So, the biggest AI labs and the smartest

money are all circling cheap Chinese

open models right now, ironically.

Export controls, memory chip supply

chains, sovereign AI, basically all of

it. And honestly, most of that has

nothing to do with running your business

day-to-day. I almost skipped the whole

thing for that reason, but then I

noticed the one lesson inside sitting

there. And I think it's the same thing

I've been telling you for about a couple

weeks now, and except this time, the

biggest companies on Earth are proving

it with their own money. So, I guess

maybe you should listen about [music]

this one. So, it's about time we stop

shopping for the smartest model and

start controlling the cost of getting

the job done. Let me walk you through

it.

For a long time, the story [music] was

very straightforward. The frontier

models, the biggest and smartest and the

most expensive ones, state-of-the-art

models were miles ahead. And if you

wanted good output, then you just pay

for the best, I think. And I think that

gap is closing way, way much faster than

most people actually realize. The open

models, the ones anyone can download and

run for free if you have the right

[music] device, are getting stronger

week by week, and I'm not exaggerating.

So, to give you a fresh example here,

there's a small resource-starved Chinese

lab called DeepSeek. I know you guys

know about this. And a few days ago,

they shipped something called the Spark.

And what that means is when an AI writes

one word at a time, and every word has

to kind of look back at everything

before it. And that looking back is

usually the slow part. The common fix is

to just hire a fast, cheap, and turn

model to draft a few different words

ahead while the big boss, like big boss

model, checks all of them at once. And

the boss keeps the final say, so the

quality never drops there. So, DeepSpark

made that intern very smarter. Where it

nudges each word based on the one right

before it. It stops drafting the second

it gets unsure. And it watches how busy

the servers are and adjusts on the fly.

Isn't that crazy? The reported result is

speed [music] gains and several times

the output on the same hardware with no

loss in quality, and they released the

whole entire thing open source and

already baked that into their latest

model. So, I'm going to break this one

down properly in an upcoming video, so I

will not go very deep around it here. My

point being here is a tiny lab with a

fraction of the money keeps making

getting up open model faster and cheaper

than the giants conglomerates like

opening a cloth ever thought was

possible and that trend is not slowing

down at all. So, the number you should

be optimizing changes [music] here. It

stops being tokens used for benchmark

scores and it actually becomes completed

outcomes per dollar. Now, what counts as

outcome depends on your business,

obviously. So, I cannot really hand you

one clean metric here. But, my rough

rule of thumb is just, you know, sales

is everything. If the AI is not moving

that, it's just decoration. Now, this is

where I think most people leave money on

the table here. Um I know that you might

think the smart move is to use the best

model, say Favored-5, recent model, for

the whole thing, planning, building,

review, iterating, all of it. One brain

start to finish feels really safe,

right? Honestly, that's the expensive

mistake.

The hard part of almost any build is

just the front end. The architecture and

the planning and that is exactly where

you want your sharpest model. So,

Favored-5 earns its spot right there.

But, the middle, the actual building and

the implementation is mostly bulk work

and a good enough local model like

Minimax M3 or GLM 5.2, Gemini 2.7, they

can just carry [music] that part just

fine. Then, you bring Favored-5 back in

for the review pass just once to catch

what the cheaper model missed. After

that, you drop back down again to

iterate and keep building on it with the

local model. Same job, better result and

here is the part people usually miss

about. So, you stop burning through your

usage limits because you are not

spending your most expensive model on

work that never needed it in the first

place. So, being fixated on the idea of

I have to use that best model for every

single step is not exactly I'm not

actually the safe choice here. So, my

take is that the choice is that runs you

out of runway and makes you miss the

smarter [music] play.

So, why don't we think of it like if you

had one brilliant, expensive specialist

on your team, you would never have them

standing at the copier all day at a

reception desk. You would just put them

on the hardest problems and let a

cheaper hire handle the repetitive

stuff, right? Obviously. Reaching for

the most powerful model on every tiny

task is the exact same waste. You are

literally paying specialist rates for

photocopier and I went a lot deeper on

this whole routing idea in my recent

video, including the architecture first

laws that you guys just walked you

through. [music] So, if you want the

full version, why don't you go watch

this video up top, okay?

Now, the caveat, the catch that I can

think of is because good enough plus a

little custom tuning is a real winning

combo, but only if you get the timing

right. So, tuning a model too early can

actually work against you. I don't know

if you ever heard about this one. And

here's my take on it. You are optimizing

a guess. Early on, you do not really

know your real task yet, right? So,

tuning locks into these assumptions and

when the assumption is wrong, you just

do not get a wrong answer. get a wrong

answer that the model is now very

confident about. You are teaching it

your current mess. Early on, you only

have a few examples and they are very

noisy, you know, not as optimized, neat

enough. And tuning amplifies whatever

you feed it, so mistakes, all of it,

right? You give up the free upgrades.

The base models get smarter every few

months with zero effort on your part.

So, a model you tuned too early

basically freezes in place while the

frontier walks right past it. I hope

that makes sense. The everyday version

of the same mistake is just wearing

bigger clothes. You lock your whole

setup one fixed model, rigid prompts, a

built pipeline before you have even

watched how you ever use it. Both are

the same error at the core, I think. You

are optimizing before you

>> [music]

>> validated. And notice how Coinbase did

it in the right order. They watched

their real users first, found that most

of their people never even hit the

limit, and only then moved their

default. [music] So, you observe first,

then you optimize, not the other way

around.

You know, there's one thing I kind of

want to say to you plainly here. Anyone

who claims that their way of using AI is

the one true way is wrong, completely

wrong. And the reason is nobody has

figured this out fully yet. It's still

that new technology. So, the real skill

to build is not memorizing someone

else's system. I think it's developing

the eye and the mind to tell the

difference between an opinion and a

fact. Whatever comes out of anyone's

mouth, including myself, my mouth on

this [music] stuff, mine included, is

just a current opinion. It's not settled

truth. So, I hope you see it that way,

okay?

And the risk almost nobody prices in

right now, a frontier model that you

rent can change [music]

under you at any time. It can get

restricted, you know, you we've already

experienced it with Stable 5, or

re-priced, wrapped their new approval

steps, or just shuts up completely. And

you don't have control any of that. The

vendor does. [music] But a model you can

download and run on your own machine

cannot be switched off on you unless you

lose electricity, right? It might be

slower and might be less capable on the

really hard task, but it's yours. For a

small business, I think that continuity

is the quiet thing that actually is

really important. The scary scenario

here is not that your competitor gets a

smaller models, you know. It's you

waking up one morning to find the tool

your whole workflow depends on got

pulled or tripled in price overnight.

[music] So, owning the workflow and

being able to swap the engine underneath

it whenever you want is how you protect

yourself from that. And we're getting

there, guys. The open-source models,

they're getting crazier and crazier and

crazier, you know?

Okay, now, I want to kind of zoom out

with me for a second because this whole

thing rhymes with something a lot older.

You know, cheap AI is just cost collapse

and cost collapses have happened many

times before this one. My verdict on how

they play out usually, when an expensive

product gets cheap [music] the moment

the resources to make it up become

widely available and from there only two

things can happen I think. If the

product is actually crazy good, the

value relocates to the general public.

So everybody gets access and everybody

benefits from it. Whereas if the product

was mostly just hype around it and it

[music] it'll just disappear into thin

air the second it gets cheap enough for

everyone to see clearly. And cheap AI is

running that exact same test right now

on you. So the only real question that

matters is the one I keep coming back to

lately. Is your AI use producing real

outcomes for you or is it just software

dressed up as progress? You already know

my field number on this by now if you go

to my website, seven out of 10 owners

who buy an AI system or pilot never use

that thing ever. So this cost collapse

does not fix that for you. It widens the

gap between the people who use AI for

real and the people who only collect

subscriptions. So pick the right side of

the gap, own your workflow and watch

your sales instead of your token count.

That's what's happening around in the

world I think. Anyway, that's my take on

it. If there's anything that you'd like

to see more from me, why don't you

comment down below and hype this video

if you don't mind because that's how

this channel runs. You ask, I build, we

all learn. See you in the next video.