Ron Brachman

The Carnegie Mellon study exposes significant stylistic differences between human writing and that produced by LLMs, revealing the limitations of these models in adapting to context. Unlike human writers, who seamlessly adjust tone and structure, LLMs maintain a rigid, information-dense style, raising concerns about their use in nuanced communication tasks. This underscores the necessity for educators and writers to critically assess the appropriateness of LLMs. Future research must explore LLMs' consistency in evaluating human writing, aiming to shape their role in educational settings and deepen understanding of their capabilities and limitations. https://lnkd.in/eQwgYaVM #education #aiwriting #languagemodels #iwork4dell

23

4 Comments

I'm excited to share three research papers from our talented team at Together AI that are advancing the frontiers of efficient and effective AI systems: ## 🚀 Accelerating Large Model Inference Our researchers have developed "Ladder-residual," a novel parallelism-aware architecture that significantly accelerates large model inference through communication overlapping. This innovation addresses one of the key bottlenecks in deploying large language models at scale, making AI systems faster and more cost-effective. 📄 Read more: [Ladder-residual: parallelism-aware architecture for accelerating large model inference with communication overlapping](https://lnkd.in/g825uRRz) ## 🤝 Enhancing Model Alignment Through Collective Intelligence By leveraging the collective intelligence of open-source models, we've developed new methods to improve AI alignment without the computational expense of traditional alignment techniques. This research demonstrates how cooperation between diverse models can lead to superior performance and better alignment with human values. 📄 Read more: [Improving Model Alignment Through Collective Intelligence of Open-Source Models](https://lnkd.in/g9FZ83P8) ## 📱 Optimizing On-device and Cloud Collaboration Our collaboration with researchers from Stanford introduces cost-efficient methods for on-device and cloud language model cooperation. This work enables more powerful AI capabilities on edge devices while intelligently leveraging cloud resources, resulting in better user experiences and reduced costs. 📄 Read more: [Cost-efficient Collaboration between On-device and Cloud Language Models](https://lnkd.in/ggM6v3R3) These papers represent our commitment to developing AI that is more efficient, aligned, and accessible. I'd love to hear your thoughts on these research directions or discuss potential collaborations! #TogetherAI #MachineLearning #AI #ResearchAndDevelopment #LLM #EdgeComputing #AIAlignment

50

At #KDD2025, I walked out of Dr. Christopher Manning’s talk feeling both grounded and optimistic. Few people have shaped NLP as much as he has (Stanford NLP Group, core textbooks/tools), and his message cut through hype with clarity. What stuck with me: 1. NLP’s “surprising victory.” We moved from formal, brittle representations to systems that learn from the richness of text. For most people today, LLMs are AI, and that didn’t happen by accident. 2. Redefining AI (goodbye “mine test”). If your definition says any system that takes input and changes the world is “AI,” a landmine qualifies. Manning’s point: intelligence = learning + adaptation + goal-directed behavior, appropriate to context in an uncertain world. That’s the bar. 3. Universal language models that generalize. His team’s Mixture-of-Experts Universal Transformers (MoEUTs) reuse learned “circuits” via shared layers + MoE in both attention and feed-forward, plus practical fixes (e.g., PeriLayerNorm, layer groups, SwitchHead). Net: better perplexity with roughly half the training compute vs. standard Transformers at similar model size, evidence that efficiency and quality can rise together. 4. From chatbots to agents. With NNetNav, they train web agents to explore unfamiliar sites, discover goals on their own, and learn sequences of actions without costly human demos. Their 8B model hits state-of-the-art among open LLMs on WebArena/WebVoyager and even beats zero-shot GPT-4 on selected tasks. The takeaway: interactive learning + linguistic priors can turn LLMs into competent doers, not just talkers. Why this matters (for leaders & builders): 1. Strategy: Don’t equate knowledge recall with intelligence. Invest in systems that learn in context and improve with interaction. 2. Product: Agentic workflows (especially on the web/enterprise apps) are moving from demo to durable capability. 3. R&D: Scaling isn’t just “more GPUs.” Architectural choices (MoE, recurrence, normalization) can deliver step-changes in quality/cost. #KDD2025 #NLP #AI #MachineLearning #LanguageModels #Agents #ChristopherMannin

61

4 Comments

It took two years to have this breakthrough: announcing our paper to appear in COLING 2025 Industry Track "(ID: 107) LLM-Friendly Knowledge Representation for Customer Support" by Hanchen Su, Wei Luo, Yashar Mehdad, Wei H., Elaine Liu, Yufeng “Wayne” Zhang, Mia Zhao, and Joy Zhang. One of the key power of LLMs are their capabilities to 'reason' intelligently. Over the past 70 years, AI researchers have been trying to encode the intelligence (or expert knowledge, business logics, rules) as code (programming languages are fundamentally state machines), workflow (a high level state machine, like your Turbo Tax flow), or rules. When GPT-4 was introduced, we thought that this amazingly powerful LLM can use its super powerful common-sense knowledge to 'reason' for different business cases, or at most use prompts to instruct it to solve business problems. It failed. Imagine hire Einstein to work as a banking clerk, Sure Einstein is super super smart, he is likely not able to follow the exact business logic to help a customer to transfer their balance from their checking account to the savings account, not to mention that using Einstein as a banking clerk is such a waste of the resources. What our paper shows is that with the innovative pseudo language called ICA (Intent, Context and Action), we can 'encode' business logic and knowledge in a way that even a weak LLM can understand and follow. ICA is less restrictive and less formal than programming language and formal rules. It is friendly for business owner to create and maintain. Through fine-tuning, the weak LLM can follow ICA to execute the business logics much more accurate compared to having the LLM to read typical knowledge articles written for human. Please come to the talk if you are attending COLING.

290

13 Comments