AIが「仕事」の意味を変える：金融リサーチ会社がサソリ毒の解毒タンパク質に挑んだ理由

Pretty much no one at this point needs to be convinced that AI is changing how work works. Certainly it changes the speed and efficiency with which we do existing tasks, and the magnitude and kind of those changes seems to grow over time. For example, while some of us may have used AI to write emails a couple of years ago, now many of us have personal agents sitting in our email perpetually, notifying us when there's something we actually have to respond to, often after having drafted the response itself. And yet, as I've always argued, that is just the tip of the iceberg. when it comes to how AI will actually change work. A great example of that comes in today's story, where a well-known financial research firm turned itself temporarily into a bio-research lab and collaborated with Anthropic to come up with some potential answers to scorpion venom. More than a novelty and more than a clickbait headline, I actually think that this can tell us quite a bit about where AI at work is headed next. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in.

First of all, thank you to today's sponsors, KPMG, Sea Data, Granola, and Blitzy. To get an ad-free version of the show, go to patreon.com slash AIDailyBrief, or you can subscribe on Apple Podcasts. If you want to learn more about sponsoring the show, send us a note at sponsors at aidailybrief.ai. And by the way, if you are just looking for all the links that I talk over on any given day, there are two ways you can do it. First, you can go into the episode companion page on the website for any episode. You're going to see everything split up by key numbers, key topics, key quotes, but you can also just subscribe to our newsletter, which is also linked from the website, which is going to have summaries of everything I discuss, as well as the key links. Check it all out at aidailybrief.ai, but with that out of the way, let's dive in. We kick off today with an update on the personal assistant theme, where Meta's Muse has reached 10 million US downloads in just 40 days. That is a very quick pace. In fact, it makes Muse one of only about a dozen apps to have achieved that milestone in a similar period of time. The other 11 are almost exclusively mobile games, specifically the smash hits of mobile

gaming like Pokemon Go, Fortnite, and PUBG. Only two non-gaming apps have grown this quickly, which is Disney Plus and MetaZone Threads. Curiously, none of ChatGPT, Claude, or Gemini grew that quickly as a mobile app, even though they eventually made it to the top of the download charts. Each chatbot was introduced as a web app first, with the mobile app coming later and missing the early phases of explosive growth. With its first month in the books, Sensor Tower had several other interesting data points about the Muse rollout as well. Meta has allocated 39% of their total advertising spend for October to Muse, which is up from 23% in September. That has boosted Muse to become the 21st largest brand by impressions for this month in the United States. Sensor Tower data showed that Muse has a 30% retention rate after 7 days and a 23% retention rate after 14 days, which compares to ChatGPT's 7- and 14-day retention rates at around 25% and the average retention rates for an iOS app at around 5% after 30 days. Interestingly, there has been some criticism around Muse's retention rate. However, that analysis compares Muse unfavorably against Instagram, Facebook, and the Google app,

which are some of the stickiest consumer apps ever created. It also looks like Muse is still early in its growth phase, with daily active users reaching 2.5 million last week, which was up 15% week over week. TLDR is, yes, they are spending a ton of advertising to push it, but Meta does seem to have a hit on their hands. The next question isn't going to be about short-term retention, but about where people ultimately find the app fitting into their life if it does. Signal on X writes, Most consumer products typically have one of two types of retention. One is desire-driven, which is just pure dopamine, like Instagram, TikTok, and games. They often don't even need notifications because you know there is a reward for your desire always waiting. The second is dependency-driven, like emails, calendars, even credit cards and banking. Most people have to do this. It's not something they necessarily desire. The challenge with personal AI agents right now is that they occupy a weird third space where they aren't emotionally rewarding most of the time to become entertainment like the first category, and they're not reliable or indispensable enough to become pure infrastructure, i.e. they fail, and there's a lot of friction. Outside of pure research-driven tasks, even one failure is often enough to not come back

for the utility value proposition. Now, he goes on to speculate a bunch about what opportunities that opens up, but I think even though we're seeing early positive signals, it is worth remembering that we are still very early days in terms of understanding the behavior surrounding personal agents. Muse also is not without its challenges. Earlier this week, researchers reported that the feature by which Muse allows users to share searches, conversations, and images had a particular challenge that were allowing other people to see those conversations. In general, those types of share features usually involve creating a unique URL that anyone who has that URL can use to load a page hosting the content. The way that sensitive information is kept from being accessed is the complicatedness of the URLs that theoretically can't easily be guessed. However, a particular design decision that Meta's engineers made around the construction of those URLs has made them easy to discover. As a result, wrote Ars Technica, researchers from security firm Cleroc said they were able to enumerate 6,800 links Muse users created for sharing conversations with each other. After opening a few hundred of them, they continue, the researchers found information that was clearly not intended to be publicly available,

including things like 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 said that they've temporarily taken down the feature and are working to harden it to make sure that doesn't happen. Next up, is another place that you're going to interact with agents your car? For almost a year, a basic version of the Gemini Assistant has been available in Android Auto. The agent mostly functioned as a voice control for maps and the stereo system, but could also send text messages while you were driving and find information on the web. It was not, however, connected to your personal context, as users have begun to expect from a personal agent. User feedback also suggested that it didn't really work that well and often failed at basic tasks. The new version adds the ability to access data from the rest of your Google-based services. Google says you can now do things like asking the assistant whether you're going to miss your flight while en route, and the agent can rifle through your inbox to find the necessary details while also checking traffic conditions in Google Maps. Another example was the ability to ask for directions based solely on a photo. Google said Gemini could now direct you to a coffee place your friend took you to last week

based only on a picture of you enjoying your espresso in a booth. Now, I'm still quite unsure about how important these features are ultimately going to be. On the one hand, I could see this type of interaction with your car becoming something that is so common that we have a hard time imagining that we didn't always have those features. And yet, at the same time, I can see a lot of people deciding that the setup and integration into another device isn't worth all the hassle when they can already just pick up their phone that's probably already sitting right there on their console and get the same information. Whatever the case, it seems that companies are going to continue to explore that, as Tesla has also integrated GrokBot into their vehicles, and you have to imagine everyone else is going to follow. Moving over to markets, an updated financial disclosure from Anthropic has leaked, sparking another round of hand-wringing in the financial press. The Wall Street Journal got their hands on Anthropic's updated prospectus draft, which included fully audited financials for the first half of the year. Their reporting focused on two big scary numbers an billion net loss for the first half of the year and up to billion in compute spending over the coming years The billion net loss is substantially larger than the billion loss for 2025

reported from the previous edition of their prospectus. But both numbers stem from the same accounting treatment of stock grants. Around $72 billion of the loss was an on-paper accounting change, leaving $8.7 billion as the actual operating loss for the first half. Responding to the claim that this is the biggest loss of any company seeking to go public, financial podcaster Jack Farley commented, this would be worrisome if the $80 billion accounting loss by Anthropic were economically real in any way. Good news, it isn't. When a company issues convertible notes and warrants and the market cap with 1,000 Xs, the on-paper loss from being short calls is massive, not fundamental. Now, for others, the bigger deal is the $750 billion in compute spending. The reporting suggested that of that, firm uncancellable commitments were around $450 billion and come due over the next 7 to 10 years. In other words, it's a very high level of spending, but one that's far more tenable spread out over a decade. There also could be an element of double ordering in the larger number, with Anthropic making soft commitments to neoclouds in hopes they can actually complete new facilities.

The journal commented that, quote, Anthropic needs more computing power than the market can provide. The one other number that people are laser-focused on is the revenue number of $16.3 billion for the first half. Anthropic had claimed a revenue run rate of $65 billion for the year up to July, and while both of those things can be true, actual booked revenue yields a lower number than might have been expected. At last week's Investor Day, Anthropic said that they added another $16.8 billion in revenue in Q3 for a 45% quarter-over-quarter growth rate, but that unaudited whisper number won't be included in the prospectus when they IPO. The rest of the reporting largely covers risk disclosures that were already in the press, things like alignment issues, the data center backlash, and customer concentration. Realistically, these leaks mostly have served to confirm pre-existing biases. The $80 billion net loss gives AI critics some new material for their next podcast, but it strikes me as fairly difficult to imagine that anyone is surprised that Anthropic issued a ton of stock options and will spend more than a half trillion on compute when all is said and done. Like the SpaceX IPO, this stock offering is not going to be about the past few quarters.

It's going to be about what people believe about the next decade. Still, market skittishness matters, and perhaps explains why some investors are pushing Anthropic to moderate their targets. Last week, Reuters reported that some investors are pushing back on the rumored $2 trillion valuation, and we'd be more comfortable at $1.5 trillion. Keep in mind, both numbers are coming from investors, and we've heard nothing attributed to Anthropic or their investment bankers. Meanwhile, OpenAI is rushing to complete their latest round of fundraising ahead of the Anthropic IPO. Reports of the round first circulated towards the end of last month, suggesting OpenAI was seeking $30 billion at a valuation of $1.4 trillion. Now, sources tell Reuters that OpenAI is trying to close the deal in weeks to get ahead of the Anthropic IPO in whatever conditions that brings to market. The sources said that OpenAI is also keen to capitalize on recent growth after the release of GPT-6 Astra, but this is an unusually fast deal even for this crazy sector. The acceleration, writes Reuters, underscores how closely the two AI developers are competing for capital as both pursue trillion-dollar price tags and growing concern about rising interest rates, significant volatility in technology stocks, and rising scrutiny of

AI valuations. These money conversations are going to get nothing but more frenetic to end the year. However, for now, that's going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always, and until next time, peace! if you're leading ai inside an enterprise you already know that the gap right now isn't capability but execution that's why kpmg's you can with ai is back with a new season featuring conversations with leaders like sarojit chatterji of emma may habib of writer mckesson cio ellery fisher and others focused on practical execution what's working what's not and what it actually takes to move from pilots to real scaled impact across strategy data readiness governance workforce and value. And of course, it's co-hosted by me, Nathaniel Whittemore. Go listen and subscribe at www.kpmg.us slash AI podcasts. That's www.kpmg.us slash AI podcasts. This episode is brought to you by C-Data. Companies rolling out AI keep hitting the same wall. Token spend is climbing, security is holding agents back, and no one knows how AI

is truly being used. The root cause? AI is cut off from the systems and context your business runs on. That's why CData built Connect AI Gateway, one control point between your AI and your systems. Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs. It governs down to the record and lets you finally see your org's complete AI usage in one place. Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction. See how at cdata.ai. When I'm in a meeting, I'm fully in it. I'm thinking about the iteration and creative back and forth it takes to actually push a goal forward. What I'm not thinking about is capturing takeaways, tracking to-dos, or any of that. And that's where Granola comes in. Granola is an AI-powered notepad that captures what happens in your meetings and turns it into clean, structured notes with the decisions and action items pulled out and easy to find. There's no setup and no configuration. It just fits into how you already work.

For me, it means I get to stay in idea mode and Granola makes sure those ideas actually become action. Once you try Granola on a first meeting, it is hard to go without. You can try it totally free at granola.ai slash brief. That's granola.ai slash brief to get your time back. Here's why most legacy modernization projects fail. The AI doing the work can't understand code bases at scale. It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem. Blitzy solves this the way it solves everything. Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph. Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head. From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith to microservices transformations, all validated end-to-end. One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents. That's 9x compression.

Retire technical debt while accelerating your roadmap. See how at blitzy.com. That's B-L-I-T-Z-Y dot com. Welcome back to the AI Daily Grief. Today we're discussing what Scorpion Anti-Venom tells us about the future of AI. Seriously, that's actually the subject of the show. Or at least a set of stories that relate to that attention-grabbing headline and the themes that they represent are the broader subject of the show. AI news aggregator Andrew Curran posted earlier this week, drug discovery, new disease cures, advancements in pharma and biomedicine, along with direct dividends or shares as is currently being discussed, these are the only viable ways to turn public sentiment around And this is an idea that you see bandied about a lot that to actually get people to stop disliking AI we got to hop to and give them some of those disease cures that they've been promised. I actually think that the new activities around biology and AI bring up a much larger set of questions than just that one, although that one certainly matters as well. So first, let's talk about a couple of the stories

at the intersection of bio and AI, particularly this crazy one with financial firms at Trini Research working on scorpion sting cures. But the first actually comes from OpenAI. The company has announced that they have acquired an AI-powered drug discovery startup to build out their biotech offering. OpenAI announced on Tuesday that they'd acquired four-year-old AI biotech startup Orchestra Bio. Terms of the deal were not disclosed, and Orchestra last raised $20 million in venture funding last year. Orchestra develops AI agents designed to assist in drug discovery workflows. Rather than exploring new candidate compounds, Orchestra has largely focused on streamlining lab tasks like monitoring the compound manufacturing process and tracking drug trials as they progress. The orchestra team will work on GPT Rosalind, OpenAI's model variant designed for life sciences work. The model was first released in April as a research preview, but has now been released as a full commercial product. Alongside the model, OpenAI offers Rosalind Workbench, which is a suite of agents to help the model connect to scientific instruments and analyze data. In their press release, OpenAI wrote,

Our vision is for AI to help scientists across the full process of discovering and developing new medicines, exploring ideas, designing experiments, interpreting results, and coordinating the work needed to advance promising treatments. Bringing Orchestra Bio together with Rosalind is a step toward that vision. Alongside researchers and our partners, we aim to make more discoveries possible and help new medicines reach patients sooner. Now, to be clear, in contrast to some things we're going to discuss a little bit later, this is not about OpenAI building their own biotech firm. It's about building better tools to help the existing pharma industry. Also from the press release, they wrote, Realizing the potential of AI in life sciences requires close collaboration with the organizations working at the forefront of drug discovery and development. Our partnership with Novo Nordisk aims to translate its scale and scientific expertise into faster R&D and commercial execution while maintaining strong governance. At Moderna, AI is helping teams translate insights from complex data into experimental workflows with the potential to accelerate R&D. In other words, OpenAI is giving you the idea that they want to better support those industries'

use cases of AI, not do that particular work with AI. Many were enthusiastic about this type of work. Ran Longevity on X writes, I really hope the big labs keep focusing on drug development. That is where we can really change the world. And Precogenetics' Parmita Mishra thinks this is part of a trend, sharing the news with the added caption, the biohype begins. However, some think that this is just an incremental step. Today, Vashista writes, The only path to profitability for the frontier labs is becoming a biotech company eventually. And that position got a bit more credence last month when Reuters reported that Anthropic had quietly set up its own biology lab. So to be clear, in contrast to OpenAI building better systems for existing labs to use, visit Anthropic setting up their own lab to do physical biology work. After the Reuters news broke, the lab's existence was confirmed by Anthropic's head of life sciences. Eric Guderer Abrams said, We believe that to do biology, the final test is still and will be for a while in real lab work. We absolutely are doing that today. and I would describe our approach as being typical of what you would see in most biotech companies,

where there's some amount of that that we're doing in our own facilities and some amount of that that we're working with external partners on. Now this news broke right at the time that the X-Risk discussion was surging once again, leaving many to wonder if it was exactly the right time for Anthropic to be building a wet lab. However, with further reporting, it became clear that it had a relatively limited remit that had pretty specific limitations in the types of research it could do. Still, this was also the time that there were reports that leaders inside Anthropic believed that at some point in the future, they would be the only company left. And given how much this felt like a move into a totally different area and a fundamentally different model for AI, i.e. not just providing the intelligence, but using that intelligence to do things, there was a lot of interest, let's say, to see how it would evolve. That brings us, however, to the story of Citrini. Back in July, Anthropic writes, a team from Citrini Research had approached us with an interesting question. Could Claude and AI drug discovery more generally enable a group of motivated non-experts to tackle a genuine problem in human health?

Now, Citrini Research, you might be saying to yourself, haven't I heard of them in the context of newsletters about financial issues? And the short answer is yes. When you go to citriniresearch.com, it links directly to their sub-stack, which it sums up as providing insights on thematic equity investing in global macro trading. So what the heck are they doing trying to design new drugs? According to Citrini's long blog post about this work, the origins of the experiment came from, as they put it, searching for opportunities outside of AI infrastructure where they were concerned about markets crowding. One area that seemed to them to have asymmetric opportunity was life sciences and healthcare. They wrote, The sector had endured the reversal of the pandemic boom in a difficult funding environment as investors found more exciting places for their money. Over the past half decade, it has been a graveyard for returns. While everyone focused on which Japanese semi-doodad producer would sell out next, the cycle had seemingly turned. However, the stocks had yet to climb in any meaningful way, and we kept hearing about AI's promise in drug discovery. A neglected sector with the chance for AI upside, it had the markings of our favorite type of setup,

a secular story at cyclical lows. And also, they said, given that their last big AI post had been effectively about how AI doing too well even could also destroy the economy, the opportunity to potentially tell an optimistic story also seemed exciting to them as well. With tongue firmly in cheek, they wrote, Finally, it was our chance to write an article offering a reprieve from the threats of a dystopian AI future dominating the news. Not that we have ever been guilty of fueling fears about the future of AI, of course. The one problem, they continued, none of us could claim to know anything about AI as it related to biology. Our usual process, they wrote, would involve deep research and dozens of calls with experts. But survey a dozen experts in biology, and you're going to hear both, AI has not created a single important drug and it won't in the future either. Two, AI will cure all diseases in the next five years. And everything in between, depending on each individual's degree of AI-pilledness. We weren't going to get a clear answer by simply asking. We needed to try it for ourselves. So we did. We used the model to design proteins, contracted a lab to synthesize them, and conduct experimental validation and again using the model to interpret the results.

Now this gonzo approach to biotech seems to have indeed yielded some results. The Citrini folks were able to use Claude to design protein binders for use against Scorpion neurotoxins. The group started with a 35-page prompt that Anthropic had already tested in a previous study, modified it to taste and set Claude on its drug discovery work. 300 subagents worked on the protein design, generating 318,000 total candidates, and selecting 211 promising options for further testing. 17 were found to bind to four different Scorpion toxins after independent testing in another lab. None of the proteins has yet been shown to stop the neurotoxin from working but it a promising initial result Also all this work came in under a couple hundred thousand dollars which while not cheap also isn prohibitively expensive for trying to go after an under on issue that affects a million people a year and actually kills 3,000. Now, what was interesting about the way that Citrini did this is that almost immediately after their ragtag group of stock analysts downloaded Clod for Science, they ran headfirst into the biological guardrails. Rather than trying to do some extensive application process, they just reached out to Anthropic and asked if they could do it sitting in the

company's offices. Anthropic was interested right away because this is exactly the type of neglected disease that they think AI has a particular opportunity to benefit. Scorpions sting about 1.2 million people per year, which leads to around 3,200 deaths, and most of those come from children in parts of the world that have extremely limited access to quality medical infrastructure. The problem, Anthropic writes, is that new treatments are expensive to make and also unlikely to earn back the cost of developing them. And this is where the potential step change in the cost of developing new treatments that AI enables could come in. There was, of course, plenty of skepticism when Citrini dropped the report. WizPlumOnX writes, So you achieved something demonstrated in the literature 15 plus years ago, proteins that bind scorpion toxins, but didn't demonstrate the thing that actually matters, neutralizing the venom. Maybe make that the headline instead of burying it in 10,000 words of Claude prose. Citrini responded to that one saying, LMAO, everyone involved in this said, Well, as soon as we publish this, there will be a handful of haters no matter what we do. We weren't going to wait another three months to publish. Still want to pursue the neutralization testing, though.

Google DeepMind's director of AGI economics, Alex Imas, wrote, Citrini has an incredible talent for getting the public to pay attention to overlooked but important issues. Last winter, it was the famous essay on how AI-driven productivity could lead to a miseration. I disagreed with Parts, but thought it did an excellent job of starting a much-needed conversation about economic risks. Today, Citrini got everyone talking about the promise of AI for science. Huge speed-ups in hypothesis generation and computation, but also illustrates the bottlenecks, verification, and testing. We need to be clear-eyed about both in order to know where to invest. And in fact, one of Citrini's major conclusions was just how much physical world inertia there is, even when AI speeds up part of the process. They wrote, Our virtual biologists may have compressed a year's worth of work into a couple days, but biology works at the speed of life. DNA had to be synthesized, proteins produced and purified, toxins obtained in usable form, and appropriate controls established. Our digital invention was now set to create work for human scientists. Former OpenAI researcher Will DePook put two of the charts from Citrini Research next to one another and said, God, these two charts really hurt my soul.

64 hours to dry lab design, a multi-binding antivenom, 39 days for results, and 4 to 7 years to release. I think that the combination of the Citrini-Anthropic collaboration and this new OpenAI deal bring up three really interesting questions. The first is whether AI labs will start internalizing scientific results rather than building technology for others to use and what it means if they do. This is the question I said really picked up steam a couple of months ago. First, when reports came out, which by the way are unverified, that Anthropic leadership thought that they would be ultimately the only private company left in the world, implying that at some point they do everything, including all the things that their customers do currently, and which picked up more steam a month later when the Anthropic wet lab was announced. Yet what this Atrini research shows is that there are so many parts of this process that even adding a wet lab to take theoretical AI-initiated experiments a little bit farther doesn't change the fact that there is an immense number of other steps and people and institutions involved in actually bringing something real into the world. I don't think this at all settles the question of whether at some point

AI customers will find themselves competing with the AI labs. In fact, that is going to be absolutely inevitable. But thinking it's going to happen all at once or across entire industries probably doesn't reflect the reality. The second question is this one about discoveries in public opinion. If AI actually does cure cancer, does that change people's perspective on AI? First of all, at least when it comes to anthropic experiments right now, they seem to be a lot less about these holy grails like curing cancer and more about addressing these medical projects which previously fell through the cracks of market incentives but which don't have to anymore thanks to AI. Anthropic might absolutely be correct that there is a lot more leverage there than the big flashy stuff, But a news headline like Anthropic Cures Scorpion Stings, which to be clear, this isn't even really close to, it's just the very, very first step in some solutions that could help this, still hits very different to Anthropic Cures Cancer. Second, even if AI is totally integral to curing major diseases, it seems unlikely that the way that that works is it just discovers a miracle compound that all of a sudden fixes everything in one fell swoop.

In fact, it seems far more likely that AI accelerated discoveries just kind of become the norm, everyone gets used to them, and they don't really have all that much bearing on whether they like AI or not. Will Manitis writes, if the AI industry thinks curing cancer is enough to turn public sentiment around, they should look at opinion polling on big pharma. To be clear, this does not in any way mean that I think that AI labs shouldn't be working on these issues. In fact, to me, my belief that very little they can do can, in the short term at least, actually really impact public opinion means they should be focused on working on the areas that they think are going to have the biggest impact, regardless of the headlines they create. As much as I would like there to be, I just simply don't believe that there is any magic bullet to the perception problems that AI has and is going to continue to have for some time. The third and most subtle, but in some ways most interesting question to me is how the Cetrini experiment specifically demonstrates the way in which AI will change not only how we work, but what we work on. Remember, the Cetrini experiment started from them doing what they normally do, which is

look for under-recognized opportunities in the markets from an investment thesis perspective. They went through their normal work process of examining different sectors in the context of returns, public perception, current market values, and basically came away believing that life sciences were undervalued, especially in the context of how AI could change those areas. The problem they ran into is that they didn't think that they could get a robust enough picture about the real potential of AI in life sciences just by asking a bunch of experts. and it was in that context that it became economically viable for them to just experiment for themselves. Now, Cetrini was very clear at the end of this that they are not about to become a biolab, that they are and remain a financial research firm. But I think that you're going to see a lot more of this type of orthogonal work, let's call it, where people and firms can do things that they couldn't do before that actually change how well they do their core work stream. This is another version of the efficiency versus opportunity AI divide that I always talk about. to go to original hands-on research deep in the biology and life science process.

If nothing else, it is a very entertaining story, and I suggest you go read the full Satrini Therapeutics research report. For now, I'm excited to keep exploring these themes with you, but that's going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always, and until next time, peace! you