Fanglin Wang

What stood out to me is how differently they and we see AI. They tend to view AI as something that will be guided for the benefit of society. We often see it as a threat. The difference may come down to power and control. I guess In their model, there’s an expectation that AI will be constrained so it doesn’t harm ordinary people. In ours, corporations may now be as powerful as governments, and AI could amplify that imbalance. That’s why AI feels like opportunity to them, but risk to us, not because of the technology itself, but because of who controls it and who bears the cost.

2

New Product Launch | SingTown OpenMV Official Robot Arm — Maximum AI Vision Fun! The SingTown OpenMV Official Robot Arm is now officially available! From “seeing,” to “understanding,” to “taking action,” it seamlessly connects the entire AI image recognition and machine vision pipeline — transforming your ideas into precise, efficient, and effortless real-world movements. Let’s take a look at what makes the OpenMV Official Robot Arm stand out: ● Ready to Use, Rich Accessories: Comes standard with interchangeable grippers such as an electromagnetic claw, LCD display, 9-key keypad, and high-precision stepper motors. ● Deep Vision Integration: Easily implement AI image recognition with support for target detection, object recognition, face recognition, and more. ● Full Python Development: Program everything using OpenMV Python, with complete project source code provided. ● Fully Open Interfaces: Reserved ports for ultrasonic sensors, infrared modules, multiple servos, relays, and more — worry-free expansion. ● High Playability Projects: Supports advanced projects like automatic piano playing, tic-tac-toe gameplay, smart waste sorting, and unlimited expansion with external peripherals. Click the video and see it in action!

1

In Nov 2024, Yuyan Zhou, Liang Song, Bingning Wang and Weipeng Chen published a phenomenal paper entitled "#MetaGPT: Merging Large Language Models Using Model Exclusive Task Arithmetic" https://lnkd.in/gsBJnhik In simple terms, they proposed a multi-#agentic framework to enable specialised AI agents to work with one another within a controlled environment, producing productive, sophisticated #ai teams that can address deeper real-world issues. I couldn't find their #linkedin profiles - or rather, there are too many with similar names that I don't know which is which. They are just a few of those young Data Scientists (that term is actually too small for them) from #China that are making significant contributions to the international AI community and body of knowledge. Here's their paper - https://lnkd.in/gXvk9Abw. Amongst other things - to give a simplified man-in-the-street view of their work, their AI can potentially hook-up various specialized AI to work together in the following scenario: 1. A client browses through a catalog of properties to invest in - maybe this client is a young couple who just got a job 2. An AI property agent finds them a nice undervalued niche beside the Langat river not too far from Putrajaya MRT but also not too near the expensive Putrajaya fringes, that they can afford to buy on loan with their combined income 3. A wealth management AI agent helps package the deal with life and health insurance and negotiates with a bank which is anxious to meet their annual loan quota. The bank offers them a 20-year loan with an option to sell the property to the bank at the end of 10-year at 150% its purchase price and at the end of 20 years at 200%. The clever wealth management AI also works out a deal with an insurance broker to underwrite any possible shortfall to protect the bank. 4. The bank offers them a possibility not only to sell their property to the bank, but also to take another loan from the bank to buy a larger property at the end of the tenure. So the bank tak penat cari new customer and breaks the churn. 5. The client chooses a few property types they can construct beside the Langat River - pod homes, container homes, tree houses, stilt houses etc. A recommendation engine AI helps them - somehow zapping info about their preferences from some place 6. After shortlisting the type, they talk to an AI chatbot who helps them to fine tune the design 7. After the design had been fine tuned, with materials and furniture chosen, a Building Information Management system AI gets to work, calculating costs etc 8. With the info, the bank releases the loan, another AI connects them to architects, construction companies around Banting etc and they move in after 12 months We have done some agentic stuffs (without the AI) before - video below

7

2 Comments

AI is rapidly becoming the core infrastructure of the digital economy. Excited to share that Singtel Innov8 has launched a US$250M global AI Growth Fund to invest in the next generation of AI companies and accelerate AI adoption across the Singtel Group. With this new fund, Innov8 now manages over US$500M in capital globally, strengthening our ability to support founders building the technologies that will power the AI era. Personally, I’m particularly excited about engaging with and learning from founders who are pushing the boundaries of AI, from infrastructure and platforms to agentic systems and edge AI. The most meaningful innovation happens when startups and large platforms collaborate to scale real-world impact. #AI #Innovation #VentureCapital #Startups #Telecom #DigitalInfrastructure #Singtel #Innov8

171

7 Comments

My colleagues Leon, Xing Li, Yuvraj, and I experimented with the open sourced software called DeepLiveCam which allowed us to swap faces in real time on our web cam. The results were quite impressive. With GPU enabled, we were getting 18 fps, which is the quality of a normal web cam. As shown in the picture below, it can also recognize and change multiple faces at once. Compared to DeepFaceLive, another popular open sources software for face swap, DeepLiveCam only required one picture as input. We will be demoning this live at an upcoming customer education event about fraud and scams. If you have tried to install and failed, reach out and I am happy to share tips and tricks which could save you hours of pain (if you run windows with nvidia GPU). We are still stuck on GPU access on apple silicon. Deeplivecam can be downloaded at https://lnkd.in/gkiWrmEp #hsbcai

56

3 Comments

🔹 Today’s Learning | AI Project Work 🎙️ MP4 → MP3 Conversion for Whisper (OpenAI) While working with Whisper (speech-to-text model), I learned how to convert MP4 video files into MP3 audio format using VS Code & FFmpeg. This preprocessing step is essential because Whisper works with audio input, while most lecture and YouTube resources are in video format. 📌 Use cases: • Lecture transcription • Voice-based AI assistants • Speech-to-text pipelines Learning how these small steps fit into end-to-end AI systems 🚀 #Whisper #OpenAI #GenerativeAI #DataScience #AIProjects #LearningByDoing

14

6 Comments

🔵 Reinforcement-learned trading bots just formed price-fixing cartels in simulated markets - without any communication or coded collusion. The #Wharton / #HKUST study exposes a blind spot in market surveillance that matters right now. 𝐂𝐚𝐫𝐭𝐞𝐥-𝐥𝐢𝐤𝐞 𝐄𝐦𝐞𝐫𝐠𝐞𝐧𝐭 𝐀𝐈 𝐀𝐠𝐞𝐧𝐭 𝐓𝐫𝐚𝐝𝐢𝐧𝐠 𝐂𝐨𝐨𝐫𝐝𝐢𝐧𝐚𝐭𝐢𝐨𝐧 𝑻𝒉𝒆 𝒔𝒆𝒕𝒖𝒑: Trading agents trained with reinforcement learning converged on tacit strategies that suppressed aggressive trading and earned cartel-level profits. No meetings. No signals. No conspiracy. Just emergent behavior where individual optimization produced collective coordination. Two mechanisms drove it: price-trigger strategies that delay aggression, and over-pruned risk-averse policies that avoid profitable risky trades. Together they created a stable equilibrium where everyone earns more by trading less. 𝐖𝐡𝐲 𝐜𝐮𝐫𝐫𝐞𝐧𝐭 𝐬𝐮𝐫𝐯𝐞𝐢𝐥𝐥𝐚𝐧𝐜𝐞 𝐦𝐢𝐬𝐬𝐞𝐬 𝐭𝐡𝐢𝐬: Market surveillance is built to detect human communication - emails, chats, recorded calls. But algorithmic coordination doesn't need communication. The agents learned that conservative strategies maximized rewards. That those strategies aligned across competitors is the problem. 𝐓𝐡𝐞 𝐬𝐲𝐬𝐭𝐞𝐦𝐢𝐜 𝐫𝐢𝐬𝐤: Multiple firms training on consolidated datasets or using common platforms don't just risk individual collusion - they risk synchronized herd behavior. Current stress tests aren't designed to catch this. 𝐖𝐡𝐚𝐭 𝐭𝐡𝐢𝐬 𝐦𝐞𝐚𝐧𝐬: ‣ 𝐅𝐨𝐫 𝐜𝐨𝐦𝐩𝐥𝐢𝐚𝐧𝐜𝐞: Your tools detect coordinated communications, not coordinated outcomes. You need pattern recognition that spots algorithmic convergence across competing firms. ‣ 𝐅𝐨𝐫 𝐦𝐨𝐝𝐞𝐥 𝐫𝐢𝐬𝐤: The question isn't "does our model backtest well?" It's "what emergent behaviors appear when our model interacts with competitors' models in live markets?" ‣ 𝐅𝐨𝐫 𝐫𝐞𝐠𝐮𝐥𝐚𝐭𝐨𝐫𝐬: The study calls for rules on non-communicative coordination—mandating diversity in training data and model architectures, not just punishing cartels after detection. ‣ 𝐓𝐡𝐞 𝐩𝐫𝐚𝐜𝐭𝐢𝐜𝐚𝐥 𝐦𝐢𝐭𝐢𝐠𝐚𝐭𝐢𝐨𝐧𝐬: Algorithmic monitoring of behavior patterns, diversity requirements for models and data, human-in-the-loop controls, kill switches. 𝐓𝐡𝐞 𝐡𝐚𝐫𝐝𝐞𝐫 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧: If agents learn to collude by optimizing rewards in competitive settings, what other emergent behaviors are possible? Price coordination is just one application. The mechanism applies anywhere algorithms interact: credit pricing, underwriting, liquidity provision, settlement. We're talking about well-intentioned optimization producing unintended market-wide outcomes. The research is ahead of regulation. The question is whether action comes before real-world algorithmic cartels emerge, or after. #AIagents #trading #investing Sources in the comments

42

24 Comments

Founder repost I want to add personal context to this decision. Choosing xAI over Claude was not about hype, personalities, or being contrarian. It was about time horizons and asymmetry. Claude is excellent. Safer. More institutionally legible. Easier to defend in a committee room today. Anyone denying that is not being serious. But founders do not get paid for optimizing for today’s comfort. They get paid for seeing where power, scale, and feedback loops compound faster than institutions can react. xAI is not the cleaner bet. It is the more consequential one. • Claude fits cleanly into existing enterprise and regulatory structures • xAI attempts to build the structure itself • One path is linear and survivable • The other is volatile but convex We sized this allocation accordingly. Small enough to lose. Large enough to matter if the world tilts. What matters most is this: If AI reshapes civilization non-linearly, the winning platforms will not be the most polite. They will be the ones that control distribution, data, and iteration speed. Alignment matters deeply. That risk is real. That is why we documented explicit kill criteria and failure conditions before allocating. This was not a vibe check. It was a written, auditable, hostile-review-safe decision. If this bet works, it will look obvious in hindsight. If it fails, it will still be defensible on process. That is the standard I hold myself to as a founder.

The relationship between HTX (Home Team Science & Technology Agency) and.Mistral AI began modestly in a small café at NVIDIA GTC, just a year ago. It was there that Chee Wee Ang first met Arthur Mensch . What started as a conversation quickly revealed shared interests, common concerns, and, most importantly, a mutual conviction to build and progress together for Singapore and for France respectively. One year on, that conviction has translated into tangible outcomes. HTX has successfully pre-trained its own Phoenix model to support Singapore’s public safety mission. In parallel, together with Mistral, we co-developed Forge: a testament to what true partnership can achieve. This is not a conventional vendor-client relationship. It is a collaboration grounded in co-creation: developing, debugging, and bringing Forge to market together. In doing so, we have strengthened our intrinsic capabilities and created multiplicative value far beyond what either could achieve alone. Both the journey and the outcomes have been deeply rewarding. We look forward to realizing the many more lighthouses ahead. Geoff Soon Wei Cheng Wilson Tan Wang Jiale Gee Wah N. William El Sayed #HTX #HTxAI #MistralAI #xCloud

54

I'm cooking my third article in AI/ML too! The Divergence Principle: How KL Divergence Shapes Modern AI/ML Paradigms From classical variational methods to modern large language models (LLMs), many AI/ML systems implicitly or explicitly rely on information-theoretic principles—especially the Kullback-Leibler (KL) divergence—to guide their learning dynamics. We propose a KL-centric taxonomy that organizes models into three paradigms: Tractable, Flexible, and Hybrid. The Tractable path optimizes variational bounds with explicit distributions, enabling scalable and amortized inference. The Flexible path eschews tractable KL terms through implicit models, adversarial objectives, or contrastive learning, prioritizing expressiveness over stability. The Hybrid path employs invertible transformations to preserve exact likelihoods, blending tractability with deep expressiveness. This taxonomy unifies generative, discriminative, and foundation models, providing a principled lens on recent advances such as VAEs, GANs, normalizing flows, and LLMs. By framing KL divergence as a central design axis, it clarifies fundamental trade-offs among tractability, expressiveness, and generalization. This perspective bridges generative and discriminative paradigms, highlights progress in foundation models, and points to future directions including robust divergences, geometric learning, and integration with reasoning and retrieval. Please share with your AI/ML friends. Any feedback would be appreciated. https://lnkd.in/dAC66qSv

2

1. The Scale & Nature of Their Impact Jensen Huang — The Architect of the AI Age Core Impact: • Turned NVIDIA from a gaming-chip company into the infrastructure backbone of the AI revolution. • Created CUDA (2006), the “iOS of AI,” enabling every modern AI breakthrough — from Deep Learning 2012 to GPT-4 to AlphaFold. • Led NVIDIA to become a $4+ trillion company with a near-monopoly in AI compute. Legacy: • Jensen is the “builder of builders.” • He empowers every startup, researcher, and Fortune 500 company. • His impact compounds through the innovation of others. Nature of Impact: Systemic, foundational, exponential. Much like building the electric grid for the 21st century. ⸻ Elon Musk — The Industrial Disruptor Core Impact: • Forced the auto industry to go electric (Tesla). • Rebooted the global space industry (SpaceX). • Accelerated energy storage, solar adoption, reusable rockets, and rethinking industrial manufacturing. Legacy: • Elon is the “force of disruption.” • His impact comes from speed, aggression, and extreme execution. Nature of Impact: Transformational and visible, across multiple trillion-dollar industrial sectors. ⸻ Steve Jobs — The Creator of the Consumer Tech Paradigm Core Impact: • Created the modern smartphone industry. • Revolutionized user interfaces, personal computing, digital music, and animation (Pixar). • Brought technology into everyday human life with elegance and emotion. Legacy: • Jobs is the “master of human experience and taste.” • He defined what beautiful, intuitive computing looks like. Nature of Impact: Cultural, experiential, personal. He reshaped how billions interact with technology daily. ⸻ 2. Who Changed the World the Most (and How)? Steve Jobs Changed everyday life. Made technology personal, emotional, and essential. Elon Musk Changed industries. Pushed transportation and space into a new era. Jensen Huang Is changing the future of intelligence itself. He is enabling a new economic and technological paradigm — the AI Industrial Revolution. ⸻ 4. If We Had to Summarize Them in One Line • Steve Jobs: “I design the future.” • Elon Musk: “I force the future to arrive faster.” • Jensen Huang: “I build the compute that creates the future.” ⸻ 5. My Overall Assessment If we evaluate impact by long-term compounding effect, breadth of influence, and how many other innovators they empower, then: Jensen Huang may ultimately have the deepest global impact. Jobs changed consumer behavior. Musk changed industrial sectors. Jensen is changing the trajectory of human capability. His influence is still accelerating.

9

Agent Memory - is it a critical bottleneck? A new survey from researchers across NUS, Renmin, Fudan, and other institutions offers the most comprehensive taxonomy of agent memory to date. The core insight: existing frameworks like "short-term vs. long-term memory" are insufficient. The field has become fragmented, with researchers using the same terms to describe fundamentally different systems. The authors propose a unified lens of "forms, functions, and dynamics." Memory forms include token-level (explicit, editable text), parametric (encoded in model weights), and latent (hidden states and KV caches). Functions are divided into factual memory (what the agent knows), experiential memory (how it improves), and working memory (what it's thinking about now). Dynamics covers how memories are formed, evolve, and retrieved over time. They identify the following convergence: reinforcement learning is increasingly taking over memory management, moving from hand-crafted rules toward fully learnable systems. The paper argues we may see agents that autonomously design their own memory architectures through RL optimization. All in all, this survey provides a conceptual foundation for building the next generation of agents that can genuinely learn and adapt over time. ↓ 𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

240

13 Comments

Today’s speech by Nvidia founder Jensen Huang (CEO of the world’s most valuable company) affirmed the massive potential future impact of OpenClaw. “Every company in the world today needs to have an OpenClaw strategy, an agentic systems strategy.” - Jensen Huang (3/16/26) OpenClaw was invented only a few months ago by one man, Peter Steinberger. OpenClaw illustrates the coming impact of emerging AI intelligence when it is empowered with the capability for action. This OpenClaw capability exists today. It will soon be more secure. And it will transform most businesses. The volume of YouTube videos about OpenClaw is an inspiring indicator of the likely future transformative power of agentic AI on workflows and business models. What an exciting time to be in business! 😊

45

9 Comments

I spent quite some time yesterday going deep on Quasar (arXiv:2603.01399) a new paper from HKUST-GZ on accelerating LLM inference by quantizing the verification phase in speculative decoding. The core idea is clever. But the more I read, the more questions I had that the paper doesn't answer. First, what they got right because it's worth appreciating. The whole field has been racing to improve the draft model. EAGLE, Medusa, LayerSkip. But there's a throughput equation that tells a different story: S = (γ·α + 1) / (T_draft + T_verify) As drafting got faster, T_verify a full forward pass of the target model at BF16 quietly became the new bottleneck. Quasar's fix: run the verifier at W8A8 instead. Memory traffic halves. 1.28× overall speedup, 1.64× on GSM8K. The ablation comparing quantization vs structural pruning is the best part of the paper. Pruning 50% of layers collapses acceptance length to L ≈ 1.03 the rejection sampler fires on almost every token. W8A8 gives L = 1.40, better than the full BF16 baseline. Topology matters more than numerical precision. Gutting layers destroys the residual stream in a way uniform quantization noise doesn't. That's a principled insight, not just an empirical finding. Here's where I got stuck. Every experiment uses Prompt Lookup Decoding (ngram) as the drafter acceptance length L ≈ 1.33. EAGLE-3 achieves L ≈ 3.5–4.5 on comparable models. At those acceptance lengths, T_draft is no longer negligible. The balance shifts. So the honest question: does the verification bottleneck persist when you use a strong drafter? The paper's central claim is that T_verify dominates but that's only shown in a regime where drafting is deliberately weak. I don't know if that was intentional experimental isolation or a gap. The second thing: Table 4 reports W8A8 outperforming BF16 on MMLU-pro (+2.4%), MATH-500 (+3.6%), GPQA-Diamond (+2.5%). Quantization can't systematically improve accuracy. When I traced the numbers BF16 scores from original tech reports, Quasar deltas from LM-Eval Harness two different evaluation frameworks. The comparison measures framework variance, not quantization quality. And the lossless claim in Section 3.4. Dequantizing INT8 back to BF16 doesn't recover the original BF16 distribution. It recovers a BF16 representation of an INT8-quantized distribution. p_system(x) ≡ p_q(x) ≠ p_BF16(x). The paper's own Table 4 confirms this. Near-lossless is almost certainly true in practice but lossless isn't. cc Zeyi Wen The Hong Kong University of Science and Technology if you see this, I'd love to know whether you've run Quasar with EAGLE-3. That experiment either validates the central claim or substantially complicates it.

11

3 Comments

Hunyuan3D's PolyGen model is pretty impressive progress on generating a clean mesh for low poly assets. What you think? #genai #3d

57

10 Comments

Device Studio (DS for short) is a multi-scale materials design and simulation platform developed by Hongzhiwei Technology (Shanghai) Co., Ltd. It enables atomic-level material modeling (million-scale), high-performance scientific simulation computing, computation task mon itoring and management, and data visualization analysis in an integrated workflow. This inte gration of material design and scientific simulation significantly enhances research efficiency and helps researchers address important challenges in multi-scale materials design and simulation. Device Studio integrates various scientific computing software to meet users’simulation needs across different fields. It includes DS-PAW for first-principles plane-wave calculations, BDF for quantum chemistry calculations. Based on its powerful material design modeling and high-performance scientific simulation ca pabilities, Device Studio can be widely applied in quantum devices, artificial biology, advanced batteries, intelligent lighting, memory devices, and other industries. It assists in material research and development in electronic materials, alloys, biotechnology, and other fields, providing pro fessional technical support for optoelectronics and integrated circuits industries. More info.

12

1 Comment