🇲🇾 Just wrapped up nearly 50 days in #Malaysia, my now second home. Here in my search to explore the supply-chain of AI and discovering how Malaysia is uniquely positioned as a chip manufacturing power-house, the next source of cheap energy for hungry data center's and a growing engineering talent pool for south-east Asia. Was a great time speaking to everyone from artists to policy makers and even global medical device companies like @dexcom. Massive shout out to @AntlerGlobal Malaysia for letting me mentor and coach their AI entrepreneur boot-camp, was an eye opening experience to hear from builders founders in the region talking complex problems. I have been writing non-stop the last 2 weeks can't wait to share very unique perspectives on East meets West, supply-chain and a global view on AI. Here's a glimpse on some of my takes so far: ➡️ The global semiconductor industry is shaping up with tariffs and the explosion in AI which is causing some shifts in where and how chips powering computers, GPUs and AI is made. ➡️ Electricity is the new oil, the chips are the machinery, this is new supply-chain race - its no longer just about "information" of training data and huge foundational models. Without the source materials this is not possible. ➡️ Like it or not AI is now a consumer product not a "technology", I was in an Apple store and consumers asking "how can I install Deepseek" at the point of purchase. What dose this mean for how AI is adopted by consumers and into everyday products. ➡️ Sovereignty with foundational models, training and evaluations are more critical than ever for counties to preserve their unique history, language and sense of identity - but is this even feasible, how is this possible? ➡️ Developing nations will continue to do more with less, expect the "Deepseek-effect" more and more with huge emphasis on open-source over closed source approach. Even speaking to colleagues of a "post deepseek" era for the next AI cycle we are going into. ➡️ Every growing automation anxiety where lay-persons sense of identity is shaken as many in the corporate world their identity's == career and now fearing potential displacement. ➡️ #Agentic is the marketing term of 2025, yet no one can clearly tell-me what and agent actually is - yet alone AI clearly. Still lots to be figured out as everyone has a view and its divergent. I'll be dropping articles and some research pre-prints over the coming weeks along with some other super exciting developments. Feel free to reach out in comments. Which statement resonates with you the most?

I just posted @ericzakariasson's /deslop cursor command to Codex :), I'm tempted to cook up some automations and symlinks to sync all my slash commands between these agentic IDEs and tools. You can grab deslop here gist.github.com/vincentkoc/d…

Offt, here we go with video llm-as-judge. Locally running @Cometml opik along with @Alibaba_Qwen 3.5-vl on @vllm_project and @ollama all on my macbook. Can't wait to ship, fully open-source multi-modal goodness

Hot take... This is what we call context engineering in 2025? 🤷‍♂️ Multiple layers of abstractions that remind me of my childhood where i would hide the sweets in the last doll in a stack of matryoshka LLMs call MCP Gateways, which call MCPs, which spawn sub-agents each layer adds context, cost, and failure modes. Although this seems logical but are we adding more and more layers of noise and complexity? This causes more failure points, more risk and more noise in to debug and evaluate.

Whos gonna finetune a vision generation model on these diagrams 🫡

@_bmacd 🫶

Shipping @Alibaba_Qwen video-to-text support for @ollama as part of my hackathon with @Cometml as I'm working on bringing video based llm-as-a-judge evals to Comet Opik. Nice to give back to opensource multimodal LLM community! github.com/ollama/ollama/pul…

Be developer wanting to use nice new text to video models with Qwen3-vl. Bruh, no frameworks support, decide to patch @ollama to bring full video support 🦙. PR incoming so i can get back to building video evals on @Cometml ;)

Who needs to go trick or treating when you can just ship stacks of PR's with your AI coding assistant. Close to 50k line changes in 72 hours across open-source projects 🚀

What happens when you forget to pay your bills on your monthly robot? 🫥

Optimizers baby! We just patched a part of @LangChainAI that is touched by nearly every AI application to force JSON. It used to only run at 10-15% accuracy, using our optimizers at @Cometml we improved to 95% accuracy (by +705% vs original). github.com/langchain-ai/lang…

So… - openai/chatgpt is down - linkedin is down - discord &stream is down What’s happening?

#shotoniphone 👌

Few of my recent snaps in Tokyo ✌️

Crazy weekend of hacking alongside with @Cometml at the largest student run hackathon with 3k+ people building with AI and robotics at CalHacks with @UCBerkley. Got the chance to sponsor a great event and event helped with the student poker tournament as dealer 🙌

Most people still in single dimension realm of AI (i.e text based LLMs) but when we start to think of intelligent systems with space (spatial), time (temporal), vision (visual) and other forms of intelligence and data need to stored, precived and understood. Bullish on multimodal

Fun stuff @Qantas put a “urgent, do not board, regrade if possible” on my next flight. Lady dismissed the message and said there is no issue with my boarding, not knowing i can see the screen. Clearly QF tried a last ditch to save me but failed 😬😂 (QF25)