An Exploded View publication
This week's theme
Anthropic, the Pentagon, the White House and the Pope each took a public position on AI risk today, with a few practical reads on building and testing underneath.
A 10-minute read · 9 storiesIn this issue
Anthropic's IPO filing (paperwork filed before selling shares to the public) reportedly warns AI may pose what it calls existential risks to humanity, per Reuters. This is the company behind the Claude models you use every day, saying in its own legal paperwork to investors that its product could be dangerous, worth watching as you build more on top of it.
Top AI company leaders are set to meet with President Trump at the White House, according to ABC7 San Francisco, right as warnings about the technology's dangers grow louder. Whatever comes out of that meeting could shape AI rules and how much the government spends on AI, both of which touch the tools and companies you already rely on.
The Department of Defense has stopped using Anthropic's AI after a court upheld a finding that it poses a supply chain risk, according to Homeland Security Today.
If you build on Claude or care about its reputation and availability, a federal agency pulling back over a security designation is worth watching.
Outside your usual reading: Pope Leo publicly said concerns about AI safety are not “fake news,” pushing back on the White House's more dismissive stance, according to AP News.
When a global religious leader with no tech industry stake weighs into the AI safety debate, it shows the argument has moved well past Silicon Valley.
This is on your desk because a genuinely top-tier open model, meaning the model's files are public so anyone can download, run, and change it themselves, changes the math on what you can build without paying for a closed API. Mistral, the French AI lab, has released a new model called Le Chonk that it says can compete with the best models out of the US and China.
"There are a lot of areas where the other labs will not focus that much," Guillaume Lample, Mistral's cofounder and chief scientist, told WIRED. "There are so many domains in which you can improve models."
Mistral calls Le Chonk the most capable open model built outside China, and "very, very close" to some of the top closed, pay-to-use models. That claim matters because of how it was built: the US government has accused Chinese labs of using distillation, training a smaller model to copy a bigger one, to catch up to OpenAI and Anthropic more cheaply. Mistral says it trained Le Chonk from scratch instead. By narrowing the performance gap and giving businesses an alternative to Chinese models, Mistral says Le Chonk removes most of the remaining reasons a company would avoid open source. "Mistral is still in the race of getting the best model," Lample says. "This is the main message."
Mistral has generally trailed OpenAI and Anthropic on performance, revenue, and release speed, with less money and fewer chips to work with. But it has had a strong run lately: in September it raised $3.3 billion at a $24 billion valuation, the largest funding round ever by a European tech company.
In June, the Trump administration put temporary limits on OpenAI and Anthropic distributing their models, citing worries they could be used for serious cyberattacks. The White House has reportedly even asked US labs to withhold unreleased models from the UK's AI Safety Institute, a group that previously helped test models for safety.
That reminder that the US government could cut off access to frontier AI on its own has opened a door for Mistral. A researcher at the Centre for European Policy Studies told WIRED in July that the EU's push for tech independence, combined with rising US hostility, unexpectedly puts Mistral in a favorable position despite performance that has not been spectacular. Lample argues the risk applies even in the US: a business that depends on a closed model risks losing access if it gets pulled. "If you use a closed model, there is no guarantee it will still be there tomorrow," he says.
This sits right where your own small business and AI interests overlap. American Express and Perplexity, the AI search company, launched tools that act like an AI chief financial officer for small US businesses, and the deal's design shows what a big financial company thinks small business owners actually need.
Late last week, AmEx (ticker AXP) and Perplexity launched AI CFO tools for US small businesses: about 10 “Skills” built inside Perplexity's “Perplexity Computer” product that turn a business's Amex Small Business Card data into cash flow forecasts, reconciliation (matching transactions to records), and tax season prep.
The catch looks deliberate: the tool can forecast a company's cash but cannot pay a single bill. The newsletter's framing is that whoever holds the actual payment authority will capture most of the value from AI CFO tools for small and medium businesses, and this deal's terms show AmEx intends to keep that authority for itself rather than hand it to Perplexity. The newsletter also promises a comparison against Anthropic's Claude for Small Business and Meta's “Muse” tool, and a signal that would show AmEx turning these forecasts into an actual lending product, but that part sits behind its paywall, so take it as teased, not confirmed.
Separately in the same edition: Stripe is buying Parafin, the lender behind small business credit for DoorDash, Gusto, Amazon, and Walmart Marketplace. The newsletter flags that Parafin's processor-agnostic model (it works with any payment processor, not just Stripe's own) is both the main thing Stripe is paying for and the part of the deal most exposed to risk. Even combined, Stripe's and Parafin's lending business would still fund only about $2 billion a year, well behind rival Square's roughly $7 billion a year.
Georgia Tech is launching a new, fully online master's degree focused entirely on artificial intelligence, according to its own news center.
It is one more sign that formal AI education is going mainstream, which matters for where the next wave of AI talent and skills comes from.
A one-table-join statistics trick can cut the number of users you need for a reliable test.
This is worth your time because it is a concrete, low-effort technique for any test you run on a product or a signup funnel, and it gets you a reliable answer faster for the same traffic. The problem: the number of users a test needs grows directly with how noisy the results are. Double the noise in a metric and you double the users you need, which can turn a two-week test into a four-week one.
Much of that noise has nothing to do with whatever you changed. Some customers spend $400 a month and others spend $4, and every test has to wade through that gap. Random assignment balances this out across many experiments, but never inside any single one, so results bounce and the confidence interval, the range your true answer probably falls in, stays wide. The usual fixes are blunt: run longer, which burns calendar time; accept a bigger minimum detectable effect, the smallest change you are willing to try to measure, which means missing smaller real wins; or cap outlier values, which just changes what you are measuring.
A technique from Microsoft researchers called CUPED (Controlled-experiment Using Pre-Experiment Data) fixes this for the cost of one extra data join. It builds an adjusted metric using each user's history from before the test started: take their in-experiment result, then subtract a weighted version of how far their pre-experiment value sat from the group average (written as Y minus theta times the gap between their past value and the group's average past value). Because the pre-experiment value was measured before anyone was assigned a group, the test cannot have influenced it, so the adjustment removes noise without changing the real answer. At a correlation of 0.7 between a user's past and future behavior, CUPED cuts variance, how spread out the numbers are, by about 51 percent, roughly halving the needed sample. At a lower correlation, the gain drops sharply. Stable, high-variance metrics like revenue per user benefit most; sparse yes-or-no metrics often barely move. The catch: CUPED needs user history, so new users and new metrics get nothing, and the pre-experiment value must be fixed before the test starts, not chosen afterward.
Etsy published a candid account of using this. In 2021 the team added CUPED to metrics like conversion rate and measured only about a 7 percent reduction in noise, shortening experiments by almost one day on average, far below the 50 percent some vendors advertise. The gap came down to the data: purchases happen rarely and correlate weakly across time, while visit-based metrics correlate much more strongly. Rather than give up, Etsy borrowed an idea from DoorDash called CUPAC (Control Using Predictions as Covariate) and replaced the single pre-period number with predictions from a more powerful statistical model, expanding from 3 input signals to more than 100. A model like that catches patterns a simple straight-line adjustment cannot.
The lesson is to measure before you promise. Before telling anyone this halves test time, replay CUPED on your last 20 finished experiments to see what each metric actually gains, and make sure your planning math uses the adjusted variance. Start with one high-variance metric, run simulated "A/A" tests, comparing a group against an identical copy of itself, which should show no real difference, to confirm the false positive rate stays in line with expectations, and check every pre-experiment input for leakage, meaning anything that could still change after someone is assigned a group. An AI model can do the grunt work: ask it to rank candidate pre-experiment inputs by how useful they would likely be, flag anything that could leak, list which metrics lack enough history for new users, and write and run the actual math rather than trusting its own arithmetic.
Here is some uncomfortable arithmetic: required sample size scales linearly with metric variance.via Business Analytics Review →
A hobbyist with no coding background has been building a game using AI only for code and art, while keeping every idea, art direction, and playtest call to himself, tracked by tagging each decision [USER] or [AI] in a file so the AI's own occasional false memories of what was decided get caught.
It is an honest account of where AI replaces skill and where it still needs you, useful for deciding what to hand off in your own projects.
Even IPO paperwork now has a section on the end of the world.
How this edition came together — from bookmarks and feeds to the page.
Where they came from
On the cutting-room floor — 31 links read but not run this week
Quality over volume: most links get a second look and a pass. The ones that made it earned their place.
The week's AI signal in 10 minutes — what happened, why it matters, and what to do with it. Curated by someone who actually builds, not a feed algorithm.