A financial research firm spun up 300 Claude sub-agents, designed 318,000 candidate proteins against scorpion venom, and shipped the results — all while remaining a financial research firm. Plus Meta's Muse hits 10 million downloads, and Anthropic's leaked prospectus gives critics an $80 billion number that isn't really a number.
AI doesn't just change how you work — it changes what you work on.
Citrini Research is a thematic equity shop. They thought life sciences were undervalued, couldn't get a straight answer from experts about whether AI drug discovery was real, and so — for under a couple hundred thousand dollars — they designed anti-scorpion-venom proteins themselves with Claude and had a lab validate them. They are not becoming a biotech company. But the fact that this was economically viable at all is the story: a new category of orthogonal work, where firms do things they could never previously justify in order to do their actual job better. That's the opportunity side of the efficiency-versus-opportunity divide.
The four-year-old startup — $20 million raised last year — builds agents that streamline lab workflows like compound manufacturing monitoring and trial tracking, rather than proposing new compounds. The team joins GPT-Rosalind, OpenAI's life-sciences model, now out of research preview as a full commercial product alongside the Rosalind Workbench agent suite.
OpenAI is explicitly building tooling for the existing pharma industry, citing partnerships with Novo Nordisk and Moderna. Anthropic, per Reuters, quietly stood up its own physical biology lab. One sells intelligence to the industry; the other is starting to use intelligence to do the work itself.
— Eric Kodera Abrams, Anthropic head of life sciences. Anthropic's head of life sciences confirmed the lab after Reuters broke the story, describing the approach as typical of most biotech companies — some work in-house, some with external partners. The news landed right as the X-risk discussion was surging again, though further reporting showed a relatively limited research remit.
— Citrini Research, in their write-up of the experiment. Citrini Research — a thematic equity and global macro shop — concluded life sciences was a secular story at cyclical lows but couldn't get a straight read on AI drug discovery, since experts range from "AI has never made an important drug" to "AI will cure everything in five years" depending on degree of AI-pilledness. So they used the model to design proteins, contracted a lab to synthesize and validate them, and used the model to interpret the results.
Starting from a 35-page prompt Anthropic had already tested, Citrini's team ran 300 Claude sub-agents to generate 318,000 protein candidates, shortlisted 211, and had an independent lab confirm 17 that bind to four different scorpion toxins. None has yet been shown to actually neutralize the venom — and the whole thing cost under a couple hundred thousand dollars.
Scorpion stings hit about 1.2 million people a year and kill around 3,200, mostly children with limited access to medical infrastructure. Anthropic's argument: treatments here are expensive to develop and unlikely to earn back their cost — exactly the category a step-change in discovery cost could rescue.
Critics noted that proteins binding scorpion toxins was demonstrated in the literature 15-plus years ago, and that the thing that actually matters — neutralizing the venom — wasn't shown. Citrini's response: they knew the haters were coming, weren't going to wait another three months to publish, and still intend to pursue neutralization testing.
Google DeepMind's director of AGI economics Alex Emoss credited Citrini with getting everyone talking about AI for science while also illustrating where the constraint sits: verification and testing. Knowing which half is which is how you know where to invest.
Citrini's own conclusion: their virtual biologists compressed a year of work into a couple of days, but DNA still had to be synthesized, proteins produced and purified, toxins obtained, controls established. As they put it, the digital invention was now set to create work for human scientists. Former OpenAI researcher Will DePue, on the two charts side by side: "God, these two charts really hurt my soul."
Whether frontier labs start internalizing scientific results instead of selling tools is a live question, fueled by the unverified reports that Anthropic leadership expects to be the only private company left. Citrini's experiment shows how many steps, people, and institutions sit between an AI-designed molecule and a real product. Competition is inevitable; the idea that it arrives across whole industries overnight is not.
The popular theory is that disease cures are the only way to turn sentiment around. But AI-accelerated discovery is more likely to simply become the norm than to arrive as one miracle compound — and as Will Manidis noted, if the industry thinks curing cancer is enough, it should look at polling on big pharma. NLW's conclusion: there's no magic bullet, so labs should work on what they believe has the biggest impact regardless of the headlines it generates.
Citrini didn't pivot to biotech — they ran a wet-lab experiment because it was the cheapest way to answer an investment question, and they remain a financial research firm. Expect far more of this: firms doing things they previously couldn't justify at all, purely because it makes their core work better. That's the opportunity side of the efficiency-versus-opportunity divide.
Only about a dozen apps have ever grown that fast, and the other 11 are almost entirely mobile-gaming smash hits like Pokémon Go, Fortnite, and PUBG. The only non-gaming comparables are Disney+ and Meta's own Threads.
ChatGPT, Claude, and Gemini all eventually topped the download charts, but each launched as a web app first and missed the explosive early phase on mobile. Muse is the first assistant to run the pure mobile playbook from day one.
Per Sensor Tower, Muse's share of Meta's total advertising spend jumped from 23% in September to 39% in October, making it the 21st-largest US brand by impressions for the month. Growth this fast is bought as much as it's earned.
30% at seven days and 23% at 14 days, versus roughly 25% for ChatGPT and about 5% at 30 days for the average iOS app. The critical takes compare Muse to Instagram, Facebook, and Google — some of the stickiest consumer apps ever built. DAUs hit 2.5 million last week, up 15% week over week.
Consumer products retain either through desire (Instagram, TikTok, games) or dependency (email, calendars, banking). As Signal argued on X, personal AI agents occupy a weird third space: not emotionally rewarding enough to be entertainment, not reliable enough to be infrastructure — and outside pure research tasks, one failure is often enough to lose the user.
A design decision in how Meta constructed share URLs made them easy to enumerate. Security firm Clarot pulled 6,800 links and found a small business owner's loan application, a real estate investor's internal performance report, and a national marketing company's internal revenue review. Meta has temporarily pulled the feature.
The Android Auto assistant can now reach the rest of your Google services — asking whether you'll miss your flight pulls details from your inbox while checking Maps traffic, and it can route you to the coffee shop in a photo from last week. Tesla has integrated Grok into its vehicles too.
NLW can see this becoming so normal we forget cars ever lacked it. He can equally see people deciding the integration isn't worth it when the phone with the same capability is already sitting right there on the console.
The Wall Street Journal got Anthropic's updated prospectus draft with audited H1 financials. Roughly $72 billion of the $80 billion net loss is an on-paper stock-grant accounting change, leaving $8.7 billion as the real operating loss. Financial podcaster Jack Farley: when a company issues convertible notes and warrants and the market cap 1000Xs, the on-paper loss is massive, not fundamental.
Firm, uncancelable commitments are around $450 billion spread across the next decade — high, but far more tenable than the headline. There may also be double-ordering, with soft commitments to NeoClouds in hopes they actually finish facilities. The Journal's line: Anthropic needs more computing power than the market can provide.
Anthropic booked $16.3 billion in H1 against a claimed $65 billion annual run rate as of July — both can be true, but the booked number lands lower than people expected. The $16.8 billion Q3 figure disclosed at Investor Day (45% QoQ growth) is unaudited and won't make the prospectus.
The $80 billion figure gives AI critics fresh podcast material, but nobody is actually surprised that Anthropic issued a ton of options and will spend north of half a trillion on compute. Like the SpaceX IPO, this offering won't be priced on the last few quarters — it'll be priced on what people believe about the next decade.
Reuters reports some investors pushing back on the rumored two-trillion-dollar valuation and signaling comfort at $1.5 trillion. Worth noting: both numbers come from investors, with nothing attributed to Anthropic or its bankers.
Sources tell Reuters OpenAI is trying to wrap a $30 billion raise at a $1.4 trillion valuation within weeks — ahead of the Anthropic IPO and whatever conditions it imposes on the market — while capitalizing on growth from GPT-6 Astra. Even for this sector, that's an unusually fast deal.