growth: Witness the exponential increase in AI biotech market size from 2015 to 2025, projected to reach $17.5 billion, reflecting robust investment and innovation. 🔬 Key Applications: Discover how AI is redefining drug discovery, genomics, and personalized medicine, driving unprecedented advancements and outcomes. ⚙️ Efficiency Improvements: See how AI-driven methods outpace traditional approaches, significantly enhancing both time and cost efficiency in biotech processes. How do you see AI impacting your field within biotechnology? Share your insights in the comments and join the conversation on the future of AI in biotech! 🔗 Integrating Futures: AI's Role in Redefining Biotechnology: https://lnkd.in/gPpGRb-Z #Biotechnology #AI #DataVisualization #Innovation

2