SPEAKER_00 0:00

I'm Mike Belcito, and this is Now Shipping, the weekly AI news show for product people. Every week I'll be bringing you three AI news stories that matter to you, the people actually building products for a living. Not the noise, not the hype, just the stuff you need to keep up with. Brought to you by the team at Mind the Product. This is Now Shipping. And we're gonna be bringing the heat this week. Some very hot news. So hot you'll literally see me dripping with sweat here because it's 100 degrees around Matt. So it's pretty darn hot in this little nook that I work from. But enough of that. Let's get right to the preview of what's happening this week. First up, Figma just put code directly on the design canvas. That handoff step that has defined how the product teams have worked for the past decade, it might be gone. We'll talk about what that means for your workflow. Also, Anthropic just made Claude Sonnet 5 the default model for every free and pro user on the platform. It performs near their flagship level at a fraction of the price. And that's not all that makes this model special. And finally, AI regulation just arrived from two directions at once. The Use AI Act has cleared its final hurdle, and Connecticut passed a law requiring employers to disclose when AI is used in hiring decisions. So if you're building products that touch employment, recruiting, performance management, the compliance clock is now ticking. Those are the three stories. Let's get into it. All right, story number one, Figma. First, you know, let me ask you something. What's the most painful part of the design to development workflow? Now, if you've been in product for more than the past year, maybe you would say it's the handoff. The moment the designer finishes something in Figma, then the engineer has to interpret it, rebuild it, ask 15 clarifying questions, figure out why this spacing looks different in production than it did in the mock-up. That handoff step, Figma is coming after it. Config just took place. That's Figma's big annual conference. It's their moment in the spotlight where they could show where their product is going. And this year was at the Moscone Center. Um, there's a lot of excitement about it. I heard it was a sold-out show. And the headline announcement, well, one of them, was about something called code layers. Now, code layers, I'll explain what it actually does. It brings executable code directly onto the Figma canvas. Now, not as a reference, not as some sort of static screenshot, but actually as live running code. So you can clone a repository, you can pull it into your Figma file, and you can see it running alongside your design layers. Designers and engineers, now they're looking at the same thing, they're in the same canvas all at the same time. So if you've spent any time in a product sprint trying to reconcile what the designer made versus what actually shipped, and you know, why is it any different? You already know why this is significant. But Figma didn't stop there. They also announced something called AI skills. Now, skills let your team package your own workflows, your conventions, your processes, your preferences into reusable agent instruction. So if your design team has a very specific way that they like to do component handoffs or a process for how specs get structured, you can turn that into a skill, the Figman AI agent. It'll run on demand, and you can connect it to external tools like Notion, GitHub, Slack, uh, products from the Atlassian suite. So the agent actually has context about what your team's working on. So think about that. I mean, now your design tool just became an agent platform, just like all the other agent platforms you're using now, right? Um, why is this all important for product people? Well, a few things. First, for years, product and design teams have built workflows around the idea that code lives in one place and then design lives in another, and someone has to sort of translate between those two places. Figma's starting to close that gap, which means the way you plan sprints, the way you run design reviews, the way you think about what done actually looks like. All of that now needs to be rethought, right? The teams that update their process now, because of all these changes, you're gonna move a lot faster than the ones that are gonna wait and see how this all plays out. So could be worth starting the experimentation with right now. Also, I'd say the fear for any design tool right now is that AI coding agents route around the canvas entirely, right? Like engineers just sort of prompt their way to a working component and maybe they never need to open Figma. Well, code layers is Figma's answer to that. If engineers can work inside of Figma, maybe they don't have a reason to bypass Figma. Maybe they'll actually do it right within the platform. So when you see a feature like this, I do think it's worth asking like, is this about helping the entire product team, designers, developers, engineers, or is this a way for Figma to stay relevant in a world where code writes itself? And what the answer is, I don't know. Probably both, really. I mean, I do think this could be a really big deal for product teams, but absolutely, this is one way for Figma to keep itself relevant. Um, finally, this AI skills announcement. I think it might be getting less attention than it probably should. I mean, the ability to encode your team's actual conventions into a reusable agent and then connect that agent to GitHub, Notion, Slack, you know, the platforms you're using. I think that's a big deal. I mean, most teams spend an enormous amount of time sort of re-explaining how they work to every new tool they adopt. Figma is saying, tell us once, we'll remember, if your processes are well defined, it turns Figma into something that actually runs them. If they're messy, well, now you have a reason to clean up those processes. So, anyway, that's story one. Let's move on to story two. All right, story number two is about Anthropic's brand new model, Sonnet 5. And yeah, I know we've done multiple episodes about Anthropic's new model releases, and we haven't even had 10 episodes of new shipping yet. But this isn't about Fable 5, which we've talked about in the past, it's about Sonnet 5. And this is the new default model for every free and pro user on the platform. Now, sometimes new model releases are more like background noise, like it's just like, okay, cool, but it's not that meaningful. This one actually matters, and it's because of the math. So let's talk numbers here. Okay, before this week, if you wanted Anthropics best reasoning, you were paying for Opus 4.8. Now you were paying $5 per million input tokens, $25 per million output. That was their flagship price. Most teams were actually being pretty selective about when they'd use that model because at scale, it adds up fast, right? A product making a thousand AI calls a day at OPUS prices. That's going to get some conversations going from the finance team. So sometimes product teams would sort of, you know, flex to a lesser model when they needed to. Um, Sonnet 5 comes in at $2 per million input tokens and $10 per million output. Now that's through introductory pricing, which holds through August 31st. More on that a little bit later. But this is the part that's really important. Sonnet 5 performs close to Opus 4.8 on most tasks. So you're not trading down here to save money. You're getting near flagship capability, but you're paying less than half the price. Now that's a very different situation than we were in just a couple weeks ago. So for most teams, what it means is that the the ceiling, it sort of stops being a constraint. We can now use Anthropic's most powerful models. We don't have to flex down, and we can actually afford it. Um, now adaptive thinking is on by default with Sonnet 5.2. I think that's a big deal too. This is the capability that lets the model actually work through multi-step problems, breaking them into steps, using tools along the way, finishing tasks end to end. Instead of, and I don't know if this has happened to you, it probably has, getting halfway through and handing back something incomplete. Now, Zapier's engineering team, they actually were playing around with Sonnet 5. They tested it early and they put it plainly. They said, look, a two-part job that used to stall halfway now finishes. And their verdict was that Sonnet 5 is a no-brainer for day-to-day automation. So, what does it mean in practice for you? Well, if your team shelved an agentic feature because the cost per task didn't pencil out at Opus prices, you might want to revisit that math. Now, today Sonnet 5 isn't a cheaper model with worse capabilities. It is, again, near flagship capability, and it's just a fraction of the cost. So if you've been getting bad feedback on your AI features and you haven't been able to pinpoint why, it might be that it was stalling halfway. Um, and now maybe that problem goes away. Um, users, they start a multi-step flow, model hits something can't handle, and now it's actually going to finish it instead of handing it to them sort of half complete. Um, I think that is a very big deal. But again, I talked about this being introductory pricing. So introductory pricing is live through August 31st. Um, then it steps up. Okay, pricing does go up, but it goes up to $3.15 per million tokens on input and output, respectively. It's a 50% increase, but it's still significantly cheaper than Opus 4.8. Um, so if your product margins depend on a specific cost per call, don't bank on the pricing that you're seeing right out of the gate. Remember that will change, but it's so much different. It's so much less in cost than what you're probably paying for with Opus. But again, same kind of benefit. So that's a big deal. That's story two. We have one more story to go. All right, story number three is about AI regulation, and it's coming from all sorts of different directions. Um, we'll start with on June 29th, the Council of the European Union gave its final green light to what's being called the AI Act Simplification Package. European Parliament had formally endorsed it on June 16th, and now it's law. Uh, and what is it actually trying to solve for? Well, it's trying to solve for the problem that AI systems are increasingly making consequential decisions about people's lives. Maybe you get a job interview, maybe you're going for a loan, uh, maybe you're trying to file an insurance claim and you end up getting flagged somehow, and you there's a negative impact. Maybe you get turned down for that loan, you don't get the job. Right now, people are being impacted by things like this happening, and they don't even know it. They don't know that an algorithm made the decision. They can't ask why, they can't appeal it because there's nothing to really appeal. Um, the company could just say, ah, actually, our system flagged you and we're moving on here. And that's it. I mean, Amazon famously had to scrap a recruiting tool that taught itself to penalize resumes that included the word women's, as in women's college, uh, women's sports. Now, hiring tools trained on historical data, learn historical patterns, and historical patterns in employment are full of discrimination. I mean, credit scoring tools have shown racial bias in the past. Healthcare AI has systematically underestimated the needs of black patients. So the EU AI Act and various US state laws coming behind it. These things exist because lawmakers looked at all this and decided all right, look, if AI is going to make these decisions, there need to be rules about how those decisions get made, documented, and challenged. So that's the backdrop of all of this. Now, the headline out of Brussels might actually surprise you. The news there isn't that there's this new deadline coming right up very soon. There's actually a delay. The simplification package actually pushes back the toughest requirements for the high-risk AI systems, the one that's covering hiring, credit, scoring, performance evaluation, healthcare decisions. That was all going to be August 26th as a deadline. Now it's December 2027. That's a 16-month extension. And the reason is that there are technical standards that companies need to actually benchmark compliance against, and those aren't ready yet. So transparency rules still kick in this August. That includes chatbot disclosures, labeling AI-generated content, but the heavy documentation and risk assessment requirements, those are what's being pushed. Now, you might think, okay, there are delays here. If this is going to affect me, maybe if you're building products that are centered around employment or, you know, healthcare, insurance claims, you might hear this and think, okay, I don't have to worry about this till the end of 2027. I wouldn't get too comfortable. December 2027 in the grand scheme of things is not that far away. Now, that's on the European side of things. Let's zoom on over to the US side of things. Earlier this month, Connecticut signed the CART Act into law that stands for the Connecticut Artificial Intelligence Responsibility and Transparency Act. It's very sweeping. It's one of the most broad state AI laws that's been enacted anywhere in the United States to date. And it actually requires employers to disclose when AI is used in employment-related decisions. So things like hiring, performance reviews, promotions, terminations. The law explicitly states that AI use is not a defense to employment discrimination claims. Now, that is a direct response to a pattern that we're starting to see emerging in employment cases. Companies arguing that, hey, they didn't discriminate, the algorithm did. Now, Connecticut is closing that door. So if your AI tool produces a discriminatory outcome, the fact that it's an AI tool, that's not a legal shield, at least not in Connecticut. That provision arrives October 1st of this year. I mean, that's three months away. But the full employer disclosure obligations, the actual requirement to tell candidates and employees when AI is influencing decisions about them, those kick in October 1st, 2027. So again, a little bit more time on that. But in the grand scheme of things, that's coming quick too. So I think it's interesting. In the same month, the EU is finalizing its AI regulation framework. Um, and now we have a state in the United States making AI hiring disclosures the law with deadlines already on the calendar. You could probably see a pattern here. We're going to start to see more and more U.S. states come up with their own uh their own requirements, their own regulations. And so if your product touches employment decisions, recruiting, performance, uh management, I any of these things in any market, I think compliance, it becomes a product architecture question, not just a legal one. Um, and again, maybe have a little more runway, but that is all coming very, very soon. And now I'd say, you know, the teams that are using all this information now to build the right instrumentation, logging what their AI does, documenting how decisions get made, uh, demonstrating non-discrimination. They're gonna be in a very different place than those that are gonna sort of punt on any sort of decisions. Connecticut's not gonna be the only state. Colorado's already passed an AI law back in 2024. Illinois has had AI hiring rules on the books for a while. California has been moving, and even the federal government has signaled that they want a national standard. So that's gonna wrap things up for this week's episode of Now Shipping, which I'm glad for because I can tell you this news is coming in hot here. I am sweating over here, but I hope that you've enjoyed this episode. And look, if you have any comments about how we can make things even better, anything that you heard here today, leave a comment below. I am checking out the comments. We want to make these episodes even better. And of course, if you are getting value in this, share it with a friend, subscribe, it would really help us out. So I appreciate you doing all of that. Once again, I'm Mike Belcito and brought to you by the team at Mind the Product. This is now shipping.