Sol, Luna, and the Math Problem Nobody Wanted

Alex Shannon: I keep going back and forth on this — I think I actually land on the side that this is a good thing. OpenAI calling in elite mathematicians to basically audit their own work? That’s humility. That’s a company saying, hey, we got this wrong, we need outside perspective, we’re going to fix it.

Sam Hinton: Really? Because I read the same story and I came out the other end deeply uncomfortable. Like, think about what you’re actually saying — the company that is selling AI as the future of knowledge work, the company that wants these models embedded in hospitals and law firms and schools, had to have a reputational crisis before they picked up the phone and called a mathematician?

Alex Shannon: OK but — isn’t the point that they’re doing it now? Better late than never?

Sam Hinton: Is it though? Or is this a PR move dressed up as epistemic humility? Because the crisis already happened. The damage is already done. And now they’re going, whoops, maybe we should have asked an expert. I just think the bar we’re holding these companies to is way too low.

Alex Shannon: I hear you. I genuinely do. But I also think if we only celebrate course corrections when they happen before anything goes wrong, we’re setting an impossible standard. We’re going to get into the full story in a bit — but that tension right there? That’s basically the theme of today’s whole episode.

Alex Shannon: You’re listening to Build By AI, I’m Alex Shannon — and yes, we are already in it. Welcome to the show.

Sam Hinton: And I’m Sam Hinton. And look — before we dive in, we actually have a quick piece of news about Build By AI itself. Starting today, we are making a format change.

Alex Shannon: Yeah, so for a while we’ve been doing daily episodes, and honestly — we love that, but we’ve heard from a lot of you that you want more depth, more analysis, more of the back-and-forth. So starting now, Build By AI is moving to one bigger weekly episode, dropping every Wednesday, covering the whole week’s AI news in one go.

Sam Hinton: It’s the same show, same energy, just more of it at once. Think of it less like a daily briefing and more like your weekly AI debrief with two people who are maybe slightly too obsessed with this stuff.

Alex Shannon: Slightly. OK, let’s get into it — because this week was genuinely packed. We’ve got new models from OpenAI, possible news from Anthropic, Meta in controversy again, Amazon throwing a punch, and more. Let’s go.

OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes

Alex Shannon: Alright, let’s start with the big model news, because OpenAI dropped something significant this week. They’ve launched two new AI models — GPT-6 Sol and GPT-6 Luna. And the headline selling points are: lower costs than previous models, and fewer mistakes. They’re also built on the same architectural foundation as a model called Astra.

Alex Shannon: Now, we don’t have a ton of granular detail on the exact benchmark numbers — but the framing from OpenAI is pretty clear. Sol and Luna are positioned as more efficient and more reliable. And that combination — cheaper plus better — is kind of what everyone’s been asking for.

Sam Hinton: Yeah, and here’s why that matters immediately — cost has been the single biggest friction point for companies trying to actually deploy these models at scale. Like, you can have the most capable model in the world, but if calling it a million times a day is burning through your infrastructure budget, you’re going to make tradeoffs. Sol and Luna potentially change that calculus.

Alex Shannon: Right. And the Astra architecture connection is interesting to me — what does it tell us that they built Sol and Luna on that same foundation? Is this a signal about where OpenAI thinks the architectural future is heading?

Sam Hinton: I think it’s really significant, actually. What it suggests is that Astra isn’t just one model — it’s a platform, a base that you can build different products on top of. Sol and Luna might be tuned differently, priced differently, optimized for different use cases — but they share a core. That’s a much more mature way to run a model family. It’s how you build an ecosystem, not just a product.

Sam Hinton: Think about it like — you know how Apple doesn’t build an entirely new chip for every iPhone tier? The A-series chip powers multiple devices because the underlying architecture scales. This feels like OpenAI is starting to move in that direction with their model stack.

Alex Shannon: That’s a great analogy. Although I’ll play devil’s advocate here — we’ve heard the ‘lower cost, better accuracy’ pitch before. Almost every major model release in the last two years has led with those same two claims. How do we know Sol and Luna are actually meaningfully different?

Sam Hinton: That’s a totally fair pushback. And look — we don’t have the independent benchmark data yet to say definitively. But there’s a difference between ‘marginally better and slightly cheaper’ and something that genuinely moves the needle for developers. I think the real test comes in the next few weeks when people start building with these things and reporting back.

Alex Shannon: So who should be paying close attention to this launch right now? Like, if you’re listening to this and you’re a developer, a product team, a business — what’s the immediate action?

Sam Hinton: Honestly? If you’re already running workloads on GPT-5 or any of the previous generation models, the answer is: go look at your pricing structure and see if Sol or Luna are worth testing as a replacement. The ‘fewer mistakes’ claim is especially worth pressure-testing on your specific use case, because that’s the thing that’s been holding a lot of people back from full deployment.

Alex Shannon: Keep an eye on the independent evals over the next week or two. That’s when we’ll really know whether Sol and Luna are a generational step or just a well-marketed increment.

Sam Hinton: Exactly. And watch how Anthropic and Google respond — because in this market, a real efficiency gain from OpenAI will trigger reactions very fast.

OpenAI wants to consult elite mathematicians about how to not fumble again

Alex Shannon: OK so this brings us back to the thing we were already arguing about at the top of the show — and Sam, I want to give you the full floor to make your case here, because I think there’s a real story under the headline. What happened is: OpenAI went through what’s being described as a reputational crisis involving flawed mathematical results. And in response, they are now consulting elite human mathematicians to help guide them and avoid similar failures going forward.

Alex Shannon: So — verified by multiple sources, this is real. OpenAI had a math problem, literally and figuratively. Now they’re bringing in the humans with the PhDs to help clean it up. Sam, you came out of reading this story uncomfortable. Explain that to me.

Sam Hinton: OK so here’s my discomfort. OpenAI has been positioning these models — and AI broadly — as tools that can do serious intellectual heavy lifting. Math is like the gold standard of verifiable correctness. It’s not subjective. There’s a right answer. And the fact that they had a public reputational crisis over mathematical results tells me the internal quality checks were insufficient for the stakes involved.

Sam Hinton: And my issue with the response isn’t that consulting mathematicians is bad — obviously that’s good! My issue is that it took a crisis to get there. If you’re shipping a model into the world that people are going to use for quantitative reasoning, financial modeling, scientific work — shouldn’t elite domain experts have been part of the process from the start?

Alex Shannon: That’s a strong argument. But let me push on it a little — because I wonder if you’re holding OpenAI to a standard that no organization could actually meet before shipping. Like, you can’t anticipate every failure mode in advance. Sometimes you find the edge cases in the wild.

Sam Hinton: I mean, maybe. But math isn’t an edge case — it’s a core capability claim. If I’m selling you a calculator and it gives wrong answers, that’s not a nuanced failure mode. That’s a foundational problem. And the scale at which these models are deployed makes the stakes really high.

Alex Shannon: OK, but here’s where I come back to my original take — I think the more interesting and actually optimistic signal in this story is that OpenAI is now explicitly building a feedback loop with human experts. Not to replace the models, but to create a calibration mechanism. That’s actually a mature, responsible thing to do.

Sam Hinton: Right, and I don’t disagree with that. I think if this becomes a structural part of how they develop and validate models going forward, it’s genuinely positive. My concern is whether this is a systemic change or a one-time PR response. Like, are these mathematicians becoming embedded in the development process, or are they being asked to sign off on something after the fact to make a press release look good?

Alex Shannon: And that’s the question we can’t answer yet from the reporting. What this means in practice — for users, for businesses — is that there’s an acknowledgment from the top that these models need domain-expert verification for high-stakes applications. Which, if you’re a company using AI for anything involving numbers or scientific reasoning, is actually a useful signal. It validates the instinct to have a human expert sanity-check the outputs.

Sam Hinton: A hundred percent. And honestly, the takeaway for anyone using these models professionally right now: do not treat AI math outputs as gospel without verification. This story is a reminder that even the companies building these things are saying — we need humans to check the work. So maybe you should too.

Alex Shannon: Watch whether OpenAI announces any formal expert-in-the-loop programs or partnerships in the next few months. If they do, that’s the story becoming structural. If it goes quiet — that’ll tell us something too.

Meta admits Muse’s likeness to OpenClaw isn’t a coincidence

Alex Shannon: Alright, let’s talk about Meta and what is, honestly, a pretty significant story that got a little buried under the model release noise this week. Meta has publicly acknowledged that its AI assistant called Muse was — and this is their word — ‘heavily inspired’ by a product called OpenClaw. And the similarities aren’t just vibes-level. We’re talking about similarities in workspace filenames and content structure. Not just like, oh the UI looks similar.

Alex Shannon: Here’s the thing though — Meta is maintaining that Muse was built from scratch. So their official position is: yes, we were heavily inspired, yes there are structural similarities down to the file naming level, and also we built this ourselves. And this is verified by multiple sources — TechCrunch and Ars Technica both have this story.

Sam Hinton: OK, I love this one because the logic Meta is asking us to accept here is genuinely remarkable. ‘We were heavily inspired by this other product, the internals look really similar, but we built it from scratch.’ Like — I’m not a lawyer, but I’m pretty sure inspiration has limits. At some point ‘heavily inspired’ and ‘built from scratch’ become contradictory claims.

Alex Shannon: Let’s be fair to Meta for a second though — can you be heavily inspired by something and still build something independently? Like in music, you can write a song that sounds like another song without directly copying it.

Sam Hinton: You can! And that happens all the time in creative fields. But in software, and especially in AI products, workspace filenames are not aesthetic choices. They’re structural decisions. If your internal file structure mirrors someone else’s internal file structure, that’s a pretty specific kind of similarity. It’s not like you both happened to pick blue as a background color.

Sam Hinton: The analogy I keep coming back to is — imagine you’re an architect and you build a house, and you say you designed it from scratch, but the blueprints have the same room names, the same dimensions, the same electrical layout as the house next door. At what point does ‘inspired by’ become something more significant?

Alex Shannon: And what’s wild to me is that Meta admitted this. They didn’t get caught and deny it — they acknowledged it. Which makes me wonder whether there’s a strategic calculation here, like maybe the legal exposure is smaller if you admit influence than if you get caught denying it.

Sam Hinton: That is absolutely a strategic legal move. And it’s not the first time we’ve seen a big tech company navigate IP questions in AI this way. The whole field is kind of operating in a legal gray zone right now where the norms around inspiration, derivation, and copying are still being worked out — both in courts and in public opinion.

Alex Shannon: For businesses and developers watching this — what’s the implication? Is this just a Meta drama story or does it have wider relevance?

Sam Hinton: It’s wider than Meta. This story is a preview of a conversation the whole industry is going to be forced to have as AI products proliferate. When does ‘inspired by’ become improper derivation? What are the legal and ethical standards for building on someone else’s design? And as more companies launch AI agents and assistants, these collisions are going to happen more frequently. This is just one of the early visible cases.

Alex Shannon: Keep an eye on whether OpenClaw — or anyone associated with them — pursues any kind of legal action. Because that’s when this story gets really interesting and potentially precedent-setting.

Sam Hinton: Exactly. And watch how the rest of the industry reacts, because if Meta gets through this without major consequence, it will send a signal about what’s permissible. If there’s real pushback, it could start to establish some guardrails.

Meta’s AI agent has been blocked from using Amazon.com

Alex Shannon: And staying with Meta for a second — because honestly their week just keeps getting more complicated — Amazon has blocked Meta’s Muse AI agent from accessing Amazon.com. Full stop. And this is verified by multiple sources — TechCrunch, Ars Technica, and The Verge all have it.

Alex Shannon: The important context here is that Amazon has no legal obligation to give Muse access to their site. Amazon runs its own AI foundation models and operates what’s described as one of the internet’s most popular inference platforms. So they’re not just a passive platform being asked to cooperate — they’re a direct competitor in the AI space.

Sam Hinton: Yeah, and I want to make sure people understand why this actually matters, because on the surface it sounds like a boring corporate dispute. But think about what AI agents are supposed to do — they’re supposed to act on your behalf across the internet. Shopping, booking, researching, interacting. If major platforms start blocking agents from access, you end up with a fragmented agent internet where your AI assistant can only go certain places. That’s a fundamental limitation on what the technology can actually deliver.

Alex Shannon: So this is less about Meta specifically and more about a structural tension in the whole AI agent ecosystem?

Sam Hinton: Exactly. Amazon blocking Muse is a single data point, but it points at a much bigger question: who controls the access rights for AI agents on the web? Right now, it’s whoever owns the platform. And that means every major website, every major marketplace, every major service can decide whether AI agents are welcome or not — and they have enormous leverage.

Sam Hinton: Here’s the analogy. Remember when Google started sending search traffic to AI overviews instead of to publisher websites, and publishers were furious? This is the flip side of that. Now AI agents want to go out and interact with platforms on behalf of users — and those platforms are saying, wait, we didn’t agree to this, and also you’re a competitor.

Alex Shannon: And Amazon’s position makes total business sense when you think about it. If Muse can help users shop on Amazon, who gets the credit? Who captures the relationship with the customer? If the agent is the interface, Amazon loses the direct customer connection they’ve spent billions building.

Sam Hinton: One hundred percent. Amazon’s whole moat is the customer relationship. The data, the Prime ecosystem, the recommendations. If a Meta AI agent sits between Amazon and the customer, that erodes Amazon’s advantage. Of course they’re blocking it.

Alex Shannon: So where does this go? Like, is this the beginning of a broader pattern where big platforms start walling off from AI agents they don’t control?

Sam Hinton: I think yes, and I think we’re going to see this become a real industry-level conflict. You’ll have AI companies pushing for open agent access, and platform companies using their terms of service and technical controls to protect their turf. And users will be caught in the middle wondering why their AI assistant can’t just do the thing they asked it to do.

Alex Shannon: The practical takeaway here is: if you’re building products or business strategies around AI agents interacting with third-party platforms — you need to be thinking about access risk. Just because your agent can technically do something today doesn’t mean the platform will allow it tomorrow.

Sam Hinton: Really well put. This is the new version of platform risk. It used to be about app stores. Now it’s going to be about which AI agents get access to which corners of the internet.

RAPID FIRE

Alex Shannon: Alright, let’s move to rapid fire — four more stories worth your attention, a little quicker but still meaty. First up: early reports suggest Anthropic has released a new model called Opus 5.5. If confirmed, this is Anthropic’s strongest-performing model to date — with lower prices than previous Opus models and performance described as comparable to something called Fable-level. Sam — quick take?

Sam Hinton: OK so according to early reports, and I want to be clear this is single-source so we’re hedging — if this is real, the combination of top-tier performance at lower prices is Anthropic directly competing on the same battlefield as Sol and Luna. The ‘strongest model we’ve ever tested’ claim is a big deal if it holds up under independent evaluation. Watch this space very closely.

Alex Shannon: Right — the timing of this, if confirmed, alongside the OpenAI Sol and Luna launch in the same week, suggests we may be hitting a new phase of model competition where both cost and capability are being pushed simultaneously. That’s actually good news for users and developers.

Sam Hinton: Genuinely exciting if true. The pricing pressure in this market is real, and the beneficiary is anyone building with these APIs. Less money per token means more ambitious applications become economically viable.

Alex Shannon: Next one — and again, early reports on this one: Snorkel AI has apparently tripled its valuation to 3.5 billion dollars with a 350 million dollar Series E funding round. If confirmed, that is a massive vote of confidence in AI training data as a business. Snorkel operates a data-as-a-service model and is capitalizing on booming demand for AI training data.

Sam Hinton: Here’s why this matters and it’s a little counterintuitive — the companies making the flashy models get all the attention, but training data is the unglamorous infrastructure that all of it runs on. A valuation tripling to 3.5 billion in this climate says that investors believe the demand for high-quality training data is not slowing down — it’s accelerating. And Snorkel has been doing this for seven years, so they have real moats.

Alex Shannon: It’s like — everyone’s covering the car manufacturers, and Snorkel is quietly making the fuel. If this funding round is confirmed, it’s a strong signal that the AI infrastructure layer is becoming extremely valuable real estate.

Sam Hinton: Exactly. And data-as-a-service is a really interesting model because it’s recurring, it’s scalable, and it becomes more valuable as your clients’ models become more capable and hungry for more training signal. Smart place to be.

Alex Shannon: Alright, third one — and still hedging since this is single-source — Qualcomm has unveiled two new smartphone chips with enhanced AI capabilities. The headline spec: the flagship chip can reportedly run a 30-billion parameter mixture-of-experts model locally on the device. On your phone. Sam?

Sam Hinton: OK so for anyone not familiar — a mixture-of-experts model is a specific architecture where instead of activating the whole model every time, you only activate the parts that are relevant to a given task. It’s efficient. And 30 billion parameters locally on a phone — if that’s real — is legitimately impressive. The story of AI moving to the edge, meaning onto your actual devices rather than in the cloud, has been building for a while and this would be a meaningful step.

Alex Shannon: And the implications of on-device AI are real — privacy, latency, cost, connectivity. If you’re not dependent on a cloud call for AI inference, a lot of things get better or at least different.

Sam Hinton: The challenge is always going to be: what can you actually do with those parameters on a phone versus in a data center? Raw parameter count doesn’t tell the whole story. But directionally, Qualcomm continuing to push this hard tells you where smartphones are going. The phone of 2028 is going to look very different from the phone of 2024 in terms of what it can do locally.

Alex Shannon: And finally in rapid fire — Rabbit. Yes, Rabbit is back. Early reports suggest they’ve developed a standalone AI agent that operates independently from their hardware device, the R1. If confirmed, users no longer need the R1 to use Rabbit’s AI agent. Sam, this is quite a pivot.

Sam Hinton: I mean — look, the R1 was not a beloved product, to put it gently. It was underwhelming by most accounts. So the fact that Rabbit is apparently moving to a software-first model makes a certain kind of sense. But I’m going to be honest — I’m a little skeptical about whether this is a strategic pivot or a survival move. The question is whether the agent itself is good enough to stand on its own merits when it’s not attached to a hardware novelty.

Alex Shannon: Fair. And decoupling from the hardware actually gives them a much wider addressable market — anyone with a phone or a computer rather than just R1 buyers. But as Sam says, the product quality question is still open. Watch for independent reviews of the standalone agent over the coming weeks.

BIGGER PICTURE

Alex Shannon: Alright, before we wrap — let’s do what we always do and zoom out. Because I feel like there’s a thread running through basically everything we talked about today, and I want to name it.

Sam Hinton: Yeah, I’ve been thinking about this as we’ve gone through the stories. And the thread I keep seeing is: we are hitting the accountability moment for AI. Like — OpenAI having a math crisis and calling in expert humans to check the work. Meta being called out for how closely Muse resembles OpenClaw. Amazon asserting that they get to decide who accesses their platform. Even the Snorkel story — the fact that training data is valued at 3.5 billion dollars — that’s partly a story about the demand for data that is actually correct and trustworthy.

Alex Shannon: That’s a really interesting frame. The story of the last two years in AI has been capability, capability, capability — what can these things do? And now the story is starting to shift to: OK, but can we trust what they do? And who is accountable when they get it wrong?

Sam Hinton: And the platforms conflict — Amazon blocking Muse — that’s an accountability story too, just at the infrastructure level. It’s about who has control, who has rights, who gets to decide the rules of the AI-enabled internet. These aren’t technical questions. They’re governance questions. And we don’t have great answers to most of them yet.

Alex Shannon: What should people be watching for in the next few weeks that connects to all of this?

Sam Hinton: Three things. First — independent evaluations of Sol, Luna, and if confirmed, Opus 5.5. Not the company benchmarks — the community benchmarks. That’s how we find out if the efficiency claims are real. Second — whether OpenAI’s mathematician consultation becomes a structural program or disappears quietly. Third — whether other major platforms follow Amazon’s lead in blocking or restricting AI agent access. If Google, or Walmart, or any major e-commerce or content platform makes a similar move, we’re looking at the beginning of a real internet access war for AI agents.

Alex Shannon: And I’ll add one more. The Meta and OpenClaw story — I think within the next six months we’re going to see real legal action in this space somewhere, from someone. The IP questions around AI products are getting too big to stay in the ‘inspired by’ gray zone much longer. When the first major verdict lands, it’s going to redefine how this industry operates.

Sam Hinton: The theme of 2025 was: look at what AI can do. The theme of 2026 is starting to feel like: OK, now figure out who is responsible when it gets things wrong. And that is honestly a healthy evolution. As uncomfortable as some of these stories are — the math crisis, the IP questions, the agent blocking — they’re signs of an industry being forced to grow up.

Alex Shannon: Growing up. In public. On the internet. While raising billions of dollars. As one does.

Sam Hinton: Normal stuff. Totally normal.

OUTRO

Alex Shannon: That’s going to do it for this week’s Build By AI — our first weekly Wednesday episode, and honestly I’m really happy with how this format feels. More time to actually dig into the stories.

Sam Hinton: Same. I feel like we actually got to finish our thoughts this week, which is novel for me personally.

Alex Shannon: If you found this valuable, subscribe wherever you get your podcasts — and if you can leave us a rating, that genuinely helps more people find the show. We’ll be back next Wednesday with everything that moves in AI this week, and based on the pace of things — it’ll be a lot.

Sam Hinton: See you then. And in the meantime — check the math. On everything.