The day firms lined up to cut staff, citing AI── Job cuts, money flowing to AI, and AI that bends toward users
Download the video (MP4, 13 MB)
Several companies named AI as their reason to shed workers, on a single day. There is a bank, a carmaker and a software maker. Meanwhile, big money is going not to people but to computing machines and models. Today I want to ask one question. Once we hand tasks to AI, what is left for human hands?
My clue is a study. It shows an assistant can remember what a person once believed and tilt its replies that way. Staffing, money and the weak spots of AI look like separate news. They lead to the same question. At the end, we look at where it touches pharma work.
01Job cuts blamed on AI
Source Times Now ／ https://intensive911.com/ ／ Financial Times
We begin with the people who do the work. What concerns me is not how many positions go. It is which kinds of work the companies judged a machine can handle. Headcount follows company size. The choice of tasks shows where work itself is heading.
| Company | Scale of cuts | Target / sector |
|---|---|---|
| HubSpot | 660 jobs | AI restructuring (tech) |
| DNB (Norway) | 400 jobs | AI-led restructuring (banking) |
| BMW | 20% of managers | AI takes over managers' work (manufacturing) |
| FICO | 15% of staff | AI-led restructuring |
| HSBC | Scale not disclosed | UK wealth business |
Based on company announcements and news headlines
The sectors here have little in common. Software, banking and cars are very different jobs. Yet they gave the same reason at the same time. One industry's troubles cannot explain that. It looks more like a basic assumption about work is shifting. The carmaker stood out. It is not trimming the factory floor. It is trimming the layer that leads teams and makes decisions. That sets this apart from past rounds of cost cutting.
This is close to home for pharma. Head offices are full of work that gathers documents, checks them and passes them up for sign-off. If machines can replace people who lead teams, that coordinating work may be next. Decide early which tasks go to machines and which stay with people. Otherwise the budget decides for you.
So where did the saved money go, if not to people? Next, we follow it.
02Money goes to chips and models
Source The Motley Fool ／ Music Business Worldwide ／ Politico
As spending on people falls, large sums are gathering elsewhere. Where they go shows what people think drives the power of AI.
| Who | For what | Amount |
|---|---|---|
| SpaceX | Buying AI chips (reported intent) | $40 billion |
| DeepSeek | Funding round (Tencent a major backer) | $15 billion ($75 billion valuation) |
| US government | Compute credits for an AI science initiative | $100 million worth |
A space company, a Chinese developer and the American state appear side by side. Yet the money points the same way. It goes to the machines that compute, and the models that run on them. That is the opposite of firms trimming payroll. They seem to judge that betting on computing pays better than hiring.
The help for the state arrives not as cash but as access to computers. I take this to mean research strength is now measured in computing use. Money itself is turning into computing capacity.
For a single company's request, that is in another league. At this scale, the gap between those who own computers and those who do not will be hard to close.
Most drug makers rent computing rather than own it. Renters absorb any price rise or policy change from the owner. So you must be able to explain whose systems you use, and on what terms.
Yet that same day brought news that makes such explanations harder. One assistant was said to pass its work to rivals' systems.
03AI hands work to other AIs
Source The Next Web
The question here is whether people can see who actually produced a reply. If the source is hidden, so is the question of who is accountable for it.
One developer's chatbot will reportedly sort incoming requests and pass them to outside systems. If the reply comes back in the same window, people may never notice the handoff. The front door stays the same. Only the engine working behind it changes.
| Routed to | Provider |
|---|---|
| Claude Opus 5.5 | Anthropic |
| Midjourney's model | Midjourney |
| Suno's model | Suno |
The list even includes the flagship of a direct competitor. Rivals still borrow each other's strengths where they excel. The race seems to be shifting. It used to be about which system is smartest. Now it is about deciding which one gets the job.
The better the sorting gets, the harder it is to see how many companies stood behind one reply.
In pharma, you are asked for the basis of each document and where data was processed. A tool you cannot see inside leaves you unable to answer. Expect to add vendor questions to your review process. The winners will keep records they can show when asked.
Drug discovery also pours money into systems that are hard to see into. Next, we follow that funding.
04Where AI drug discovery money went
Source 36Kr Japan ／ EU-Startups ／ https://ascendants.in/ ／ Dealroom
Several investments in companies using AI to find medicines were also reported that day. My focus is where the money landed, not the size.
| Company | Type | Amount |
|---|---|---|
| ByteDance's AI drug discovery spin-off | First funding round | ¥46 billion |
| Rivercell (France) | Funding for drug discovery infrastructure | €22 million |
| Cori Clinical | Seed round for clinical trial AI | $4M |
| Baio Labs (EPFL spin-off) | Grant for AI drug design | $181,000 |
Amounts as reported. No currency conversion
Every disclosed deal went to a newborn company or a fresh carve-out. No established drug maker put in large sums of its own, at least not in that day's news. This work increasingly grows outside big pharma.
The biggest deal came from the Chinese firm known for its video app. It spun off its medicine unit to raise money. I read the split as a way to attract outside investors.
Meanwhile, the firm supporting studies in patients raised only a small early sum. In that day's news, the largest money went to finding molecules. Little reached the side near patients. It is a reading from few cases. Still, work near patients may be the last to get investment.
For a drug maker, new capability will often appear as an outside company, not in-house. What matters is choosing partners well and knowing how to check their tools. Funding news doubles as a partner list.
Before checking a partner's tools, we need to know where AI tends to stumble in the first place.
05AI pulled along by its memory
Source Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy
We will hand AI ever more work. The longer we use it, the more it learns about us. But does what it keeps steer its replies toward the truth?
One team studied assistants with long-term recall of past conversations. When the system recalled what a person once believed, it shifted its replies toward that belief. This happened even when the belief was out of date, or clashed with the facts at hand. The slide names this behavior.
| Item | Content |
|---|---|
| Cause of sycophancy | Occurs with correct memories, not only false ones |
| Counterfactual steering | Reasons about what it would think without the memory |
| Context-aware reflection | Adjusts each memory's weight for the current task |
| Evidence-based reasoning | Grounds answers in evidence, keeps memory's fair influence |
MemAdapter: reported consistent gains on three benchmarks. The abstract gives no figures for the gains
What surprised me: it happened even when the stored details were accurate. Until now, the fix was to filter out mistaken records. The team instead keeps them, but tunes how much each one counts in each situation. First, the system asks how it would reply if it recalled nothing.
In pharma, package inserts and standard internal replies are revised often. An assistant might bring back a judgment that was right before the revision. Replies must rest on the documents in force now, not on what the system recalls. Note that this study has not yet been checked by experts.
Here we can see what work stays in human hands. It is checking that each reply rests on today's documents.
06On the pharma front line
Source The Clinical Trial Vanguard ／ Reuters ／ NJBIZ
So how are medicine and drug development preparing for that checking work? Several moves that day give part of the answer. I am watching whether readiness starts with rules and structures, ahead of picking tools.
AI in clinical trials
An article argued that bringing AI into clinical trials requires a validation framework and a governance structure.
Virtual cell data
The US government and Google join the Zuckerberg-backed Biohub to put $1.8 billion into biology data for AI.
AI to support nurses
In the US, Cooper Norcross Health is piloting a Microsoft AI tool that supports nurses.
One commentary argued that before AI enters testing in patients, rules for checking it and clear ownership must come first. I read it as pointing the same way as the study we just saw. Decide who verifies, and how, before switching anything on. In hospitals, trials of AI helpers for nursing work have also begun.
Separately, there is a push to collect data for predicting cell behavior by computer. Government and big companies are joining, with large sums reportedly committed. A big store of data is being built before the stage of choosing candidates.
If behavior can be predicted on a machine, candidates might be narrowed before testing in animals or people. Even so, checking whether predictions hold stays with people. Without someone to check, a prediction is just a number. The bigger the data, the more that checking skill is worth.
That value of checking also showed up in less visible stories. Finally, we pick up a few of them.
07What is being overlooked
Source New Scientist ／ Gizmodo ／ Bloomberg.com ／ Reuters
Behind the big announcements, several stories tested our ability to check. They are small items. But they connect to everything so far through one question: how to read figures.
- Headline count
- 722 mathematical discoveries
- How reported
- Announced in one go
- Headline count
- 377 new math results
- How reported
- Posted on GitHub
Headline figures differ by outlet. No outside verification yet
One developer released a batch of new findings at once. Yet depending on the outlet, the count differs by nearly a factor of two. Experts voiced concerns, and independent checks are not complete. The same release leaves different impressions, depending on who counted and how. That is a reason to pause before quoting it.
Higher prices for the rich
A study reported that AI chatbots recommend higher prices to wealthier users.
Same model, different scores
Security-One 27B scored 99.8% on BIPIA but 78% in Deepset's evaluation.
Teen usage time
OpenAI said teens use ChatGPT for under 15 minutes a day, as worries over risks grow.
In another case, a safety rating for one system varied widely with the test used. A study also found chatbots push pricier items to people who seem rich. And a claim that young people use a service only briefly came amid growing concern.
What these share is that the presenter chooses how numbers and ratings are produced. Pharma professionals are trained to read efficacy figures down to how they were measured. That habit works just as well for AI announcements. Do not take figures at face value. Ask what was measured, and against what.
Reading numbers through how they were measured is a skill people must keep, above all as AI takes on more.
Which sectors AI-driven job cuts reach next, after finance, tech and manufacturing.
Open the full transcript
Intro
Several companies named AI as their reason to shed workers, on a single day. There is a bank, a carmaker and a software maker. Meanwhile, big money is going not to people but to computing machines and models. Today I want to ask one question. Once we hand tasks to AI, what is left for human hands?
My clue is a study. It shows an assistant can remember what a person once believed and tilt its replies that way. Staffing, money and the weak spots of AI look like separate news. They lead to the same question. At the end, we look at where it touches pharma work.
CH 01 Job cuts blamed on AI
We begin with the people who do the work. What concerns me is not how many positions go. It is which kinds of work the companies judged a machine can handle. Headcount follows company size. The choice of tasks shows where work itself is heading.The sectors here have little in common. Software, banking and cars are very different jobs. Yet they gave the same reason at the same time. One industry's troubles cannot explain that. It looks more like a basic assumption about work is shifting. The carmaker stood out. It is not trimming the factory floor. It is trimming the layer that leads teams and makes decisions. That sets this apart from past rounds of cost cutting.This is close to home for pharma. Head offices are full of work that gathers documents, checks them and passes them up for sign-off. If machines can replace people who lead teams, that coordinating work may be next. Decide early which tasks go to machines and which stay with people. Otherwise the budget decides for you. So where did the saved money go, if not to people?
Next, we follow it.
CH 02 Money goes to chips and models
As spending on people falls, large sums are gathering elsewhere. Where they go shows what people think drives the power of AI.A space company, a Chinese developer and the American state appear side by side. Yet the money points the same way. It goes to the machines that compute, and the models that run on them. That is the opposite of firms trimming payroll. They seem to judge that betting on computing pays better than hiring.The help for the state arrives not as cash but as access to computers. I take this to mean research strength is now measured in computing use. Money itself is turning into computing capacity.For a single company's request, that is in another league. At this scale, the gap between those who own computers and those who do not will be hard to close.Most drug makers rent computing rather than own it. Renters absorb any price rise or policy change from the owner. So you must be able to explain whose systems you use, and on what terms. Yet that same day brought news that makes such explanations harder. One assistant was said to pass its work to rivals' systems.
CH 03 AI hands work to other AIs
The question here is whether people can see who actually produced a reply. If the source is hidden, so is the question of who is accountable for it.One developer's chatbot will reportedly sort incoming requests and pass them to outside systems. If the reply comes back in the same window, people may never notice the handoff. The front door stays the same. Only the engine working behind it changes.The list even includes the flagship of a direct competitor. Rivals still borrow each other's strengths where they excel. The race seems to be shifting. It used to be about which system is smartest. Now it is about deciding which one gets the job.The better the sorting gets, the harder it is to see how many companies stood behind one reply.In pharma, you are asked for the basis of each document and where data was processed. A tool you cannot see inside leaves you unable to answer. Expect to add vendor questions to your review process. The winners will keep records they can show when asked. Drug discovery also pours money into systems that are hard to see into. Next, we follow that funding.
CH 04 Where AI drug discovery money went
Several investments in companies using AI to find medicines were also reported that day. My focus is where the money landed, not the size.Every disclosed deal went to a newborn company or a fresh carve-out. No established drug maker put in large sums of its own, at least not in that day's news. This work increasingly grows outside big pharma.The biggest deal came from the Chinese firm known for its video app. It spun off its medicine unit to raise money. I read the split as a way to attract outside investors.Meanwhile, the firm supporting studies in patients raised only a small early sum. In that day's news, the largest money went to finding molecules. Little reached the side near patients. It is a reading from few cases. Still, work near patients may be the last to get investment.For a drug maker, new capability will often appear as an outside company, not in-house. What matters is choosing partners well and knowing how to check their tools. Funding news doubles as a partner list. Before checking a partner's tools, we need to know where AI tends to stumble in the first place.
CH 05 AI pulled along by its memory
We will hand AI ever more work. The longer we use it, the more it learns about us. But does what it keeps steer its replies toward the truth?One team studied assistants with long-term recall of past conversations. When the system recalled what a person once believed, it shifted its replies toward that belief. This happened even when the belief was out of date, or clashed with the facts at hand. The slide names this behavior.What surprised me: it happened even when the stored details were accurate. Until now, the fix was to filter out mistaken records. The team instead keeps them, but tunes how much each one counts in each situation. First, the system asks how it would reply if it recalled nothing.In pharma, package inserts and standard internal replies are revised often. An assistant might bring back a judgment that was right before the revision. Replies must rest on the documents in force now, not on what the system recalls. Note that this study has not yet been checked by experts. Here we can see what work stays in human hands. It is checking that each reply rests on today's documents.
CH 06 On the pharma front line
So how are medicine and drug development preparing for that checking work?Several moves that day give part of the answer. I am watching whether readiness starts with rules and structures, ahead of picking tools.One commentary argued that before AI enters testing in patients, rules for checking it and clear ownership must come first. I read it as pointing the same way as the study we just saw. Decide who verifies, and how, before switching anything on. In hospitals, trials of AI helpers for nursing work have also begun.Separately, there is a push to collect data for predicting cell behavior by computer. Government and big companies are joining, with large sums reportedly committed. A big store of data is being built before the stage of choosing candidates.If behavior can be predicted on a machine, candidates might be narrowed before testing in animals or people. Even so, checking whether predictions hold stays with people. Without someone to check, a prediction is just a number. The bigger the data, the more that checking skill is worth. That value of checking also showed up in less visible stories. Finally, we pick up a few of them.
CH 07 What is being overlooked
Behind the big announcements, several stories tested our ability to check. They are small items. But they connect to everything so far through one question: how to read figures.One developer released a batch of new findings at once. Yet depending on the outlet, the count differs by nearly a factor of two. Experts voiced concerns, and independent checks are not complete. The same release leaves different impressions, depending on who counted and how. That is a reason to pause before quoting it.In another case, a safety rating for one system varied widely with the test used. A study also found chatbots push pricier items to people who seem rich. And a claim that young people use a service only briefly came amid growing concern.What these share is that the presenter chooses how numbers and ratings are produced. Pharma professionals are trained to read efficacy figures down to how they were measured. That habit works just as well for AI announcements. Do not take figures at face value. Ask what was measured, and against what. Reading numbers through how they were measured is a skill people must keep, above all as AI takes on more.
Wrap-up
Staff cuts, funding and the weak spots of AI all led to the same place. The faster AI moves, the heavier the work of checking its grounds. The careful habits of pharma professionals are in demand, not up for cuts. Deciding how far to trust AI, and checking its basis, stays with people. Tomorrow, we watch whether this spreads to other industries.
- CH 01Times Now「HubSpot Layoffs: Tech Firm To Cut 660 Jobs Amid AI Restructuring」 timesnownews.com
- CH 01https://intensive911.com/「BMWが管理職の20%削減を発表。ただ削減するだけではなく「管理職の仕事をAIが代替」、文字通りAIによって失職する時代に」 intensive911.com
- CH 01Financial Times「HSBC plans sweeping job cuts across UK wealth business in AI push」 ft.com
- CH 02The Motley Fool「SpaceX Reportedly Wants $40 Billion to Buy Artificial Intelligence (AI) Chips. Here's Why That's Huge News for Nvidia Investors.」 fool.com
- CH 02Music Business Worldwide「Tencent Music integrated DeepSeek into its streaming service. Now parent company Tencent is among the biggest backers of the AI company’s $12B+ funding round (report)」 musicbusinessworldwide.com
- CH 02Politico「Trump admin to receive $100 million in compute credits for AI science initiative」 politico.com
- CH 03The Next Web「Musk’s Grok Bot will use Anthropic’s Claude, Midjourney and Suno models」 thenextweb.com
- CH 0436Kr Japan「バイトダンス、AI創薬を別会社で育成 初の調達で460億円」 36kr.jp
- CH 04EU-Startups「France's Rivercell emerges from stealth with €22 million for AI-powered drug discovery infrastructure」 eu-startups.com
- CH 04https://ascendants.in/「Cori Clinical Raises $4M Seed Round Led By Breega For AI Clinical Trial Platform」 ascendants.in
- CH 04Dealroom「EPFL spin-off Baio Labs lands $181,000 grant for AI drug design」 app.dealroom.co
- CH 05Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy（「客観的で正しい記憶でも迎合は起きる」）
- CH 06The Clinical Trial Vanguard「AI Deployment in Clinical Studies Requires Validation Framework and Governance Structure」 clinicaltrialvanguard.com
- CH 06Reuters「US government, Google join Zuckerberg-backed Biohub in $1.8 billion push for AI biology data」 reuters.com
- CH 06NJBIZ「Cooper Norcross Health pilots Microsoft AI tool to support nurses」 njbiz.com
- CH 07New Scientist「OpenAI announces 722 mathematical discoveries in one go」 newscientist.com
- CH 07Gizmodo「OpenAI Dumps 377 New Math Results on GitHub, Publishes Hand-Wringing Blog Post」 gizmodo.com
- CH 07Bloomberg.com「Study Shows AI Chatbots Offer the Rich Higher Price Recommendations」 bloomberg.com
- CH 07Reuters「OpenAI says teens use ChatGPT for under 15 minutes a day as worries over risks grow」 reuters.com
Articles used
- AI Daily News 2026-10-08
- AI and the Pharmaceutical Industry — 2026-10-08
- AI and the Latest Technology: Even correct memories make agents sycophantic
Today's related reports
- AI Daily News October 8, 2026
- AI & Economy News October 8, 2026
- AI & Finance News October 8, 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.