Jigyasa Grover

Soumith Chintala, 𝗰𝗿𝗲𝗮𝘁𝗼𝗿 of PyTorch and one of the most respected builders in 𝗺𝗼𝗱𝗲𝗿𝗻 𝗔𝗜, has joined Mira Murati ’s fast-rising startup, Thinking Machines Lab. After 11 years at Meta (Facebook), where PyTorch grew from a 𝗻𝗶𝗰𝗵𝗲 𝗽𝗿𝗼𝗷𝗲𝗰𝘁 𝘁𝗼 𝗽𝗼𝘄𝗲𝗿𝗶𝗻𝗴 𝗴𝗹𝗼𝗯𝗮𝗹 𝗔𝗜 𝗶𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻, Chintala’s move signals a new wave in human–𝗔𝗜 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻. Thinking Machines Lab isn’t just gathering top talent from Meta, OpenAI, Anthropic, and leading universities—they’re offering 𝗴𝗮𝗺𝗲-𝗰𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝘀𝗮𝗹𝗮𝗿𝗶𝗲𝘀 (up to $500K), 𝗮 𝗺𝗮𝘀𝘀𝗶𝘃𝗲 $𝟮 𝗯𝗶𝗹𝗹𝗶𝗼𝗻 𝘀𝗲𝗲𝗱 𝗿𝗼𝘂𝗻𝗱, and are aiming for a $50 billion market valuation. Their new product, Tinker, is already in use at Princeton, Stanford, and by major enterprises. Meta, meanwhile, is regrouping under the 𝗻𝗲𝘄 𝗦𝘂𝗽𝗲𝗿𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗟𝗮𝗯𝘀 division led by Alexandr Wang, with other AI leaders reportedly on the move. Chintala’s reasoning: time to pursue “𝙨𝙤𝙢𝙚𝙩𝙝𝙞𝙣𝙜 𝙨𝙢𝙖𝙡𝙡, 𝙨𝙤𝙢𝙚𝙩𝙝𝙞𝙣𝙜 𝙣𝙚𝙬, 𝙨𝙤𝙢𝙚𝙩𝙝𝙞𝙣𝙜 𝙪𝙣𝙘𝙤𝙢𝙛𝙤𝙧𝙩𝙖𝙗𝙡𝙚.” His legacy—the PyTorch tool now thriving in classrooms from 𝗠𝗜𝗧 𝘁𝗼 𝗿𝘂𝗿𝗮𝗹 𝗜𝗻𝗱𝗶𝗮—𝗶𝘀 𝘀𝗲𝗰𝘂𝗿𝗲. #AI #TechNews #PyTorch #Meta #ThinkingMachines #Startup #AIJobs #TalentWar #Innovation #BigFunding

8

Research Paper Highlights: "When Retrieval Succeeds and Fails: Rethinking Retrieval-Augmented Generation for LLMs", By Yongjie Wang Large Language Models (LLMs) have transformed language understanding and generation, enabling numerous real-world applications. Yet because they are trained on static datasets, LLMs struggle with rapidly changing information and niche domain-specific tasks. Retrieval-Augmented Generation (RAG) emerged to bridge this gap by supplying LLMs with fresh, relevant external knowledge. As LLMs grow more capable, this paper reevaluates the importance, relevance, and future direction of RAG systems. Challenges in RAG: - Contextually irrelevant retrieval can mislead LLMs and undermine answer quality. - Dependency on external retrieval systems adds architectural complexity and latency. - Difficulty in aligning retrieved content with LLM reasoning, leading to hallucinations or mismatched answers. - Retrieval bottlenecks, especially when queries are ambiguous or poorly formed. Key Takeaways: - RAG remains highly valuable in scenarios where LLMs alone fail, especially in domains requiring up-to-date, factual, or specialized knowledge. - The field needs re-examination and innovation, as conventional RAG designs may not keep pace with rapidly evolving LLM capabilities. - A deeper understanding of RAG objectives, architecture, and limitations is essential for building the next generation of knowledge-augmented systems. - Effective RAG can significantly boost reliability, factual grounding, and domain expertise when combined with powerful LLMs. This work highlights the need to rethink the role of RAG in an era where LLMs are rapidly advancing. It stresses that while LLMs are improving, RAG remains crucial for dynamic, specialized, and fact-sensitive tasks. Further reading-https://lnkd.in/da5_QtP6

3

Another interesting new RL perspective. They propose Reinforced Attention Learning (RAL) to address the difficulty of applying RL techniques that work well for LLMs but tend to struggle more with multimodal LLMs, although I think the idea is more general than that. They argue that this approach is more robust to reward hacking and less prone to diversity collapse. In practice, they add to the standard RL objective an additional term that looks like: E[ A_t · D(p_t^θ || p_t^old) ] Here, A_t is the advantage, while p_t is the attention distribution policy at generation step t, defined from the final attention layer (averaged across heads): the current one (θ) and the one that generated the batch (old). p_t^θ(i) is simply the attention weight from the generated token at position t to a previous token at position i < t, normalized by the sum of attention weights before t. D is the Jensen–Shannon divergence (JSD), a symmetric and bounded divergence defined via the mixture of the two distributions, and less aggressive than Kullback-Leibler divergence (KL). JSD(P || Q) = 1/2 · KL(P || M) + 1/2 · KL(Q || M) With M = 1/2 · (P + Q). What is happening here is that we incentivize attention patterns associated with high advantage and discourage those associated with low advantage. As the authors put it, when A_t < 0 the model is explicitly pushed to explore new reasoning paths. The full objective becomes: L_RL + λ · L_AttnRL. They then introduce an extension that combines RAL with on-policy knowledge distillation and attention distillation. The idea is similar to on-policy distillation: you generate with the student (the current network), compute the teacher outputs on the same generation, and align them using a reverse KL on the logits. In addition, they also align the attention distributions between teacher and student, again using JSD. Unlike RAL, this attention distillation term does not use an advantage, it is pure structural alignment. Takeaway: RL and on-policy distillation do not have to operate only at the token level. Extending the training signal to internal structures like attention can be quite effective, and I expect this direction to be explored much more in the future. Thanks Petru-Daniel (Dan) Tudosiu for suggesting the paper!

85

1 Comment

Just wrapped up a deep read of NaturalThoughts: Selecting and Distilling Reasoning Traces for General Reasoning Tasks from FAIR at Meta. It’s one of those papers that doesn’t just report numbers - it invites you to rethink how we approach reasoning in LLMs. What makes this work stand out is its clarity of focus: not all reasoning traces are created equal. The authors move beyond the traditional "scale everything" mindset and ask better questions. Which types of reasoning examples actually make models better at general reasoning? What kinds of traces help them think - not just produce answers with lots of tokens? NaturalThoughts is built on the insight that performance gains come not just from more data, but from smarter data curation. The team shows that selecting examples based on difficulty - like when multiple teacher models disagree - or based on diversity in reasoning strategies leads to stronger students. In some cases, these strategies outperformed popular distillation datasets like OpenThoughts and LIMO across STEM reasoning benchmarks. They also explore mixing System-1 and System-2 style reasoning traces, allowing student models to be both more efficient and more adaptable at test time. That part struck me, because it feels closer to how humans actually reason - we switch between quick intuition and deliberate analysis depending on context. If you're working on instruction tuning, alignment, or data-centric model improvement, this is worth your attention. It’s not just about finding answers, it's about learning what kinds of reasoning help models grow. #AIResearch #ReasoningInAI #InstructionTuning #ModelInterpretability #DataCentricAI #FutureOfAI #LinkedIn #connections #AI #ML #DL #NLP #GenerativeAI #PromptEngineering paper link : https://lnkd.in/e85qQ2KJ https://lnkd.in/enu6EF2R

22

4 Comments

Stanford HAI on Generative AI: Promise, Perils & Pathways Generative AI is moving faster than any technology shift we’ve seen in years. What excites me about the latest perspectives from Stanford HAI is how clearly they show both sides of the story - the immense potential and the very real risks. ** Why this technology feels different It’s not just about efficiency - GenAI is opening the door for anyone to create. Text, code, designs, even art - all accessible in seconds. For knowledge workers, it’s like having a research partner who can summarize, analyze, and surface insights instantly. Entire industries are rethinking their business models, from healthcare and education to finance and media, because GenAI changes how value is created and delivered. ** But the challenges are just as real These systems can unintentionally reinforce bias, and that risk grows as adoption spreads. The same tools that inspire creativity can also fuel misinformation at scale. Behind the magic lies a heavy environmental footprint - training and running large models requires significant resources. And while innovation is racing ahead, governance and accountability are still struggling to keep pace. ** Why this matters for us This isn’t just another wave of tech disruption. GenAI will shape careers, companies, and societies. And if we want the outcomes to be positive, the focus has to be on responsible adoption - building with ethics, transparency, and long-term sustainability in mind. Stanford HAI puts it beautifully: The goal isn’t simply to build more powerful AI, but to make sure it serves humanity in ways that are fair, trustworthy, and sustainable. If you want a thoughtful, research-backed perspective, I highly recommend reading the full report here: https://lnkd.in/dqEPUKFw Do you want me to also shorten this into a crisp 5–6 line version (for when you want a quicker, more engaging post), or should we keep it detailed like this one? #StanfordHAI #GenerativeAI #AIethics #ResponsibleAI #FutureOfWork #ArtificialIntelligence #AIInnovation #AIforGood #AITrends #TechLeadership

10

By building a taxonomy for compound AI systems, researchers are showing that when agents are optimized using structured evaluation frameworks, downstream task accuracy improves. Better evaluation → better agents. Featuring Melissa Pan and Mert Cemri of University of California, Berkeley / CSGE Berkeley Interviewed by Jessica Dai and SAIL Media Thank you to ‪Lambda‬ for making these interviews possible! 👉 Subscribe to our calendar for priority invites to upcoming events: lu.ma/twill #AI #ArtificialIntelligence #MultiAgentSystems #LLM #LLMAgents #CompoundAI #MachineLearning #AIResearch #BerkeleyAI #CSGEBerkeley #TechEvents #StartupCommunity #SiliconValley #FutureOfAI #AI #ArtificialIntelligence #MultiAgentSystems #LLM #LLMAgents #CompoundAI #MachineLearning #AIResearch #BerkeleyAI #CSGEBerkeley #TechEvents #StartupCommunity

34

2 Comments

🌟Supervisor Spotlight!🌟 Introducing the wonderful Bonan Zhao from School of Informatics: "My work in computational cognitive science explores how people learn, form hypotheses, and communicate. To me, responsible NLP means grounding NLP systems in how human cognition works. When people interact with language models, they rapidly form beliefs about what these systems can and cannot do—often inaccurately. My research on human learning and hypothesis formation can help us understand how users develop these mental models, and design systems that actively shape accurate expectations, preventing over-reliance, misuse, and harm. People also learn language in remarkably efficient and environmentally sustainable ways, building rich understanding from limited examples. By understanding and leveraging human inductive biases, we can create NLP systems that develop more like people do—requiring less data and fewer computational resources while achieving more meaningful generalization." Bonan Zhao's website: https://lnkd.in/e8Py9X73 Register for Applicant Webinars here: https://lnkd.in/e8B9zFC9 Find more information about the CDT ➡️ https://lnkd.in/duGN2b4Y #ResponsibleNLP #ResponsibleAI #ComputationalCognitiveScience #PhDResearch School of Informatics, The University of Edinburgh | The University of Edinburgh | Edinburgh Futures Institute | Edinburgh College of Art | 🌟

31