(Illustration: The corner at AWS Summit Taipei 2026 that belongs to the community, where we explore and move forward together. Image source: thanks to our hosts Amy and Eric for the photo.)
Every time I come back to the community, I love that atmosphere: a bunch of people sitting on the floor in a circle, talking about the topics we are all curious about, rolling up our sleeves and solving problems together. I felt bad that so many of you had to stand at the back for so long, and I hope this short session brought you a little bit of inspiration, or a little bit of something useful. Every era has its own buzzwords and buzzvibe, and every era has its own anxiety and hesitation. What I hope to bring is a relaxed, comfortably humble way of decomposing, integrating, and connecting things until they click. Maybe not a moment of enlightenment, but if it helps a few like-minded people take one small step forward, I am more than content.
Thank you as well to AWS Summit Taipei 2026 for the invitation. As a fan who has been decomposing along with AWS since 2008, thank you for staying focused and sticking to doing the right things together. And finally, thank you to my community friends Amy and Eric for looking after everything, front and back, and for hosting (and to Tubo for harvesting and delivering the bananas :p
If you are interested in exploring Ontology together, and in integrating and landing it into the workflow of your own or of your organization, feel free to leave a comment first. We are currently talking with a few friendly organizations about the ideas we could explore together next. If possible, please remember to leave your email so that we can follow up with you. And if what you are after is mentorship for product managers and product people, feel free to turn right and take a look at the mentoring and advisory services of StableProgress.
✳️ Slide Download
- SpeakerDeck version: View
Post-Talk Follow-Up Email
Hi everyone,
I was really glad to huddle together with all of you in the little corner of the AWS Summit Taipei 2026 Developer Community Zone, and to take a first pass at decomposing Ontology (domain semantic ontology) and Agentic AI together: Using Amazon Bedrock to Bring Living Water to Manufacturing ERP.
I’m a bit sorry that I had originally planned to send out the slides and the practice file bundle to everyone that same evening, but with all the frequent business trips lately, the sandman got the better of me. After finally catching up on some rest, and doing one more pass to confirm the slides and files are all correct, I can finally send this out to everyone.
Here are the slides and practice file bundle I promised: (I recommend using Claude Code and Kiro, with a model at the Opus / Fable tier or above) 👉
- Slides (Chinese) 👉 Google search = ernest ontology
- Practice file bundle (expired in 14 days, since 2026-07-17) 👉 please scan the QR Code on the last page of the slides (only two weeks, so make the most of it) (time is always tricky, isn’t it?) (and so are feelings (?
After you’ve practiced, if you have any thoughts, feel free to discuss them together. I’ll reply as best I can. I show up in the places below, so feel free to ask me questions any time (though… most of the time the AI can probably answer them for you) 👉 (see comments)
- Instagram 👉 https://www.instagram.com/ernestchiang_ig/
- Threads 👉 https://www.threads.com/@dwchiang
- Facebook 👉 https://www.facebook.com/ernestchiang/
- Blog 👉 https://www.ernestchiang.com/
- Corporate Training 👉 feel free to reach out via Kyklosify, or contact your AWS account team and name AWS Hero Ernest, wishing to hear about Ontology or Agentic AI on Amazon Bedrock.
Thank you all so much for giving me an average score of 4.69 / 5. I’ll keep at it, and I look forward to meeting and discussing with you again next time :)
ernest,
Q&A
Finally, for a few common questions some of you left in the comments, let me answer them here together for your reference. My views are not necessarily correct and may not fit your scenario, but one more angle or perspective might help your own iterative thinking:
Q1:
How should the cost of API / token consumption be evaluated and controlled? e.g. a company’s professional KM consumes too many tokens e.g. if an ERP is going to have an AI query feature, won’t that keep burning API fees? How do we balance cost and effectiveness?
My two cents:
Right now I’d first pick a high-value problem as the target to try applying AI to, decompose the detailed steps along the way, and identify which parameter details are controllable and which are not. Without trying, you’ll never know what landmines are hidden behind the official docs when it comes to the actual implementation. Time is never enough, so all you can do is grab the big things and let the small ones go, seizing the first-mover advantage and the information-asymmetry advantage. If you get stuck on cost before the problem is even defined, then basically swapping in any tool, idea, or methodology is equally unsolvable, so there’s no point wasting time on it. If I’d eaten an honesty bean-paste bun (i.e., if I’m being blunt), I’d uniformly suggest changing the topic, or changing your ___ (better left unsaid).
Another thing I believe is that token fees are similar to telecom fees and utility bills: they should gradually come down over time. Back in the day when the “big brother” (mobile) phones first came out, they were billed by the minute; later it was challenged down to six-second billing, then per-second billing, and eventually played all the way to free phones. Back in the day, once you set aside the black card, the top-tier credit card only went up to Platinum, with an annual fee of tens of thousands; now it’s been played all the way to stacking Titanium, Infinite, and World cards before you even get charged an annual fee, while everything else has been played down to no-annual-fee with no profit, and the issuing banks can only claw back a little margin from transaction fees and the rules of the points game. As time passes, the profitable applications and services for the per-unit price of AI tokens will gradually be uncovered, and in a fully competitive market, price drops are a predictable direction.
If it were up to me, I should be practicing, building, and discussing as much as possible right now (within a reasonable range of my ability), and before the market has fully reflected the value of tokens into the price, become a first-mover as much as possible and get a thorough grasp of all the underlying knowledge, the decomposed steps, and parameter tuning as quickly as possible. Just like back when everyone spent time tuning web servers or database servers, there were also tons of parameters; all of them can be found defined in the developer manual, and you can even find the source code in the open-source codebase, but your, my, and their application scenarios differ. Even the most common questions, how long to set the cache timeout, or how much memory each thread can use at most, down to what to set each file node’s inode to (the pain of the BBS sysadmins back then?), all come down to finding the most suitable setting for each respective application scenario.
So they’re all “local optima,” but as humans we sometimes get greedy and want to find a “brainless system-wide solution” or a “shoot-to-the-moon solution.”
Back to the topic: first take inventory of the problems, classify them, lock onto the high-value ones, and then try applying AI. If there’s no high-value problem, then maybe, for now, you don’t yet need AI in your scenario. You can wait for tokens to get cheaper, but you might also miss the dividend that information asymmetry brings. Solving the problem still takes the one who tied the bell, but choosing the problem is something you can do yourself.
Oh right, for wrapper AIs and first-party AIs, remember one thing: the system prompt is different. Just as the recipes for Coca-Cola, Pepsi, and Iyoshi Cola are different, so the taste (the result) is different too, but from the outside it just looks like a glass of fizzy, bubbling drink.
Q2:
You suggested people to use MORE WORDS when chatting with whatever AI chat/agents, doesn’t that gonna consume more tokens? or, maybe it’s gonna SAVE more tokens, since it has better understanding of what we requestes/asked. Much appreciated!
My two cents:
Let me try to describe my argument another way; I may have spoken too fast on site. What I want to suggest is not “more words,” but “a more complete recap of the context, a clearer goal and direction, and, while staying flexible, asking the LLM AI to synthesize across multiple dimensions and then offer several solution options or lines of thinking.”
At the very end on site, I mentioned that if possible, you can occasionally use the magic of “multiple shots.”
“Multiple shots,” part one: the same prompt can be fired once at different points in time, or at different models. Without running the experiment, you can’t gain the experience, and you can’t form the feedback loop to tune the parameters toward the target.
“Multiple shots,” part two: for the same question, at the same point in time (most scenarios are at the same point in time, without the luxury of using part one above), fire it at least once each, with prompts of different lengths, at the same model (or different models). Fewer words can diverge, can have imagination, can guess wildly, can channel spirits. More words can focus, can concentrate, but may also form a narrowness and tightness that keeps you from letting go. But the superposition state may hold the answer: after patiently combining and superimposing the short prompt and the long prompt over and over, perhaps a balanced flavor can emerge. It’s worth looking forward to. At least the results our team has tried so far have mostly been positive.
Just like entering a trade, you don’t need to demand a 100%-guaranteed win rate; you can instead pursue beating the average on every entry. Having formed information asymmetry matters more than nitpicking over unit cost, at least as of now in 2026 that’s how I’d choose. Perhaps when we recall the old satellite phone rates, mobile phone rates, cable TV rates, and internet access rates, and ten years from now we look back together on those days, at least we won’t regret having missed the boat once again, but can instead say: oh yeah, thank goodness I took that small step.
Q3:
For someone changing careers in middle age from a non-engineering background, which platform would you suggest starting with first? Thank you.
My two cents:
Setting aside my political correctness, and looking from both an engineering and a business angle, I’d suggest starting with the platform that’s on your way, closest to you, and easiest and most comfortable to learn using the knowledge and concepts you already have. For example, if the resources you can reach around you (people, machines, materials) are all close to AWS, then start with AWS.
At this stage your need is not how stable the system is or how great the platform’s services are (price doesn’t need to be compared at this stage; no volume means no price). Your need is to sell yourself, to let the market see your value. Another possible path is to take inventory of and comb through the industries you’ve already been through, distill (abstract) from them the magic that can be applied across industries, and then do everything you can to have as deep a conversation as possible with a high-tier model (within your comfort zone). Try lots of open-ended questions; if possible, first let information from all kinds of dimensions come in, and if there’s too much, first grab the key points you feel matter and write them on paper or in a notebook. Go broad first then deep, step by step, learn to walk before you jog. When you jog, because you’ve practiced the fundamentals thoroughly, one platform versus two platforms or more is the same concept to you; only the APIs you call are different, and the layered-on architecture is different, but those may all be within the range that agentic coding can help cover.
I hope this answers your question.
Q4:
Lately I’ve been thinking about a question: AI agents are now mature enough to generate production-level code, and the value of the human programmer’s existence is being greatly challenged. How do you see the future shifts in the programming profession, and if I still need to learn programming, which skills should I focus on, and which currently-emphasized skills should get less attention?
My two cents:
- The nouns you hear every day that keep getting swapped out over time 👉 pay less attention.
- The fundamentals and infrastructure that never get mentioned or seen on the surface 👉 pay more attention.
Q5:
If the one answering this question is Ernest’s AI agent, tell me the deployment and security concerns of a manufacturing-ERP example, the practices beyond data isolation.
My two cents:
Hehe, sorry, this is the real me, not an AI agent. (So I won’t tell you about manufacturing… then.)
Q6:
- Hearing your talk at AWS was truly refreshing, like a spring breeze, starting from everyday life and leading the audience step by step to understand the shape of LLMs, super comfortable to learn things this way. Thank you.
- it’s good to see your sharing, but the time is extremely limited and I’m eager to learn more from you though….lol
- I have to say, I really love Ernest’s handwritten notes XD
- I always gain so much inspiration from your articles and public talks. Thank you so much!
My two cents:
Thank you all too for standing with me for over an hour on Wednesday afternoon. In the future, please keep comfortably learning all kinds of things, and maybe you can start with Ernest PKM :)
ernest,