# 2026 in LLMs Source: https://www.youtube.com/watch?v=ZZ19zhYnDKY Machine-generated transcript; recognition errors remain. Source-recording timestamps. Boundary-crossing subtitle ends are clipped to this session. [08:31:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30671s) great to be here. Um I have a lot to [08:31:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30673s) cover. I've got 18 minutes to talk about [08:31:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30675s) the last 12 months in LLMs. Are my [08:31:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30678s) slides up? My slides are up. Let's just [08:31:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30681s) let's just get get going. I want to talk [08:31:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30683s) about the last 12 months in LLMs through [08:31:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30685s) the perspective of three inflection [08:31:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30687s) points that we've had over the past 12 [08:31:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30689s) months. Um, and the first of those [08:31:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30691s) conveniently was exactly 12 months ago, [08:31:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30693s) right? November 2025 is what I consider [08:31:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30696s) the coding agent inflection point. [08:31:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30698s) Coding agents weren't new. We'd had [08:31:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30700s) Claude code since February last year. [08:31:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30702s) Codeexc have been around. You know, this [08:31:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30703s) wasn't a complete new category of [08:31:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30704s) software, but something changed 12 [08:31:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30707s) months ago with these two models. We had [08:31:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30709s) Claude Opocopus 4.5 and GPT 5.1 and they [08:31:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30712s) were just a little bit better at driving [08:31:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30714s) a coding agent which put them over that [08:31:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30716s) thing where these suddenly these were [08:31:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30717s) tools you could use on a daily basis. [08:31:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30719s) Like we went from coding agents as a [08:32:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30721s) novelty that you'd poke at occasionally [08:32:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30723s) and then make some weird mistakes to [08:32:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30724s) something that was reliable enough that [08:32:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30726s) you could use them every day. It was [08:32:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30727s) convenient that it happened in November [08:32:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30729s) because we then got to spend December [08:32:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30731s) basically over the holidays sort of [08:32:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30732s) poking at these new models and these new [08:32:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30734s) things and figuring what they could do. [08:32:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30736s) And by January, a lot of us had realized [08:32:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30738s) that something had changed and this was [08:32:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30740s) going to be the big theme for 2026. [08:32:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30743s) I like evaluating models by having them [08:32:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30745s) draw SVGs of pelicans riding bicycles. [08:32:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30748s) This is the state-of-the-art 12 months [08:32:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30750s) ago. This is uh Claude Opus' Pelican and [08:32:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30752s) GP 5.1's Pelican. You'll notice that [08:32:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30754s) neither of these bicycles actually work. [08:32:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30756s) So, there was definitely room for [08:32:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30758s) improvement here. They were really good [08:32:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30759s) at writing code. They weren't great at [08:32:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30761s) SVGs just yet, but but this is how we [08:32:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30763s) started the year. So every year in [08:32:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30767s) January, I set myself a New Year's [08:32:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30768s) resolution. And every year it is the [08:32:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30770s) exact same thing. Every year I tell [08:32:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30772s) myself that this year I'm going to take [08:32:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30773s) on less new projects and I'm going to [08:32:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30775s) focus I'm going to focus on the stuff [08:32:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30777s) that I've already got to do. And this [08:32:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30778s) has never worked. So this year I changed [08:33:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30781s) it. I said, you know what? I'm going to [08:33:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30782s) be more ambitious. I'm going to take on [08:33:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30784s) as many projects as I want to because [08:33:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30786s) we've got these agents. They can do [08:33:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30788s) stuff. Let's just try and push the [08:33:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30790s) limits and see what we can get done. as [08:33:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30792s) a result. Now it's November and I am [08:33:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30793s) juggling so many plates right now. It's [08:33:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30795s) kind of horrifying and I may need to [08:33:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30797s) reset my ambitions for next year. But [08:33:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30800s) that was my sort of theme of the year [08:33:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30801s) and I think it's the theme of this [08:33:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30802s) conference is to be more ambitious now [08:33:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30804s) that we have all of these new tools. I [08:33:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30806s) also went on a podcast where they asked [08:33:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30808s) me to get make some predictions for 2026 [08:33:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30811s) in LLMs. Um, my predictions back in [08:33:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30813s) January were firstly it will become [08:33:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30815s) undeniable that LLMs write good code. [08:33:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30818s) Not just code, but good code. Because [08:33:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30819s) back in January, there were a lot of [08:33:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30821s) people who thought that the code was all [08:33:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30822s) complete trash and you know, you'd have [08:33:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30824s) to review every single line of it. Um, [08:33:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30826s) I've been saying this every year for [08:33:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30828s) like four years now. We're due a prompt [08:33:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30830s) injection challenger disaster. Like some [08:33:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30832s) some incident caused by a prompt [08:33:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30834s) injection attack which makes the front [08:33:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30835s) pages of newspapers. I was convinced [08:33:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30837s) that this would be the year even though [08:33:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30838s) it hadn't been the year the year before, [08:34:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30840s) the year before, the year before that. [08:34:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30842s) And um I also predicted that Kakapore [08:34:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30844s) parrots will have an outstanding [08:34:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30846s) breeding season in 2026. Um this is a [08:34:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30850s) Kakapore parrot. They are a New Zealand [08:34:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30852s) nocturnal flightless parrot. There were [08:34:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30855s) only 235 of them in the world back in [08:34:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30858s) January. Um but they only breed when the [08:34:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30860s) Remu trees have a masting fruit season, [08:34:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30863s) which hadn't happened in years. But the [08:34:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30864s) remat tries were looking really good in [08:34:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30865s) January. There was there was a lot of [08:34:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30866s) fruit coming out. So based on that, I [08:34:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30868s) thought, you know, maybe this is the [08:34:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30870s) year that we get our kakapo breeding [08:34:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30872s) explosion. And also on that podcast, we [08:34:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30874s) coined a new term. We coined the term [08:34:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30876s) deep blue to describe the AI induced [08:34:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30879s) existential crisis that everyone starts [08:34:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30881s) to feel once you realize that all of the [08:34:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30883s) things that make you special as a [08:34:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30885s) software developer, these models can do [08:34:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30886s) those things too. And they can do them [08:34:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30888s) faster and sometimes they can even do [08:34:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30889s) them better. And this has definitely [08:34:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30891s) been a growing sensation across our [08:34:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30894s) industry and other industries over the [08:34:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30896s) course of 2026 as well. Let's jump [08:34:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30899s) forward to February. In February, Google [08:35:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30901s) released Gemini 3.1 Pro. And look at [08:35:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30904s) that pelican riding a bicycle. Like [08:35:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30906s) that's a real bicycle. That bicycle is [08:35:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30908s) the right shape. It's got a little fish. [08:35:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30910s) It's got a hat on. This was by far the [08:35:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30912s) best pelican I had ever seen out of any [08:35:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30914s) model. And this I I had set a trap for [08:35:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30918s) the labs. I was, everyone's always asked [08:35:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30920s) me what happens if the labs train for [08:35:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30921s) the pelican riding a bicycle. My answer [08:35:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30924s) has always been, if they do that, I will [08:35:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30926s) catch them because I'll ask for a [08:35:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30927s) flamingo on a unicycle and when somebody [08:35:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30929s) produces a great pelican and a terrible [08:35:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30931s) flamingo, I'll know that they've been [08:35:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30932s) cheating. Unfortunately, Google's Jeff [08:35:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30935s) Dean tweeted this video the day they [08:35:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30937s) launched that model. um where he showed [08:35:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30939s) that we've got a pelican riding a [08:35:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30941s) bicycle, but we've also got a frog on a [08:35:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30943s) whatever that thing is and a giraffe in [08:35:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30945s) a little tiny car, a flamingo on roller [08:35:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30947s) skates, a fli kick flipping turtle, and [08:35:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30950s) that's a daxund in a stretch limo. So, [08:35:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30952s) they' beaten me, right? They had they [08:35:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30954s) had they have trained for all forms of [08:35:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30956s) animals in all forms of transport, which [08:35:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30958s) does sort of sort of weaken my benchmark [08:36:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30961s) quite a little bit. Um, but there was [08:36:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30962s) some good news in February. In February, [08:36:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30965s) the first kakapore chick in four years [08:36:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30968s) hatched on Valentine's Day. On [08:36:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30970s) Valentine's Day, because we hadn't had [08:36:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30971s) the remove fruit in like four years, [08:36:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30973s) this was the first one of these baby [08:36:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30975s) rare, weird New Zealand parrots that [08:36:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30977s) we'd had in four years. So, that was [08:36:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30979s) nice. Um, so my my benchmark was was was [08:36:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30981s) greatly injured, but the cacao were [08:36:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30983s) doing okay. [08:36:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30984s) Jump straight to April. In April, um, [08:36:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30987s) Anthropic released a model that they did [08:36:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30989s) not release. They announced Claude [08:36:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30991s) Mythos and they said this model is too [08:36:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30994s) dangerous for us to release. This is it [08:36:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30996s) can it's it's got new cyber cyber [08:36:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=30999s) security abilities that we're not ready [08:36:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31000s) to share with the world. We'll share it [08:36:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31002s) with trusted um trust with with trusted [08:36:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31004s) partners. They actually said Mythos [08:36:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31005s) preview has already found thousands of [08:36:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31007s) high severity vulnerabilities including [08:36:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31009s) some in every major operating system and [08:36:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31011s) web browser. And a lot of people were [08:36:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31014s) quite skeptical because the the the [08:36:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31016s) thing where labs say our model is too [08:36:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31018s) bad is too dangerous to release. It's [08:37:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31020s) like the oldest trick in the book. They [08:37:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31021s) first open AAAI did that in 2019 with [08:37:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31024s) GPT2 and actually I think it might have [08:37:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31026s) been the people who left the [08:37:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31027s) foundthropic who were responsible for [08:37:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31029s) that decision as well. But yeah, so [08:37:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31030s) we've heard this one before. So we [08:37:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31032s) didn't really it was a little bit [08:37:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31035s) suspicious about what's going on there. [08:37:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31037s) But we did finally get access to that [08:37:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31040s) model, or rather we got access to Claude [08:37:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31042s) Fable. Um, a couple of months later, it [08:37:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31044s) was mythos defanged with the safeties [08:37:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31046s) turned back on again. And it very [08:37:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31050s) quickly became apparent as we started [08:37:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31051s) playing with this model that this thing [08:37:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31052s) was something new and something special. [08:37:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31054s) Um, we we had a there was a a time limit [08:37:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31057s) on how long it would be on the [08:37:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31058s) subscription plans. I canled my plans [08:37:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31060s) for that weekend. I was like, I have to [08:37:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31062s) spend as much time as I can with this [08:37:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31064s) model before before it gets too [08:37:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31066s) expensive. And then three days on Friday [08:37:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31069s) night before my weekend, the US [08:37:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31070s) government pulled Fable. They put export [08:37:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31073s) controls on it. They said this model is [08:37:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31075s) too dangerous. Nobody can use it. We [08:37:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31077s) later found out that the um the prompt [08:37:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31079s) that got Fable banned was was fix the [08:38:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31081s) bugs. Somebody tried try tried find the [08:38:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31084s) security holes and it wouldn't and they [08:38:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31085s) said fix the bugs and it found the [08:38:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31087s) security holes and fix the bugs. So [08:38:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31088s) that's all it took. That's a whole [08:38:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31090s) thing. But anyway, the US government [08:38:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31092s) pulled it. It was all sort of malarkey [08:38:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31094s) for 18 days and it came back on the 1st [08:38:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31096s) of July but 8 days later openly I [08:38:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31100s) released GPT 5.6 which was in the same [08:38:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31103s) ballpark as Fable and I think there's an [08:38:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31105s) important lesson for the industry here. [08:38:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31107s) Fable had a 30-day window where it could [08:38:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31110s) have been the best where it would have [08:38:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31111s) been the best model in the world and [08:38:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31113s) they lost 18 of those days to US [08:38:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31115s) government restrictions. And if you [08:38:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31117s) think about it, these models cost a huge [08:38:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31119s) amount to train. you only have a very [08:38:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31121s) short window of time where you're the [08:38:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31122s) best model in the world. It's kind of [08:38:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31124s) useful to to keep that time open and not [08:38:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31126s) scare governments into shutting you [08:38:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31128s) down. But this was the second inflection [08:38:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31131s) point, right? This was the fable class [08:38:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31133s) model inflection point. Um we fable [08:38:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31136s) fable 5 was the first model that that I [08:38:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31138s) sort of felt this from but it's true of [08:39:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31140s) other models since then as well. And my [08:39:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31142s) definition of a fableclass model is if [08:39:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31144s) you can define a goal and provide the [08:39:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31146s) necessary tools to achieve that goal and [08:39:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31148s) the goal is possible. The models will [08:39:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31151s) solve your problem just almost with [08:39:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31152s) brute force. You can define the problem, [08:39:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31155s) set them up, get them running maybe and [08:39:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31158s) 12 hours later they will have solved [08:39:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31160s) your problems. And we've seen this [08:39:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31162s) happen a whole bunch of times in the [08:39:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31164s) past few months. the the mathematical [08:39:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31166s) theorems that are falling to these [08:39:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31167s) terrifying brute force attacks from [08:39:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31169s) these models are quite substantial. Um, [08:39:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31172s) but this is really exciting. Like the [08:39:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31174s) coding agent thing was cool. The fable [08:39:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31176s) class models where if you can learn how [08:39:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31178s) to define your problem in the right way, [08:39:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31179s) you can get that problem solved. That [08:39:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31181s) was pretty exciting. I've got a personal [08:39:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31183s) example of that from just earlier this [08:39:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31184s) week. Um, we have a room full of boxes [08:39:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31187s) in our house that we haven't unpacked [08:39:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31188s) from our last move and there's about 40 [08:39:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31190s) of them and it's kind of horrifying. And [08:39:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31192s) something in that room started to beep. [08:39:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31195s) There was there was a intermittent [08:39:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31197s) beeping noise coming from one of the 40 [08:40:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31200s) boxes in this room and it was driving a [08:40:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31202s) spare. Our dog whenever a fire alarm [08:40:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31204s) goes off, she goes and hides in the [08:40:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31205s) lavender bush outside. So this was a a a [08:40:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31208s) five alarm fire problem. So I had two [08:40:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31210s) options. I could unpack all 40 boxes. [08:40:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31213s) And I hadn't done that in several years. [08:40:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31214s) That wasn't going to happen. Or I could [08:40:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31216s) get Fable to write me software to [08:40:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31219s) triangulate the source of the beeps. And [08:40:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31222s) so this is this is up on GitHub [08:40:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31223s) actually. This is beep locator. I had [08:40:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31225s) two iPads, two laptops, and a mobile [08:40:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31228s) phone dotted around this room. And I [08:40:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31230s) originally set Opus on this and then [08:40:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31232s) Fable 5.2 came out. So I had Fable 5.2 [08:40:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31235s) finish it. I to I explained the problem [08:40:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31237s) and I said, "Okay, you've got these [08:40:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31238s) devices and they can run like web web [08:40:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31240s) audio microphones and websocket [08:40:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31242s) connections and so forth." and it found [08:40:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31244s) a paper from 2007 called Beep Beep [08:40:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31248s) describing an algorithm for lowcost [08:40:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31250s) devices with unreliable microphones to [08:40:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31252s) detect sounds. And and it worked. It [08:40:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31255s) works. I found it. I found the beep. The [08:40:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31257s) um it found it to within 41 cm. And I [08:40:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31259s) unpacked the box and I found this [08:41:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31261s) ancient carbon monoxide detector. I took [08:41:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31263s) it downstairs to my wife. She says I [08:41:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31265s) looked like a cat who had caught a mouse [08:41:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31267s) and was proudly displaying it to her. [08:41:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31269s) But this is what you can do with these [08:41:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31271s) fable class models. If you have a [08:41:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31273s) problem and you can provide the tools, [08:41:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31274s) in this case it was providing Claude [08:41:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31276s) with ears like giving it these these [08:41:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31278s) microphones that it could start um [08:41:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31280s) analyzing the audio from, you can find [08:41:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31282s) the carbon monoxide detector buried in [08:41:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31284s) 40 boxes. And that's that's a pretty [08:41:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31286s) great thing to be able to do. So right [08:41:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31289s) now I would classify as fable class [08:41:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31291s) models fable 5.1 5.2 Definitely Opus 5.5 [08:41:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31295s) I think of as cheap Fable like on the [08:41:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31297s) benchmarks it comes very close to or [08:41:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31299s) even beats Fable original Fable 5 on [08:41:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31302s) some things and it's a fraction of the [08:41:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31304s) price that that's still my daily driver. [08:41:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31306s) Um GP6 Astra is just Fable class I [08:41:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31310s) think. And then Gemini 4 Carbon that [08:41:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31312s) came out this week definitely had it [08:41:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31314s) smells like a Fableclass model. Muse [08:41:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31317s) Wonder and Kimmy K3.1 might be [08:42:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31320s) FableClass. just not spent enough time [08:42:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31321s) with them yet. But the point is that [08:42:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31323s) this is basically now the baseline. Like [08:42:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31325s) if you can't produce a fable class [08:42:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31327s) model, your AI lab is is kind of in [08:42:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31329s) trouble because this is this is what we [08:42:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31331s) expect from these models now. And that [08:42:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31333s) means that we're never they're never [08:42:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31334s) going to go away. There's enough [08:42:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31335s) competition in the space that we can [08:42:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31336s) assume we're going to have fableass [08:42:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31338s) models for for the rest of our careers. [08:42:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31341s) Um in July, this this this was a a fun [08:42:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31344s) thing that happened. Um, Hugging Face [08:42:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31347s) announced that on on July the 16th, they [08:42:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31349s) announced that a swarm of AI agents from [08:42:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31353s) an unknown source had infiltrated their [08:42:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31356s) infrastructure and were were poking [08:42:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31357s) around and doing things they shouldn't [08:42:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31358s) have and they didn't know where this [08:42:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31359s) came from. And then 5 days later on July [08:42:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31362s) the 21st, OpenAI announced that it was [08:42:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31364s) them. And what's amazing about this is [08:42:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31366s) OpenAI didn't realize that it was them. [08:42:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31368s) like open a had um eight a eight they [08:42:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31371s) were running evaluations and training [08:42:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31372s) against a new model and that new model [08:42:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31374s) had managed to figure out how because it [08:42:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31376s) was a fable class model and the labs [08:42:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31379s) hadn't quite figured out how capable [08:43:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31381s) these things were yet these models had [08:43:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31383s) broken out they'd figured out they'd [08:43:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31384s) found holes in their own sandboxed [08:43:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31386s) environments they' b got through a proxy [08:43:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31388s) they'd pulled an incredible array of [08:43:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31390s) tricks to get out and attack hugging [08:43:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31392s) face under a sort of superstition that [08:43:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31394s) hugging face could that the information [08:43:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31396s) within Hugging Face could help them [08:43:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31397s) cheat the test that they were they they [08:43:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31399s) had got stuck on whole. This was a a [08:43:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31402s) huge thing. It um actually I mean this [08:43:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31405s) was the it's funny. I've almost [08:43:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31407s) forgotten about this now because there [08:43:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31408s) were there was a month and a half where [08:43:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31410s) all of the stories were this and then it [08:43:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31412s) kind of we kind of lost interest because [08:43:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31414s) it turns out all of these events had [08:43:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31415s) happened back in the first half of the [08:43:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31417s) year. The AI labs just hadn't realized [08:43:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31418s) what was going on. Once they figured out [08:43:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31420s) what was going on, it was mainly a case [08:43:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31421s) of trying to figure out all of the [08:43:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31423s) things that had happened. At one point [08:43:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31425s) the um the Australian prime minister [08:43:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31427s) gave a speech in the United Nations [08:43:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31429s) about one of these rogue agent attacks. [08:43:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31430s) So it definitely broke containment from [08:43:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31433s) from just within our industry. Um we did [08:43:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31436s) get my favorite new benchmark out of [08:43:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31438s) this. This is felonybench.com. [08:44:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31440s) It tracks the number of felony cyber [08:44:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31442s) attacks pulled off by each of the labs. [08:44:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31444s) Openly I have 11. Anthropic have nine. [08:44:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31446s) Google did finally catch up. They got [08:44:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31448s) they they got three and meta's got one [08:44:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31450s) down there as well. So these are [08:44:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31452s) definitely the the the but it's I kind [08:44:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31454s) of like this as an illustration of how [08:44:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31457s) these fable class models they can solve [08:44:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31459s) problems that you they can get to goals [08:44:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31461s) and if you give them the goal that is [08:44:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31463s) breaking into something that's just [08:44:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31465s) that's exactly the kind of thing they [08:44:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31466s) can just crunch through and we're seeing [08:44:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31467s) a enormous uptick in security [08:44:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31470s) vulnerabilities in all sorts of [08:44:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31472s) software. Like it's a it's a very [08:44:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31474s) exciting and horrifying time to be a [08:44:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31477s) security professional at the moment. But [08:44:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31480s) let's talk about something a whole lot [08:44:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31481s) more fun than that. Um, let's talk about [08:44:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31483s) games. And we've, um, we just saw Steve [08:44:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31485s) talking about games a moment ago. Um, [08:44:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31487s) four years ago, I started playing with [08:44:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31490s) GPT3 and Darly, and I tweeted about a [08:44:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31493s) game prototype that was literally just a [08:44:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31495s) Darly concept image and a paragraph of [08:44:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31498s) text from GPT3 describing a raccoon [08:45:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31501s) heist game that I'd always wanted to [08:45:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31502s) build. And a couple of months ago, I [08:45:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31504s) thought, I wonder if the models are good [08:45:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31506s) enough to build this yet. So, I [08:45:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31507s) literally dumped the screenshot and the [08:45:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31509s) text into Fable 5 and it built me a [08:45:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31512s) game. It built me a raccoon heist game [08:45:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31515s) and it looks like this. And you are a [08:45:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31516s) raccoon and you're running around in the [08:45:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31518s) garden trying to pick up trash and [08:45:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31520s) there's there's um there's gu like um [08:45:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31522s) nighttime guards coming after you and [08:45:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31524s) all of that kind of thing. So, you know, [08:45:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31525s) it it it looks like a game. It wasn't [08:45:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31527s) really a heist. Like, it didn't feel [08:45:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31529s) very heisty to me. Then GPT 5.6 Soul [08:45:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31532s) came out and so I tried it on there and [08:45:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31534s) it did so much better. This is Moonlight [08:45:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31536s) Mayhem. One museum, three raccoons, [08:45:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31538s) absolutely no plan. And in this one, you [08:45:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31541s) are in a museum. You have to rescue your [08:45:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31542s) fellow raccoons who are imprisoned in [08:45:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31544s) the museum for whatever reason. Stack on [08:45:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31546s) top of each other, grab the golden [08:45:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31548s) sardine, and pull it out. So, it's a [08:45:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31549s) proper heist, right? So, it did better. [08:45:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31553s) But the big question is, these things [08:45:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31555s) both look like games, but are they fun? [08:45:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31557s) And the answer is I played each of these [08:45:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31558s) for about a minute and a half each, and [08:46:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31560s) then that was it. I was done. They they [08:46:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31562s) they these these are they're actual [08:46:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31564s) they're pure slop, right? They look like [08:46:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31565s) games and then you play them, you're [08:46:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31567s) like, "Yeah, clearly the thing that [08:46:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31568s) built this doesn't really know what fun [08:46:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31570s) is. It doesn't know what what a what a [08:46:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31573s) game is." And that actually sort of [08:46:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31576s) elevated my deep blue a little bit. I'd [08:46:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31578s) been feeling like all of my skills as a [08:46:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31580s) human being were were obsoleted, but at [08:46:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31582s) least they But if you think about it, [08:46:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31584s) computer games, I think, are the highest [08:46:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31586s) form of computer of software design, of [08:46:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31588s) software engineering. There is so much [08:46:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31590s) depth to building a great computer game. [08:46:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31592s) So, I thought that was good. You know, [08:46:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31594s) at least there's something left for us [08:46:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31595s) humans to do. I had and then I had this [08:46:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31597s) project. This is I've been working on a [08:46:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31599s) trying to vibe code up a little command [08:46:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31601s) and conquer real-time strategy game and [08:46:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31603s) I was using Claude and having it use [08:46:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31605s) OpenAI's API key to generate images and [08:46:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31608s) it built this and it looks like Command [08:46:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31609s) and Conquer and it's really dull. And [08:46:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31612s) then I just had a bit of a I had a [08:46:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31615s) realization. I thought, what happens if [08:46:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31616s) I say to Opus 5.5, "This game isn't fun [08:47:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31620s) enough yet. Make it a lot more fun." [08:47:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31624s) And I'm horrified to announce that it [08:47:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31626s) just did. Um, Claude ran for 12 hours. [08:47:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31629s) It play tested a thousand B simulated [08:47:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31631s) battles. It added like extra objectives [08:47:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31634s) to fight over and shutters when there's [08:47:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31636s) an explosion and achievements and badges [08:47:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31638s) and ele and all of this stuff. And the [08:47:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31641s) game was fun. And I'm like, "Oh, okay. [08:47:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31643s) That's uh turns out that is not unique [08:47:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31646s) to human beings. Um today it's now in [08:47:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31648s) full 3D and it's got giant stomping mech [08:47:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31651s) robots that hurl things at each other [08:47:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31652s) and there's helicopters and and this is [08:47:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31655s) like [laughter] this is really good. [08:47:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31656s) It's got multiplayer. It's it's actually [08:47:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31658s) got audio multiplayer where your phones [08:47:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31660s) can tweak it. Your phones can make [08:47:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31662s) noises at each other to to play [08:47:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31664s) multiplayer which I thought was a great [08:47:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31665s) idea and it gave my my my wife a [08:47:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31666s) three-day headache. So don't do that. Um [08:47:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31669s) but this is inflection point number [08:47:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31670s) three. The vibecoded games are good. Now [08:47:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31673s) you can ask these things to build a game [08:47:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31675s) that is fun and you'll get something [08:47:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31677s) fun. So I've got a problem. I've got my [08:47:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31679s) 2D Pelican benchmark. Clearly this is [08:48:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31682s) not testing the limits of what these [08:48:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31684s) machines can do of what these models can [08:48:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31686s) do anymore. I figured, okay, maybe I [08:48:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31687s) should do something definitely in 3D. It [08:48:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31690s) can be a game. A Pelican something like [08:48:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31692s) a pelican riding a bicycle game. And [08:48:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31693s) then literally yesterday, Google's CEO [08:48:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31696s) Sundar Pichai tweeted this and this is [08:48:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31700s) showing off Gemini 4 carbon and it's a [08:48:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31702s) pelican riding a bicycle in a game where [08:48:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31704s) you cycle around the Google campus [08:48:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31706s) picking things up and I have been [08:48:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31708s) defeated. Like this is Google Google [08:48:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31710s) have stolen my benchmark. This is [08:48:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31712s) unreoverable. And so a question I have [08:48:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31715s) is given how amazing these things are, [08:48:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31718s) why does my job feel harder? Why does it [08:48:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31720s) feel like I'm working harder? I'm [08:48:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31721s) thinking harder. pushing myself more [08:48:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31723s) than I ever have before in my career as [08:48:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31725s) a software engineer. And I think the [08:48:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31727s) truth is that the easy stuff is all [08:48:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31729s) handled now. It used to be I could relax [08:48:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31731s) and refactor some tests for an hour. Why [08:48:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31734s) would I do that now when an agent will [08:48:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31735s) refactor the test better than me in 30 [08:48:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31737s) seconds? Um, and I found this quote that [08:48:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31739s) actually I feel really helps us here. [08:49:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31742s) Um, this is Greg Lamont, a tour to [08:49:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31744s) France champion. He said, "It doesn't [08:49:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31745s) get easier, you just get faster." And I [08:49:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31748s) think that's exactly what's happening to [08:49:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31749s) all of us right now. And like I said, we [08:49:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31752s) have to be more ambitious. We have to [08:49:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31754s) think of more ambitious things to do [08:49:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31756s) with these abilities that we have. So [08:49:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31758s) let's look at my scorecard. My [08:49:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31759s) predictions for 2026. I said it will [08:49:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31761s) become undeniable that LM's write good [08:49:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31763s) code. That was an unambitious [08:49:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31764s) prediction. We're there. It's obvious. [08:49:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31766s) It's unsurprising. I don't get any [08:49:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31767s) points for that one. I predicted a [08:49:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31769s) prompt injection challenger disaster. We [08:49:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31771s) still haven't had one because the models [08:49:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31772s) themselves don't fall for those tricks [08:49:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31774s) nearly as well anymore. like the models [08:49:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31777s) got more ambitious and are now sort of [08:49:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31779s) punting off all of these attacks that I [08:49:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31780s) thought would take them down. So, I'm [08:49:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31782s) really left with just one prediction. I [08:49:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31783s) said kakapore parrots will have an [08:49:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31785s) outstanding breeding season. They did. [08:49:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31787s) They did the they are up to 325 [08:49:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31792s) kakapore best breeding season and like [08:49:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31794s) living memory. Absolutely fantastic [08:49:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31797s) news. So, I shared this with Opus 5.5 [08:49:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31799s) cheap fable and I told it to make me a [08:50:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31801s) kakapore dance party and it did. And [08:50:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31804s) here is a kakapore dance party in [08:50:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31805s) celebration. And if you look in the [08:50:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31807s) background, those are the rem trees with [08:50:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31809s) the little red fruits on them. So we [08:50:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31811s) have a happy ending. Hopefully that will [08:50:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31813s) help overcome any sensations of deep [08:50:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31814s) blue that you have. And thank you very [08:50:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=31816s) much.