Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, pull up a chair, I'm your boy Tony DeLuca, and today's menu is heavy on the money side of the AI business. We got chip valuations doubling in two months, a Brooklyn lab trying to plant a Western flag in open-weight models, a nineteen-year-old selling you a thirty-five-hundred-dollar box, and a VC catfight that would make a daytime talk show blush. We also got some three-thousand-year-old skulls to wrap up with, because sometimes you need to step away from the token economy and look at what people actually did with sharp objects. Let's have at it.
Let me start with the one story on today's menu that I think a builder should actually chew on, because it tells you something about where the ground is shifting under your feet. And that's the open-weight model race, specifically a company called Reflection AI.
Now, I want to be careful here, because the reporting on this one straddles a couple of dates. The underlying buzz around Reflection and its model goes back to this summer and the weekend before, when there were reports it was close to launching. What's in front of me now is the official confirmation: Reflection AI has unveiled Beam, its first frontier open-weight model. That's the piece from Rebecca Bellan.
Here's the shape of it. Reflection is a two-year-old, Brooklyn-based startup. Founded in 2024 by two former Google DeepMind researchers. And the thing I want you to notice first is not the model, it's the money. This company has raised roughly four-point-seven billion dollars. From Nvidia, from Sequoia, from Lightspeed. Its last round valued it at a twenty-five billion dollar pre-money valuation. And it locked up compute, signed deals collectively worth more than seven billion with SpaceX and Nebius to get access to Nvidia's GB300 chips through 2029.
So this is a company that raised billions and inked billion-dollar compute contracts before it had a public model out the door. Now the model's here. Beam. The company calls it a text-only mixture-of-experts model. Five hundred and one billion parameters, twenty-three billion active. Pretrained on nearly twenty-four trillion tokens, with a million-token context window. The pitch is that it matches the leading Chinese open models on advanced reasoning benchmarks while using, in their words, three to four times less inference compute.
Now, kiddos, I'm your skeptic in residence, so let me say the thing the article says plainly: Reflection's performance claims haven't been independently verified. Those are the company's own benchmarks. When a lab tells you its model is as good as the other guy's while being cheaper to run, you nod politely, you write it down, and you wait for somebody who doesn't have four billion dollars riding on the answer to check the math. The weights and full technical details are supposed to come this month. When they do, we'll know more.
But here's why I'm leading with this instead of the raw benchmark numbers. The interesting part is the business model, and it's a word you're going to hear a lot: AI factories. The pitch is not "come use our chatbot." The pitch is to let enterprises and sovereign nations build their own customized, local AI systems by training Reflection's models on their own proprietary data. They're already piloting a sovereign AI factory partnership with Shinsegae Group in South Korea. And the reporting says hedge funds and trading firms are among those eager to build these systems.
Think about who benefits here and why. Nvidia backs Reflection. Nvidia's CEO Jensen Huang has been beating this AI factory drum for a while, and the article says it plainly: it's a vision that would also benefit Nvidia, whose GPUs would power those systems. So you've got a chip company funding an open-model company whose whole go-to-market sells more chips. That's not a conspiracy, that's just how the pieces fit. But as a builder you should see the shape of it. The open-weight story in the U.S. isn't really a hobbyist story or a cheap-model story anymore. It's an enterprise control story. It's about companies that do not want to ship their proprietary data and their trade secrets through somebody else's closed API and effectively teach a vendor how their business works.
And that connects to something I want to flag from continuity. We talked back on the third about Anthropic spending a hundred million dollars on a Claude Frontier Academy to train engineers for the enterprise. Different tactic, same war. The closed labs are trying to lock in the enterprise by training the people. The open-weight players are trying to win the enterprise by handing over the keys, the weights, the control. Reflection is positioning itself against the closed labs, against the Chinese open models, and against Western open players like Mistral, Meta, and Cohere. Its most direct U.S. rival, the piece says, might be Inkling, the open model out of Mira Murati's Thinking Machines Lab from July. Reflection's own benchmarks show Beam outscoring Inkling on four coding tests, but — and this matters — Inkling is multimodal and Beam is text-only. So it's not apples to apples.
Here's my read for you, the builder. If you run anything touching regulated data, financial data, legal data, patient data, you should be paying attention to the open-weight enterprise pitch, not because the models are magically better, but because "your data never leaves your trust boundary" is becoming a real competitive axis and not just a compliance checkbox. The question that'll determine whether Beam matters is boring and unglamorous: when the weights drop this month, are the independent benchmarks anywhere close to the company's slide deck, and is the fine-tuning story actually smooth enough that a mid-sized enterprise can run it without a research team? Watch for the independent numbers. Don't buy the press release. That's the whole game.
Now let's move from the model layer down to the silicon, because the same money is pouring into the picks and shovels, and it's pouring in fast enough to make you dizzy.
Here's the one that made me raise an eyebrow. Etched, the AI chip startup, is fielding funding offers at a forty-billion-dollar-plus valuation. And I want you to hold the timeline in your head while I say this, because the timeline is the story. A couple of months ago, Etched raised seven hundred million at a twenty-one billion dollar valuation. That was September. This is reporting that incoming bids now range from forty billion from top-tier investors to fifty billion from lesser-known backers. These talks are early, terms can change, Etched declined to comment. But sit with it: the valuation roughly doubled in a couple of months, on offers, before the ink dried on the last round.
And it's not even the first time for them. They did three hundred million at a ten-point-three billion valuation in July, led by Sequoia. Then seven hundred million at twenty-one billion in September. Now forty-plus on offer. The article has a nice dry phrase for this — back-to-back funding rounds that essentially act as a single financing split into two tranches with separate valuations, increasingly common among the buzziest startups. Translation for the kitchen table: they're raising money in installments and marking the company up between each spoonful.
So why are the VCs this hungry? There's actually a real business under the froth, which I appreciate. Etched is a four-year-old company building full AI hardware systems powered by its own proprietary chips, specifically designed to speed up inference — that's the computing that happens after you hit enter on a prompt. Quant trading firm Jane Street led the last round and is also a customer, took delivery of an early system. For a trading firm, a microscopic advantage in speed turns directly into money, so they'll pay. Etched said in July it had already secured a billion dollars in customer orders. They've pulled roughly fifteen percent of their four-hundred-person workforce out of Nvidia. They've got a test chip made at a TSMC factory, a ten-megawatt data center in Silicon Valley, and a facility in Taiwan to coordinate production.
Co-founders Gavin Uberti and Chris Zhu met in an advanced math course at Harvard, the COO Robert Wachen was Uberti's roommate, and they dropped out to build it. Fine. Good story. But here's my protective-of-your-time take. When a chip company's valuation doubles on offers in eight weeks, that is not primarily a statement about the chip. That is a statement about how much capital is chasing anything that sounds like a credible challenger to Nvidia. There is genuine customer traction here — a billion in orders is not nothing — but the valuation is running way out ahead of the revenue. If you're a founder trying to raise right now, understand the climate you're in: there's a flood looking for a channel. If you're trying to read the health of the industry, remember that marking a company up every two months is the kind of thing that looks brilliant on the way up and gets a very different name on the way down.
And speaking of challengers to Nvidia, there was a nice little detail buried in a different piece worth a quick mention. At the PearX demo day — that's Pear VC's accelerator, caps its cohorts at twenty startups, the kind of thing investors actually watch — one of the companies that got buzz was a chip startup called Saia. Twenty-year-old founder, Ayaan Govil. The claim is a chip that bypasses traditional memory by running AI directly out of flash storage. They say it delivers way better speeds, eight times the capacity, four times less power than Nvidia's Jetson, which is a leading local AI chip. Already talking to Samsung about memory integration, test chips next year, mass production aimed at 2028.
Now, I'm not going to pretend I can validate a flash-storage inference chip from a demo day pitch. Hardware is notoriously brutal, and that's a direct quote from the piece, not me being cranky. But I flag it because there's a through-line today: everybody and their roommate is trying to get between you and the memory, which, as we've noted on this show before, is in a genuine shortage expected to last through 2028. When the bottleneck is RAM, the money flows to anybody who claims they can route around RAM. Watch that space, but watch it with your hand on your wallet.
Alright, let's shift from the chips in the data center to the box somebody wants to put on your desk. Because the other hardware story today is a nineteen-year-old and a thirty-five-hundred-dollar computer.
The company is Ghost. Founder Zain Javaid, nineteen years old, used to be a quant, got bored, then ChatGPT came out in 2022 and he told the reporter, Dominic-Madori Davis, that he felt fundamentally excited in a way he never had before. Fair enough. Ghost came out of stealth with an eleven million dollar seed round led by Andreessen Horowitz, with Abstract, Audacious Ventures, SV Angel, and Nova also in. Founding team's young — a couple of nineteen-year-olds, a twenty-three and a twenty-four-year-old.
The product is called Core. It's a personal computer designed specifically for AI agents that act on your behalf. The idea is to centralize your data — desktop, apps, smart home devices — into one piece of hardware that continuously runs agents. Javaid calls it a "brain in a box" without a screen; you talk to it through a phone or an app. It's got an Nvidia Blackwell GPU bundled in, its own software layer, browser, filing system. Comes with a few open models preinstalled — Qwen and Gemma variants — and you can load others from Hugging Face or your own. Thirty-four ninety-nine. Preorders open, shipping end of October.
Now here's where I give the kid credit, because the pitch is actually coherent, and it ties right back to the Reflection story. The whole premise is ownership. Javaid says the models and personal memory run locally, everything's encrypted so no cloud provider or third party — including Ghost — can meaningfully access it. He says the device is programmed to keep working if Ghost itself shuts down. His quote: "All of the source code, logic, and model weights live on the device, not the services." They're building a firewall that monitors every outgoing network request and, if the agent tries to exfiltrate data it shouldn't, rejects it and flags it for your review.
That last part is the genuinely interesting bit, and it's the part nobody talks about enough. If you're going to give an agent your logins and your passwords and let it take actions, the scary failure mode isn't that it books the wrong flight. It's that it reaches out to the whole internet with your credentials in hand. A box that watches its own agent's outbound traffic and can slam the door — that's a real idea, whether or not this particular startup is the one that nails it.
But let me put on the skeptic hat, because that's my job. He described Core's magic use case like this: Core proactively tells you that you might be getting sick because it sees your resting heart rate is elevated, sees from your computer history you were up late on a project, and knows it's flu season where you are. And look — that's a cute demo. That is also exactly the kind of thing that sounds delightful in a press release and lands somewhere between "mildly helpful" and "why is my computer diagnosing me" in real life. When the reporter asked how this is different from a Mac Mini — which plenty of people already set up to run AI tools — his answer was basically: ours is purpose-built, models preinstalled, no setup. Which is a fine answer. It's also an answer that Apple can give too, with a supply chain and a billion devices, the second it decides this category is worth owning.
Javaid's not worried. His line was that people give incumbents credit for already having built the product because they have the ingredients, and he said if they fail to build a product people love, "Apple isn't the reason we failed." I actually respect that framing. It's the right founder mindset. But as a buyer, here's my plain advice: a thirty-five-hundred-dollar screenless box from a one-year-old company with a shipping date three weeks out is a bet, not a purchase. The ownership and the outbound-firewall ideas are the real signal here. Whether the first batch of hardware from a San Francisco facility delivers on them, that's the thing to watch. Don't be the beta tester for your own password vault unless you like excitement.
Now, that theme — an agent that lives with you, acts for you, reaches into your accounts — that's not just Ghost. That's the whole consumer AI land grab right now, and there were a couple of moves on that board today worth a fast look.
Instinct — the AI agent valued at ten billion as of its latest round — is bringing its agent into group chats. And here's the clever hook: it works even if your friends haven't joined Instinct. So you can drop the agent into a thread to plan a trip, snag event tickets, run a fantasy league, figure out who's bringing what to Thanksgiving. Founder Noah Shinn laid out the privacy plumbing: your personal agent asks permission before connecting to the group's agent, the group agent is siloed from your personal account, and if new people join the chat, any pending replies from your personal agent get held back before they're shared. That last detail is the thoughtful one — because the obvious disaster with a group-chat agent is it blurts out something personal to the wrong room.
This puts Instinct a step ahead of Meta's Muse, which doesn't have the group-chat feature yet, though Meta's older Meta AI can already live in WhatsApp and Messenger and Instagram threads. And it all sits against OpenAI launching its personal agents — the Dots — last week, which can work in collaborative spaces alongside humans. So the picture, and we touched on this theme the other day with the whole "how to choose your personal agent" conversation floating around: everybody's racing to the same form factor. Connect to your apps, use a browser for you, reach out proactively, now show up in your group chats. They're all converging. The differentiation is shrinking to privacy defaults and which messaging app they live in — and as one commentator put it, within six months they'll probably all just work in every messaging system. So the moat there is thin.
One more in this bucket, and it's a smart niche play. Two sisters launched a startup called Hot Girl Hotline — relationship and dating advice for young women, pitched by the reporter, Sarah Perez, as Dear Abby for the AI era. Now, your first instinct might be to roll your eyes. Mine was. But read the reasoning and it gets sharper. The founders, Baila and Sumrin Mudgil, start from a real and dark observation: people are already pouring their personal lives into general-purpose chatbots, and ChatGPT has faced a wave of lawsuits over alleged manipulative conversation tactics that plaintiffs say contributed to serious mental health harm, including, in some cases, suicide.
So their whole product design is a reaction to that failure mode. Baila's line: "We're not trying to be an AI companion. A companion is really going after solving loneliness, escapism, boredom — and you can rely on a lot of emotional dependency from that AI companion." Instead, they built something deliberately the opposite. The conversation's Socratic — it asks you questions to help you get to your own answer. If flags come up, it refers you to a crisis line. And here's the part I think every builder working on anything emotional should underline: the AI knows when to end the conversation. It doesn't keep asking follow-ups to drag you down a rabbit hole. It pushes you to go take action in the real world rather than sit and ruminate.
That's product design as harm reduction. Every engagement-maximizing instinct in the industry says keep them talking, keep them in the app. These two built a thing whose stated goal is to get you off the app and out the door. Nine ninety-nine a month, hundreds of active paying users already, bootstrapped. Will it be a giant business? Who knows. But it's the rare case where "we care about safety" is baked into the mechanics instead of stapled on in a blog post. I'll take it.
Now let's shift from how people talk to agents to how the industry gets paid, because two moves from OpenAI today tell you where the money machine is heading.
First one: OpenAI is building out advertising inside ChatGPT. The announcement introduces a new visual ad format in ChatGPT and expands measurement tools, attribution partnerships, and brand suitability for advertisers. That's the whole summary I've got, but you don't need more to see the direction. The company that trained a billion-plus people to ask it questions is now building the apparatus to sell answers-adjacent ad space. This is the oldest story in tech with a new coat of paint — build the audience first, monetize the attention second. The question that matters for you, whether you're a builder relying on ChatGPT or a brand thinking about where your dollars go: once there's ad inventory inside the assistant, the incentives around what gets surfaced start to bend. Keep an eye on it. "Helpful answer" and "advertiser-friendly answer" are not always the same sentence.
And it dovetails with a piece we touched on in the resurfaced reporting today — TikTok rolling out an AI shopping assistant and one-click checkout. Now, I'll be straight with you: the TikTok Shop launch itself dates back to 2023, so I'm not going to pretend the whole thing is brand new. But the AI shopping assistant and the in-app one-click checkout are the fresh moves, built with partners like Shopify and Stripe. Same logic as OpenAI's ads play from the other direction. TikTok wants to turn its impulse-driven feed into a place where you ask the product questions and buy on the spot, so you never leave to go ask ChatGPT. Everybody's trying to be the last app you touch before your credit card comes out. The assistant is the new storefront. For founders building commerce tools, the lesson is that the discovery layer and the checkout layer are collapsing into one conversational surface, and the platforms want to own both ends.
The second OpenAI move is the one with teeth for anyone who ships AI-generated text, and this one is current. OpenAI is going to start watermarking ChatGPT's and Codex's text in the European Union to comply with the EU AI Act. The Act's transparency rules took effect August second and require AI companies to mark AI-generated content in a way other systems can identify.
Here's how it works, and it's clever. It's not a visible symbol. It subtly shapes the model's word choices — nudges the next-word predictions using a secret key — so it leaves a statistical pattern a reader can't see but a detector can pick up. And because it lives in the words themselves, it travels with the text when you copy and paste. OpenAI published a technical report on the method, called textGrain, co-written with researchers from the University of Pennsylvania and Yale. The company says it doesn't identify the user and didn't meaningfully change model performance.
Now the honest part, and credit to OpenAI and to the reporter, Aditya Mehta, for putting it right out front. Can you strip the watermark by editing? Yes. In one test, replacing ten percent of the words with synonyms dropped detection from about ninety-two percent down to sixty-six percent. Short passages, math answers, and translated text are harder to detect at all. OpenAI even cautions that a missing watermark does not prove a human wrote it. Their words: watermarks can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.
So let's be clear-eyed about what this is and isn't. It's rolling out over the coming weeks to EU users on all plans. Developers using the API anywhere in the world can switch it on for select models starting now, but it's off by default, and OpenAI is explicitly not making it a global default. And why not? The paper trail answers that. The Journal reported back in 2024 that OpenAI had built a text watermark before and sat on it, partly worried users would just switch to rivals who didn't watermark. That's the whole tension in one sentence. The first company to watermark globally and by default risks losing the customer who doesn't want their output flagged.
This comes two months after Anthropic said it would watermark Claude's text worldwide — and that drew backlash from some Claude users who argued they supplied the instructions, the context, the decisions, while Claude was just the tool. And that's a fair argument, honestly. If I use a word processor, nobody watermarks my essay as "machine-assisted." But the regulators in Brussels have decided provenance matters, and so here we are, with a watermark that the company itself admits a determined person can scrub with a thesaurus.
My take for builders: don't architect anything around text watermarking as a reliable signal, in either direction. A watermark present means an AI probably touched it. A watermark absent means nothing. If your product's trust story depends on detecting AI text, you're building on sand, and OpenAI just told you the sand is wet. What this really is, is a compliance artifact — a way to satisfy a European transparency rule — not a working truth-detector. Treat it accordingly.
Alright, let me give you the palate cleanser of the day, which is pure industry theater, and I mean that as a compliment because it's genuinely instructive about how this ecosystem polices itself. This is the Factory story, reported by Julie Bort, and you have to follow the bouncing ball.
Factory's an AI coding startup, recently valued at five billion. Last week its co-founder and CEO, Matan Grinberg, got into a very public airing of dirty laundry. He alleged he'd fired an investor, Chris Degnan, from a board advisor role, because he feared Degnan may have shared confidential information with Factory's biggest competitor, Cognition. The trigger was Degnan moving to Cognition as its chief revenue officer. Degnan disputed the whole story, said he resigned, and said he'd actually turned down a competing job offer from Grinberg.
So far, so soap opera. But then a who's who of the industry piled in, and the most surprising entrant was Vinod Khosla. Now here's the beautiful conflict: Khosla's firm is an investor in both Factory and Cognition. And instead of staying out of it like a diplomatic man with money on both horses, Khosla called Factory a desperate, "struggling second tier competitor," and accused Grinberg of lying and impugning Degnan's character. That's your own portfolio company you just called second-tier in public, Vinod.
And here's where it gets interesting as a business signal rather than just gossip. Menlo Ventures then came out and announced, proudly, that it had invested in Factory — part of that five billion dollar round. A source said it was significant, not a token gesture, and Menlo would've put in more if there'd been room on the cap table. Menlo's Matt Murphy is the guy who famously bet on Anthropic when it looked like an also-ran to OpenAI, and he's been on a tear since. And notably, Menlo is not an investor in Cognition.
So what you're watching is a reputational market correcting in real time. One powerful investor trash-talks a company; another powerful investor steps up with a full-throated, money-where-my-mouth-is endorsement to counteract it. Some VCs are now calling Factory "the next Anthropic." Sequoia's Shaun Maguire, also a Factory backer, called Menlo "on a tear."
Here's the useful lesson under the drama, and it's for founders especially. In this market, your cap table is a public reputation signal, and the people on it will fight about you in the town square. Khosla being in both Factory and Cognition is exactly the kind of overlapping bet that looks smart in a spreadsheet and gets ugly the moment the two companies collide. If you're raising, understand that an investor who's also in your competitor is not neutral, and may not act in your interest when the knives come out. And if you're just watching, understand that a loud endorsement right after a loud attack is itself a trade — Menlo is talking its book, Khosla's talking his. Nobody in that square is a disinterested referee. Read all of it as positioning.
Now let me pull us out of the money stories for a minute, because there's an infrastructure story that got resurfaced today that I don't want to pass without comment, even though the underlying event isn't brand new.
This is reporting on Russian drones striking Ukraine's data centers. I want to be careful on timing — the piece references events going back through the year, including an incident from February, so I'm framing this as a recent report pulling together a campaign, not a bulletin from this morning. But the substance matters for anyone who thinks about where their compute physically sits.
The reporting, from Jeremy Hsu, describes a surge in Russian jet-powered drone strikes on Ukrainian data centers and telecom infrastructure, concentrated on Kyiv. Ukraine's digital ministry reported a hundred thousand households in the Kyiv region temporarily losing internet from the initial strikes. And here's the technical detail that should sober you up: Zelenskyy said Ukrainian air defenses intercept more than eighty-five percent of the conventional Shahed-type drones, but that interception rate falls to just fifty or sixty percent for the newer, faster, jet-powered versions. Faster drone, lower intercept rate, and suddenly the data center that was behind the air defense umbrella isn't fully behind it anymore.
Ukraine's response tells you the playbook: they've moved some data centers out of the country, invested in underground telecom, and officials are urging people to get Starlink terminals before winter. But as the former deputy minister put it, you cannot move everything underground. And the piece notes this isn't a one-off — when the U.S. and Israel struck Iran earlier this year, Iran's retaliation hit multiple Amazon data centers in the Gulf, and Amazon acknowledged the permanent loss of some customer data.
I bring this up on a show for builders because we spend all day talking about data centers as financial objects — valuations, compute deals, AI factories. This is the reminder that they're also physical objects in a physical world, and that world now includes cheap, fast, one-way drones that can take out a communications backbone. "Where does your compute live, and what happens to your customers' data if that building is gone" has quietly become a real business continuity question, not a hypothetical. Amazon losing customer data permanently to a missile strike is the kind of line that should make every founder with a single-region deployment think twice. Resilience isn't just redundant servers anymore. It's geography.
Okay. We've done money, silicon, agents, ads, drama, and drones. Let me close somewhere completely different, because you've earned it, and because it's a good story.
New forensic evidence supports Egyptian retainer sacrifice. This is Jennifer Ouellette's piece, and it's about as far from token economics as I can get you, which is the point.
Here's the setup. Egyptologists have argued for a century about whether the earliest pharaohs of the First Dynasty were buried with servants and officials — killed to serve them in the afterlife. Human sacrifice, essentially. But there's been almost no hard evidence, because the oldest tombs were plundered or dug up hastily a hundred years ago. A new paper in the journal Frontiers in Environmental Archaeology reassesses a collection of skulls excavated in the 1920s and concludes that over a third show evidence of blunt force trauma.
The author's a bioarchaeologist at Purdue named Roselyn Campbell, and I love how she frames the state of the debate. Her line: "People tend to fall into one of two camps. Either retainer sacrifice definitely happened, end of discussion, or it definitely didn't happen, end of discussion. But nobody's actually looked at the bones in a very long time." That's a sentence that applies to about half the arguments in tech, by the way, but I digress.
So she goes back to forty-six skulls, originally dug up by the famous British archaeologist Flinders Petrie, stored at Cambridge, from a cemetery full of subsidiary burials around the royal tombs — graves that look like they were filled and sealed all at once, which is unusual. Petrie, for what it's worth, kept meticulous notes, which is the good news. The bad news, and Campbell doesn't soften it, is that Petrie was a deeply racist man, a proponent of so-called dynastic race theory — he literally believed native Egyptians couldn't have built the monuments and some outside culture must have come in. He also coated the skulls in paraffin and dumped the rest of the bones mixed together in a pit. A product of his time, as she puts it.
And here's the part I find genuinely moving, because it's about how knowledge actually works. Petrie didn't note any fatal trauma on these remains. Campbell doesn't think he missed it because he was sloppy — she says he was a careful observer. He just, in her words, didn't know what he was looking at. It's only in the last few decades that forensic science learned to tell the timing of a bone injury — how dry bone versus fresh bone versus six-day-old bone breaks differently. If you don't know that, you see a crack and you assume it happened during the dig. She found linear fractures in eighteen of the skulls, about thirty-nine percent — the kind of blunt-force fracture that would've knocked someone unconscious and likely killed them. Some showed sharp-force trauma too. Not from a mace, she says, because a mace makes a different kind of dent — more likely a flat, wide, lighter weapon. A board. Maybe a sandbag, like Petrie guessed. A makeshift club.
And then the reason, which is the chillingly rational business logic of it. Campbell: "When you establish a new form of government, especially one that relies on a single divine or semi-divine ruler, you have to convince a huge population that you're in power, and control over life or death is one of the best ways to do that." But — and this is the kicker — she notes it dies out very quickly, because it's economically unsustainable. Her words: "It is economically very costly to kill off hundreds of people every 20 to 40 years. That's not really sustainable." Even human sacrifice, it turns out, has a burn rate.
The reason I wanted to end here isn't just the macabre fun of it. Campbell said something about her own field that I think every one of us building anything should tape to the wall. She said: "I'm sure in 100 years people will be shocked by the things that we have done. This is just part of science. It continually self-corrects. Our methods may be the best we have now, but if science works as it should, in 100 years, they'll have much better methods. We want to preserve as much as we can for those future scientists."
A careful man a hundred years ago looked right at the evidence and couldn't see what it was telling him — not because he was dumb, but because the tools to read it didn't exist yet. Hold that next to everything else we talked about today. The benchmark claims we can't independently verify. The watermark that admits it can't really detect anything. The box that promises it'll protect your data. We are all Petrie right now on something. We're looking at the evidence with the best tools we've got, and some of what we're confidently concluding is going to look like a crack we assumed came from the shovel. A little humility, in a week where everything's getting marked up every two months, is not the worst thing to carry out the door.
That's the menu, kiddos. Verify the benchmarks, read the cap table as a loaded document, keep your hand on your wallet when a valuation doubles in eight weeks, and don't let your computer diagnose your flu just yet. This has been Barely Possible, I'm Tony DeLuca, and I'll see you back here next time. Be good to each other.