David Zeng

This post continues my takeaways from AI Dev 25 x NYC  by DeepLearning.AI & Andrew Ng. The conference overview highlights "breaking the limits of #AI growth" and next-gen SW development as focus areas. I was very interested to hear the speakers' thoughts on handling "large, complex, long-horizon" models. Even before words were spoken, I was struck by who was speaking. On one panel, I met three professionals from different organizations who all had the same title: "Head of Developer Relations." They are "Connectors," and their job is to “bring the engineers' expertise out in the community," said Josh Gordon (Google). His answer to how we move forward is all about utility and #collaboration. A product becomes "useful" when its general-purpose "capabilities" are connected to "specific verticals, environments, and needs."  While LLM builders are reaching out to the community, the community is being empowered by #VibeCoding. AI enables everyone to "develop adjacent roles" and participate in processes that were previously out of their reach. The most powerful outcome is not just that they can build something from scratch. It's that they can "#iterate much faster than before." Coding skills are becoming accessible and integrated into many other professions besides software engineering. On the topic of scaling AI, Michael Kearns, Amazon Scholar, comments that the #supplychain is immature. He defines this as the entire ecosystem downstream from the model—#cloud development, deployment, and the connections between services. “There are so many inefficiencies in the supply chain exacerbated by GenAI. We went from incredibly targeted applications for which we knew exactly what we were trying to predict to an incredibly general-purpose model. I personally think it went too far. The connections between those supply chains are the most immature,” said Kearns. There are reasons for these inefficiencies, including contractual reasons. How do we move forward? Once again, “We need permissions and cooperation.” An emerging trend is the "fork in the road" between large, centralized models and the rise of "#SmallAI". With over 90% of AI projects reportedly failing to produce business impact, the high cost, latency, and closed ecosystems of large models are starting to be a barrier. The alternative is a move toward smaller, open-source, domain-specific models that enable rapid, low-cost prototyping. This is the hardware-level argument for the democratization and "vertical-specific" AI that others were talking about. Instead of all AI living in a few data centers, Arm is pushing for a future where it lives on your phone, in your car, or in a specific medical device. The unifying thesis? "The greatest value creation will come from developers who can tailor AI for specific vertical industries, embedding domain expertise directly into models". In my next posts, I will discuss two of these key verticals: finance & #healthcare. #wearenina Laurence Moroney Eric Sondhi

17

1 Comment

John Snow Labs #Healthcare #NLP 6.0.4 is out! Updated terminology mapping models, zero-shot entity recognition, and massive speed improvements for private inference at scale: • ONNX model optimizations and batch inference improvements (GPU & CPU) • Batch inference and #OpenVINO support for PretrainedZeroShotNER, our most capable and preferred medical NER framework • New MedS-NER #LLM for structured medical entity extraction via 3B size small LLMs • Advanced clinical one-liner pretrained pipelines for #PHI detection and extraction • Clinical code resolution via one-liner pretrained pipelines for medical #terminology mapping • Clinical document explanation, one-liner pretrained pipelines for comprehensive #clinicalentity analysis • New NER model and pipeline about #vaccinations and #infectious diseases ◦ Improved memory management for #ONNX in ML-based transformers ◦ Enhanced #geoconsistency algorithm in #deidentification to reliably process previously skipped geographic entities • Updated notebooks and demonstrations for making the software easier to navigate and understand Overall, the library provides 2,500+ #SmallLanguageModels that run privately and at scale within all the major #data and #ai platforms.

21

2 Comments

Excited to join AMI’s $1B seed round! Most AI today predicts the next token. But the real world runs on cause and effect. Generative architectures trained with self-supervised learning have unlocked extraordinary capabilities such as search, summarization and coding. But factories, hospitals, and robots operating in open environments need something different. To operate in the real world, AI must understand how environments evolve and how actions shape outcomes. That’s the promise of World Models, the next chapter of the AI revolution. By learning abstract representations of the real world and predicting consequences of actions, World Models enable systems that can reason, plan, and operate safely. AMI is advancing this frontier across robotics, industrial automation, healthcare, and other real-world systems where reliability and control matter. Congrats to the incredible AMI - Advanced Machine Intelligence team. Yann LeCun Alex LeBrun LG Technology Ventures Dong-Su Kim Krishna M. Ramachandran Taejoon Park Robert McIntyre https://lnkd.in/gSBz3NqU

82

3 Comments

You don't need a separate reference model for RL. We built an RL pipeline built on a multi-turn-focused modification of Self-Distillation Policy Optimization (SDPO) that uses the same model in two roles: student and teacher. The difference? Context window. → The student generates actions with no knowledge of the outcome. The teacher re-evaluates the same tokens with hindsight: the observation from each action, terminal test output, and a successful rollout from the same prompt. The per-token advantage is just the logit gap between these two views. → This gives you dense, per-token credit assignment instead of a single scalar reward for the whole trajectory. Think tokens get 0.25× weight (loosely guided, not tightly constrained). Action tokens get full gradient. System/user/tool result tokens get nothing. → The "teacher" is not a frozen checkpoint or a separate reward model. It's your current policy, conditioned on richer information. This is well-defined under MoE because same parameters with different routing is normal operating behavior. → We use a tiered reward system with hard gates only for truly unrecoverable failures: sandbox escapes, test corruption, harness crashes, timeouts, infinite loops. Everything else like wrong directory, failed commands, incorrect approaches are learnable signal, not wasted compute. The key insight: most of the information you need for credit assignment is already inside the model. You just need to ask it the right question with the right context. On LiveCodeBench v6 SDPO reached 48.8% vs GRPO's 41.2% beating Claude Sonnet 4 with an 8B model! That's bonkers.

5

A targeted approach to fine-tuning AI Models: Learning from failure, not just successes Some time ago, we presented our MedPi research, which evaluated medical AI model performance across a range of multi-turn patient conversations. Building on that work, we've now developed a "medical mistake benchmark" - a dataset focused specifically on the types of errors and failures we observed in those interactions. The key insight is that by fine-tuning your model on this targeted dataset of known mistakes, you can dramatically improve its performance in real-world, multi-turn settings. It's no longer about teaching the model what it already knows, but addressing its true weaknesses head-on. We're excited to share the results of this research in our new paper, which shows exactly where the mistakes were made and how this approach can lead to better-performing medical AI. Take a look, and let us know if you have any questions! We're committed to helping companies build AI assistants that can truly excel in complex, multi-turn conversations.

19

1 Comment

Ex-DeepMind researcher Misha Laskin on building superintelligence, coding as a shortcut, reinforcement learning, launching ReflectionAI's first product, Asimov, and competing with the big AI labs. LINKS: YouTube: https://lnkd.in/e78FrF-X Spotify: https://lnkd.in/eJ-4hmbz Apple Podcasts: https://lnkd.in/ekTizGXs 00:00 - Intro 01:42 - Reflection AI: Company Origins and Mission 04:14 - Making Superintelligence Concrete 06:04 - ASI vs. AGI: Why the Goalposts Moved 07:55 - Organizational Superintelligence as an Oracle 12:05 - Coding as Shortcut: Hands, Legs & Brain for AI 20:55 - Capturing Tribal Knowledge in Organizations 26:31 - Introducing Asimov: Deep Research for Code 28:44 - Team-Wide Memory 33:07 - Multi-Agent Design 34:48 - Data Retrieval and Integration in Asimov 39:41 - Reinforcement Learning in Asimov 41:04 - Misha's Journey: From Physics to AI 42:06 - Growing Up in a Science-Driven Desert Town 53:03 - Building General Agents at DeepMind 56:57 - Founding Reflection AI After DeepMind 58:54 - Product vs tech in AI 1:02:22 - The State of Autonomous Coding Agents 1:04:26 - What's Next for Reflection AI

82

10 Comments

All of this is wonderful news but it will always be held back by "leaders" who don't understand it. For example, in 2011, Unilever hired me to help them finish building a Machine Learning platform and commercialize it. Yes, you read correctly, Unilever. While flying the airplane while it's being built, I devised a business model using the AI within a Wellness vertical. The CTO at the time, Neil M., may he RIP, empowered me. In under 3 months we had partnerships in place to give Unilever access to 1.3 billion lives. This was needed to prove the Machine Learning elegance of the system. FYI, I accomplished this single handedly at a cost of just my salary, email and a PowerPoint deck. Neil empowered me... only to have the Chief Marketing Officer kill the whole project because, "he didn't understand the use of Machine Learning." Of course I was not included in that dialogue, being an outsider/contractor. If I was included in that critical meeting, we would have captured 1.3 billion people ➕️ one. Neil's contract as Unilever's first CTO was not renewed and he went back to J&J and brought me with him on a newly created Innovation team. I felt like I arrived. It was heady. Neil challenged me to build an enterprise wide AI system for J&J, as well as commercialize it. While I assembled a team of both insiders and outsiders to build the AI system for J&J, I drafted a business model for J&J Consumer, Medical Devices and Diagnostics, and Pharma. This time, I leveraged the massive power of the J&J brand and devised a business model for "Whole Country Capture." Yes, you read correctly, whole, entire countries. I leveraged a healthcare vertical for a common globally shared issue, obesity, in order to open dialogues with governments that supported government sponsored healthcare. Again, Neil empowered me. Keep in mind that I was representing one of the most powerful corporate brands in the world, J&J. In less than a month of work, again single handedly and with a massive budget given to me, all of $2 million dollars, J&J had agreements in place with the Central government of China (ten cities signed up in one week = 133 million people), the UK captured via the NHS, the Middle East via the Royal family, Disney wanting to use all of their IP and Animation Studio, the BCBS Association and many others. All done in one week and because I was empowered by Neil. Back to that critical, higher up, "leader" element... A new Chairman of J&J Consumer was brought in. You can probably guess what happened. All of my work needed Chairman level approval. The new Chairperson, having zero background in both the Consumer space and technology, decided to drop the AI project. She didn't see the value. So... without leaders, true visionary leaders, not just leaders who know how to play the corporate game but rather leaders who are visionary, of the student mind, and can empower subordinates, company AI use will progress slowly. God bless the Neil's of the world.

8