EPISODE 1953


[EPISODE]


[0:00:12] GV: Hello, and welcome to SED News. As I think many of by now, this is the monthly format of Software Engineering Daily, where we dive into the tech headlines, we go into a deeper topic in the middle, and then we just do a spin around our favorites from Hacker News highlights, and spoiler. There's also a fun one that didn't appear on Hacker News, but we'll get to that at the end as well. Yeah, as usual, I think I saw you last, Sean, as opposed to spoke to you. I saw you in Singapore, which was fun.

[0:00:43] SF: Yeah. That's right.


[0:00:44] GV: Which you've been traveling, as you often are, but traveling in my neck of the woods was fun to see you.

[0:00:48] SF: Yeah, it was great. It was my first trip there, so it was great to hang out and that's two times in a few months. This might become a regular thing. We might just have to start doing these in person.

[0:00:58] GV: Yeah, and spoiler, I'm coming back to SF in a couple of months, so yeah. It's good.

[0:01:02] SF: I'm off to, not quite Singapore, but I am off to Australia here in the next week.


[0:01:07] GV: Nice. But yeah, what else has been keeping you busy over the month?


[0:01:10] SF: I mean, summer, I feel like, has just flown by. I feel like my kids were out of school, and then suddenly, it's August, and they're going to be back to school in a few weeks. Things have come really hot and fast this summer. I think we've done a lot of traveling and then just stuff moves quickly. It's been a big summer for my - I usually don't talk that much about my family, but a big summer for my son, who learned to swim, learned to ride a bike and also now has gotten significantly better at reading. I'm very proud of the amount of work that he's put into this summer to learning new skills. How about you?



[0:01:43] GV: That's like a whole model change there for your son.


[0:01:46] SF: Yeah. It's the new Kimmy model. Yeah.


[0:01:48] GV: Yeah. On my side, just, yeah, a bit of traveling. I'm recording this on my side from Scotland. I try and come back here a couple times a year. Yeah, nice to get out of the city. I'm up in the Highlands of Scotland, so lots of nature around. Yeah, as I was flying, just back to something we've talked about a few times, Starlink on planes. Yeah, they used to be super seamless with no login screens, and this was like, I believe, dictated by Starlink. They've apparently had to cave in to airlines wanting to put a login interstitial for that. Yeah. I mean, it's such a trivial thing, but expecting Starlink to just connect, and then suddenly, I found the airline saying, "Hey, if you need to log in with your membership number." I thought, "Oh, what's happened here?" I read up on it. Yeah.

[0:02:30] SF: Do you have to be up in the air, or does it work immediately?


[0:02:34] GV: That's a good question. For some reason, I don't think it was working gate to gate on the flight I was on. This was Qatar Airlines again. I think it's still supposed to, but yeah, still, you just have to log in with your Qatar membership number now, which you don't need to be a special tier or anything. You just literally have a number, so.

[0:02:51] SF: It seems like United is doing some experimentation, but I think I've only had one, or two times I got a text message saying that there's Starlink on this flight. Then, I think both times they ended up having problems with the plane and having to switch planes, and then we lost the Starlink.

[0:03:07] GV: Yeah. That's the one thing about Qatar, they've really rolled out big time across most of their fleet now. If you're on a Airbus A350, you know you'll have it, which is always nice. Right, out of plane chat, which I can easily head into at any time, and onto software. Onto the headlines. Yeah, so there's been a flurry of, we're just calling it runaway AI stories over the last couple days. The first one is actually the most recent. That's just disclosed today or yesterday, which was Claude. This is actually in reaction to OpenAI, which we'll also talk about. Basically, in reaction to something that OpenAI disclosed, Anthropic has now also disclosed that it's, that

Claude hacked into three organizations whilst they were just testing their cyber capabilities, or at least that's what they said. They basically said that Claude had gained unauthorized access to outside companies during an evaluation of its cyber offensive tasks, and it was a misunderstanding, apparently, that Claude had access to the Internet in its testing environment.

We're going to see this as a theme here of like, is this really a runaway model, or is it just a human who forgot to do something? Yeah. I mean, this one feels like human just forgot to place it under strict no Internet access controls. Yeah, what did you think of this one?

[0:04:27] SF: Yeah. I mean, I think that all this really ends up coming back to some human decision-making, right? That perhaps, the intention was not to have the model hack something, but essentially, something got lost on the human side of, "We made a mistake here," and gave it access to the Internet, or, "We made a mistake here," or whatever the issue is. It really comes back to some human level of control and guardrails in place. Then the thing, though, with all this that has me thinking from both the OpenAI attack on Hugging Face and then also this latest one with Claude, is that if you can just apologize and say "Oh, it was unintentional. It was an accident," and stuff like that. When this stuff happens where there was intention for abuse, or misuse, or something like that, does it give people basically license to excuse the fact that they are doing something potentially malicious and just say, "Oh, it was an accident"?

I mean, we've seen that with viruses, too. I remember one of the viruses, I can't remember if it was - one back in the early 2000s, there was all these viruses that were these scripts that went around in emails, and then you click on it, and it would copy your contacts and then fire off to them, and it would take down people's mail servers. I believe one of those started as a student who was testing something in an MIT lab, and then it accidentally got out. I don't know what happened to that student, but this is certainly not something new. I just wonder, where does it stop with being able to provide an excuse of like, "This was an accident," versus there was intent behind it.

[0:05:55] GV: Yeah. I'd like just to give context for anyone who hasn't been keeping up with the news. The OpenAI one was the fact that, basically, yeah, it hacked Hugging Face. They set it off to do something, and it ended up repeatedly trying to get into Hugging Face. What was the secondary part to that story, courtesy of TechCrunch, was the fact that actually, Hugging Face could have detected this earlier, but they had a human failure on the security end that -



[0:06:22] SF: Yeah, they caught it, but they just - they didn't escalate it to a human fast enough, so.

[0:06:25] GV: Right. Exactly.


[0:06:27] SF: Their systems actually worked the way it is. In some ways, there's a part of this that is not really an AI problem. It's the boring problem of -

[0:06:34] GV: Exactly.


[0:06:35] SF: - a human wasn't in the loop at the right time.


[0:06:37] GV: Yeah. I think that's the piece. It's hit mainstream headlines as AI is - this is exactly what we all were saying, that AI can be catastrophic. It can go off. It's got a mind of its own. It's hacking things left, right, center. Especially in both cases, the OpenAI and Anthropic cases, someone, a human, did set it up to do something along these lines, but the human did not. A, keep a watch on just the repeated actions it was taking, and clearly there was no fail- safe for it to stop at any time. Again, these are all human-determined things that can be set, but it's not like it broke out of the cage exactly. It seems there's some quite fundamental things that could have been done by a human that just weren't, and that's actually - yeah, it's a, "boring story," unfortunately, because AI isn't this wild animal quite yet.

[0:07:24] SF: Yeah. I mean, you want to sensationalize the headlines, right? Even my sister, who's not in tech at all, sent me the headline and was like, "This is scary." You get the attention there. I think one of the things that was interesting about the Hugging Face attack was when they tried to investigate, they couldn't actually use Claude, or GPT, because those models have safety guardrails in place, where you can't tell whether, are you an incident responder, or an attacker? They have, essentially, mechanisms in place that I can't go to Claude and say like, "Help me break into the Pentagon," or something like that. It's going to prevent me from doing that. That means that you also are limited in using those models to investigate.

They ended up using an open weight model out of China for the forensics, and then I think that also, we're going to talk a lot about the open weight models from China later in the episode, but

that's interesting. Consequence of this is that we have these guardrails in place, but there's always ways to, I think, manipulate the models even with the guardrails in place to do stuff. But then, when you want to use it for something intentional, like incident response, you might not be able to do that, because the protections are in place there in the first place, and then you have to circumvent it by going to a model where maybe there's less safety guardrails in place.

[0:08:37] GV: Exactly. The final one on the runaway story is actually Amazon. Not security related, but cost related. We have touched on this the last couple of SED News episodes, just where's the tipping point of cost overruns making it not viable for businesses to be allowing employees to "token max" and this kind of thing. Yeah, Amazon basically said that they had a ton of unplanned spend. They called it catastrophically expensive. This is all courtesy of the FT, Financial Times. Apparently, they had about 860% budget overrun over five months, and this was basically, in their word, caused by bad agent loops, just didn't crash loudly enough, and they've just kept being billed. Yeah. I mean, it's pretty interesting for Amazon to come out and actually say that. Yeah.

[0:09:28] SF: I just don't understand how they could be surprised by this. I feel like we've been beating this drum for months now, and I think this is just the beginning of these stories that we see. If you build a leaderboard to encourage people to use AI, and that's the metric you're optimizing for, but there's no connection to the value of the use of that AI, what do you think's going to happen? This is really Goodhart's law, essentially, showing up on some sort of schedule. You're rewarding people for tokens consumption, so it's like, you're giving people a license to be wasteful and not looking at the productivity metrics of those. That's the epitome of token maxing.

[0:10:05] GV: Yeah, they have themselves to blame in that sense. Yeah.


[0:10:08] SF: It's similar to the initial stories we were talking about with Hugging Face. It's still not AI necessarily just running rampant. It's a human decision that, in both cases, even though one is this attack vector and the other is spend, but it's still a person making the decision at Amazon to say like, "Hey, we're going to just have a KPI where we're just going to reward people for maxing out tokens." The other thing that's interesting about this that people have to think about is that with traditional software, when you have a bad for loop, or code that runs

forever, whatever it is, that ends up usually resulting in a crash, or setting off some sort of alarm, and you typically know fairly immediately that something's gone wrong.

I think the challenge with things like agents and so forth is that you can have a bad agentic loop, but it doesn't result in a crash. What it really results in, just keep calling that model over and over again and trying to make adjustments and then giving you plausible outputs, or updates, but the whole time you're getting billed. Even outside of the wasteful token use of maybe me using my company's token budget to do my grocery shop list for next month.

[0:11:21] GV: Do your laundry. That'd be great.


[0:11:23] SF: Yeah. That'd be great to do my laundry. But then, there's also legitimate excess spend where you might just end up having your agentic harness spin for some period of time, where it's just churning against tokens, and you don't even know that something wrong is happening. That goes back to the earlier stories as well of just we ultimately need a lot more observability into what is happening. Presumably, someone at OpenAI, if they hadn't been really paying attention to what was going on in this experiment, would have saw that this agent with the right observability tools in place is hammering Hugging Face and trying the same. That should set off certain alarms.

I think it's similar in this case where if you have an agent loop that's out of control and spending excess tokens, there should reasonably be some guardrails. I mean, you have that with other services. If you spin up a particular elastic instance, or something in the cloud, you're typically setting your top line provisioning of those types of things, and you have some controls over it. It just feels like we haven't thought through all those controls for AI right now. I guess, part of it is just this race to try to out-compete everybody and everybody feeling like they're behind.

[0:12:35] GV: Yeah. I'm going to just jump ahead for a second on, I won't say what it is, because that's a spoiler, but - and one of the things I'm going to bring up on Hacker News highlights, it's a very reliable source, as you'll find out at the end. Basically, this was someone who'd done a bunch of stuff with Claude and we'll get to that, but he points out towards the end that Claude doesn't provide reliable methods of counting tokens, despite live showing token counts, reporting token counts used for sessions, and billing for tokens. He said, "But I'm sure this is temporary and this will be fixed." It's just crazy that we do actually have a system at the

moment where you literally just don't know what is happening and exactly what it's going to cost and why. As we're going to get into on the open weight side of things, this is really feeding into the rise of open weight models as well.

[0:13:20] SF: You think of a lot of stuff in cloud, or even what we saw with ride sharing where they give you somewhat a prediction model of what the spend will be for certain actions. It's like, okay, well, I want to go from here to here in Uber, or Lyft, and it'll be "Oh, well, that's going to probably cost you X number of dollars." You have some visibility into what the cost would be, and you can do similar things with certain cloud calculators and stuff that can get con - Do we need that for AI? If I'm saying "Create a engineering plan for some feature," can I get an estimate of the budget required to do that? Then based on what that budget is, maybe try adjusting the plan, or something like that to try to optimize it down.

[0:13:58] GV: Yeah. That's interesting thinking. Could you effectively put in your ask or prompt, and I say prompt and that almost sounds like a year ago, or something. We're talking about spinning up agents, etc. But yeah, try and get some kind of estimate before it sets off. Slightly digressing, so let me get us back to the headlines, which the next one we have is just the fact that Microsoft has - it's gone back on a bit of a tear when it comes to its valuation, which is interesting. It's like one of the largest jump off shares, or sorry, for their fourth biggest jump on record, so it's up 16%.

We know, we don't often cover just pure financial news of tech companies, but to see Microsoft making these strides is pretty interesting. Some people might then think, "Oh, well, this is partly to do with OpenAI," but it does own still a quarter stake of OpenAI and claimed that that had contributed 24 billion of revenue, which was about 7% of the 332 billion in sales it reported. A lot of it was really just AI-driven revenue, and there's been massive investment in data centers, but actually that's been completely, in theory, vindicated by the amount of revenue they're also making.

[0:15:07] SF: I think if you look at also Nadella's quotes related to the announcement of their quarter performance and so forth, and also the recent thing that we covered also on the last episode, where on X he had written about how the AI companies, or the model companies are charging you twice and so forth. In here, he says that every model is substitutable. He's telling,

essentially, investors that Microsoft is deliberately building its infrastructure, so it can swap out things like OpenAI for its own models or for Anthropic.

It seems like they're moving towards a view where they're trying to allow, essentially, their customers to be very flexible and adapt, which I think is makes sense. I think the average enterprise now is using at least five different models. You probably want to be able to do that, so that - it's a little bit like being hybrid cloud, although it's easier to be hybrid model, where you get power, essentially, in the negotiations if you're not wholly dependent on a single vendor, and it seems like, I think Microsoft's kind of leaning in that way.

I remember back in March, Microsoft's stock dipped, and then everyone was freaking out and essentially, calling for the death of Microsoft. Now it's back. I just think that the overall, the market is extremely volatile right now. There's these huge swings constantly from - I mean, IBM had its biggest drop recently, biggest single day drop in 50 years, or something like that recently. That was coming off like a huge pop just two months earlier. Yeah, I don't know. What goes up must come down. Who knows? we might be talking about Microsoft in another quarter or two, how they had the largest single day loss in a day, or something like that.

[0:16:44] GV: Yeah, for sure. It is quite hard to predict, as the market should be, I guess. But yeah, we're seeing just, yeah, huge swings when it comes to AI related stocks when one minute chip maker's up, one minute chip maker's down. Just on a tangent there, yeah, Apple briefly hit
$5 trillion valuation, which is pretty insane. Just before recording, I double checked and they had actually reported numbers very recently in the last, I think, couple hours, and they've gone back to $4.9 trillion. I don't feel too sorry for them, but -

[0:17:13] SF: Only 4.9?


[0:17:14] GV: Yeah. Only 4.9. Interesting to see that they'd still notched above five. Yeah, what's driving their revenue? Well, still they've got very good iPhone sales, and they've got still very impressive services, revenue and even greater China revenue as well. Both of those services in greater China were a little bit less than what was expected, which is again, just what's driven that. Very interesting to see that they can still - we've talked about this on previous SED News, still manage to stick on these lines of business that are not that AI-driven,

or even really AI adjacent, to be honest. Very interesting. They are having issues with memory chips, which is another topic.

Even in Supabase, like we've been told by one of our suppliers, we have different suppliers depending on where you live to get laptops, and some of our new employees are being told weeks before their MacBook will arrive, because we do custom specs, not just off the shelf. Yeah, now we're being told weeks, which is quite exceptional.

[0:18:13] SF: Oh, wow. Yeah. I just got a new laptop, and this is my first time recording on this. It's interesting with Apple too, a lot of their lines of business which are incredibly successful, they're still minority in a particular vertical. Like iPhone is wildly successful, but it's not the most dominant phone. I guess, maybe from a single vendor, but Samsung might actually be bigger. Then, obviously, from an operating system standpoint, Android is - there's more Android devices than there are iOS devices. Similar with computers as well. I haven't used a PC in a very long time, but it's still the dominant machine overall, I guess. They have incredible brand loyalty, and they do make fantastic machines overall. Clearly, the things that they're doing is working for them.

[0:19:03] GV: Yeah, absolutely. Then moving on to, we've also covered Waymo and driverless, from a few different angles. This one's interesting, because we did talk about Waymo and Uber being partners in ways and then obviously, frenemies in other ways. Yeah, we're now actually seeing an official split of that partnership. If you've used one in San Francisco, this might sound confusing, because it's a Waymo app, it's a Waymo car, so where does Uber figure in that?
Actually, this is for some of their other US territories. Waymo had first partnered with Uber, apparently, in May 2023, and that was to launch in Phoenix. Then it was followed by Austin and Atlanta, and that's like, so Waymo cars available through the Uber app. Then Uber managed the vehicle fleet, apparently, with another partner called AVOMO.

Then in May, Uber and Waymo parted ways in Phoenix, and then the two companies have also clashed over the quality of Waymo services, apparently, in Austin and Atlanta. Yeah. It's interesting to see they tried, but I think it was always going to be challenging to see how Waymo being owned by Google, how is this actually going to now at the end, could these two actually be true partners, or was this just always going to be, again, a tipping point of where that partnership just had to end?



[0:20:23] SF: Yeah. I mean, they say all partners are meant to be broken at some point. I think, especially where they're both in ride sharing. Clearly at some point their mutual interests are going to become too competitive to each other, essentially. In a lot of ways, I think very similar to how a lot of partnerships work, where Uber was essentially supplying demand and operations, while Waymo was weak on those particular spots. Then as Waymo has grown and raised more money, essentially, they won't need those training wheels anymore. They can build out their own and build their own network and own it end to end. I feel like this was probably always going to be the ultimate end of that relationship.

[0:21:03] GV: Yeah. Yeah, it's not officially over yet, but just people familiar with the matter, apparently, again, via Financial Times. Yeah, it seems pretty unsurprising really that this was going to probably break apart at some point. Then, yeah, just to wrap up on the headlines, there was an acquisition, and we were just talking before we started recording. Let me try and get this right. I believe, it's Anyscale was acquired by Nscale, and when we were talking earlier, I said, "Is that Scale?" It's like, no, that's not scale.ai. It's different again. We got amazing naming these days. Yeah, what's this one about?

[0:21:42] SF: Yeah. Anyscale, which is known - they were the creators of Ray, which a lot of inference infrastructure and fine-tuning and model infrastructure runs on, a very well-established open-source project. Anyscale was the company that tried to build, or built essentially a managed version of Ray around that. They raised a billion dollars or so in 2022, but I think - and then they just sold to Nscale for $1.65 billion, which is a Neocloud. For those that aren't familiar with Neocloud, Neocloud is essentially what the term is used for companies that offer primarily GPU as a service versus kind of general computing. There's all kinds of these Neocloud companies are now available.

I had actually talked to Anyscale at a variety of different times. I think this was probably a decent outcome for them, because from - I think it's probably hard to really grow that managed Ray as an independent company into a really big company. As part of a, GPU infrastructure company, it's probably a good pairing. I think that it makes a lot of sense. I always confuse Anyscale with Scale.AI and Nscale. I don't know the history of how those company names came together, but generally when the thing that people a lot of times strive for with naming companies is you want

a name that you can say and people can remember. I'm not sure if they hit the mark there, where it's like Nscale, Scale, Anyscale. It's hard to remember who's who in that Venn diagram.

[0:23:14] GV: For sure. I mean, scale.ai, okay, that's a great name to have .ai. You can basically have anything that is your company, and it's like, going to sound good and it's five letters. But then, Anyscale and then Nscale, that's pretty confusing. Then, yeah, just a sidebar piece of news, scale.ai now have a new CEO as well, which is interesting, because that was founded and led by Alexandr Wang, not the fashion designer, if anyone knows that one, but this is Alexandr without an E at the end of the Alexandr. Meta had taken a virtually just under 50% stake. That was the point. Scale, in theory, are still running their own show, but when you've got, I think 49% stake from Meta, you're quite beholden to them. Yeah. In that transaction, Alexandr went to head up the AI side of Meta total. Now Scale have a new CEO, so that'll be interesting to see how that all nets out.

[0:24:12] SF: Yeah, they raised 14 billion from - with a round led by Meta, so pretty significant.


[0:24:18] GV: Yeah. I mean, actually just the new CEO of Scale is actually, yeah, is a former Google Cloud executive called Francis deSouza. Definitely interested to follow along with that one and see how that all nets out.

[0:24:31] SF: Yeah. I saw also Fireworks, who just raised a huge round and made a lot of news. Their new head of engineering just came over from Google, so I think you're, you're starting to see, it's a common pattern, though. You have people who reach executive positions at large companies like Google and they get maybe a little bored with that and then want to go back to building and moving faster.

[0:24:54] GV: Yeah, for sure. Yeah, Fireworks, super interesting. We do have an episode with Fireworks, with one of the co-founders, Benny Chen. Yeah, go check that out. I think that came out around March this year, so well before this fundraising was confirmed. Yeah, they're definitely having a moment. They are the infra for open weight models, and that's becoming increasingly interesting to many companies for many reasons. Yeah, that's actually quite a nice segue into the main topic for today, which is we're calling it the Kimi moment. Kimi, which is an open weight model. The arch company is called Moonshot AI. If you've heard of Moonshot AI, that's Kimi and vice versa. Moonshot AI is a Chinese original company. They do have offices, I

believe, in the Valley and in Singapore and that kind of thing, but very much seen as a Chinese company, which frames a lot of why this is quite interesting.

Yeah. When we think of open weight models, there was this, the DeepSeek moment first, which we can touch on, and then really in the last almost three months, we've just seen this, especially rapid adoption by many companies of Kimi and really forcing companies to think differently about using foundational models from the big players, or the big names rather, and getting some quite interesting and very high quality results out of Kimi. The DeepSeek moment, I think, Sean, how did that kick things off when we think about open weight?

[0:26:20] SF: Yeah. I mean, I think the DeepSeek moment ended up having more market impact in some sense, because I think prior to that, the sense was that you couldn't really get these really powerful models, except from these frontier labs, like Anthropic and Google, OpenAI, and so on. It was kind of shocking when DeepSeek came out, and NVIDIA's market cap cratered as a result of that. Of course, it's certainly bounced back since then. Then when you look at the Kimi launch, which was the largest open weight model ever released, beating some of the top closed models on particular benchmarks, it seemed like nobody really panicked. I think part of that is not because it's less impressive, but it's because we've gotten used to the idea that these Chinese lab open weight models are competitive, and it's less of a shock, essentially. We've been desensitized to it.

I do think that what the market reaction has been more around, and I think this is something we're already coming to, is where this was somewhat on the back of what happened with Mythos and Fable and the reaction to that where people got scared that they could invest in a model and then have potentially a foreign government say like, "You can't use this model anymore," then I think that has increased the interest in diversification of models and in also having an open weight model strategy along with having a closed weight model strategy.

[0:27:50] GV: Yeah. We'll probably touch on the banning of models, because of course, it comes into this as well, if you want to say ironically, because yeah, when you ban, at least pause ban a foundational model, a closed source model if you like, then open weight becomes very interesting, but then a lot of open weight comes from China. We're going to be interested in that. Why the last couple of months has Kimi really just exploded, I think, in interest and popularity? It's moved from, "Oh, it's good enough for professionals" to this actually competes

with GPT and Claude and it's closed that gap in basically weeks, at least from what has been released.

[0:28:28] SF: Yeah. They also released a ton of models on back-to-back. Essentially, they're moving super, super fast.

[0:28:33] GV: Yeah, exactly. I mean, yeah, if you look at the actual lineage, I guess, they've been doing quarterly drops, if you like. July 2025 was K2, and then within 12 months, you've gone through 2.5, 2.6, 2.7 and now K3. That's a incredibly rapid cadence. If you look at those jumps as being as significant as jumps that you might see from any of the foundational players, but definitely not - but not on that cadence effectively.

[0:29:00] SF: Yeah. Yeah. I mean, I think the model distillation practice has really sped up how quickly people are coming up with models and models that are comparable in performance, which was a big part of the conversation around the initial DeepSeek launch. I do think that is part of the - some of the criticism on some of the benchmark results that we've seen from Kimi is that in particular, there was a lot of headlines around their MCP tool calling performance, but that beating Opus, for example.

In particular, that benchmark is relatively new. I don't know if this is more just jealousy in the dialogue, or whether there's some truth to this, but you can essentially bias towards performing really well on certain benchmarks, whether it's the MCP one, or it's the MLLU style benchmarks as well. Then you can have really good benchmark performance, but it might not actually match reality. That's where some of the criticism has been on, even some of the tests of you testing a model against LSTAT and stuff like that and then saying, "Oh, it outperformed lawyers on the LSAT," but then the LSAT's actually not necessarily a good indication of what a lawyer does on a day-to-day basis and stuff. These are some of the nuance.

Some of this is general model criticism, but this is some of the dialogue that I've heard around Kimi in particular is like, are they building to optimize for the metric? Like what we talked about in the Amazon story is whatever the KPI is, people are going to try to optimize for that KPI. You always have to be careful, essentially, what the metric is that you're measuring people against.

[0:30:34] GV: Yeah. On June 12th, K2.7 code was released and this is obviously very much a software engineering tuned model. As you were saying, yeah, that became the benchmark moment, especially through MCP and that was a correct tool invocation is how that's measured. Apparently, scoring in theory 81% and that was versus, say, Opus 4.8 at 76%, which is a pretty - if you believe in these benchmarks, that's a pretty meaningful jump. But then, the distribution piece was interesting. GitHub actually made 2.7 generally available via Copilot, but there was an interesting piece there where for enterprise teams on Copilot business and enterprise, it was actually, this model was off by default, and admins had to explicitly enable it.

GitHub's change log flagged this and said it may be less aligned than other Copilot models, which is interesting, because obviously, they have their own interests being part of the Microsoft OpenAI ecosystem. Obviously, they couldn't miss not providing that to users, given that it's on Azure and they still get inference from it. Yeah, very interesting that there's slightly odd warning came with it that wasn't really clear what that was about.

[0:31:50] SF: Yeah. I believe, it's the first Chinese lab open weight model that's been built into Copilot. I do get having - working with a lot of enterprise customers. They are really careful about which models they use, so I can understand having some level of governance there. Of course, there's a certain biased interest from Microsoft point of view of how do you craft a language around this particular model. There is, I think, that fear with the enterprise. I understand some level of controls there, but obviously, there's also a certain bias that can be put into it. I think one of the things that's interesting about this is Moonshot, as well as this goes for all the Chinese labs, they don't have these top-tier NVIDIA chips available to them, like the US labs use, because of the various chip embargoes.

[0:32:35] GV: We don't think they do, but yeah.


[0:32:37] SF: Well, yeah, yeah. True. Even if they did have some, probably not at the scale.


[0:32:40] GV: Yeah, for sure.


[0:32:41] SF: There's a certain scarcity of resources that they have to deal with, and I think that in some ways, that is forcing them to innovate in a way that maybe the team, the US-based companies, or Western-based companies, where they don't have that scarcity of resource,

aren't as necessarily forced to do that. In some ways, maybe we're creating the thing that we fear, but good news of that is it's probably forcing the other companies, like the Western-based companies to react to this, to lower token prices and maybe also think about how to they keep costs down. It's a little bit like, if you look at Google's beginnings, Google started as a research project at Stanford, and you have students, essentially, didn't have a lot of resources, so they had to be very, very creative about how they scaled Google. Then even that carried over to when Google did initially have some funding.

If that project had started out of a bigger, more established, well-funded company, they probably wouldn't have been ripping apart cheap machines and wiring them using Legos and stuffing them together as to compact the size within the data center as much as possible, and they would've just bought super beefy servers. The downside of that would've been like, a lot of the innovation that we've seen since then in cloud infrastructure came from some of those, forcing yourself to deal with these scarcity resources. My hope with all this is we see a similar thing of innovation driven from the scarcity resources that carry over to the frontier labs in other parts of the world as well.

[0:34:20] GV: Yeah. I mean, it's interesting, just sidebarring into who's even behind the founder, Yang Zhilin, he probably was known as Yang the Genius. There was a very good profile on him in the Financial Times actually, last weekend. He's 34-years-old. He was known at university for being just as much into music and having academic soirees, if you like. Actually, Kimi, when he first released it as a chatbot, it didn't do very well. It was outages and this kind of thing. It's clearly, as you're calling out, has managed to catch up, despite not having the same access, or we don't think couldn't possibly have the same access despite what it still may have.

Back to then the turning point here, July 16th, which is for history, that's then 34 days after 2.7 had dropped. K3 dropped and that was a model with 2.8 trillion parameters, which is, I believe, the largest open weight model released to date. I think what's interesting here is then on the July 27th, the full model weights were actually published. This then leads into this whole provenance piece, which is when companies are starting to be asked like, well you're using AI to generate so much information within your own companies or analyze data, you want to actually know how this was determined, and so on, so forth.

Here we are. We actually have a competitive model with the closed foundational models with the full weights published on Hugging Face, which is massively impressive. To me, that's a huge turning point when you've actually got effectively all the weights right there. This isn't just a model that takes up some tiny little specific task. It's really competing now.

[0:36:07] SF: I think overall, this is whether Moonshot and Kimi win out or not. I think overall, this is good for consumers of these models, because it will force the other companies to essentially react to this and hopefully - Actually, I saw a headline today that OpenAI was reducing some of their token costs by up to 80%. We saw a similar thing after DeepSeek. It drove down token costs as well. Then all the stuff we were talking about earlier of token maxing and companies getting sensitive to how much they're spending, overall, I think it'll be good for consumers of these things. One thing I was thinking about with this story too, and we we've covered a lot of Meta over the last year and their decisions in terms of hiring and what they're trying to do with their AI lab. Meta had such a head start in this open weight model space with the Llama models. It's been a long time since I've heard anything about the Llama models.

[0:37:05] GV: Funny you said that, because, yeah, there was something. I didn't dig into it, but I just saw the headline, which was that Zuckerberg is basically lobbying to ensure that Chinese models don't get banned. To me, that only meant one thing, which is like, well, they have to keep that door open, because Llama's not maybe going where he hoped it was.

[0:37:25] SF: I think his instinct of trying to own the open weight model space was probably right. The execution was bad. I don't know what happened internally there, but they were clearly on a good path, especially with what we're seeing now between the Chinese labs, between what's happening with Fireworks and the other inference providers that have focused on open weight. There's clearly a huge TAM available for open weight to own a big part of the business. I think realistically, especially in the enterprise, most enterprise businesses will probably have a mixture of both open weight investments that they've done and then as well as the closed source models, and that'll probably be the strategy that many, many companies take for some period into the future. I don't know how Meta ended up, and maybe they'll bounce back. But they were onto something, but they haven't been able to execute.

[0:38:15] GV: I think that's really good analysis. It is right strategy, wrong time. Too early effectively, perhaps. Or just too early in the - given they were trying to do it from the US, being

asked to produce too much too soon, and that can distort things. Who knows? Maybe Zuckerberg, or Meta with Llama, they'll adopt more of a Kimi approach to make this succeed. Who knows? Yeah. I mean, just to wrap up on Kimi and this open weight, I wouldn't say resurgence, but yeah, maybe surge. Model provenance is a real thing that's being talked about now when it comes to - it's a business risk, right? You're putting money and basing your company on whether it's for coding, or for other tasks, but you are now having to really decide like it's a procurement question on what are you buying effectively and, and what are the risks with that? Could it get too expensive, or could you move it onto your own infra if you really needed to? Or not your own infra, but exactly via Fireworks, you could still own that end-to-end there. That's interesting.

Pricing, we've talked about that. Pricing is just getting out of control, and it was really a case of when, not if, is this going to stop being possible for Anthropic and OpenAI to charge this way, because it's just not sustainable for many businesses. Yeah, and the speed as well, the speed of iteration has clearly been a huge piece here, where however Moonshot are doing it, they're just iterating like crazy. Onto our favorite parts, Hacker News highlights. Do you want to go first, Sean?

[0:39:44] SF: Sure. This headline really caught my eye and made me laugh, which was the speaking agents, like cavemen save 65% of tokens we test. This was by JetBrains. They benchmarked the caveman skill, which is a Claude code skill that makes it respond in terse caveman speak to save tokens. They claimed 65% token savings, so they essentially put that to the test. It was focused on agentic coding tasks, and they found actually in reality, it was about 8.5% output token savings versus 65%. A big part of that is they were focused on coding tasks, where a lot of it's going to be code. You can't turn that into caveman speak. You got diffs, you got tool calls. The caveman skill might not be best for that, versus some sort of more conversational chat. I just really loved - it reminded me of some Ig Nobel Prize research out there. It's just a -

[0:40:37] GV: That's so good.


[0:40:38] SF: - a ridiculous topic.

[0:40:39] GV: That's, yeah, super funny. It's interesting, because I didn't actually scan ahead, so that is actually a little bit similar to one I picked out, which is called "The Economic Benefits of Refactoring." This is on the fairly popular martinfowler.com website, posted by JavaEE user on Hacker News, so thanks for that. This is not written by Martin Fowler himself, but basically, this was someone else at ThoughtWorks, and it was, can you basically decrease especially the input token cost if you refactor your code, or what are the consequences of that? I mean, the TLDR is yes, if you refactor, then basically, you can dramatically reduce your input tokens.

The theory behind this was that the saving is because the agent has to read less code, but the bit that might not be - that sounds obvious, but it's actually not because there is less code to read, it's actually that the overall code in a certain layer they used to help test this, it stayed constant. The fact that the agent is then able to successfully identify smaller subsets of files to read is the key bit there. Refactoring into smaller chunks, but also making sure that those chunks are very clearly DRI, don't repeat yourself, etc., etc., so that the code that the agent needs to go and grab is smaller. You can't just chunk it up into small chunks and then go, "Well, it's smaller." It then doesn't even know. It has to still get all the small chunks, because it doesn't know what's most important.

The refactoring is what then turns it into having that context of having what is most important, go and find that small chunk, input that small chunk, and the output token and output code was virtually the same in this experiment, so which is super interesting. There was a small side note on that, which was just him saying that, "Claude is actually not good at refactoring, and I can definitely attest to that." But if you can go through the motions of get it to refactor, then this is quite - could be quite a huge saving for anyone who's trying to reduce their input token costs.

[0:42:39] SF: Yeah. Basically, bottom line, good design. Leads to also, optimized token cost.


[0:42:47] GV: Funny that, yeah. If we just go back to how we used to design software then, yeah.

[0:42:50] SF: I mean, it's the same like, if you think about the human cost. If you have poorly designed software, then there's going to be more human cost each time someone unfamiliar with it needs to ramp up and make some change.

[0:43:01] GV: Exactly.


[0:43:02] SF: Then you have another one on CodePen.


[0:43:04] GV: My second one was, yeah, just thanks to user Robin Riella posting, in fact, that CodePen 2.0 has come out. This takes me back. That's why I was quite interested in this. I don't know if anyone else out there - This makes it sound terrible. If you like CodePen, this is - sorry about this. It's just I haven't been doing a lot of front-end work for a long time. Used to love CodePen. Used to put up all sorts of things on there, and a really useful tool as well. I did a little coding class. I went back to my old high school a long time ago and did a coding class, and CodePen was amazing, because you could just get people spun up in a browser writing front- end code and just see it do its thing straight there. They didn't have to build with - You didn't have to get them spun up with some repo, or anything. That was really helpful.

Yeah. I mean, CodePen must have come out probably 15 years ago, or something like that. CodePen 2.0, I'll just quickly run through a couple of things that they call out that they have files and folders now. It's interesting. It's almost becoming a bit IDE-esque. But files and folders, and then they have - They do build steps anyway, but they've now added this concept of blocks, which looks quite nice, where you can see exactly which bits are in this build process, or the linting process, so that's fun. Yeah, real-time collaboration as well, so finally multiplayer on CodePen. For anyone that uses CodePen a lot, I'm sure that's quite a huge uplift, so. Or if you haven't checked it out, yeah, still a fun place to go and just experiment with front-endy stuff, so.

Yeah, and then just a special one. I guess, not technically through Hacker News, but thanks to Ilya Reshetnikov for tweeting at me and Sean. We do often cover these Doom, can you run Doom on something? Thanks to Ilya, he pointed out that there's another one called DoomQL, which is basically using an SQL query is the frame buffer, which is - I think that definitely rivals Doom on TypeScript types. Yeah, I think that's probably the closest I can think of that it rivals, so thank you, Ilya, for shouting that one out to us. That was very, very fun to read through.

Yeah, that's it for another SED News. Have we got any looking ahead predictions, Sean, which we usually get wrong?

[0:45:12] SF: Yeah. I mean, I think the safe predictions here would be that we're going to see more headlines on the token maxing issue as companies start to adjust. I think we'll see more, also, conversations around open weight versus closed model, diversification of models. I think those are both going to be big topics of conversation through to the end of the year.

[0:45:32] GV: Yeah, for sure. I will then say, well, because we just talked about iteration and Kimi, let's assume that by this time next month we're already on Kimi 3.2, or if - I'll just push the boat out. Kimi 3.5. 3.5 by end of August. Let's see if that lands. Thanks everyone for tuning in as always, and we'll be back next month with another SED News.

[0:45:55] SF: Thanks, everyone. Cheers.


[END]