Meta's AI Backfires, GLM 5.3 Flash, OpenAI's Jalapeño Chip & the End of Programming

Shimin (00:00) Hello and welcome back to Artificial Developer Intelligence, a weekly conversation show where three software developers try to sort out the hype around AI and what it actually delivers. My name is Shimin Zhang, and with me today are my co-hosts, Dan Organization Transformation, Lasky, and Rahul. Call him Odysseus because he just returned from an island full of boobies. Yadav!

Rahul Yadav (00:26) Ha ha

Shimin (00:27) Pew pew

Welcome back, Rohul. We missed

Rahul Yadav (00:31) Thanks.

Shimin (00:31) ya.

Rahul Yadav (00:31) Yeah, glad to be back. missed you Guys.

Shimin (00:34) How how are the sirens and the shipwreck and the hydras and various other monsters?

Rahul Yadav (00:42) they well, I’m hard of hearing, so I wonder if people really needed to tie themselves if they had any defluence on the ship But yeah,

Shimin (00:51) Yeah.

Rahul Yadav (00:52) it was fun. Saw a lot of wild unique wildlife for the first time and everyone talks in geological time in Galapagos.

Shimin (01:00) Mm-hmm.

Rahul Yadav (01:00) They go, This island is

Pretty young and then you say, Well, how young and they say, a couple of hundred thousand years old and that really puts things in perspective, on how slow Earth time moves and evolution is you know, works but is painfully slow. You get to see it there.

Shimin (01:18) And AI is the complete opposite of that, where two weeks

Rahul Yadav (01:20) Yeah.

Shimin (01:21) is like four years in normal time. So

Dan (01:24) Yeah. Do you even know what’s happened in the past three? We’re actually

Shimin (01:26) Ha ha ha.

Rahul Yadav (01:27) I don’t know.

Dan (01:29) not using AI anymore. We’ve moved on to something new, which we’ll talk about later on in

Rahul Yadav (01:34) Yeah.

Dan (01:35) the show.

Shimin (01:35) Absolutely. Okay, so on this week’s show, we are going to start as per usual with the news threadmill. we’re gonna talk about the latest happenings at Meta as well as some SpaceX open AI drama.

Dan (01:50) Then considering how many times we talk about models on this show, we have decided apparently, apparently, to split out models

Shimin (01:57) Apparently.

Dan (01:58) into its own model review section. So we’re gonna talk about what’s new in models and share a little bit of our own experience in using them.

Shimin (02:06) then we’re gonna move on to the hardware hut where Dan is gonna tell us a little bit about jalapeno.

Dan (02:12) The spicy chip.

and then we’re gonna be moving right along to post processing where we’re gonna talk about the end of programming. Period.

Ha ha

ha.

Shimin (02:25) and last but not least, we’re gonna go to our two minutes to midnight segment, where as always we’re gonna talk about the financial side of the AI. Is it a bubble? Is it not? We shall find out. Okay.

Our first news segment is brought to us by Dan.

Dan (02:42) Mm-hmm. Who’s in turn brought to us by our old friend Mark Zuckerberg. We talk a lot about yeah,

Shimin (02:47) And our technica.

Dan (02:49) it’s true. We talk a lot about meta on the show because they just are so good at making controversial AI headlines. So here we are with another one, which is that AI agents meant to replace meta workers made large scale disruptive actions.

Here, but that headline actually kind of buries the the real headline here, which is that early in 2026, Meta created what they called the AI native plan. So, you know, some fun sort of corporate buzzwords for you, which was a project which they codenamed OT, which is organizational transformation, which is apparently is where I get my middle name from.

Shimin (03:27) No.

Dan (03:28) The most creative name that you can come

up with for something like that. And apparently what it was meant to do was cut headcount by as much as 60% on some teams in two rounds of layoffs with about 25% overall headcount being cut.

Shimin (03:44) Mm-hmm.

Dan (03:44) so Zuck himself apparently set this project OT in motion and directed the execs to proceed.

And so that first round, I think we may have even talked about on the show, that happened back in May, actually happened. And then nearly immediately after they canceled the second. and so kind of the rest of the article digs into a little bit about like, why might you cancel something like that after you know, making such drastic changes? So the the sort of positive signal that I think it it sort of

made them think that maybe this is going to be successful was like code changes made to internal software platforms and infrastructure employees using the job were up 220 year over year. So good, yay, exciting. And that’s according to Meta CTO, Mr. Bosworth in early June. But the changes that actually led to new or upgraded features

reaching meta’s users were only up 36%. So pretty big disparity between like the, you know, code changes versus like what’s actually like being shipped feature-wise. the other sort of interesting thing that came out of that

internal posts also reportedly pointed to AI agents making large-scale disruptive actions that humans are unlikely to execute. So that led to overall a 40% increase in major technical and security incidents compared to the prior year. And employee time spent resolving those problems increased by as much as 70%, which is quite a bit of firefighting.

Meta, of course, declined to comment on the record on those internal posts. and, you know, what are like what’s the conclusion we can draw from this? Well, like the most AI forward, well, maybe not the most, that might be anthropic or open AI, but like one of the most AI forward companies out there is struggling to actually replace humans, right? Like real human workloads

Shimin (05:32) Mm-hmm.

Dan (05:32) in what’s arguably the most sort of

replaceable field out of all the fields it’s made inroads in. and it also kinda highlights the risks of being a little too AI forward, right? Without having some some controls.

So

Shimin (05:45) so what happened to the rest of the a hundred and seventy percent of code changes that meta pushed up that didn’t, you know, cause a security issue? I don’t know.

Dan (05:53) Maybe maybe all those had to just be removed because of the forty percent security issue.

Rahul Yadav (05:58) Yeah.

Shimin (05:58) That’s likely.

Rahul Yadav (05:59) I think a lot of it probably cause this is the also referring to the token maxing times. So

Shimin (06:06) Mm.

Rahul Yadav (06:07) a lot of that internal stuff was probably just here’s a fun thing that I thought I would spend my tokens on. and maybe just you know, helps that one person. and I think there is some good friction to like

Even if you write two hundred and twenty percent more code, do you really wanna overwhelm your users with tons of features? So you don’t want every single thing going out to them. So it’s probably also a lot of that just got dropped because of that. Because it might have sounded fun, but maybe doesn’t fit nicely.

Shimin (06:39) Yeah. And this also reminds me of the conversation we’re having last week. when Nick described giving every single developer an AI agent as like giving everybody a really nice supercar. but then just ‘cause you have a supercar for everyone or maybe even a rocket car, like that doesn’t mean your old highway system isn’t gonna be able to handle all of it, right? Like without

Dan (07:00) Mm-hmm.

Shimin (07:01) reconstructing

how the organization functions. You can’t just go in and cut people ‘cause the cars are just gonna crash and lean into more security incidents.

Dan (07:09) Yeah.

Shimin (07:10) and maybe fewer people will actually get you where they wanna go.

Dan (07:13) And even I would argue like to some degree I think we might need to even shape reshape how we do development. Like not just like life software lifecycle stuff, but also like, you know, is it worth having like fifty microservices versus one monolith, you know, depending on which one performs better for in the long term for agents and stuff like that.

Shimin (07:33) Right. And like, what what does our repo even look like going forward? all sorts of interesting questions. I’m glad Meta decided to like I have to give Zuck props for trying to do something really you know, quote unquote brave and r quote unquote revolutionary. It just didn’t work out, ‘cause I don’t expect anyone to know at this moment what the right answer is. But

Some organizations are just kind of not doing much, whereas at least meta

Dan (07:59) Yeah.

Shimin (08:00) tried, even if it didn’t work out.

Dan (08:02) Me the I

I find it hard to give people props for like the plan to cut twenty five percent of your workforce though regardless of why. I think like you

Shimin (08:09) that’s true. Yeah. That’s the

Dan (08:14) could do a organizational change like that by like I don’t know, like actually taking some abrupt measures.

That didn’t involve firing everyone. But what do I know? That’s why I’m not a fancy CEO sitting in

Shimin (08:24) Yeah. That’s the

Dan (08:28) a supercar somewhere.

Rahul Yadav (08:29) Not with that attitude Dan

Shimin (08:31) We all have supercars now. Okay. yeah, let’s go to our next news item. this one brought to us by OpenAI.

Rahul Yadav (08:39) so SpaceX acquired Cursor recently for close to sixty billion dollars or something.

Dan (08:46) huh.

Shimin (08:46) Mm-hmm.

Rahul Yadav (08:47) And obviously that means and cursor uses open AI as one of its models or the users can use it. but open AI and SpaceX are competitors and so open AI announced this was

Two or three days ago, August 28th, that they’re going to stop OpenAI models from you know showing up in cursor. They’re they’ll cut cursors access. And they’ve given them until November 12th, because that’s the maximum time that the contract allows. So they’re trying to do it in good faith so that the at the end of the day, the users who use open AI models get impacted.

They cited the lawsuits that they had with SpaceX and they cited how Elon in the past has not complied to open AI’s terms of service and under oath had yeah, had admitted

Dan (09:36) He he stated under oath that he didn’t.

Shimin (09:37) That’s that’s that’s that’s that’s that’s

Rahul Yadav (09:40) that XAI, which is now, you know, SpaceXAI, the it’s the the whole big thing.

they had violated OpenCI terms of service. So they’re just trying to prevent that from happening from cursor. And the I I think the, you know, the implication without saying it out loud is we don’t want you to be distilling our models through cursor and we’d rather just cut access. one interesting thing I realized after when I was reading this was

we’re gonna see more and more of these consolidations and it’ll be and as soon as, you know, when w any of your competitor consolidates something, you almost immediately you can see this type of reaction ‘cause everybody is gonna try and protect the distillation of their models and everything. so I wonder any of the big popular apps like Cursor that are probably gonna get you know, or might get acquired in the future, how

their user base would get impacted and to what extent. on the other hand, if switching models is easy and if it doesn’t make as big a difference to you, it might not be a big deal.

Shimin (10:46) Right, like the the old question is always like what is cursors moat that, you know, open router doesn’t doesn’t already have, right? ‘Cause if if your harness is gonna be so special that especially if your harness should be tailored to your organization, why use cursor? And I I really can’t blame open AI for not wanting to be like man in the middle by their competitor. so

Rahul Yadav (11:11) Yeah.

Shimin (11:12) I I get it. I I I have to say I am kind of in agreement. And OpenAI is already known to be the like frontier US model that is most amenable to having its models used by other harnesses. Right? Like unlike Anthropic that does not allow Claude to be used by PyAgent, for example, right? Like OpenAI is like

Dan (11:33) I think they do know.

Shimin (11:34) you cool. I still think it requires tokens.

And it doesn’t work with subscription. I haven’t double checked, so but that was my gist last time I looked at it.

Dan, what do you think?

Dan (11:44) check

I still have my pie hooked up to local imprints so I’m not surprised either especially with sort of the recent product announcements too which we haven’t really talked about but like also there like Grok is by is through that acquisition trying to take on I forget what GPT’s thing is called like work or whatever

Rahul Yadav (12:03) yeah. Chat open air work? Chat GPT work? Just just work. Some like

Dan (12:05) It’s not cowork. Yeah. Something like that. Yeah. So like

Shimin (12:07) I think chat GPT work, yeah.

Dan (12:10) they just recently announced that, which in turn I think is trying to take on like, you know, Anthropics Cowork. And then now

Grok had just released their own very similar kind of like, you know, it’s a standalone agent that runs in the cloud and you know goes off and does tasks for you. thing is, you know, it’s kind of squarely aimed at the business market. So I think like

Well Grok was busy staying on Twitter and or like you know, kinda carving out their own little niche so to speak. OpenAI probably didn’t care too much. But now that they’re like, yeah, we’re gonna go squarely against you in terms of business targets, then I could see them kinda doing this as you know, just a gentle punitive measure. So

Shimin (12:48) Yeah, we’ll we’ll see if the consolidation continues. And my bet is yes, and my bet is we’re gonna see something later on in this episode that talks about this vertical integration of Frontier Labs. But first,

Dan (13:00) It’s almost like you’re the one that put the outline together. Whoa.

Shimin (13:05) let’s go to our new segment.

Rahul Yadav (13:05) Mind blown.

Shimin (13:07) Model review. I think it’s time that we you know give a little more dedicated time to how we feel about the latest models.

And this past week we got a little bit of interesting news, which is there was a dark horse model on open router called Ox Alpha. came out like a couple of weeks ago and it was all the rage. Everybody was talking about it. This this this model was super cool and it was being given out for free. so that also gave everyone

a lot of interest, like this almost frontier level model that is just free to use. and last week or two weeks ago it came out that aux alpha is actually ZAI’s GLM53 flash and what ZAI says about the model is that it is frontier level intelligence at the cost of a flash model. two other

really big thing about it is it is it was completely served by Chinese made chips during that free access period, which I think is is huge. and also that it is a big jump in the Pareto frontier on the artificial analysis index. And you can kind of see it here, right? Like we’re looking at a cost task versus cost.

frontier chart and you see that mimo v2 was the the only model cheaper than GLM five three flash and then we saw a huge intelligence index jump up to GLM five three and the fact is at least according to the artificial analysis their benchmark GLM5.3 flash is

Better than the likes of Deep Seek V4, OpenAI’s Luna at max level. It even beats out Sonnet 5 at max thinking. It beats out Qwen3.8 and it also beats out GROK 4. It’s around the same level as Grok46, which is all really, really, really impressive. The other thing, speaking of it being flash cost, the price of

GLM53 right now is 0.075 cents per million input and only a quarter per million output. it is 50% off. But even if you double that, right? Like compare it to Sonnet 5, which is at $2 per 1 million token input and $10 per million token output, or 2 and 6 in the case of Grok 4.6, like this is a a a really great deal.

when it comes to just like not frontier level intelligence, but good enough intelligence potentially with basement bottom prices, I wanna say. what else? yeah, so they also talked about how the this model was completely served on

Chinese chips, but that also GLM53 self-improved its inference on those chips. similar to kind of this hardware level agentic self-improvement, recursive self-improvement, where

Dan (16:05) Mm-hmm.

Shimin (16:06) the model itself improved the serving stack, its memory and bandwidth bottleneck by by a good deal. and it is also vision enabled.

So it can be used to do more tasks, right? Like front end design. which is what we probably use the vision for for the most part. So all you know, this appears to be like a solid four five to five ish level model at like one tenth the price, I wanna say.

Dan (16:33) That’s pretty amazing.

Shimin (16:35) Yeah. And so as a part of model review, I did run the model through my standby benchmarks, which includes asking it what the model should do with two acres of land somewhere in Washington to make money.

Dan (16:49) Yeah.

Shimin (16:52) didn’t do so good, I’ll be honest. The the two acre benchmark, middling, maybe subsonnet slightly less than sonnet the most

The most damning flaw is it expects all the agricultural work jobs to pay like twenty to fifty dollars an hour, which is absolutely not the case, guys. You you get like you get you get no money, you get minimum wage doing agricultural related stuff. then I ran it through the given what you know about AI, what are the second, third, fourth, fifth, and sixth order effect.

benchmark that I’ve been doing. And recall what happened with GLM52. When it got to the fourth, fifth, and sixth order effects, it will start spewing things like, we are talking about cosmic level of effects. It is

Rahul Yadav (17:39) Ha ha ha.

Shimin (17:41) about like what does it mean to be a human? GLM53, at max thinking, did not fall into the same trap.

So even at fourth, fifth, and sixth levels of impact, it gave like reasonable chains of causations. So if we take a look at this particular example on the screen, fourth level, if you can’t sue an agent, you should be able to insure one. So then there will be an insurance industry that’s about liabilities for agentic behavior. And then the fifth level, legal systems will then formalize this insurance. So then

You have some idea of what a machine personhood is, not like based on sentient personhood, but liability personhood, because agents are gonna make

Rahul Yadav (18:22) Mm-hmm.

Shimin (18:23) make mistakes and somebody’s gotta be responsible for it. And then the sixth level is we’re starting to talk about you know human dignity and legal capacities when it comes to AI agents. this is like a little bit closer to the what does it mean to be a human?

stuff but it’s still very narrow and and limited to the legal area for agents. so I think this is a very good result on this particular set of questions. And what is interesting though is it did talk about Freedom of Information Act requests and how agents were able to quickly enable that. Which

For a Chinese model, it’s interesting that it jumped to Freedom of Information

Rahul Yadav (19:01) Yeah.

Dan (19:01) Yeah.

Shimin (19:02) X immediately. Huh. I suspect there’s some distillation happening. Or at least there’s a lot of that in their initial training corpus. yeah. All in all, a really strong model, and especially since the chips are made in China, it reminds me of what back in like, I don’t know, 2000

twelve thirteen when you started hearing about these Chinese smartphones, like the Xiaomi’s of the world. And at first it was like, this thing sucks, but then it was so cheap and it was good enough that like I think Xiaomi is the third largest smartphone producer in the world, right after Apple and Samsung. So this potentially has the first inkling of something like that for the AI world.

Dan (19:44) I had to ask deep seek for what it’s what it would do with two acres in Washington. I was just curious, like my self-hosted

Shimin (19:50) Ha ha ha.

Dan (19:51) version. So I said, here are the realistic plays, roughly by effort slash return, high margin crops.

Shimin (19:59) Mm-hmm.

Dan (20:00) Best on the east side, which is sunny. So it’s at hops, but it knows that hops you know, are only on the east side. So

Shimin (20:08) yeah, that’s

Dan (20:09) Not bad.

Shimin (20:11) okay.

Dan (20:11) But they’re saying you only get ten to twenty K per acre at harvest tops.

Shimin (20:15) There there

there goes your Hobwater tycoon dreams, then. Maybe.

We can grow some for ya.

Dan (20:21) You’ll have to send me the exact prompt that you use with that. I wanna see how well my my Q two

Shimin (20:26) well, yeah.

Dan (20:27) deep seek does. That’s gonna

Shimin (20:29) So yeah, give it a shot. GLM five three flash seems to be a really strong model that is also super cheap.

Dan (20:36) And it looks like I can run the two bit version of it on my hardware.

Shimin (20:40) Well

Rahul Yadav (20:40) Yeah.

Shimin (20:41) that’s really nice. Okay, why don’t we move on to the hardware hut where Dan is gonna get spicy.

Dan (20:47) Speaking of hardware. Mm-hmm.

Yeah, so there was recently a forget the name of it, hot chips, I think is what it’s called. Trade show, somewhat ironically. This with the name jalapeno. and OpenAI did a whole spiel on on stage where they were talking about their their new

A jalapeno chip and so a bunch of details came out of that. And then they also did a presser on their website about it. So it’s a little sparse on the stuff that I would find super interesting, but their numbers are pretty compelling. So just a little bit of background first, if I think we covered this a little bit, but in case you didn’t listen to that one, OpenAI, like many other Frontier Labs, is trying to make their own hardware now.

And they were able to create essentially a brand new chip from the ground up in less than nine months from the start of R D to tape out, which is pretty crazy. and they’re also working with shoot, I forgot the arm vendor’s name.

Anyway, one of the more popular arm vendors to actually like do the the fab and everything. Hmm?

Rahul Yadav (21:51) Was it cerebrus? Cerebrous or no?

Dan (21:56) No.

it’s like a router company that you’ve heard of like a million times. so the thing that they’ve really been focusing on with this new chip is like cost per token per second per kilowatt, which is kilowatt hour. Yeah, or kilowatt. Yeah, no not no hour, which is kind of fascinating. So they ran a bunch of like off the shelf open

weights models on the new hardware, which is not even optimized for those models, right? It’s just like you know just the inference chip. So they GPT OSS like 120 or no, not even 120. It’s like 538 billion or something like that. Like the by far the biggest one got 22K tokens per second per kilowatt.

and running the equivalent workload. It’s interesting. If you look at the appendix, they tell you it’s actually being run on GBM 200, so like in video

Shimin (22:45) Mm-hmm.

Dan (22:45) chips, but they’re very, very careful to avoid that anywhere outside of the appendix in the article. only got 427 tokens per second per kilowatt. deep seek deep seek R1, similarly high, like 12k versus 118.

And then Kimi K two point five six thousand seven hundred and forty-four versus a hundred and twenty tokens per second per kilowatt, which is pretty crazy. so that’s in aggregate against sort of like randomized batch jobs. if you look at like per user decoding throughput, it’s not quite as drastic.

Shimin (23:18) Mm-hmm.

Dan (23:19) so it drops down with across the same three models from like 1459 tokens to 535.

700 versus 169 on Deep Seek and 694 versus 182 on Kimi So it’s still impressive, but like, you know, for just a single user, it’s not quite as crazy. But obviously, when you’re a big company, that single user case doesn’t matter as much. but the thought the thing that I did find pretty interesting that they don’t go into a ton of detail here, but I found a couple other articles were

They’d covered the actual presentation they gave at Hot Chips is how they’re actually achieving some of these numbers.

Shimin (23:53) Mm-hmm.

Dan (23:54) So in a more traditional like GPU-based setup, I guess it’s pretty common to stage the workflow into three chunks. So you’ve got like pre-fill that’s happening on one set of cards, generation on another, and then like they’re doing like multi-token prediction.

stuff too. So it’s also running like the forecasting models. but the interesting thing about that is they shuffle the KV cache around between those three sets of machines every time they do that. and so the one thing the jalapeno did that actually like drastically increases its efficiency is they don’t ever shuffle the KV cache. It always stays on the same die in terms of like how it runs operationally.

But what they have is sections of the chip that are optimized for each of those tasks. And they’re behind like essentially a power gate. So if it’s in the pre-fills section and it’s not doing anything else, it literally shuts off the rest of the chip. And that’s how they’re able to like get such, you know, crazy power gains. because they’re just kind of shutting it down. Plus the the KV shuffling and

not having to deal with like, you know, potentially bandwidth limitations there, means that their overall like head tail latency went way down. it’s

Shimin (25:09) Yeah.

Dan (25:10) like one point five times less latency than running on Nvidia hardware. So

Shimin (25:14) We’ve spoken about like the you know, Google’s tensor processing units on

Dan (25:19) Mm-hmm.

Shimin (25:19) the show before and compared to that, this feels like more of a optimizing the chip just for this large language model, this particular class of operations, right? It’s almost

Dan (25:30) Yeah.

Shimin (25:31) like OpenAI thinks the ecosystem or the workflow is mature enough that it is it’s gonna harden or calcify enough that it makes

Rahul Yadav (25:38) Mm-hmm.

Shimin (25:39) sense to

to do this kind of optimization.

Dan (25:42) Yeah, and the other part that’s wild about the optimization is they’re all those numbers that we read off are all running on single token prediction. And the NVIDIA numbers are all running on multi-token prediction with the forecast models.

Shimin (25:55) Mm

Dan (25:55) So it’s actually winning without even having like the, you know, the MTP models in place, which is kind of wild. Which, like, if you haven’t heard of that or haven’t played around with it, it’s essentially like a smaller copy.

that attempts to predict what the next token is, like more cheaply than the big model running it, and then it either validates or it doesn’t. So like every now and then you get a hit and it paid off that extra CPU that you burnt to to try to get it.

Shimin (26:23) Yeah, and they are comparing with GB two hundred and G B three hundred, which are like top of the line Nvidia chips too. Like they’re they’re not, you know, comparing this to my RTX forty ninety, you know, not not consumer

Rahul Yadav (26:34) Mm-hmm.

Shimin (26:35) grace stuff, right? This is like current generation NVIDIA chips. Meant for data center workflows.

Dan (26:40) Yeah. but yeah, it’s pretty pretty interesting to see them doing this. And it’ll be supposedly their first generation is ready and it’s already being deployed. Like on the roadmap document they showed, it’s actually kind of like some of the basis for their like GPT work stuff.

Shimin (26:54) Mm-hmm.

Dan (26:55) the rollout of that is like pinned, you know, not pinned, but it’s being like held up by some of the extra compute provided by these chips. And then

generation two of these is already in tape out, so it’s like pretty much ready to be produced and gen three is kicking off. So they’re they are not joking around about this. that yeah.

Shimin (27:14) A lot of secrecy, yeah. And

they’re they are using they’re also using AI to help them design and optimize the chip.

Dan (27:20) That yeah,

so that was the last thing I was gonna mention. So not only are they using it to do that, but I it seems like there’s at least some element of like software programmable part to the chip that’s in place that it they’re using AI to like optimize it. So very similar to what we actually just heard about from ZAI, yeah, is

Shimin (27:37) Is the AI? Yeah.

Dan (27:39) like the the apparently the design of the chip and the like

system language or whatever that’s used to program it is all intended to be very easy for both humans and AI to understand and model. So it’s supposedly very easy to program and you know totally optimized for for AI improvement. So

Shimin (27:57) Yeah, the thing that I kind of there’s a lot of secrecy involved with this whole project to begin with. it caught my eye that they’ve been using GPT Astra, their unreleased latest and bestest model, on this chip for at least two months as part of the final kind of optimization program. So they’ve had Astra for at least 60 days, and they just haven’t even considered releasing it to us.

Miffed about that. Alright.

Dan (28:22) ‘Cause who knows what it could act.

Shimin (28:27) Yeah. They were they were doing this and also I think Astra was also the model that that hacked hugging face.

Busy. Busy I open AI.

Dan (28:35) So yeah, that’s jalapeno

Shimin (28:36) All right.

Dan (28:36) in a nutshell. It’s seems a little spicy, but you know, there’s more shrouds than spice right now. It’d

Rahul Yadav (28:42) Yeah.

Dan (28:43) be very interesting to get like a full architecture deep dive and compare that to something like the

Rahul Yadav (28:47) Mm-hmm.

Dan (28:47) TPUs. But

Shimin (28:49) Let’s go to something that’s closer to our everyday developer workflow.

Rahul Yadav (28:55) Mm-hmm.

Shimin (28:56) for post processing this week. we’ve got a title an article titled The End of Programming. You just you have to you have to include an article with that with that title.

Dan (29:05) The end of programming.

Rahul Yadav (29:07) the end of programming brought to you by Paul Dix. He’s a founder and CTO at Influx Data, which is the creator of InfluxDB. so Paul recently started writing this blog and in other articles too. He has little like hints of this that where we’re going is you know.

potentially AI is going to do most of the work. And this article, he cites the recent release of Bun 1.4. this is the one where they were they had announced that they’re gonna rewrite it from Zig to Rust and then it wrote it had a mil over a million lines of new lines of Rust code.

The interesting piece is it seems like it was written by one person using Fable V. the they used it to write the initial version, and then even more impressive, over the course of two or so months, the agents were continuously they wrote it over the course of about eleven days. it was close to 7,000 commits and

it would have cost roughly like hundred and sixty-five K in API pricing. and then after those eleven days, before releasing it, they the agents spent close to two months in, you know, tweaking it, fixing all the bugs and testing it out before it got released. and that’s the impressive part here is that

Dan (30:33) Well also

the the node compatibility between well like the previous the the hand coded, partially hand coded whatever previous release that was in Zig. I forget what the node compatibility was. Meaning like compatibility with the Node.js API.

Shimin (30:48) Mm-hmm.

Dan (30:48) it was like forty percent test coverage or something like that. And I think now it’s up to like seventy or eighty in the new version too. So it’s actually like not only did it work, but

It’s actually more compatible than the old one was.

Rahul Yadav (31:01) Yeah.

And I guess you kinda need more and more thorough testing to be able to use A AI, right? Yeah. Without it.

Dan (31:09) To do that. Yeah. But it’s also like it’s it’s

the best case scenario. You have a well-documented, well-tested code base, and you basically want the exact same functionality in a different code base, you know?

Rahul Yadav (31:23) Ha ha ha.

Shimin (31:24) Yeah.

Dan (31:25) Like, yeah, is like made for that task, I feel like.

Rahul Yadav (31:30) Yeah.

Shimin (31:30) Yeah,

but at the same time, this latest version of Bun is being used by millions of people, right? It is what is running.

Dan (31:36) Yeah, it’s underpinning

Claude code. Yeah.

Shimin (31:39) Exactly.

Rahul Yadav (31:40) Yeah.

Shimin (31:40) So if there is some actual issue that the tests didn’t cover, we’ll discover that by.

Rahul Yadav (31:46) Yep. Yeah. and so then the this is the first example and then Paul has a couple of things that he tried on like w how much of fable usage he could get on a couple of projects at InfluxDB and he does know that they’re not production ready, but he was able to mostly give oversee the architecture.

but didn’t actually look at the code or anything. and then it took about twenty-eight hours to get to the get to a working implementation of one of the projects. and so the point that Paul makes is the per the person who used Fable Five to rewrite Bun and took about eleven days, let’s say they’re a year

ahead of where you know we peasants are with our access and our limits

Dan (32:35) Ha ha ha.

Rahul Yadav (32:36) and and the yeah,

Shimin (32:37) We don’t have a hundred thousand to spend on tokens. Yeah.

Rahul Yadav (32:40) the inability to burn 165K in 11 days or so. but but the interesting thought experiment that he’s proposing is if you look about a year ago from when he wrote this blog last week

we had early versions of Opus and GPT, and compared to that, how great things have gotten a year later. And so if you play that out a year from now, we would have very cheap intelligence, but a very like you know, smart intelligence that people are using today to rewrite these things. And so if you keep playing that out, at some point we would have agents that would be writing all the

code because it would be so cheap and they would be able to move so fast and everything. and we would basically be overseeing, which to me is also maybe at that point because if agents are writing and moving so fast, at some point either you you know, accept that you’ll be the bottleneck and you’re continuously overseeing everything or you just end up having you play down the road and

Agents are rewriting things, agents are testing things, everything’s happening for agents. It’s like, you know, of the people for the people by the people, but substitute people with agents.

Shimin (33:56) Nice reference. Yeah.

Rahul Yadav (33:59) yeah, because more and more of the internet is being written for agents now as well, where you know, you wanna make sure that it shows up in AU and things are more parsable by agents and everything. and if we go down that road,

then Paul said that will be the end of programming. So

Shimin (34:17) Yeah. And

then we’re gonna need to have insurance for our agents, right? And then also

Rahul Yadav (34:20) Yeah.

Shimin (34:22) legal liabilities for agents. I I I think this article does bring up good point that if your day-to-day programming feels more like translation than it does like truly building, it’s probably true that, you know, the the role that you’re currently doing, the days are numbered. but I also wanna point out that, you know, it took eleven days to write the initial version

But it took a couple of months to refine it. Now the refining part probably won’t be automated away just as quickly. And it’s still very much needed. So we’re not talking about a lights out AI dark factory for code. Right. We’re still talking about human in the loop. it’s not magic and refinement still requires a lot of skill.

so if you’re going to go where the puck is going to, like that is an area that you should focus on.

Rahul Yadav (35:12) But if you get much better agents that can work on a long horizon where, you know, let’s just go with this example where you already have the end state that you’re aiming for. I could see a year from now where you need less human involvement even in refinement, ‘cause it already knows and it can keep running the loop until it hits that end state.

Shimin (35:36) Yeah, absolutely. So you need to build a verification system and you need to be able to give proper directions. And I don’t think you can just take anyone off the street and be like, hey, go translate Bun to Rust, right? Like like that’s not

Rahul Yadav (35:50) Yeah.

Shimin (35:50) gonna happen. So you still need someone who is skilled enough to give those directions and skill enough to know w what verification looks like.

Dan (35:57) Yeah, I mean I think I could see that being a a plausible future case, but like

I mean, I guess to some degree it’s already here, right? Like that there’s I don’t know how much code I’ve written with Claude in the past six months that I’ve spent a lot of time personally reading. I usually try to do the teams a favor and do like a code review of it myself before I force another human to review it. But that’s largely because I feel guilty having a human review it without

Without doing that first, you know.

so I don’t know.

Shimin (36:29) This this is why you’re

a senior, yeah. ‘Cause you you have good manners.

Dan (36:33) Yeah.

Rahul Yadav (36:34) Also the i y then going off of what you were saying, I was thinking about how, you know, we use Claude code or cursor or pick whatever ID

most of the people who work in those companies, from what I can tell, are also not reviewing that code. And yet, most of the time they don’t have issues and it works fine. And same with Bun how Shimin was saying, i it’s being used by millions of developers. part of the solution seems like if you have a large

base, then you would be able to rely on them to test things and then you can do your, you know, blue-green deploys and you can even slowly roll some things out and test them and easily roll back. And you can automate all of that. So it almost feels like it would be hard to have a small base and then do AI generated stuff because your base would not be large enough to test

Shimin (37:31) Mm.

Rahul Yadav (37:31) all the different variations.

And then if you keep making if you keep shipping buggy stuff to a small base, it would never get to a large base. But you already if you already have a large base, then you can keep just iterating on it and and at a any given time all you need to do is test on a very small percent of that large base and you can automate a lot. So it it feels like this chicken and egg problem a little bit on on how AI agents might be used.

Shimin (37:58) Maybe it’s true in the case of bun but like we just talked about how Meta saw a seventy percent increase in security incidents. So it’s it’s not foolproof, right?

Rahul Yadav (38:06) That’s sure. Yeah.

Dan (38:08) But I mean again, like that’s because

Like you’re you’re not starting from a well documented thing and then just transitioning it to another, right? It’s more hand wavy what you’re doing for like traditional product product development. Like so I I mean I guess my personal example of this is I I’ve probably talked about on here. I have a a whole like home lab setup and play around a lot with like home assistant too on the home lab. So it’s like automating stuff in my house.

And there is a really good Python library for managing this like third party like smart home system called Tuya. It’s like really common in like Chinese manufactured stuff. like in you know, light controllers and stuff like that. A lot of them have like Tuya chips in them. And but all of the automation stuff that I

wrote, which is like a whole bunch of custom automations running on that system are all in TypeScript. And so I was like, I wonder how much it would take to just port the core of that module over to TypeScript. And Claude did it in like an hour, complete with all the tests. And I think the thing that made it easy was the fact that the test suite had payloads in like hex already for the actual different like messages that that thing

like speaks in. So it was able to basically a hundred percent test all of the like binary math to make sure that, you know, like it’s XORing things and stuff like that to make sure that it all complied. So it was actually like a pretty easy port. So that’s like that’s what I’m saying is like that kind of thing. That is the you know, the name of the game, I feel like whereas like as soon as it’s

Shimin (39:38) But how different

how different is that from a user story that says as as a user I wanna be able to do that and and get why and and have the output like listed out in some sort of payload format?

Dan (39:49) There’s a lot less that can go wrong with the validation, I feel like, in something like that. You know, it’s sort of like the difference between like unit like testing on the front end and the back end, right? Like you’ve done a lot of front end work. So you know, like the front-end testing story, even in 2026, even with AI, is still kind of which shall we say a shit show? Right.

Shimin (40:08) Yeah, it’s kind of it’s kind of crappy. Yeah.

Rahul Yadav (40:09) Ha ha ha.

Dan (40:12) Whereas on the back end, you can write tests in

Mostly they work, you know, and you don’t

Shimin (40:17) Yeah.

Dan (40:17) have things constantly just randomly breaking because they decided to. and so I f I feel like that same well, I feel like that same correlation

Shimin (40:21) Maybe f yeah, maybe front end developers will have a job after all. I don’t know.

Rahul Yadav (40:25) Ha ha ha.

Dan (40:27) kind of applies, right? Where it’s like the acceptance criteria for what makes a good like whatever for distributed systems cool, but like the acceptance criteria for like what is the

User’s perception of that distributed system hooked up to another one, hooked up to a UI is maybe different, right? And that’s a harder thing to like quantify and harness. I’m not saying it’s impossible. I just think we haven’t really solved that yet as well.

Shimin (40:51) Mm-hmm. Yeah.

Dan (40:52) And I’m not and obviously it’s working for a lot of people, but like that until that is like really well solved, I think that’s why when you’re gonna continue to see like a little bit of a disconnect. And that’s why something like Bun is

Perfect, right? Because it’s like you could literally just like take code and run it in both environments. And if it runs perfectly, then you know like that was a good test, you know.

Rahul Yadav (41:14) Yeah. Okay.

Shimin (41:15) Yeah. And speaking of

the disconnect, Paul did mention in this article that and I quote, organizational inertia will likely mean that there’s another decade of human writing code by hand and having their colleagues review every line of it. Many, if not most, companies will continue to develop software as they have before. and and this goes up to our previous discussion too. Like the technology is here, we just haven’t figured out how to harness it correctly yet, and we are seeing glimpses of it.

there might be a way to have, you know, better specifications and better refinement processes even when

Dan (41:49) Yeah.

Shimin (41:50) the spec is like not well defined. We just haven’t been able to get there yet.

Dan (41:55) It’ll be really interesting to see how that changes with world models, right? Because I think I think we haven’t really like touched on that too much. Like we’ve talked about like, you know, world models are, you know, people are focusing on them and you Jan Lacoon’s talking about it and everything. But like if the model’s understanding of reasoning isn’t built purely on language, all of a sudden

those sort of UX tasks that we were just talking about or like UAT type tasks might make a lot more sense because like the reasoning would be built more around like an understanding of the physical world and like UIs after all are essentially like meant to be a stripped down version of the physical world where like things like buttons are very literally buttons, right? You know, it was like

Rahul Yadav (42:34) Yeah.

Dan (42:34) so

Yeah, we’ll see. It’s exciting time to be alive as as per usual.

Shimin (42:40) And GLM five three did also predict that world model and robotics would be a big thing. So we’ll we’ll see. We’ll see if that happens.

Dan (42:46) Yeah.

Shimin (42:48) Alright, let’s move on to two minutes to midnight, where we borrow the analogy of the Armageddon clock by the bulletin of atomic scientists where midnight is Armageddon and also in our case when the AI bubble will burst. we are at four minutes as of this week, as of last week. So Dan, you’ve had our first article of the week.

Dan (43:14) Yeah, so NVIDIA released some numbers and they’re doing pretty good, it seems like. So they have you know, as the headline says, forecast a 70% sales growth next year, which is you know just forecasting. but their actual numbers, they reported ninety-six point two billion for the quarter. and 108.

billion for their current quarter, which beat the streets expectations. So that’s exciting. and then their CFO is on the record expecting the strong growth to continue. And that’s where that 70% number comes from is is collect crest’s numbers. And she said customer demand was set to double, but revenue will be limited by supply constraints. So that’s the second half of the the the story overall with this is

Their chip group, which as you can imagine is driving a lot of this, is has declining margins. And that is purely caused by which so their gross margin was 75% in the quarter, but they said it’s gonna drop as low as 71% by early next year. And that’s being driven by the cost of high bandwidth memory.

Shimin (44:16) Mm.

Dan (44:17) So the chip crunch is even hurting NVIDIA, which is pretty

Shimin (44:21) Yeah.

Dan (44:21) wild.

Shimin (44:22) But

they’re not taking, you know, competitors like jalapeno into account there. Or the Chinese chips that we’ve talked about.

Dan (44:28) Right. Yep.

And then the other thing I thought was kind of odd is that you know, there are all these big numbers, but then they’re sort of hand waving a little bit and they’re like, It’s commitments with suppliers increased to two hundred and seventy nine billion from a hundred and nineteen billion quarter over quarter.

Shimin (44:45) Mm-hmm.

Rahul Yadav (44:45) Cheese.

Dan (44:46) So like, talk about margins, but the the and they said that was pri primarily rel related to the procurement of memory, which is kind of nuts. And the other thing that’s kind of concerning is their free free cash flow dipped from to twenty one point three billion last year, which is down almost half from the previous three months. So that was forty

Shimin (45:05) Alright.

Dan (45:05) eight point six billion. So you can look at that, I think two ways. I mean, again, I’m not a financial analyst, but like

Everybody’s always talking about free cash flow. and the

Shimin (45:14) We just play one on TV.

Dan (45:15) Yeah. Well, like I you could read it either way, right? But isn’t that also related to their debt raising too, potentially? So it’s like, you know, they’re spending there’s more money’s going out to vendors and then they you know, these circular financing deals could be eating into that too. I don’t know. Again, not an analyst.

Shimin (45:33) Potentially. Yeah. my article this week, it’s actually just a stat, is also related to Nvidia. and this I think it’s the first time we’ve quoted something from from a tweet or a blue sky, a skeet on this show. Nvidia has ninety six point two billion dollars in revenue and Amazon has two hundred billion in revenue. So Nvidia has half the revenue of Amazon.

And like also it has twice the market cap, which is pretty crazy in my opinion. I I know the video is large, but like so is Amazon. and and to have that much money coming in, and still be double the the Amazon market cap is something that like it’s kinda hard for me to comprehend. Yeah.

Dan (46:17) Hype, hype, hype, hype, hype. Yeah.

Rahul Yadav (46:19) I p

Shimin (46:23) Hype? Okay. Rahul what do you got?

Rahul Yadav (46:25) mine is not related to NVIDIA, directly at least.

Dan (46:28) Woohoo.

Rahul Yadav (46:29) This is by Marina Temkin at Tech Tech Crunch Anthropics annualized revenue is surges to sixty-five billion dollars. So d just in May they were at forty seven billion. Back last year they were at nine billion, and at the end of July,

their revenue based on the recent period is projected to sixty five billion dollars. and they’re the company the investors are expecting that it’s gonna keep growing and that they’re gonna end the year somewhere between six a hundred billion and hundred and twenty billion dollars. also they are planning to IPO soonish, sometime

Shimin (47:06) Mm-hmm.

Rahul Yadav (47:07) this fall, and fall is here.

It’s September first, today. So yeah, we’ll we’ll we’ll find out what the market thinks of Anthropic pretty soon.

Shimin (47:17) Yeah, sixty-five billion dollars. I asked AI to I asked Claude to pull up some numbers of similar companies that have sixty five billion dollars in annualized revenue. Boeing is one such player, IBM, Morgan Stanley, BASF, the world’s largest chemical company, Unilever, LG Electronics, just to get like a diverse range of revenue numbers from all around the world.

This it’s it’s kinda madness how quickly those numbers came up, right? Like four years ago it was

Rahul Yadav (47:44) Yeah.

Shimin (47:45) nothing, now it’s as big as BASF and Unilever.

Dan (47:47) Yeah.

Shimin (47:48) And they’re gonna go to a hundred and twenty,

Rahul Yadav (47:49) Same.

Shimin (47:50) like oof.

Dan (47:51) It’s pretty wild.

Shimin (47:52) Yeah.

Rahul Yadav (47:53) Yeah.

Shimin (47:53) Alright.

All that said, how do we feel about the clock and where it should be for

Dan (47:58) I’m actually

Shimin (47:59) This week.

Dan (47:59) fairly optimistic about the clock right now. Like I think what we’ve pulled up this week is like, you know, a ton of evidence that we’re clearly in a bubble, like in case there was any doubt. But

I don’t think some like if Nvidia’s driving the circular financing and they’re still doing quite well, I don’t think we’re in any immediate danger of it going anywhere.

Shimin (48:19) Yeah, I’m with you. Like Anthropics revenue number is as large as any large company in the world. So

Dan (48:25) I I did read a really interesting article, which maybe we can throw on to a a future one, which is basically saying that the receipts will come due for a lot of this when the sort of data center build out that’s been like pushed, right? Actually hits the books, ‘cause right now it’s all sort of like not really on the

Shimin (48:43) Obligations.

Dan (48:44) books. Yeah. and apparently that’s due to hit sometime next year. So

That also makes me think that it’s kinda like, you know, we’re gonna see this kind of go happily along for a while and and even

Shimin (48:55) Yeah.

Dan (48:56) even like these IPOs may not change things all that much, like in the way that I’d sort of earlier been thinking, like, we’ll know a lot after these IPOs, right? We may not.

Shimin (49:04) Right. I mean

Rahul Yadav (49:04) Yeah.

Shimin (49:04) six sixty five billion dollars is a lot, but not when the denominator’s two trillion, which is how much

Dan (49:10) Mm-hmm.

Shimin (49:11) we are projected to invest over the next two years in Data Centers. So

Rahul Yadav (49:13) Yeah.

Shimin (49:14) I’m with you. I can do four fifteen, four thirty even.

Dan (49:17) Okay. yeah. I’m not I’m

Shimin (49:18) No for fifteen or thirty.

Dan (49:21) not like super pushy about pushing back. I just don’t want to go forward right now.

Shimin (49:25) No.

Dan (49:26) It’s kinda where

I’m at.

Rahul Yadav (49:27) Anthropic has according to Google, Anthropic has around five thousand employees. And so five sixty five thousand divided sixty five billion divided by five thousand, you’re getting about thirteen million per employee and by when you talk about efficiency. Yeah.

Dan (49:45) It’s

Shimin (49:46) That’s a healthy per employee revenue number.

Dan (49:48) Yeah.

Shimin (49:49) Okay, let’s do

Rahul Yadav (49:49) Every job interview is

can this person make us at least thirteen million dollars in a year or more?

Dan (49:54) Wow.

Shimin (49:57) okay, four fifteen it is, and with the setting of the clock, we have come to the end of the show. thank you for joining us for our little debate and study session this week. If you like the show, if you learned something new, please share the show with a friend. You can also leave us a review on Apple Podcasts or Spotify, or subscribe on YouTube. it helps people to discover the show and we really appreciate it. If you have a segment idea, a question for us or a topic you want us to cover

shoot us an email at humans at adipod.ai. We’d love to hear from you.

You can find the full show notes, transcripts and everything else mentioned today at www.adipod.ai Thank you again for listening and we’ll catch you on next week’s episode.

Rahul Yadav (50:38) See ya.