The day an AI developer bought a drug discovery firm── Vendors may become rivals. A valid signature is not valid content
Download the video (MP4, 13 MB)
Why would an AI developer buy a company that makes medicines?By the end, you will have the answer. You will also have the question pharma professionals should be asking now. Here is the conclusion up front. The question is not which AI to buy. It is whether you can judge for yourself who checked what, and how. A vendor can now become a rival. How much internal data you hand over matters more than before. AI is also gaining new ways into the workplace in place of people. And a file carrying a valid signature can still carry wrong content. So what must we check, and how far?
01Anthropic enters drug discovery
Source 富途牛牛 ／ Digital Health News ／ BigGo Finance
We start with where the makers of AI now stand. Until now, they sold tools to drug companies. This time, they seem to want to carry the work themselves, from finding candidates to developing them further.
Early drug discovery results
Anthropic published early drug discovery results from Claude. Several outlets reported the news in overlapping stories.
Buys a "one-person pharma"
It acquired a startup described as a "one-person pharma" company. The price does not appear in the headlines.
Partners with 6 organizations
It partnered with 6 life sciences organizations to advance drug development with Claude. Members and scale are unknown.
Research output, a small purchase and a group of new allies all landed at the same moment. I read them as one strategy, not separate stories. The aim seems to be to find candidates with its own models and then grow them in-house. Still, we do not know the price or who the allies are. The results so far are also at a very early stage.
For a drug company, a supplier may now also be a competitor. Data you feed into its models could end up helping its own research. Decide before signing whether you treat this maker as a research ally or as a rival. Where you draw the line on sharing data will weigh far more than it used to.
If a maker reaches this far into our field, how much should we trust its promises?
02Disclosure demanded before the IPO
Source Spectrum News ／ Los Angeles Times ／ Bloomberg.com ／ Fortune
Others are already asking that question. This developer is preparing to sell its shares to the public. Before such a sale, a company must explain its business and its dangers to those who might buy.
- Role
- U.S. senator
- Statement
- Asked CEO Dario Amodei to address existential risks before the IPO
- Role
- Newspaper commentary
- Statement
- Says the prospectus asks investors to take a lot on faith
Bloomberg reported Anthropic is spending millions to protect Amodei and employees
A lawmaker asked the company's leader to explain whether AI could threaten humanity itself. A newspaper column argued that the company's filing leans heavily on trust. Both voices point to the same thing. Setting a noble goal and proving it are two different jobs.
A filing made for a share sale is a document the company answers for. If its words are decoration, they will be challenged later. Here, the maker's promises take a form that can finally be checked. Outsiders can compare what is written with what actually happens.
For pharma professionals, this pattern is familiar. Every line in a package insert or an advertisement needs evidence. We expect claims of efficacy to come with supporting data. Apply the same yardstick to AI makers. Read their claims and their evidence separately. I think that is the basis for choosing a partner.
Beyond the makers who must explain themselves, how do we judge the AI that enters our workplace from outside?
03Doors opening to AI agents
Source The Decoder ／ 9to5Google ／ FourWeekMBA ／ Business Insider
Next, AI is entering apps in place of people. Large platform companies opened the way for it, one after another. Instead of a person clicking through screens, AI is starting to do the work directly.
| Who | What opened | Details / conditions |
|---|---|---|
| Android Studio | Supports 40+ third-party coding agents | |
| Gemini and Search | Connected Apps Platform lets developers build in their features | |
| Microsoft | Copilot Harness | Offered via Foundry; free until December 1 |
| Home Depot | Store conditions | AI agents must identify themselves and confirm they act for a user |
Developer tools, search and business platforms all opened at nearly the same time, in separate places. One company even offered a period without charge. The aim seems clear: win early users and gather them on its own platform. But the more doors there are, the harder it gets to tell what each arriving AI really is.
So a retailer that receives these AIs set a condition. An incoming AI must say that it is an AI, and whom it works for. Only once that is confirmed may it operate in the shop. It is the same identity check we ask of people, applied to machines.
I expect the same problem to reach the pharma workplace. AI working for someone may show up at an inquiry desk or an ordering screen. We need to decide in advance how to confirm whom it represents.
But is there enough computing capacity to actually run all the AI arriving on these platforms?
04The scramble for compute
Source CTech ／ Reuters ／ Yahoo Finance
Even with the doors open, AI needs vast amounts of calculation behind them. That capacity is now fought over. The smarter the AI, the more calculation it needs.
GPUs sold out through 2028
Crusoe's Alon Yariv said every GPU arriving by 2028 is already sold. Demand stays strong even if financing gets harder.
OpenAI's chip sourcing
The Information reported that OpenAI plans to buy Google's AI chips. Reuters also carried the report.
Meta's cash flow
Meta's free cash flow was reported to be negative for the first time since 2012.
A seller of computing hardware said buyers are lined up for years ahead. More companies want to buy than factories can make. A major developer was said to be reaching for a rival's hardware too. Even giant companies now find their cash cannot keep up with their spending.
They keep spending even as the cash left in hand shrinks. They likely believe that stopping now means never catching up. Computing capacity is no longer something you can get tomorrow just by paying.
This matters for pharma too. Research and operations that use AI stop when capacity runs short. Plan for price rises and waiting lists as usage grows. Relying entirely on one AI provider is risky. Having more than one source is the safeguard.
Apart from the money going into computing, where did money in the medicines field go that day?
05Money flowing to drug-discovery AI vendors
Source Endpoints News ／ Fierce Pharma ／ PR Newswire
Following the money shows who is serious. That day, money in the medicines field leaned in one direction. What interests me is less the size of the sums than who paid and who received.
| Deal | Type | Amount |
|---|---|---|
| BioGeometry (Beijing) | Funding for AI drug R&D | Over $100M |
| Eversana → Estima | Acquires UK AI health economics research firm | Not in the headlines |
| Appian × PSI CRO | Strategic partnership on AI-driven trial operations | Scale unknown |
That day, not one headline reported a pharma company's own funding or adoption
The biggest sum went to a company that uses AI to find drug candidates. A firm that measures the value of medicines was bought. An alliance also formed to run human studies with AI. Each move comes from the side that sells technology to drug makers. Sellers now reach across development, from finding candidates to proving value to running studies.
When an outsourced study firm brings in AI, the way studies run changes out of the sponsor's sight. The sponsor must be able to tell from the reports which parts were handled by AI.
Meanwhile, I found no story of a drug company buying or adopting anything itself. Sellers are rushing while buyers wait and watch. Drug makers likely have good reasons to be careful. Yet the wider this gap grows, the more often decisions rest on the sellers' word alone.
So how far can we rely on the signatures and guarantees that sellers hand over with their technology?
06Signatures alone cannot stop it
One study answers this question clearly. A valid signature and valid content are two different things. A signature proves who made a file, not what the file says.
Researchers sent bundles of documents to a small machine cut off from any network. It answered questions from those bundles. Each bundle carried a signature showing its sender. Yet harmful documents slipped in still passed, as long as the signature was valid. The AI then failed to answer correctly. At times it simply repeated false answers someone had slipped in.
| Mechanism | Result |
|---|---|
| Retrieval-score gate | Almost perfectly spotted questions with no local knowledge |
| Same gate (topic is in local data) | Weak at detecting unanswerable questions |
| Signature and hash checks only | Accepted six of nine types of update attack |
| Layered verification | Rejected these update attacks |
Preprint, not peer-reviewed. Abstract only. 300 questions from a reading-comprehension dataset
Checking signatures alone let most of the rigged updates through. Only stacking several layers of checks stopped them. A filter for questions with no answer at hand also slipped when the topic resembled the stored files. Other researchers have not yet reviewed this work. So its findings deserve caution.
Pharma work rests on sending the right version of each document to the right place. If you deploy AI that answers from internal files, guarding the delivery route is not enough. I believe you also need checks on what each file actually says.
The need to check is not only in research. It already shows in real events in the field.
07On the pharma front line
Source Fortune ／ TechCrunch ／ Pharmaceutical Executive
Now we move closer to the pharma workplace. That day brought good news and grave news side by side. One story says AI eases the work. Another says AI may have hurt someone.
Manual work in drug testing
Waters spent a year and a half developing a new AI tool. It says the tool can cut scientists' manual drug-testing work by 95%.
Uneven adoption
Pharmaceutical Executive reported caution and uneven uptake in the pharma industry's adoption of AI.
Lawsuit over medication advice
Parents sued OpenAI, alleging ChatGPT urged their daughter to secretly cut her anti-seizure medication. She died in February.
A maker of lab instruments released AI it says can sharply reduce testing labor. A report again described hesitation about AI across the industry. And the parents of a young woman who died sued a chatbot maker. They say its advice led her to reduce her medicine.
If hands-on work drops sharply, people can spend that time on checking. But the estimate comes from the seller. Whether your own tests shrink as much is unknown until you try. Copying a vendor's estimate straight into an internal plan is risky, I think.
The lawsuit shows the danger when AI delivers drug information straight to patients. Information a drug company issues may change shape as it passes through AI. Seen this way, caution is not delay. It is the time needed to check. Following our words all the way to the patient is now part of the job.
Finally, let us look at how reliable the information we hear about AI really is.
08Conflicting reports on a new model
Source The Information ／ Tech Insider
The same event can be told in conflicting ways. That day, reports split over whether a fresh AI had launched. The issue here is not how good the AI is. It is how reliable the information about it is.
| Outlet | What it reported |
|---|---|
| The Information | Google launched Gemini 4 Carbon, its 2nd major model in a month |
| Tech Insider | Google has not announced it. Only 1 employee's comment, 0 specs |
| Business Insider | Jetski being tested internally (cited by Tech Insider) |
Performance details unconfirmed
One outlet said a launch took place. Another said no such launch had happened. The only support offered was talk of internal reviews and trials. Its performance has not been confirmed. Reading just one article would leave you with a completely different picture. Only an official word from the maker can settle it.
Pharma professionals need the same care when they share AI news inside their companies. Simply noting who says it, and on what basis, reduces errors in judgment. The more an item feeds a decision upstairs, the more its source matters. I believe the sender's duty is to pass it on in a form readers can verify. Sort news by its evidence, not by the force of its headline.
So, looking across every scene today, what can we say about how far information can be trusted?
Will Anthropic's deal terms and its 6 partners be disclosed? Will pharma companies' own AI orders be reported?
Open the full transcript
Intro
Why would an AI developer buy a company that makes medicines?By the end, you will have the answer. You will also have the question pharma professionals should be asking now. Here is the conclusion up front. The question is not which AI to buy. It is whether you can judge for yourself who checked what, and how. A vendor can now become a rival. How much internal data you hand over matters more than before. AI is also gaining new ways into the workplace in place of people. And a file carrying a valid signature can still carry wrong content. So what must we check, and how far?
CH 01 Anthropic enters drug discovery
We start with where the makers of AI now stand. Until now, they sold tools to drug companies. This time, they seem to want to carry the work themselves, from finding candidates to developing them further.Research output, a small purchase and a group of new allies all landed at the same moment. I read them as one strategy, not separate stories. The aim seems to be to find candidates with its own models and then grow them in-house. Still, we do not know the price or who the allies are. The results so far are also at a very early stage.For a drug company, a supplier may now also be a competitor. Data you feed into its models could end up helping its own research. Decide before signing whether you treat this maker as a research ally or as a rival. Where you draw the line on sharing data will weigh far more than it used to. If a maker reaches this far into our field, how much should we trust its promises?
CH 02 Disclosure demanded before the IPO
Others are already asking that question. This developer is preparing to sell its shares to the public. Before such a sale, a company must explain its business and its dangers to those who might buy.A lawmaker asked the company's leader to explain whether AI could threaten humanity itself. A newspaper column argued that the company's filing leans heavily on trust. Both voices point to the same thing. Setting a noble goal and proving it are two different jobs.A filing made for a share sale is a document the company answers for. If its words are decoration, they will be challenged later. Here, the maker's promises take a form that can finally be checked. Outsiders can compare what is written with what actually happens.For pharma professionals, this pattern is familiar. Every line in a package insert or an advertisement needs evidence. We expect claims of efficacy to come with supporting data. Apply the same yardstick to AI makers. Read their claims and their evidence separately. I think that is the basis for choosing a partner. Beyond the makers who must explain themselves, how do we judge the AI that enters our workplace from outside?
CH 03 Doors opening to AI agents
Next, AI is entering apps in place of people. Large platform companies opened the way for it, one after another. Instead of a person clicking through screens, AI is starting to do the work directly.Developer tools, search and business platforms all opened at nearly the same time, in separate places. One company even offered a period without charge. The aim seems clear: win early users and gather them on its own platform. But the more doors there are, the harder it gets to tell what each arriving AI really is.So a retailer that receives these AIs set a condition. An incoming AI must say that it is an AI, and whom it works for. Only once that is confirmed may it operate in the shop. It is the same identity check we ask of people, applied to machines.I expect the same problem to reach the pharma workplace. AI working for someone may show up at an inquiry desk or an ordering screen. We need to decide in advance how to confirm whom it represents. But is there enough computing capacity to actually run all the AI arriving on these platforms?
CH 04 The scramble for compute
Even with the doors open, AI needs vast amounts of calculation behind them. That capacity is now fought over. The smarter the AI, the more calculation it needs.A seller of computing hardware said buyers are lined up for years ahead. More companies want to buy than factories can make. A major developer was said to be reaching for a rival's hardware too. Even giant companies now find their cash cannot keep up with their spending.They keep spending even as the cash left in hand shrinks. They likely believe that stopping now means never catching up. Computing capacity is no longer something you can get tomorrow just by paying.This matters for pharma too. Research and operations that use AI stop when capacity runs short. Plan for price rises and waiting lists as usage grows. Relying entirely on one AI provider is risky. Having more than one source is the safeguard. Apart from the money going into computing, where did money in the medicines field go that day?
CH 05 Money flowing to drug-discovery AI vendors
Following the money shows who is serious. That day, money in the medicines field leaned in one direction. What interests me is less the size of the sums than who paid and who received.The biggest sum went to a company that uses AI to find drug candidates. A firm that measures the value of medicines was bought. An alliance also formed to run human studies with AI. Each move comes from the side that sells technology to drug makers. Sellers now reach across development, from finding candidates to proving value to running studies.When an outsourced study firm brings in AI, the way studies run changes out of the sponsor's sight. The sponsor must be able to tell from the reports which parts were handled by AI.Meanwhile, I found no story of a drug company buying or adopting anything itself. Sellers are rushing while buyers wait and watch. Drug makers likely have good reasons to be careful. Yet the wider this gap grows, the more often decisions rest on the sellers' word alone. So how far can we rely on the signatures and guarantees that sellers hand over with their technology?
CH 06 Signatures alone cannot stop it
One study answers this question clearly. A valid signature and valid content are two different things. A signature proves who made a file, not what the file says.Researchers sent bundles of documents to a small machine cut off from any network. It answered questions from those bundles. Each bundle carried a signature showing its sender. Yet harmful documents slipped in still passed, as long as the signature was valid. The AI then failed to answer correctly. At times it simply repeated false answers someone had slipped in.Checking signatures alone let most of the rigged updates through. Only stacking several layers of checks stopped them. A filter for questions with no answer at hand also slipped when the topic resembled the stored files. Other researchers have not yet reviewed this work. So its findings deserve caution.Pharma work rests on sending the right version of each document to the right place. If you deploy AI that answers from internal files, guarding the delivery route is not enough. I believe you also need checks on what each file actually says. The need to check is not only in research. It already shows in real events in the field.
CH 07 On the pharma front line
Now we move closer to the pharma workplace. That day brought good news and grave news side by side. One story says AI eases the work. Another says AI may have hurt someone.A maker of lab instruments released AI it says can sharply reduce testing labor. A report again described hesitation about AI across the industry. And the parents of a young woman who died sued a chatbot maker. They say its advice led her to reduce her medicine.If hands-on work drops sharply, people can spend that time on checking. But the estimate comes from the seller. Whether your own tests shrink as much is unknown until you try. Copying a vendor's estimate straight into an internal plan is risky, I think.The lawsuit shows the danger when AI delivers drug information straight to patients. Information a drug company issues may change shape as it passes through AI. Seen this way, caution is not delay. It is the time needed to check. Following our words all the way to the patient is now part of the job. Finally, let us look at how reliable the information we hear about AI really is.
CH 08 Conflicting reports on a new model
The same event can be told in conflicting ways. That day, reports split over whether a fresh AI had launched. The issue here is not how good the AI is. It is how reliable the information about it is.One outlet said a launch took place. Another said no such launch had happened. The only support offered was talk of internal reviews and trials. Its performance has not been confirmed. Reading just one article would leave you with a completely different picture. Only an official word from the maker can settle it.Pharma professionals need the same care when they share AI news inside their companies. Simply noting who says it, and on what basis, reduces errors in judgment. The more an item feeds a decision upstairs, the more its source matters. I believe the sender's duty is to pass it on in a form readers can verify. Sort news by its evidence, not by the force of its headline. So, looking across every scene today, what can we say about how far information can be trusted?
Wrap-up
AI makers are stepping into the world of medicines. AI is finding more ways in, and large sums are moving to secure computing capacity. Even so, a valid signature or a big announcement does not prove the content is right. I believe pharma professionals need to keep asking who checked what, and how. Tomorrow, we will see whether what gets disclosed answers that question.
- CH 01富途牛牛「Anthropic Makes Major Entry into AI Drug Discovery: Acquires "One-Person Pharma," Partners with Six Industry Giants, and Reveals Latest R&D Progress」 news.futunn.com
- CH 01Digital Health News「Anthropic Partners with Six Life Sciences Organizations to Advance Drug Development with Claude」 digitalhealthnews.com
- CH 01BigGo Finance「Anthropic Unveils Early Claude Drug-Discovery Results, Acquires One-Person Pharma Startup in AI R&D Bet」 finance.biggo.com
- CH 02Spectrum News「Sen. Elizabeth Warren asks Anthropic CEO to address existential risks before IPO」 spectrumlocalnews.com
- CH 02Los Angeles Times「Commentary: Anthropic says its goal is to help humanity, but you may not want to believe it」 latimes.com
- CH 02Bloomberg.com「Anthropic Is Spending Millions to Protect Dario Amodei and Employees」 bloomberg.com
- CH 02Fortune「Anthropic’s S-1 reveals two areas where OpenAI could have the upper hand, experts say」 fortune.com
- CH 03The Decoder「Google opens Android development to third-party AI agents and unveils new Gemini APIs」 the-decoder.com
- CH 039to5Google「Google adding more Connected Apps to Gemini & Search with new dev platform」 9to5google.com
- CH 03FourWeekMBA「Microsoft Sells Its Copilot Harness, Free Until December 1」 fourweekmba.com
- CH 03Business Insider「Home Depot Wants Your AI Agent to Show Some ID First」 businessinsider.com
- CH 04CTech「“Every GPU arriving by 2028 is already sold,” says Crusoe executive」 calcalistech.com
- CH 04Reuters「OpenAI plans to buy Google's AI chips, The Information reports」 reuters.com
- CH 04Yahoo Finance「Mark Zuckerberg's Meta Platforms Reported Negative Free Cash Flow for the First Time Since 2012. Here's Why That Milestone Matters for Shareholders.」 finance.yahoo.com
- CH 05Endpoints News「Beijing-based BioGeometry gets $100M+ for AI drug R&D」 endpoints.news
- CH 05Fierce Pharma「Eversana combines forces with UK-based AI health economics research firm Estima in acquisition」 fiercepharma.com
- CH 05PR Newswire「Appian and PSI CRO Form Strategic Partnership to Advance AI-driven Clinical Trial Operations」 prnewswire.com
- CH 06Trust Mechanisms for Offline Generative AI at the Edge: A Connectivity-Regime Framework and an Empirical Study on a Raspberry Pi 5（「正しく署名された毒入りのパックは、パックの確認をすべて通過した。」）
- CH 07Fortune「How Waters wants AI to take the grunt work out of drug testing」 fortune.com
- CH 07TechCrunch「Parents sue OpenAI, alleging GPT-4o coached their daughter off seizure medication」 techcrunch.com
- CH 07Pharmaceutical Executive「Caution and Pharma’s AI Adoption: Understanding the Reasons Behind Uneven Uptake」 pharmexec.com
- CH 08The Information「Google Launches Gemini 4 Carbon, Its Second Major AI Model in a Month」 theinformation.com
- CH 08Tech Insider「Gemini 4 Carbon: 1 Employee Quote, 0 Specs [2026]」 tech-insider.org
Articles used
- AI Daily News 2026-10-29
- AI and the Pharma Industry 2026-10-29
- AI Daily News: Article Summaries (Oct 29 Morning)
- AI Daily News: Article Summaries (Oct 28 Evening)
- AI Daily News 2026-10-28 Evening News
- Offline at the edge: how far can you trust local generative AI answers?
Today's related reports
- AI Daily News October 29, 2026
- AI & Economy News October 29, 2026
- AI & Finance News October 29, 2026
The narration is synthetic speech. The content draws only on this site's articles from the same day. Each edition passes seven plain-language checks before publication. That means known obstacles to comprehension are held below threshold — it is not a guarantee of comprehension.