George Hotz | Programming | how I actually use agentic coding | Agentic AI

Did I set it up correctly?

Cool.

Um, so I want to show you guys how I actually use aentic coding.

Let me see if I can get that to be a little quieter.

tape.

Tape. Can you guys hear that? Is that

loud and annoying?

Um, is that loud and annoying?

No, I don't think it's audible at all.

Okay great.

>> [snorts] >> Um, so I mean yesterday is a bad example of how to actually use those things. Uh, in

reality they're useful for certain things, but you have to stay super carefully uh on top of them.

Um, so like all right, you can see it added two helper functions here. It looks like

there's a ton of duplicated code. Do not

duplicate code.

Is this used?

This fallback that I don't really like.

I had it add some assertions.

you have to carefully read any of the code that it writes.

Um, This is the other thing that I'm working on. Uh

on. Uh I'm trying to write firmware uh for the eGPUs.

Uh so I got some eGPUs in the mail. Uh

they have a nice board where we set up a really nice uh debug loop. This is an FTDI that can unbrick the chip no matter what. So, I'm having uh Claude try to

what. So, I'm having uh Claude try to enumerate the uh USB. Uh I got the UART working.

Um again, these things aren't that good and you have to hint them uh you have to help them a lot. Uh like I had to tell it to disable parody. It couldn't figure out why the things weren't working, but

that was discussed in a Slack message.

Um, so yeah, I'm working on three PRs right now. Uh, I'm working on this DS perm

now. Uh, I'm working on this DS perm instructions one.

Um, it's so annoying if I'm not logged into GitHub. I guess I can see it. Okay,

into GitHub. I guess I can see it. Okay,

that's fine. As long as I can see if it gets a green check. Uh, you know, I don't do any logged in stuff uh on stream. I'm also working on this RDNA4

stream. I'm also working on this RDNA4 PR Um, so you can see that this one unfortunately got an X on the llinter.

Uh, we can check that with precon.

Um, so I'm sshed here into a tiny box. Oh,

and also I hope this stream is working because I do have tail scale up. Um, I

need to have my tail scale up. I'm using

an exit node in order to access uh Claude from Hong Kong.

Yeah. No, I mean yesterday's stream uh sets a bad example. Uh you shouldn't actually be using agent coding anything like that. Um and there's a lot of stuff

like that. Um and there's a lot of stuff that I still do by hand uh if you want to see some of the by hand stuff, right?

But it's useful for like figuring things out. Um so I'm not working in tiny here.

out. Um so I'm not working in tiny here.

So I have an extra tiny grad to work on.

I can manage about three or four of these.

Um, but if I try to manage too many, it gets uh I just get overwhelmed.

Command to test plus on RDNA4.

add it to the AMD IDE test in GitHub actions.

Uh, and then here, so we can see this on GitHub. I mean, sometime I'll just push stuff to GitHub. Um, and I'll look at the diff on GitHub if I feel

like I can. Uh some diffs I really have to stay on top of uh by hand because you really can't start to let the slop in.

People have this mythical idea that agents can somehow manage their own uh slop.

Um fear and loathing in Gast Town. That

would have been a good name.

Uh okay. So oh just

okay. So oh just test test tiny is fine.

So this is adding RDNA4 support to the emulator. There's one new part of tiny

emulator. There's one new part of tiny grabb that I've entirely uh coded in agents. But yeah, so the only question

agents. But yeah, so the only question is whether the zero is there by default.

These are variables like I they're all assigned before use. The zero is finally replaceable.

Why would it declare if they already

exist in varss?

Show an example of this.

All right. So you can see that this thing is working. It's flashing

firmware. Whenever you can give these things loops to close themselves, they work pretty well. Um, sometimes they can't do it if they can't get any

incrementalism on the problem. Uh, like

I tried to get it to write an emulator uh that was cycle accurate and it doesn't try to like make a model of the GPU. It just adds hacks. If it can get

GPU. It just adds hacks. If it can get closer by adding hacks, it generally doesn't work. But if the only way to if

doesn't work. But if the only way to if it can get closer by like doing the right thing if like the gradient flows

um then uh it works uh correctly, right? So

see like here we can see some tests failing. See what's going on here. Um

failing. See what's going on here. Um

and then you have to kind of decide which ones you have to stay on top of.

Right? So look at my Yeah. So instead of duplicating that

Yeah. So instead of duplicating that code, it listened to me and it dduplicated that code. Um,

see this is not a good this is not a good fix. I don't understand this fix.

good fix. I don't understand this fix.

Oh, that lost mass equals true. It

didn't update that correctly.

Um, so like I might want to complain and say use set default here, right? like use

set defaults here also.

I mean, I could also have it check the type uh maybe if I want to add more assertions, but you still nothing about agentic coding prevents you from having

the entire state of the system in your head. And it also can hurt you because

head. And it also can hurt you because when you write the code yourself, it's much harder to delude yourself into

thinking that you understand it. Um,

but when the agent writes it, you can like look it over and be like, "Yeah, I understand all of that." And then you read it more carefully. Uh, so I think Carpathy has a lot of good things about

this. He talked about how

this. He talked about how you uh you start to atrophy your code writing skill, but you improve your code reading skill.

All right, so that uses set default now.

All right, I'm pretty happy with that diff. Um, [snorts]

diff. Um, [snorts] let's run Python path equals this precomand like mosh or something set up. So, this

just works better.

So, it's adding some timing and fireet.

So, this is reverse engineering um firmware. These names are terrible.

firmware. These names are terrible.

I'm going to interrupt it and say uh do you have better names slashmore

info from registers.h.

registers.h.

This will help you debug.

You must understand all registers, not just copy.

So yeah, I mean it just it makes the state uh a little bit more annoying.

to like track.

Uh, all right. So, let's see if this works.

No, that's not um Oh, AMD and MPU are already set to one.

So, we just set them to one here. When

we run the test, uh, debug equals 4.

This is wrong.

It's just annoying.

Uh, so I never have it commit.

That stuff's just broken.

So, it's it's a lot of state that you sort of have to to keep in your head.

Um, run my llinter in free commits passing.

uh copy the names from registers.h. In

fact, we can include registers.h

instead of redeclaring everything in agency.

So this is that one I switched to set default.

Um it's not really Oh no no no. This isn't

the problem.

I want that to start to add a ton of crap.

Maybe it is that. I don't know.

Um I'm not sure if I really have less uh like bandwidth to talk. I I feel like it requires really the exact same sort of

focus as programming. This three is a little much for me.

Um I don't always do it as aggressively as this.

I'm kind of showing off on the stream a little bit.

Um but this basically looks like what my uh like agentic workflow looks like. So

let's put this over here and let's take a look at this code.

um pass vars to functions that need it.

Like why is this there? I don't like that. Why am I not passing varss to

that. Why am I not passing varss to other functions?

And I don't think that is access for FPS adder like meanings. Yeah,

I don't like that. Can we make it explicit?

So like I don't love how it's handling this. We have different enums for RDNA3

this. We have different enums for RDNA3 and RDNA4.

Um it's importing them as this and it has this really copy and pasted alias list.

Um get picodict is okay.

Some of these changes to like use inst are just good generic changes. Um

yeah, move this.

add to if LDS with an L if branch.

I don't understand this. I don't

understand this code. So, you have to stop and wait until you understand the code. And then here's all that RDNA4

code. And then here's all that RDNA4 stuff repeated again.

Um don't need to check these instance here.

Um, remove the helper one line call.

It added some new debug crap. I mean, I like that it made that change. I don't

understand why that print needs flush equals true.

I don't really like that.

Remove this. You don't need flush equals true.

Now we're back to these stupid functions. I don't know if they're

functions. I don't know if they're actually correct. Like I don't like that

actually correct. Like I don't like that there's a hack giving the name of the function, right? That's the kind of crap

function, right? That's the kind of crap that like you end up with vibe coded stuff. So, we're going to have to fix

stuff. So, we're going to have to fix that. Uh this stuff all seems okay.

that. Uh this stuff all seems okay.

Okay, I already fixed that to use the uh helper from Tidyrad. Uh putting that as a property on the GPU setting that that seems okay.

Um as GPU props, it's like something that's like returned from the fake file system. They added the word RDNA4 here.

system. They added the word RDNA4 here.

They added arch and they added an arch call. That all seems good. Um they set

call. That all seems good. Um they set the arch there. That's good.

So like this is this is the workflow.

you really have to uh stay on top of these sort of things.

Um, and it's not it's not sexy.

All right. Now, I push this as cleanups.

I probably shouldn't abuse the GitHub runners so much, but whatever. That's

kind of what they're there for.

Um, okay. There's no diffs to this one.

I'm not sure what I have to still do on this one. Let's look it over. Let's see

this one. Let's look it over. Let's see

how this one's going. It's still

running.

Uh, the USB file registers are not being written to properly. That seems fine. I

like that it's including registers.py. I

assume it updated the make file, so we can read the diff for that.

Um, throw that over. I should go there. I

should go there. That's also not right.

This needs to go to that goes there.

That goes there. I'm not perfect with this new setup yet.

Um, all right. So, read and capture boot

all right. So, read and capture boot messages. Come in. Great. I don't know

messages. Come in. Great. I don't know how much I like that expected stuff from the claw MD. So, we're just going to go in there and we're gonna

uh yoink that.

Okay, you see it added it modified make file to add a C flag. So, that's good.

Like, no amount of AI prevents you from understanding.

I guess that's kind of the the moral of this story. Um, if you don't fully

this story. Um, if you don't fully understand the code that your agent is writing, you're backing yourself into a corner. Um, and then you ask, well, I

corner. Um, and then you ask, well, I mean, yeah, but someday these agents are going to become superhuman. I can trust a chessbot without uh understanding what the chessbot is

doing. Uh, and yes, because you and the

doing. Uh, and yes, because you and the chessbot have already completely agreed on what you want to do. You want to win the game.

Um, a lot of this stuff is just kind of agreeing on what you want done. There's

things that agents still do that are universally bad, and I think those things will be fixed, but all right. So, let's go back here. I

all right. So, let's go back here. I

don't like that it's using this X data stuff.

Are these not some special hardware initialization don't accept rights a pull up on D+ line that makes sense see like when you read this no pull up on D+ line if you don't

already understand that that's how USB signals that a device attaches right like you have to understand that um And it's so interesting how the skills required to use these are the identical

skills from before. You just have to be able to sort through a lot of information a lot of information quickly understand what's relevant, what's not relevant. [clears throat]

relevant. [clears throat] Um AI isn't some magical uh

trick. Yeah, probably the clock

trick. Yeah, probably the clock tree.

Why are you using X data here?

Not in registers.h.

Uh okay.

So, I already pushed that change.

Uh, let's check on these. You see that I have a typo there. So, it's kind of annoying that I'm not actually logged into GitHub here. I don't log into anything on stream, but normally I'd

just be doing this logged in. Like, I'm

still not thrilled by overall how this stuff looks, but I don't really have a better way to write it. So, I'm not sure what I'm going to do yet. Um, this pull

request I'm pretty happy with. I'm glad

that that passes all the checks. So

let's see if this is good. Uh let's read every line now. Okay, it adds source two support there. That's fine. Um it had to

support there. That's fine. Um it had to add this hack for noop because noop has a variable loop. It's actually the only pode with it. So that's okay. That's

fine. Those are fine. I complained about this. It said it couldn't fix it. I

this. It said it couldn't fix it. I

don't love it, but we'll say that that stuff's fine, too.

Um, let's just kind of look at that one in context, right? Then I can use GitHub here to look at context. Okay, compile

me up. It quickly exits. It's still

using the pcode parser. That's pretty

good. Okay, so it added this new function, sum of absolute differences.

It added a masked flag. It calls it here with mask. That's fine. Oh, yes.

with mask. That's fine. Oh, yes.

Um, is this used? It added this hack for knop and I think it should just be removed.

Uh, this seems okay. I complained about that. I couldn't fix it. I like that it

that. I couldn't fix it. I like that it added assertions there. That's a

simplification.

Uh, it's removing those things. That's

fine.

Um, I don't really love that.

like it should be using ul to determine if it's a Uint64. Let's see. Um,

can you use ul to determine if it's 64?

Uh, okay. That's See, that's fine. It it

added a bunch of these hacks where it returns zero. We're slowly removing

returns zero. We're slowly removing them. It's interesting how code

them. It's interesting how code production kind of looks more like diffusion now. Like you can still get

diffusion now. Like you can still get quality code out of these things, but your first thing is very fast and then your refinement stage is longer. Um it's

unclear if these things are overall a net productivity win. All right, so that just changes that to set default. I'm

totally fine with that.

Uh, other used show.

That's it. Runs my test suite for me. It

knows how to run it. There's just a test failure in test test hex literals. But

that's it's probably good. Now I have to pay attention here to make sure that it knows that's probably a test bug in hex,

right? Because that's the test for the

right? Because that's the test for the pcode parser, not actually the test of any of the pode that's parsed. So it'll

do that. It'll check that. We'll come

back. We'll see what it's doing. I have

to keep that context in my head. Um, now

we're also working on this RDNA4 one. I

still haven't solved that problem there.

Uh, okay. So, again, this has not figured out how to get it to enumerate a USB device, but it has a complete loop where it can run on its own.

Um, oh, okay, good. Things are working, maybe. Oh, they're not working.

maybe. Oh, they're not working.

Um, did it remove a lot of those stupid X data things? Good. It removed the X data

data things? Good. It removed the X data things. I didn't like the X data things.

things. I didn't like the X data things.

There's still three of them there.

All right, it looks like it's still working. Let's let it

working. Let's let it um keep trying it get out of this dump regs thing, which is kind of cool.

That's fine. Like this is a good This is fully like agentic because it has a loop that it can close on itself.

Um so yeah, we see that it dumps the eggs here. That's pretty cool. So let

eggs here. That's pretty cool. So let

that keep going.

Uh, let's read this diff and see what it did.

Okay. If is U64. That's fine. We

stripped that. Cool. I like that mostly.

Um.

Oh, interesting. If it has the F. Oh,

that's cool. See, it automatically started doing that for all of them. I

don't really understand why that's duplicated. I guess this is duplicated

duplicated. I guess this is duplicated logic to deal with the non-hex ones. It

could be cleaner, but whatever.

Um, all right. So, it fixed that.

Uh, oh, great. And even better, that seems to have fixed the other problem.

The LRA syntax is used internally during loop unrolling.

Uh, you shouldn't need that in the P code.

parser forg okay um I haven't dealt with this one yet so like this one's kind of on hold you're see my limit of what I can manage

um think it found something there that seems okay I don't need to feed that one uh

see like okay where am I now I'm complaining about that so After loop unrolling, it becomes a number.

Um, but that shouldn't be used in the parks, should it? I don't like that.

should it? I don't like that.

I don't like that. Right.

Can we use the loop unroll instead?

Why is the loop unroller calling back into the parser?

Um, someone described this well when they said it was like interacting with the code at a higher level.

All right.

Like that one's not going to run. We'll

let that one run for a minute. Now,

let's think about what we want to do for this.

So, yeah, this this hack needs to go.

Um, this needs to be removed.

You cannot have special handling for some functions.

seems okay.

I mean, I just like that this one's removing crap. I didn't really realize

removing crap. I didn't really realize that that was being called back into.

We'll let that try.

I'm not really thrilled. See, there was something I didn't understand. It's

calling back into the PCODE parser after it does the loop unrolling, right? like

that shit.

Um, all right.

What change did this make?

I don't love that it did that, but at least it removed the hack.

Okay.

That stuff really shouldn't be in the peer. I'm unhappy about that.

peer. I'm unhappy about that.

So, it looks like it's trying to write to something and it's failing.

All right.

Whatever. That can run. Tiny empty. Um,

I probably I'm not going to bring it up on this stream, but usually I would have

uh well, sometime I'll tunnel into uh I'll SSH tunnel so I can see the edits that it's making. Um, so I'll show you one of my goals for this week is to

unify. There's an op called define

unify. There's an op called define global in Tiny. So, we're going to start here. Um,

here. Um, replace define global with param for now.

Uh, just alias.

We should also see what tests pass. Oh,

this is going to this is going to slow my computer down.

Um uh define global

2B param and see what test tiny tests All right.

I complained on Discord. Some tests

broken.

Uh So the AI is not able to reference any data sheet because there is no data sheet. Um all of these registers were

sheet. Um all of these registers were reverse engineered by Claude from the firmware.

Uh which was kind of my first foray into these into these vcoded things. Um

so let's see how it did this aliasing.

Okay, that aliasing looks nice.

Uh, all right. So, all the tests just pass.

all right. So, all the tests just pass.

It's now able to use define global as the thing. Let's see if it's doing what

the thing. Let's see if it's doing what I expect browser here.

This is viz and tiny. You can see it's now using param uh instead of define global which is kind of nice because eventually I also

want to replace these kernel ops with call ops.

Um but that's a different change. I

first want to get rid of define global.

So you can see here in the output of kernel split um it's switching those things to params instead of to uh

yeah. So unfortunately the tiny test

yeah. So unfortunately the tiny test suite's a little bit too slow to run uh in the agentic loop. So usually what I'll do is I'll like have it make a

change like that. It's a very small change. I can easily look at it. Um then

change. I can easily look at it. Um then

I run the tests. Then the tests that I see that are broken uh that are bad I paste into the thing and I I ask it why they're broken. So on and so forth.

they're broken. So on and so forth.

Close the loop. All right. This is still trying crap with uh let's move that over there. We don't need that for now.

there. We don't need that for now.

This is still trying to get rid of and clean up the thing. Like this gets into where I don't fully understand the code.

Um no, see this is very bad. It

shouldn't be doing this. This is like a type confusion kind of thing, even though it did succeed. So like I now have like a knot

uh I have like a knot in my code. Um

should never be S0. It should only be don't check both fix the bug. Right. But

now I have like a knot that I don't quite understand. And I I didn't really

quite understand. And I I didn't really realize when I started with this there's this absolutely massive cost to doing that. Okay, so this removed that hack

that. Okay, so this removed that hack that looks like a strictly winning change. Go over here.

change. Go over here.

Should manage my windows a little bit better. You know, this is hard for me to

better. You know, this is hard for me to do on stream to like move windows around. Um, okay. So, unfortunately, all

around. Um, okay. So, unfortunately, all of these all of these test config viz things were already failing. I should

probably just fix this. Maybe someone

already fixed it. I complained about it.

Make an alias for that.

See if anyone fixed anything.

Someone fix that.

It's just some like I don't know. We

should see probably what's going on here.

Maybe it's just easier for me to fix it.

I don't know about that.

It's some like I'm getting sniped by this now because this was a pre-existing bug that I knew about just this morning.

Um, throw another one to open code. Y is

failing.

Okay. Now, why is this failing? Um

probably has to do with some param defined global dduping thing. You can

see that also all the test calls are failing. So, there's something that's

failing. So, there's something that's trying to call pointer dtype on them because it's a it's a U.

Uh yeah. Okay. The she was in shape.

That seems uh okay. Remove the define global case.

uh okay. Remove the define global case.

Add a check for is pointer d type to param and returng thing if

it's probably enough for it to get it.

It might not be. You always have to make a trade-off like do I want to type it out better? Do I want to type it out

out better? Do I want to type it out less? I don't hit 100% with keeping all

less? I don't hit 100% with keeping all the state in my head.

Um but uh okay. So that's good. Now

okay. So that's good. Now

does look a lot more reason number one.

Why these funks as lambdas a big difference here.

Let's check in on this guy.

Uh, I see.

Ah, cool. Uh, yes, that's an easy fix.

Great.

And I kind of don't like that that's all over the place, but whatever. It's not

like I can set kernel to default to that. It's gonna open the stupid device,

that. It's gonna open the stupid device, too.

Let me just say in Discord I have the fix for that so nobody wastes time.

Whatever. It's a tiny fix.

I guess cutting out. If you change it to M496, I can get mine to brown out. Oh,

interesting. Unrelated to the Discord thing.

So, I should probably uh like it would be cool if there was just one I posted about that if there was like some bot in our Discord that would do that automatically.

Uh, all right. So that's fixed.

That alias works.

Let's control C that. And let's do this.

All right, we're back.

Yeah, you're seeing one of the downsides of my real workflow. Uh I don't know why the Wi-Fi dropped. That's just annoying.

I can't wait till I have an office with like a real fiber connection soon. Okay,

all the tests passed there. Uh, all

right. Um, remove define global. It is

param now.

Let it run the I'll do the replace later.

All right. Is the stream back?

I got no viewers on this one. So you

guys like when I do crypto scans, right?

Like welcome to a real workflop.

Um name param is not defined. Why not?

Oh, it's annoying.

Don't let it waste time.

See if that works.

Uh yeah, so it's much le much much less exciting, but welcome to a real agentic coding workflow.

Uh I wanted to demonstrate this in contrast to yesterday. Uh the top programmers in the world are not uh spamming gas town and claws and all

of that stuff. They're using these things in a very uh restrained and tasteful way. Uh the

slot apocalypse is not coming. If you

post PRs on Tiny Brad that are clear that they have not been reviewed by a human, uh not been carefully like to the same level of care that I'm using here,

uh reviewed line by line by a human, um I will disclose them. Don't waste my time. I have the same AI as you do.

time. I have the same AI as you do.

Uh yeah, that was one thing I wanted to to show and that's my real new workflow.

Uh it's not a lot of spamming.

Uh I'll turn on nonsubscriber and we'll actually only do questions for five minutes. Look, we're we're disciplined

minutes. Look, we're we're disciplined today because this is actually a work day for me.

Uh you know, it's Monday. I'm at work.

Um, uh, what chips in the eGPU? It's a 9060

XT8 gig. Oh, good. That's still trying.

XT8 gig. Oh, good. That's still trying.

Uh, using AI agents makes one dumb. I

think you have to stay more vigilant.

because you can just drift into the slop and drift into the autopilot.

Uh I view AI agents very similar to GPS for cars. Uh so I don't use nav systems

for cars. Uh so I don't use nav systems when I drive unless I really have to get somewhere. Uh but I do use a map and it

somewhere. Uh but I do use a map and it is a magic map that has a dot on it.

Uh oh. Brad, sorry I missed your earlier

oh. Brad, sorry I missed your earlier thing. Do I use check untapped defs?

thing. Do I use check untapped defs?

We're moving to that. Um, Chenu is Chenu is working on it. Uh, you can, uh, if you're interested in helping with Tiny Grid development, you see kind of what it looks like. Um, you can join our

Discord. Uh, it's like this kind of

Discord. Uh, it's like this kind of stuff. If someone wants to work on

stuff. If someone wants to work on adding cna support to the emulator or finishing the RDNA RDNA 4 support for the emulator, I'm finding bugs in the emulator.

Oh, look at this. Oh, it got something.

Oh, hit boys.

Um, yeah. No, using I mean using the LLM as

yeah. No, using I mean using the LLM as a teacher is very good and it comes down to how people use them. Uh, the

Anthropic released a study uh this is like a hype study. it they

actually here how AI assistance impacts the formation of coding skills. Um, and

their end conclusion, uh, if I recall correctly um, on average, participants in the AI group finished about 2% faster, although

the difference was not significant. Uh

there was however a significant difference in uh difference in test uh scores. The AI group understood way less

scores. The AI group understood way less than the hand coding group.

Um the largest gap was on debugging questions suggesting the ability to understand when code is incorrect and why it fails. Maybe a particular area of concern if AI impedes code in

development. Um, so yeah, I mean there

development. Um, so yeah, I mean there really are bad ways to use these things and it's unclear if there are such good ways to use these things. Uh, this is

consistent with a lot of results I've seen about AI that it does not actually make anything faster. Um, and I'm still playing with this new workflow for

myself. I don't know if it's making me

myself. I don't know if it's making me faster. Um, I think there are some

faster. Um, I think there are some tradeoffs. I think I'm really trying my

tradeoffs. I think I'm really trying my best to avoid this downside.

Uh the other alternative is to say that I'm not going to use these tools. And I did feel that way until Opus 4.5. Um because

it is clear that you're going to have to like use these things. Software

engineers are going to get more productive. It's unclear if the tools

productive. It's unclear if the tools today actually make people more productive. Um, I wrote a post about

productive. Um, I wrote a post about this uh a while ago.

Um, and I think something that I'm mostly I stand by most of the things in this post except for

uh this one.

I do think that these will work interactively. Um, so I stand by

interactively. Um, so I stand by everything in this post except for this.

Uh, and then this I mean this kind of like because the interactive thing works. I do think that refinement flows will start to work. Um,

I do still find that the first thing that it gives you is generally its best result. And this is kind of how you can

result. And this is kind of how you can fall into this trap with AI assistance where it looks like it's 98% of the way

there, but it's actually like 20% of the way there. The thing that's cheap and

way there. The thing that's cheap and expensive has changed.

Um, I fully stand by this. Uh, is

anybody still using cursor today? And

like cloud code and open code are barely doing anything. Um I just prefer the

doing anything. Um I just prefer the open code interface. I see a lot of the same vibecoded slop patterns in it. Uh

so you know all software engineers need to be careful about this but it's better than cloud code which is also vibecoded slop and it's vioded slop that sends data back to anthropic which again like I'd be fine with if the workflow is

better. I'm not trying to hoard my data,

better. I'm not trying to hoard my data, but like I said in, you know, my post to Andropic about this, uh, you're not going to get me to go back to I would

prefer to use open code with Kimmy, uh, than to use claw code with Opus.

Uh, so, and I also like that I can run Kimmy, especially I'm a little upset. I

wish that it was faster. Um, I wish that the the cloud code was faster.

Uh, I don't know why I didn't just put those in the direct like you just see a mis pattern like that. It's like why didn't I put it

like that. It's like why didn't I put it there? It probably has some, you know,

there? It probably has some, you know, who knows?

Uh yeah CEX CLY is better than Claude. I found

Codex to be extremely slow. Um,

the register still reject Rio investigating this works on the real so it is doing

something you aren't.

Uh, the study was so bad. Why? And

actually that study also didn't even talk about claude code like things. It

talked about the old workflow of copying and pasting from chat GPT.

Um should we have AI generate documentation? I haven't found that much

documentation? I haven't found that much use for that. I find that it mostly writes slop.

Um Kimmy is cheaper. I mean I'm not that worried about the price. All of these things are a, like I said, my salary is $100,000 a year. These things are a

uh very small fraction of my salary. I

would be happy to pay up to $20,000 a year for AI assistance. And I would do the same thing for uh you know, everyone who works in my companies. Um, if AI

assistance actually made people 20% more productive, I'd absolutely spend 20% of their salary on on LMS.

I'm not sure that it does. And I think that these these tools can be very dangerous if you don't have good

judgment and taste and if you don't like understand like know yourself really well. Um, so I think that these things

well. Um, so I think that these things are going to help certain people and hurt other people, but I don't think that the overall change to the rankings

of programmers will be that high because the skills look very similar, right? The skills of programming look

right? The skills of programming look like I imagine, think about it when uh we

switched from manually uh adding up stuff in the accounting department to the accounting department using spreadsheets.

I think that for the most part even though workflows changed a lot the same basic skill of accounting was

there uh the same basic [clears throat] like there there's a there's an old saying in machine learning it's uh garbage in garbage out and it becomes

the same thing is also true about spreadsheets uh you know like like what did it just do there? Why did it add

do there? Why did it add whatever?

Um, yeah, they they did Kimmy for free. Uh,

if you don't have anything that you pay for, Kimmy free Kimmy's pretty good. Um,

I found it slightly worse than Claude, but like I say, if Anthropic ever tries to uh lock down this stuff again, then I

will just as happily uh switch back.

Um, yeah, when you code in languages you don't know, well, I tend to let AIS take the wheel. I think that this is like

the wheel. I think that this is like concerning.

If you're using it, there's a way that you can use it for syntax. Like there's

something that they're just so incredible at, which is writing these little oneoff scripts. Um, and it's not that they're better at the scripts than you. It's like that script would take me

you. It's like that script would take me 15 minutes and it takes the AI 15 seconds, which is really cool. And it

can write all of these one-off scripts to just like debug something where you're like, it would just take you so much longer.

Um, what do we think? We think it's going to show up on USB. Like, I'm so excited. It's going to eventually just

excited. It's going to eventually just show up on lsusb uh when we do this as an emulator. No,

shows that the right still fails. Let me

check if it enumerated. It did not enumerate.

Um see now it's doing everything in this emulator. I gave it an emulator.

emulator. I gave it an emulator.

Um Gemini 3 is more cogent unlimited use of code examples. Yeah, I don't know. I I

code examples. Yeah, I don't know. I I

feel that they're all like pretty close except Codeex is actually slow.

I don't know how many tokens per second I'm getting from uh Anthropic.

Uh that looks better.

Weird. Really?

Oh, that's cool.

Oh good.

Uh, this is the session that I have my SSH keys forwarded on. That's why I come to this one to get push.

I don't like giving the agents the SSH forward stuff. Obviously,

forward stuff. Obviously, [snorts] it's almost lunchtime for me. Um, the

main takeaway from this stream is that you need to be the Earth and the AIS need to be the moon. You are a senior who's reviewing junior dev code. Yeah,

they fool you by being fast.

Um, and they will get better, but I think that the workflow is I don't know. It's going to be interesting. It's going to be

interesting. It's going to be interesting.

This this is something that I didn't predict.

Uh and I think that other people predicted it uh before me. I also didn't predict chatting.

Uh and I had like I used GBT3 for stuff.

Um I I worked on the compression the language compression stuff really early.

I knew what these I knew that like predicting and compressing uh could in theory do this stuff. I just

didn't predict the productization.

And here I didn't expect these loops. I

mean, every time I tried one of these before Opus, it was always kind of uh just bad.

like I would I would run it and it would uh go off on some tangent. It would put crap everywhere.

Uh so I mean it's great that these things are becoming more intelligent and we all have to learn how to use them. Um

so my advice for the day is uh don't fall for everything you see online. Uh

you know uh the creator of uh of Gastown, Steve Jay, uh is a is a is a famous guy. uh yegeay Google platform

famous guy. uh yegeay Google platform rant I this is one of his works that uh I reference a lot you know it's great to

see him back with uh with gas town but uh we joked on the uh the first day we came back to to to come Harold was like

uh I [snorts] read something on LinkedIn that said this is the most important day for CTO's of uh you know their career Um, you know, we're we're joking about

that. You know, I told every everyone a

that. You know, I told every everyone a comment that I expected everybody to be 5x more productive uh now that we have agentic coding. Um, tongue and cheek,

agentic coding. Um, tongue and cheek, obviously. And if you're at the kind of

obviously. And if you're at the kind of place that isn't saying this stuff tongue and cheek, if you're at the kind of place that has like weirdly embraced

some allin on AI uh thing where you basically again have like nonsoftware engineers driving policy at your company, uh you should

quit. Uh that company is is going into

quit. Uh that company is is going into the ground. Uh so yeah, I mean this is a

the ground. Uh so yeah, I mean this is a time when when fortunes are made and fortunes are lost.

um companies will end up in the slot apocalypse.

You know, there is always the Deosex Machina. Don't worry, Opus 5.0 is going

Machina. Don't worry, Opus 5.0 is going to come in and clean up all the slop.

And I don't know, maybe it will. Like I

my my prediction is as good as yours. I

I I'm not uh historically great at predicting when these when these things become good enough to use. Uh even when I first saw Chat GBT,

use. Uh even when I first saw Chat GBT, I'm like, "Oh, that is a cute art project." I did not expect at some point

project." I did not expect at some point everyone kind of realized that this was going to replace Google. Uh but the first time I saw it, I I still didn't

didn't expect it.

Uh I like, you know, I knew about tool use. I had early stuff doing it, but I

use. I had early stuff doing it, but I I' I've in some ways underestimated the progress of these things. uh but in other ways

also realize that they're kind of fooling me. Um oh data efficiency is

fooling me. Um oh data efficiency is super important. Uh we talk about data

super important. Uh we talk about data efficiency all the time at comma because um data efficiency, you know, it's not only that we're we're out of data,

right? This is a this is a the scaling

right? This is a this is a the scaling laws didn't stop because we couldn't continue to actually scale the weights.

The scaling laws stopped because we ran out of data. Um, but yeah, then you're left with with data efficiency. And then

there's an interesting I don't know if you guys saw the Dorcash uh Schmidt Huber take on LLMs where Schmid Huber said basically that no LLMs are not the bitter lesson and

these scaling laws are not the bitter lesson because these things are not being uh being learned from from data.

These things are simply compressing the internet. They are not using search and

internet. They are not using search and optimization to to infinitely scale.

They've reached a point, a scaling point. Um,

point. Um, waiting for Tiny Grad to get good enough that I can run Kimmy on my cheap home lab. Yeah, I'm very interested in this

lab. Yeah, I'm very interested in this too. Uh, so I have I tried to get Kimmy

too. Uh, so I have I tried to get Kimmy running on uh one of my uh AMD boxes and I got VLLM up and it was running at three tokens per second. Uh, and then

like you needed to enable some stupid flags.

Uh, and the flags didn't work. You know,

this VLM rock muse IDER and then you enable these things and it gives you errors with MX FP4. So, I still haven't run Kimmy locally at more than three tokens per second.

Um, the problem I mean basically no matter how you slice this, you're going to need

uh like 768 gigs of RAM to run it. Uh,

so I like I don't know how you're going to get 768 gigs of RAM.

Um, the Macs are something, but that RAM isn't very fast.

Yeah, I know. RAM's not cheap, right?

Um, you know, it's cool how all these things are are mixture of experts models. What mixture of experts does is

models. What mixture of experts does is it divorces the amount of RAM from the amount of compute you need. But I'm

thinking about it like I really want to make a a Kimmy product uh for Tiny Corp.

What would you be willing to pay?

Uh what would you be willing to pay for a Kimmy box and how many tokens per second does it need?

Let's read this over again.

Still kind of working. I got to go to lunch, too. I'm hungry.

lunch, too. I'm hungry.

$3,000 for 30 tokens per second. Yeah, I

just don't think that's going to happen.

Um, I'm in Hong Kong.

No, I don't think there's a learning curve in using AI.

Um, I think that every time someone talks about how some new workflow is better or some system prompt is better, I never actually see it. Uh, I think

that there is some like getting used to working with these dumb people who are fast.

Um, yeah. No, China making cheap RAM.

I'm very interested in I'm going to be in Shenzhen next week, by the way. Uh,

if anyone is there and uh is making like Yeah, I'm interested in RAM. I'm

interested in compute chips. I'm

interested in like what I can buy uh 5K. I don't think you're going to be

uh 5K. I don't think you're going to be able to run. I don't think you can even like what's the cost of RAM now? Even if

uh so like well there's a question of how fast you want the RAM to be, right?

But um cost per GB of DDR5. And DDR5

still isn't fast enough. You need faster RAM than this.

Uh I mean maybe you could bank it and it'd be fine, right? So, we're talking about $10 to $20 per gigabyte, right?

So, let's even say we could get that for $10 a gigabyte. You're spending $7,000.

You're spending $8,000 on RAM alone, right? And that's not even server RAM.

right? And that's not even server RAM.

Like, server RAM here is even more. Um,

so the RAM is probably going to be more like $20. So, you're talking 15K just

like $20. So, you're talking 15K just for the RAM.

Um, do I check InstaDMs? No, only from girls. Uh,

girls. Uh, send me an email. Like the way to actually get in touch with me, I do read all of my emails. I don't reply to a lot of them because I'm just like, why would I reply to this, right? Like what am I

going to what do I get from this?

Um, I mean, the other thing too is like this would enter a global market pretty fast, right? If there was some way that the

right? If there was some way that the Chinese could make like normal ram chips and just stick them on sticks of RAM, they'd be doing it already.

Uh so there's no there's no like like alpha here really.

Uh Oh, interesting. The original firmware

Oh, interesting. The original firmware does enumerate. Cool.

does enumerate. Cool.

Okay. Uh

I don't know. I think RAM will eventually become cheap. Um I do think boxes to run these things. The question

will be is kind of is this where weight scaling is going to stay? Does weight

scaling stay at something that's like somewhat approachable like a terabyte or are we going to start getting into things that are distributed across you

know 10 or 100 machines and are these are these um remember that the human brain uh has about um

well so let's synapses in human brain uh count.

Yeah. So, like those are the numbers we're looking at for human brains. Um

we're still uh like if Kimmy has one trillion, so we're still two to three orders of magnitude off from the synapses in the human brain. So, it's possible that

human brain. So, it's possible that scaling is going to continue to that.

Um, I mean it's possible scaling will even continue beyond that, but we're still we're still we're still off from the human brain. I think our

mixtures of expert ratios are pretty similar to the human brain.

It's done in a very like rudimentary way but um what do you mean we haven't even figured out what 2% of it does? I think we

mostly mostly know. How long am I going to be in Shenzhen? Couple days.

All right, we're getting into slop questions. Thank

you for watching the stream. Uh,

look, real firmware works. You make your firmware work. I will give you a cookie.

work. I will give you a cookie.

Um, yeah. No. And again, like I I read all the emails, but I don't reply to most of them, right? Like what's the like why?

I'm not interested in like having lunch.

I'm just I'm not interested in that. Um

uh yeah, but I wanted to do this stream today just to show you what real agent coding workflows look like. They do not look like spamming on a digital ocean box. That is for partying on Sunday.

box. That is for partying on Sunday.

Mondays are for real work. You guys

should be out there doing real work. If

you're not doing real work today, get a job uh and go do real work and figure out how to produce more than you consume. And as long as you continue to

consume. And as long as you continue to do that, you'll be fine in the future.

Happy Monday, everybody.