Subrata Das

I’ve noticed that many GenAI application projects put in automated evaluations (evals) of the system’s output probably later — and rely on humans to manually examine and judge outputs longer — than they should. This is because building evals is viewed as a massive investment (say, creating 100 or 1,000 examples, and designing and validating metrics) and there’s never a convenient moment to put in that up-front cost. Instead, I encourage teams to think of building evals as an iterative process. It’s okay to start with a quick-and-dirty implementation (say, 5 examples with unoptimized metrics) and then iterate and improve over time. This allows you to gradually shift the burden of evaluations away from humans and toward automated evals. I wrote previously in The Batch about the importance and difficulty of creating evals. Say you’re building a customer-service chatbot that responds to users in free text. There’s no single right answer, so many teams end up having humans pore over dozens of example outputs with every update to judge if it improved the system. While techniques like LLM-as-judge are helpful, the details of getting this to work well (such as what prompt to use, what context to give the judge, and so on) are finicky to get right. All this contributes to the impression that building evals requires a large up-front investment, and thus on any given day, a team can make more progress by relying on human judges than figuring out how to build automated evals. I encourage you to approach building evals differently. It’s okay to build quick evals that are only partial, incomplete, and noisy measures of the system’s performance, and to iteratively improve them. They can be a complement to, rather than replacement for, manual evaluations. Over time, you can gradually tune the evaluation methodology to close the gap between the evals’ output and human judgments. For example: - It’s okay to start with very few examples in the eval set, say 5, and gradually add to them over time — or subtract them if you find that some examples are too easy or too hard, and not useful for distinguishing between the performance of different versions of your system. - It’s okay to start with evals that measure only a subset of the dimensions of performance you care about, or measure narrow cues that you believe are correlated with, but don’t fully capture, system performance. For example if, at a certain moment in the conversation, your customer-support agent is supposed to (i) call an API to issue a refund and (ii) generate an appropriate message to the user, you might start off measuring only whether or not it calls the API correctly and not worry about the message. Or if, at a certain moment, your chatbot should recommend a specific product, a basic eval could measure whether or not the chatbot mentions that product without worrying about what it says about it. [Truncated due to length limit. Full text: https://lnkd.in/gygj3y7w ]

4,070

178 Comments

Excellent piece by HuggingFace CEO Clément Delangue on the necessity of a robust open source / open weight / open research approach to AI to keep the pace of progress. Closed/proprietary strategies slow down overall progress. I have made that point many times over the last dozen years, and became more vocal about it over the last 3 years, as prominent American AI companies started clamming up. https://lnkd.in/euTH8qwk

2,725

135 Comments

A good article by Cade Metz at the NYT about the limitations of current AI paradigms and how human-level AI (or AGI) is not "just around the corner" and will require new concepts and a new paradigm. I'm quoted at the end the article. There is no question in my mind that, at some point in the future, machines will be as smart as or smarter than humans in all domains. But contrary to what the Elons, Sams, and Darios of the world claim, it won't happen within two or three years. https://lnkd.in/ehp-nzfa

1,381

160 Comments

An excellent piece in Newsweek by former Bell Labs President Marcus Weldon on "the 8 principles for the future of AI." In a remarkably clear and concise manner, the piece distills some basic predictions from a series of interviews with roboticist Rodney Brooks, neuroscientist David Eagleman, and myself. https://lnkd.in/dsJSSvGM

859

49 Comments

An interview with Jeff Jarvis and Jason Howell for the AI Inside podcast. We discuss the limitations of current large language models, why human-level AI is still years away, Meta's open-source approach with LLAMA, and the vision for AI assistants that understand the physical world. https://lnkd.in/e8-umf4m

300

24 Comments

Today Michael Pfeffer, Percy Liang and I are delighted to introduce a benchmark to test LLMs on real world clinical tasks. MedHELM comprises a taxonomy -- validated by 29 clinicians from 15 specialties -- and 31 datasets mapped to tasks spanning 22 subcategories to evaluate 6 models. 🔹 GPT-4o leads in 2/5 categories. 🔹 Llama-3.3-70B-instruct in Patient Communication and Education. 🔹 Gemini 1.5 Pro in Clinical Note Generation. The best model achieves a 0.72 mean win rate across healthcare tasks! Leaderboard: https://lnkd.in/gZ7CkNqc Blog: https://lnkd.in/gsVQUc2P MedHELM is collaboration of Technology & Digital Solutions - Stanford Medicine, Microsoft HLS and researchers in the Stanford University School of Medicine and Human-Centered AI Matthew Lungren MD MPH, Eric Horvitz, Alyssa Unell, Hejie Cui, Suhana Bedi, Miguel Ángel Fuentes Hernández, Nikesh Kotecha, Michael Wornow, Juan M. Banda, Shrey J., Hao Qiu, Mert Öz, Yifan Mai, Akshay Swaminathan, Mehr Kashyap, Philip Chung, Roxana Daneshjou, Jonathan H. Chen, Emily Alsentzer, Keith Morse, Vanessa Kennedy, Sanmi Koyejo, Kevin Schulman, Jennifer Lee, MD, PhD, MBA, Timothy Keyes, Nerissa Ambers, Carlene Lugtu MCiM, BSN, RN, Aditya Sharma, Bilal M., Alex Alekseyev, Vicky Zhou, Vikas K., Jarrod Helzer, Jason Fries, Anurang Revri.

922

42 Comments

Similarities and differences between human brains and GPUs. Both are about the same size/weight. So the internal processing is considerably different. GPU's draw much more power under load, brains no so much. Brains last much longer and require vastly much less infrastructure to produce. #genai #gpu

193

26 Comments

Interesting paper alert - a commentary in Nature Medicine highlights a pivotal tension in healthcare today: while #AIagents promise transformative impacts — from autonomous triage to multi-modal clinical decision support — regulatory frameworks remain built for a different era (a recurring scheme in the world of Health AI which I do not see going away anytime soon: technical progress in AI will always outpace regulatory actions). Here are the take aways from the paper: 🔹 Current state of regulation of AI agents in health and medicine: Most regulatory pathways, like the EU Medical Device Regulation and the U.S. FDA’s 510(k), are optimized for narrow, static, non-adaptive tools. Approval of AI-driven devices with broad scope and high autonomy remains rare and challenging. 🔹 The problem (why existing regulatory frameworks fall short of capturing agentic AI applications): AI agents orchestrate multiple tools, access sensitive data, and make goal-directed decisions autonomously. Their adaptive nature, coupled with non-deterministic outputs and complex multi-component architectures, pushes them outside the assumptions underpinning today’s medical device regulation. 🔹 Proposed solutions (how can we explore and experiment safely with AI agents in clinical settings in the absence of comprehensive regulatory guidance): The paper explores several emerging regulatory strategies, including: Voluntary alternative pathways → tailored approvals for high-risk AI agents Adaptive oversight → dynamic, real-world performance monitoring Regulatory sandboxes → controlled environments to experiment safely Outcome-driven evaluation → focusing on patient-centered results rather than static technical compliance 💡 The takeaway (so where does this leave us): While regulation is still playing catch-up with the risks and complexities of AI agents in healthcare, it is possible to safely experiment and sandbox their use — provided there are clear guardrails, strong oversight, and rigorous evaluation frameworks. This balanced, sandbox-first approach could enable innovation without compromising patient safety. For first hand insights on how this is all playing out in the real world right now, have a read of these two recent papers from Eric Topol, MD (https://lnkd.in/gEycp__S) and Harsha Nori (https://lnkd.in/gKNxpxsA) John Lim Christopher Bain ZongYuan 宗元 Ge 戈 Deval Mehta #AI #Healthcare #Regulation #AIagents #DigitalHealth #NatureMedicine

136

13 Comments

🚀 Just published a practical introduction to LiteLLM — a lightweight but powerful tool to manage multiple LLMs across vendors. In many GenAI teams, costs spiral and wrappers multiply when you start working with many different LLM providers. LiteLLM (YC W23) provides: ✅ A unified OpenAI-compatible API framework ✅ A self-hosted proxy server with built-in monitoring, quotas, and API key management ✅ Out-of-the-box integration with LLM orchestration libraries such as LangChain and LlamaIndex. In this guide, I walk through a full setup — from Helm Charts to code — and share a few things I learned along the way.

13

Quantum needs data—not just qubits. aqora’s new public datasets hub marks a practical step forward for the field. By creating a central space to upload, access, and collaborate on quantum-specific datasets, they’re addressing a quiet but critical challenge in the ecosystem: reproducibility and shared benchmarks. Why does this matter? Quantum machine learning and benchmarking rely on consistent, high-quality data—but until now, much of it has been scattered, inconsistent, or locked in silos. This hub offers: ➡️ Versioned, well-structured datasets ➡️ Integration with tools like pandas and polars ➡️ Options for both public and private data sharing ➡️ A base for teaching, testing, and real-world use cases As the field matures, tools like this can help shift the conversation from isolated experiments to collaborative progress. 🔗 Full article in the comments. #QuantumComputing #QML #OpenData #QuantumResearch #TQI #TheQuantumInsider #Aqora #EcosystemBuilding

79

5 Comments

Gemini 3 is here! It was incredible reading about Google's new model. It's feels far more than a simple performance bump—it signifies a real shift where AI moves into being a genuine “thought partner”. Few features that stood out: • State-of-the-Art Reasoning: I'm excited that this model is built to genuinely grasp depth and nuance, which means it's phenomenal at "reading the room" for complex problems and is setting a new standard for intelligence. • Advanced Agentic Power: What really caught my attention is Google's claim that this is their most powerful agentic and vibe-coding model yet. It enables planning and executing complex, multi-step workflows, which is a massive leap forward for productivity. • Broad Multimodality: I was amazed to read it's claimed as the best model in the world for multimodal understanding. It can seamlessly synthesize information—vision, code, and text—helping us learn, build, and plan complex tasks in totally new ways. Looking at these features, we are entering a new era of intelligence. I absolutely cannot wait to see what everyone builds with this! #Gemini3 #GoogleAI #AIInnovation #FutureOfWork #LLM #AI https://lnkd.in/guMDTxAU

114

4 Comments

🔍 Why the Future of AI Research Isn’t Just About Retrieval—It’s About Deep Reasoning Large Language Models (LLMs) are no longer satisfied with simply fetching information. Enter Deep Research (DR) Agents—a new class of autonomous AI designed to tackle complex, multi-step research workflows that blend: ✅ Dynamic Reasoning (thinking step by step) ✅ Adaptive Planning (adjusting strategies on the fly) ✅ Tool Use (coding, analyzing data, browsing the web) ✅ Structured Report Generation Unlike traditional retrieval-augmented generation (RAG), DR agents don’t just retrieve and repeat—they synthesize, validate, and plan. What sets them apart? • Hybrid Search: Combining API calls with live browser exploration to handle both structured and messy, real-time content. • Multi-Agent Collaboration: Specialized agents work together, mimicking research teams. • Non-Parametric Learning: They evolve capabilities without retraining, learning by storing and reusing prior workflows. • Dynamic Workflows: From clarifying your intent to planning and executing tasks, all in real time. This paradigm shift has big implications: 🏥 Healthcare: Faster synthesis of the latest studies. 💼 Finance: Real-time risk assessments with live market data. 🧪 Scientific Discovery: Automated literature reviews and hypothesis generation. The paper also highlights key challenges—like aligning evaluation benchmarks with real-world needs and optimizing multi-agent coordination. How soon do you see DR agents becoming your co-pilot in complex decision-making? #AI #DeepResearchAgents #LLMs #FutureOfWork #AIResearch

41

12 Comments

Get faster and smarter MOE inference straight out of the box. 👇Deep dive on scaling expert parallelism with TensorRT-LLM. LLMs with MOE promise higher model capacity without linearly increasing compute costs - but they introduce new challenges -- more conditional computation, dynamic routing, and non-uniform GPU utilization -- solved by TensorRT-LLM. ✨ New in TensorRT-LLM has native support for expert parallelism—designed for fast, efficient inference with MoE models like Mixtral (Mistral AI) and DeepSeek (DeepSeek AI). This gives you: ✅ Dynamic expert routing: Automatically route tokens to the top-k experts with minimal overhead. ✅ Efficient expert scheduling: Balance expert loads across GPUs using smart sharding and token bucketization. ✅ Memory-aware execution: Maximize hardware utilization while respecting memory budgets. ✅ Drop-in support: Use @HuggingFace models with minimal code changes via TensorRT-LLM's #Python API. 🧠 How it works: MoE models activate only a subset of "experts" for each token. This dynamic nature is powerful—but hard to optimize. It’s all done under the hood using custom #CUDA kernels and NCCL-based communication primitives—giving you low latency, high throughput, and better GPU scaling. ✨ TensorRT-LLM handles: ✅Token-expert mapping using the gating network. ✅Token sorting to batch same-expert tokens together. ✅Expert parallel execution across GPUs. ✅Merging outputs for final predictions. 🛠️ Developer workflow - here is the code to get started. # Clone the repo git clone https://lnkd.in/g-GiDX23 # Use included examples to load and run a Mixtral model cd TensorRT-LLM/examples/mixtral From there, the Python API lets you load the model, convert it with TensorRT, and run expert parallel inference—all with a few lines of code. Results? 📈 Performance at scale. Tests show up to 2.3x faster inference throughput compared to standard tensor parallelism when using 8 GPUs and top-2 experts per token. Even better—TensorRT-LLM keeps efficiency high across increasing batch sizes. Want to see it in action or contribute? 👉 Read the full tech blog: https://lnkd.in/g_7YV3vV 👉 Explore the code on GitHub: https://lnkd.in/gNjQ5W2U 👉 Follow updates in the TensorRT-LLM repo: https://lnkd.in/gqSHYQ4u Share your experiences with us.

78

3 Comments

I shared some thoughts with Ars Technica on why the next phase of AI is about efficiency, not size: https://lnkd.in/dwwrFxwp Smaller, fit-for-purpose models like IBM Granite empower small businesses by opening the door to more accessible, customizable, and cost-effective AI.

137

1 Comment

Introducing Generative Adversarial Distillation (GAD): a novel GAN-style formulation and framework that facilitates both on-policy and black-box distillation of large language models (LLMs). GAD is the first technique to enable block-box on-policy distillation from proprietary teachers where internal logits or parameters are inaccessible, or distillation between teacher and student LLMs with incompatible vocabularies. GAD expands our prior work on white-box on-policy distillation (i.e., MiniLLM), pioneering block-box on-policy distillation for LLM training. Specifically, GAD frames the student LLM as a generator and trains a discriminator to distinguish its responses from the teacher LLM’s, creating a minimax game. The discriminator acts as an on-policy reward model that co-evolves with the student, providing stable, adaptive feedback. Experimental results show that GAD consistently surpasses the commonly used sequence-level knowledge distillation.  In particular, Qwen2.5-14B-Instruct (student) trained with GAD becomes comparable to its teacher, GPT-5-Chat, on the LMSYS-Chat automatic evaluation. The results establish GAD as a promising and effective paradigm for black-box LLM distillation. Our team has been conducting fundamental research in knowledge distillation with wide adoptions across the industry. -         MiniLM: We introduced multi-head attention distillation, establishing the most effective distillation method for BERT-style models. The open-source MiniLM models (e.g., 6x384) have become the most widely utilized small encoder models on the Hugging Face. -         MiniLLM: Our proposed Reverse KLD is recognized as one of the most effective, de facto on-policy distillation approaches for modern LLM training, which has been widely used by Thinking Machines, Gemma, and many other teams and models. -         BitDistill: We proposed BitNet Distillation to finetune off-the-shelf full-precision LLMs (e.g., Qwen) into 1.58-bit precision (ternary weights {-1, 0, 1}), achieving performance parity with the full-precision counterparts on specific downstream tasks. -         GAD: The development of Generative Adversarial Distillation (GAD) now allows for black-box on-policy distillation, overcoming two major prior limitations: (1) Distillation from proprietary teachers where internal logits or parameters are inaccessible; (2) Distillation between teacher and student LLMs with incompatible vocabularies. https://lnkd.in/gMaP2c7w

125

2 Comments

Excited to share another entry to the Fin AI blog - this time a deep dive on getting LLM inference times down, to investigate putting reasoning models in latency critical tasks. There's a lot you can do to get inference time down, compared to a naive approach. https://lnkd.in/ej8VbSkg

91

4 Comments

This Forbes article (https://lnkd.in/eBxYcBK3) from last week about the hallucination rates of the lastest LLM based reasoning models raises a few questions: Are we going in the right direction if the new models show hallucination rates of up to 79% in simple tests? Should we rather think about combining a set of LLMs and rules based solutions to build in check-points that identify and flag errors? Then there is this article: https://lnkd.in/eyWh9NDA A few months old already, and even at that time 60% of physicians reported that they use LLMs for decision making. To support diagnosis, identify drug interactions as well as treatment planning. Almost all of the physicians reported that they use other tools to check the output of the LLMs. Lets think about this: If the new LLMs have hallucination rates of up to 79% and physicians check their output for correctness, will they soon stop using these models? Since there would be no time saved if a lot of the results are just made up by an LLM? In my opinion, we need to think much more in specific use cases and build more than one LLM that is supposed to answer all questions in the world. LLMs are probabilistic models, while in healthcare we have a lot of fixed rules and foundational biological principles. Bringing these two things together is crucial. There are solutions emerging from startups and companies that have domain expertise and focus on a specific problem to be solved. These approaches will bring the real value to the healthcare industry! If you have a specific use case or example that you want to discuss please comment on this post or reach out.

21

1 Comment