Longen Lan

Why I Left Meta to Focus on AI Safety TL;DR: Most AI companies inevitably continue to prioritize growth over safety. The risks from AI are however already growing for consumers and enterprises. We are launching: 1. SEVzero.ai (https://www.sevzero.ai) to build safety infrastructure for AI, and 2. The AI Risk Report (https://lnkd.in/ghkxGPXH) to systematically analyze AI risks and solutions. GenAI is changing the world for the better (Nano Banana FTW! ;) ) Unfortunately, like the internet and social media before it, AI also carries significant risks with it. Are we aware of the risks and doing enough to address them? After a decade of working on Facebook Messenger and GenAI at Meta, I've seen a consistent pattern emerge in the overall industry: growth trumps safety. It's a playbook most tech companies choose in a winner-take-all world (until safety issues inevitably come to bite us). Unfortunately, this has serious negative consequences. We saw, for example, the impact of social media on mental health, misinformation, and privacy, and we're seeing it again with the AI boom. * In the enterprise, most AI deployments are dying as POCs, often due to inadequate data quality and risk controls. The often quoted study from MIT reveals that as many as 95% of enterprise AI projects fail to reach successful production. AI model hallucinations also cost businesses over $67 billion in 2024.  * Our own experience over the last year developing workflow automation using LLMs confirmed firsthand the real obstacles CXOs face in taking AI beyond POCs (more about it later). * For consumers, unchecked use of AI is already suspected to contribute to cognitive dependency, psychological harm, and misinformation. A recent lawsuit alleges that a chatbot acted as a "coach" in a teenager's suicide :( BTW: kudos to OpenAI for launching parental controls, super quick follow-up! The risks due to AI are no longer something to worry about in the future. Neither were the risks of social media, but IMHO we got to that too late. Let’s not make that mistake again. That's why I quit Meta: to singularly focus on AI Safety. We are launching the following today. * SEVzero.ai: Novel opensource solutions and consulting to help enterprises deploy safe & dependable AI. Our flagship product-in-development is a safety platform called Model Metastore, with support for policy enforcement and root-causing SEVs. * The AI Risk Report: Objective analysis of AI safety risks for consumers and enterprises. We'll dive into the technical & human causes behind these risks and point out promising solutions from industry and academia. The need for safer, more dependable AI is urgent. Let's make it happen! 👉 Please subscribe to The AI Risk Report 📢 Stay tuned for product and open source announcements from SEVzero.ai. 🤝 DM me if you are also passionate about AI Safety! Thank you, Vishy #AISafety #GenAI

184

23 Comments

Unlocking LLMs' Graph Reasoning Potential Through Cognitive-Inspired Collaboration 👉 Why This Matters Large language models often falter when analyzing transportation networks, social connections, or citation graphs—not due to lacking intelligence, but because of working memory constraints. Imagine solving a 1,000-node shortest path problem while simultaneously memorizing every connection. Like humans juggling too many thoughts, LLMs lose accuracy as graph complexity increases. 👉 What GraphCogent Solves This new framework addresses three core limitations: 1. Representation confusion: Mixed graph formats (adjacency lists, symbols, natural language) 2. Memory overload: Context window limitations for large-scale graphs 3. Execution fragility: Error-prone code generation for graph algorithms Drawing inspiration from human cognition's working memory model, GraphCogent decomposes graph reasoning into specialized processes mirroring how our brains handle complex tasks. 👉 How It Works Sensory Module - Acts as an LLM's "eyes," standardizing diverse graph inputs through subgraph sampling - Converts web links, social connections, or traffic routes into uniform adjacency lists Buffer Module - Functions as a "mental workspace," integrating graph data across formats (NetworkX/PyG/NumPy) - Maintains persistent memory beyond standard LLM context limits Execution Module - Combines two reasoning modes:  - Tool calling for common tasks (pathfinding, cycle detection)  - Model generation for novel problems using preprocessed data 👉 Proven Impact - Achieves 98.5% accuracy on real-world graphs (social networks, transportation systems) using Llama3.1-8B - Outperforms 671B parameter models by 50% while using 80% fewer tokens - Handles graphs 10x larger than previous benchmarks through efficient memory management The framework's secret sauce? Treating graph reasoning as a team effort rather than a single AI's task—much like how human experts collaborate on complex problems. Key Question for Discussion As multi-agent systems become more sophisticated, how might we redesign LLM architectures to better emulate human cognitive processes for specific problem domains? Paper: "GraphCogent: Overcoming LLMs’ Working Memory Constraints via Multi-Agent Collaboration in Complex Graph Understanding" (Wang et al., 2025)

65

2 Comments

I attended the DeepLearning.AI and AI Fund buildathon yesterday, and Andrew Ng gave a fascinating talk about how AI is fundamentally changing software development. Andrew's core message is that code has become disposable. With AI assistants and massive context windows, we can rewrite entire systems in hours. What's scarce now is judgment and the ability to validate ideas rapidly. His new team mantra captures this: "move fast and break as little as possible." His economic argument was compelling: "drive the cost of doing a proof of concept so low, you just don't care." Winning teams now ship twenty prototypes, watch eighteen fail cheaply, and scale the two that work. Andrew stressed that staying current with AI tools isn't optional. "Being half a generation behind makes a very material difference right now," he said. Your toolchain determines what you can explore in a weekend! He identified the "product management bottleneck"—when code becomes easy, product decisions become the constraint. His advice: use data to train your intuition. When survey data contradicted his preference, he interrogated why his gut was wrong and upgraded his thinking. Andrew was emphatic: "We'll look back on 'don't learn to code' as some of the worst career advice." It's about fluency in telling computers exactly what you want - "the most important skill in the future." He made a fascinating reference to the past, mentioning that back in the punch card days, people thought COBOL would destroy software as a craft, but when it became easier to write software, more people could do it. I was inspired by that insight, and I worry about this generation of graduates who got caught in all of this. But overall, it was a hopeful message that I believe in, but I think the seasoned professionals out there need to help the younger generation now. He highlighted a labor market mismatch. We need AI application engineers who can orchestrate agents, manage context, and iterate with users. The best engineers are becoming lightweight product teams. His most provocative insight is that if AI can regenerate implementations on demand, code as a permanent artifact loses value. You can try three schemas before lunch and migrate by dinner. Teams should treat decisions as reversible and prioritize learning velocity. His closing charge: Stay current with tools, instrument feedback loops, write reversible specs, and choose momentum while "breaking as little as possible" when choosing between pristine plans and empirical momentum. The key takeaway: quality emerges from quantity. Teams that learn fastest win. Also - In case you're wondering what happened to me, I absolutely got my a** handed to me but I had a blast.

41

1 Comment

Multi-agent AI isn’t missing capability. It’s missing structure, and the machine teaching methodology is the missing ingredient. Right now, everyone’s talking about what multi-agent AI systems could do. What’s missing is a conversation about how to actually build them so they work at scale, in the real world. We’ve seen this pattern before: ✅ Formulas made Excel usable ✅ SQL made databases accessible ✅ Prompt engineering gave LLMs structure None of those shifts were about better features. They were about better methods: ways to apply enabling technologies consistently and effectively. --- 🧠 We’re now at the same point with multi-agent AI systems. Engineers need a structured way to turn domain expertise into modular systems that can learn, adapt, and perform under pressure. That’s what machine teaching provides. Not hype. Not theory. Just methodology that works in production. ✅ This is what I post about. For more discussion about #machineteaching and how to build multi-agent #ai systems that perform high-value tasks in real life, consider 👍, ✍️, ♻️ and following me on Linkedin: https://hubs.ly/Q03zSFbc0, then activate the 🔔

30

2 Comments

Introducing the AI Engineer World's Fair multi-turn conversation benchmark. I wrote a post for the AI Tinkerers newsletter about: - why multi-turn is hard for LLMs, - the particular challenges of audio/voice, and - a benchmark I use to think about LLM performance for conversational agents. https://lnkd.in/ecEfvQBs

48

6 Comments

In agentic AI, experience matters more than hype. I’ve seen firsthand how the conversation around AI often focuses on its promise rather than the practicalities of deployment. Richard Li and I explore the implications of agentic hype in a new blog. While agentic AI is compelling, getting it to work at scale is a fundamentally harder problem than many realize. It's not about clever prototypes or your development framework; it's about building hardened, distributed systems that are durable, observable, and scalable. Success with agentic AI won't be achieved through simple orchestration. It depends on a deep understanding of infrastructure and a commitment to engineering excellence. We need to move beyond the hype and focus on building a foundation of stability and trust. Are you ready for your agentic transformation? We'll help you move AI from concept to reality. #AI #AgenticAI #Innovation #Technology #Leadership https://lnkd.in/gbvSNuZy

52

10 Comments

Ever fallen into a “vibecoding doom loop” with LLMs i.e., get stuck tweaking prompts endlessly, hoping it ‘just works’?  I came across this short blog post with some good advice on avoiding this. 👉 Work in small steps and check the output after each step. This sounds obvious (and boring) but it is very effective if you want to build something reliable AND if you want to learn along the way. I have been doing this for a while, especially for high-stakes use-cases, and highly recommend it. 👉 Show the LLM what you want by editing its output. Don’t just adjust your prompt—show the LLM what “good” looks like by directly editing its response. This works surprisingly well. (ChatGPT and Claude don’t let you do this in their web apps yet AFAICT, but Gemini does via AI Studio. If you are using the API, you have full control.) The blog post uses a tool called SolveIt (from Jeremy Howard + Answer AI) to illustrate, but the advice is generally applicable. Worth a read.

129

11 Comments

Anyone interested in playing around with gpt-oss as the llm for a voice agent? Here’s an example for getting started. The gpt-oss model runs on Groq. For both the 20B and 120B models, I’ve noticed a lot of variability in response time and in the responses themselves. It doesn’t like following my instructions about not using asterisks. https://lnkd.in/gXK5i3Gy

40

14 Comments