One afternoon we went to AI Update Bangkok 2026, at the invitation of Dr. Jim (Jimmy Tejasen), who runs the event. We expected to come home with a list of new model names. What we came home with was not a model name at all.
You have probably heard "a new era of AI" three times this year. Generative, then agentic, now physical. Every round ends with the same advice, which is to adapt quickly, and nobody ever says adapt to what.
The event gave a straighter answer. These are not three waves to pick from. They are three steps that follow each other. And when you move from the step where AI works inside a computer to the step where it moves real objects, the question changes too. It stops being how capable the model is, and becomes whether you can undo the damage when it gets something wrong.
Productize covers AI Update Bangkok 2026 here, keeping only what you can use, not a roll call of who spoke.
Part 1What Physical AI is, and how it differs from agentic AI
Pathom, from the AI hardware distributor SVOA, walked the three steps in the shortest form possible. Generative AI makes things for you. Agentic AI decides and then does the work itself inside a computer. Physical AI adds eyes, ears and sensors, and sends it out to handle real objects.
He summed it up in one line: generative AI understands information, physical AI has to understand the world.
Say the words and most people picture a walking robot. Pathom pushed back on that himself. It means anything that can connect to the internet. Cars, factory machines, or a small box with a sensor in it. Robots are only the most visible face of it.
One more thing worth saying. The new step does not replace the old ones. The part that thinks is still the same language model we all use every day. What got added is hands and eyes.
So what do you do with that? Pathom turned the question around on stage.
The business question is not "How much is the robot?" It is "Which jobs in your organization still require a human mainly because the world was designed for humans?"
On the microphone he put it more briefly, saying we are moving into an era where the question is no longer how much robotics you need, but whether there is any work left that genuinely needs a person. What makes that question useful is that it separates work that truly needs a human from work that needs one only because the stairs are this high, the door is this wide, and the parts sit at this height.
Try it on your own workplace. The answer is usually not the one you would have guessed.
Part 2Three model groups sit on one curve, not three quality tiers
Before the robots, a detour through model selection, because it is the floor everything else stands on.
On stage Somchai put model capability against price, then drew a curve through the best value option at each price point. That line is called a Pareto curve. Three groups sit on it, cheapest to dearest: compact (small and cheap), commodity (mid range, good for general work), and frontier (top end, most expensive). Another dozen or so points scatter below the line, labelled not recommended.
We first read it as three quality tiers, cheap up to good. It is not that. All three groups are equally good value. They differ only in what you can afford.
You do not lose money by picking the wrong point on the curve. You lose it by not noticing you have fallen off the curve. A point on the line can be moved later. A choice off the line costs you on every single call.
Read it that way and the question changes. Instead of asking how capable a model this work needs, you ask what we can afford here, and whether the model we are on is the best value at that price. The first question walks you up the price list. The second sends you back to check whether something cheaper does the same job.
One small detail that matters. The price axis he used is cost per task, not per token. They sound alike and are not. A cheap model you have to run three times before the output is usable is not cheap.
Part 3Why Physical AI has to start in simulation
This part never mentions model intelligence once.
When an agent gets work wrong, you tell it to do the work again, take some mild criticism, and move on. When a robot picks something up and drops it, the object is already broken. There is no running that again.
What decides how you build is not how accurate it is. It is whether a mistake can be undone. Once it cannot, the whole method has to change. Run it in simulation as much as possible, harvest as much data from there as you can, and only then let it out into the real world.
We think this reaches well past robots. Try the same test on the work you already hand to AI. Drafting an email, summarising a meeting, writing a first pass: all of that can be fixed after the fact. Work that touches the live database, work that sends something out to a customer, work that deletes files: that is a different category entirely.
In practice that means work you can undo is work you can let AI try. Work you cannot undo needs somewhere to rehearse, and a person to confirm before anything real moves.
Part 4A local LLM that competes is not one that speaks your language
Pathom asked the room whether a country like Thailand could build a large model to compete with China or the United States, then answered it himself: no. But a local LLM, yes.
The line after that is the one worth sitting with. Stop treating "local LLM" as a claim about being good at Thai. What it should mean is a model that is genuinely good at one field, expert in what a particular group of people know deeply, not fluent in the language they happen to speak.
His example was Thai medical schools building their own models to read X-rays, MRIs and tissue samples. The interesting part is not the model. Software like that has always been sold bundled with the machine. Once you own the model, you can buy a bare machine from anyone. What locked you in was never the hardware. It was the software that came with it.
For anyone building a product, that is a map for finding your own opening. The opening left for you is not the one where you have to be bigger. It is the one where you know more.
Part 5What still sells once clients can build the prototype themselves
Pathom told a story from his own software business. Clients now arrive with something already built. Not a pretty mock-up, but a working app with a full interface, roughly 80 percent done.
So why keep hiring anyone? He answered plainly: if you build it yourself and it breaks, who do you shout at? Hire a company and there is still someone to hold responsible.
Which means what still sells is not the coding, it is the person who stands behind it when it breaks. He also said he has no idea whether that will still be true in five years.
Both speakers closed on the same worry, arriving from different directions.
Somchai said an employee had asked him directly what she would do once her process no longer needed people. He said he genuinely could not answer. The only thing he could answer was that if you do not use it, you cannot compete with the people who do.
Pathom had seen something a step further. He described a writer who stopped writing after reading what AI produced and deciding he could not match it, and a composer who stopped for the same reason. What worries him is not people being replaced. It is people deciding for themselves that they cannot compete, and walking away while nobody has asked them to.
He said he wants them both back at work, because what a person puts in and a machine still cannot is feeling, and what they have actually lived through.
What to take away
- Stop asking how capable a model this needs. Ask where this work sits on the value curve, and whether you have fallen off it.
- Count cost per task, not per token. A cheap model you run three times costs more than an expensive one that lands first time.
- Sort work by whether a mistake can be undone, before sorting it by difficulty. Reversible work, let AI try. Irreversible work needs a rehearsal space and a person to confirm.
- Your opening is depth in a field, not language. If you are building something that competes, start from what you know better than anyone else.
- What still sells is accountability, not speed of production. Design your service from there.
If you take only one, take the third. It changes the first question you ask before handing work to AI, from can it do this, to if it breaks this, what do we lose that we cannot get back.
That second question is far easier to answer, and you can answer it before you start.
Source
Everything here comes from AI Update Bangkok 2026, held on 21 August 2026 at The Cloud Bangkok by Jimmy Tejasen, who invited us along. Thanks to him, and to the two speakers, Somchai and Pathom, for the talks.
- Event livestream: watch the recording
- Jimmy Tejasen's work: aiserver.in.th, news and reviews of AI server hardware, and agentic-press.com, a publishing house for AI books
Both diagrams were drawn by us to explain the principle. They are not copies of anything shown at the event.