The day DeepSeek showed its autonomous agent── Delayed fixes, hidden orders to a court's AI, a power plant lawsuit
Download the video (MP4, 11 MB)
When AI starts acting on its own, what actually runs short?I do not think it is intelligence. Reading today's reports together, I see a shortage in the capacity of people and organizations to absorb what AI produces. The speed of making things has gone up. But the people who check, the money that pays, and the electricity that runs it all do not seem to be growing at the same pace. This is not a distant story for those in pharma. Decide on adoption by looking only at what a tool can do, and the receiving side may stall first. So where should we begin?
01DeepSeek's autonomous agent
Source Pasquale Pillitteri ／ Global Times ／ Bloomberg.com ／ BusinessToday Malaysia
Until now, a person asked and then the machine acted. If AI stops waiting for instructions, that order changes. What happens then to the role of the people doing the work?
What was released
The DeepSeek Harness team released dsh-superagent, an open source prototype.
Multiple tasks
Global Times reported a "Super-agent" design that handles multiple tasks in parallel.
Who it competes with
Described as a rival to Claude Code and Codex. Bloomberg's headline: "Aim to Rival Anthropic."
A prototype that shows its insides is now challenging the big players. This kind of tool may no longer belong to a few companies.
What I watch is not how clever the machine is, but how its limits get set. It offers ideas unprompted. Carrying them out needs a range set by people. So I see the human job moving from asking to permitting.
For now, these tools write instructions for computers. This is my guess, but the way of working could carry over to document work.
A tool anyone can get may reach the workplace outside official channels. Review may stop being only a purchasing task.
In a pharma workplace, you need a chart of what goes to AI and what stays with people. This prototype's real ability is still unknown. Even so, I think setting the range early means less rework later.
Set a range and delegate, and AI does more work. The next question is how much to pay for that work.
02Changes in usage and billing
Source Business Insider ／ The Motley Fool ／ The Decoder
The harder AI works, the more the shape of the invoice changes. If you have planned adoption as a set sum, that assumption is now in doubt.
| Who | Move | Conditions and source |
|---|---|---|
| Anthropic | Internal AI use exceeds employee pay | At customer prices. From its IPO filing |
| Microsoft | Copilot moving from flat fee to usage-based billing | The Motley Fool article |
| OpenAI | Device- and card-level limits on paid ChatGPT accounts | The Decoder report |
To me, these moves point the same way. Paying one price for unlimited use is fading. Paying for what you consume is taking its place. I think heavy users weigh too much on the seller. Cost may come to arise at each use, not at the moment of the decision.
At the maker itself, spending on machines has passed spending on people. That looks like a further stage of this shift. Note that this is calculated at selling prices. It is not actual spending.
The same change is reported to be reaching tools already in daily work. As more people use them, bills get harder to forecast. Budgeting per head may stop working.
For pharma budget holders, AI may become a monthly moving cost, not a piece of equipment. Without measuring use by task, you cannot explain the benefit either. A way to watch costs needs building alongside adoption.
Costs show up in numbers. So what burden is AI adding that numbers hardly show?
03Bugs outrunning the fixes
Source Startup Fortune ／ Forbes ／ The Times of India
One less visible burden is the time of those doing the repairs. Finding a defect and repairing it do not move at the same pace.
Machines search without tiring. Repair, though, presumably needs human judgment, plus a check that nothing else breaks. So the reports say more discoveries mean a growing backlog.
- Role
- xAI Founder / Tesla CEO
- Content
- Warned AI agents could be the next security threat to crypto assets
- Role
- Product vendor
- Content
- Added AI, ID management and quantum resilience to security products
Voices of alarm and sales of protective products are arriving together. It is natural for protection to become a product. But the stalled repair work remains after you buy one. Buying does not by itself bring safety.
Even after a remedy is shipped, the receiver has to apply it. The problem does not vanish on the day of release. Choosing when to apply it and checking afterward remain the receiver's job.
In a pharma company, systems that handle records likely have the most checks before any update. Faster discovery could stall those steps. So I see value in naming an owner and setting priority criteria before trouble starts.
So far, this was about finding weak spots and repairing them. Next comes people deliberately targeting a machine's weak spots.
04Hidden orders to a court's AI
Source Tom's Hardware ／ New York State Bar Association ／ The Business Journals ／ Yellow Scene Magazine
When a machine becomes the reader, a document changes meaning. Characters people cannot see can reach the machine as commands.
The core point is this. Legal professionals tried to steer a machine reader by tampering with a document. The party they wanted leniency from was a system, not a person. And a machine caught it.
I think the penalty itself matters more than the amount. It shows that deceiving a machine is no mere prank.
Bar association guidelines
The New York State Bar Association's House of Delegates approved AI guidelines. One aim: keep legal judgment in human hands.
Insurance claims
The Business Journals reported that insurance claims caused by AI are already arising.
Publishers' lawsuit
The Boston Institute for Nonprofit Journalism and 59 publishers sued OpenAI and Microsoft.
I read these moves as a search for who holds responsibility. Judgment stays with people, damage becomes a money claim, and those who create articles go to court. AI problems seem to be moving from technology to who bears the burden.
What makes this trick troublesome is that it only takes having a machine read someone else's document. The concealed characters get read too. A person rereading the page may not notice.
In pharma, machines increasingly summarize papers and contractor reports. Treat the content of outside documents as information, not as commands. Build that setting and a human check into your intake procedure.
Document problems can be handled through procedure. But the electricity and parts that run the machines cannot be solved by procedure alone.
05Power, components and money
Source Open Magazine ／ The Korea Times ／ EnergyNow.com ／ Wowtale
AI looks like something inside a screen. In fact it is an industry that needs land, electricity and parts. That fact is now surfacing.
| Area | Who | Move |
|---|---|---|
| Power | xAI | Neighbors sue over gas turbines in Mississippi |
| Components | OpenAI | Altman met Samsung executives in South Korea to secure RAM |
| Local | Virginia, US | AI backlash in "Data Center Alley" |
| Funding | Cohere | Series E initial close announced. Targeting up to $5.1 billion |
What I see here is technical competition widening into a contest for resources. An executive travels abroad to visit a parts supplier. Residents take generating equipment to court. Huge sums are being raised as well. Securing electricity and parts, more than being clever, may increasingly set the speed of development.
Without enough computing parts, even the finest design will not run. The effect of a shortage may eventually reach users as usage fees or waiting time.
It also matters that the dispute rests on an environmental law. A project can be stopped as a question of air, without anyone debating AI itself. Winning the technical argument may not be enough without local consent.
Pharma professionals contract for a service behind a screen. Its stability can depend on consent from residents far away and on parts deals. If long-term work runs on it, naming a fallback for outages may become a procurement condition.
That is how outside conditions are moving. So what moved today inside the pharma industry itself?
06In pharma practice
Source Yahoo Tech ／ The Business Journals
Pharma readers may stop first at the figures comparing doctors with a machine. The wider the gap, the more it feels like we could hand things over.
| Subject | Result on complex cases | Conditions |
|---|---|---|
| Microsoft's diagnostic AI | Identified 83% | Simulated test in a study |
| General physicians | 36% correct | Same study |
The gap looks more than double, yet it comes from a constructed test. Real care involves a patient in front of you and someone answerable for the result. Several stages of checking remain between winning a test and being usable in care. Pharma people already know not to compare numbers from trials run under different conditions.
Collected and selected
AI and pharma section: 241 items collected, 0 selected. No reports involved amounts or capital moves.
Breakdown of exclusions
Out of date 124, out of scope 53, duplicates 33, low-quality outlets 20, no investment signal 6, and others.
Access rights before rollout
The Business Journals headline: the data access problem easily missed before rolling out Copilot.
On the same day, pharma investment was quiet. Plenty of reports came in, yet none deserved keeping. Automated tallies can miss things. Still, on a day of flashy numbers, no report showed money moving in pharma. Keep buzz separate from the movement of funds.
So I see no reason today to rush a change in pharma decisions. What you can do first is check who can see your internal documents. I have read only the headline on this. But it deserves attention before letting a machine read widely. This work helps whether or not you adopt.
If decisions need not be rushed, the next question is how solid the decision material itself is.
07Headline-only reports
Source Fortune ／ New York Post ／ The Times of India ／ Startup Fortune
The bigger the topic, the more easily it spreads unchecked. Today's reports include some whose substance we cannot confirm. Headlines are built to catch the eye, so impressions stick before content does.
| Report | Outlet | Status |
|---|---|---|
| Anthropic valuation on a $2 trillion scale | Fortune | Headline-only information. Details unverified |
| AI giant's donation of about $1 million | New York Post | Headline attribution needs checking |
| Remarks by Nadella | The Times of India | Headline cut off. Content unverified |
| Small model with 109M parameters | Startup Fortune | Link to LeCun unclear |
Several headlines are single-source or lack detail. Figures and valuations need checking
Huge sums and famous names stay in memory. But the items here are ones we have not traced to their substance. Some headlines break off midway. With others, we cannot tell whose story it is. This screen exists to show what we do not know as unknown. Judging from headlines alone can lead you to misread the story.
A report from one outlet is not bad in itself. But until separate reporting confirms it, all we know is that a report exists. This matters when you speak inside your company. "Reported" and "found" are different words.
Pharma professionals do this work daily in promotional review. They trace each claim to its basis, go to the original, and separate what can be said from what cannot. Apply the same eye to AI news, and unverified numbers stay out of internal briefings.
The power to verify is something pharma work already has. It applies to today's whole story as well.
Watch the real performance of DeepSeek's prototype and further news on Anthropic's listing terms.
Open the full transcript
Intro
When AI starts acting on its own, what actually runs short?I do not think it is intelligence. Reading today's reports together, I see a shortage in the capacity of people and organizations to absorb what AI produces. The speed of making things has gone up. But the people who check, the money that pays, and the electricity that runs it all do not seem to be growing at the same pace. This is not a distant story for those in pharma. Decide on adoption by looking only at what a tool can do, and the receiving side may stall first. So where should we begin?
CH 01 DeepSeek's autonomous agent
Until now, a person asked and then the machine acted. If AI stops waiting for instructions, that order changes. What happens then to the role of the people doing the work?
A prototype that shows its insides is now challenging the big players. This kind of tool may no longer belong to a few companies.What I watch is not how clever the machine is, but how its limits get set. It offers ideas unprompted. Carrying them out needs a range set by people. So I see the human job moving from asking to permitting.For now, these tools write instructions for computers. This is my guess, but the way of working could carry over to document work.A tool anyone can get may reach the workplace outside official channels. Review may stop being only a purchasing task.In a pharma workplace, you need a chart of what goes to AI and what stays with people. This prototype's real ability is still unknown. Even so, I think setting the range early means less rework later. Set a range and delegate, and AI does more work. The next question is how much to pay for that work.
CH 02 Changes in usage and billing
The harder AI works, the more the shape of the invoice changes. If you have planned adoption as a set sum, that assumption is now in doubt.To me, these moves point the same way. Paying one price for unlimited use is fading. Paying for what you consume is taking its place. I think heavy users weigh too much on the seller. Cost may come to arise at each use, not at the moment of the decision.At the maker itself, spending on machines has passed spending on people. That looks like a further stage of this shift. Note that this is calculated at selling prices. It is not actual spending.The same change is reported to be reaching tools already in daily work. As more people use them, bills get harder to forecast. Budgeting per head may stop working.For pharma budget holders, AI may become a monthly moving cost, not a piece of equipment. Without measuring use by task, you cannot explain the benefit either. A way to watch costs needs building alongside adoption. Costs show up in numbers. So what burden is AI adding that numbers hardly show?
CH 03 Bugs outrunning the fixes
One less visible burden is the time of those doing the repairs. Finding a defect and repairing it do not move at the same pace.Machines search without tiring. Repair, though, presumably needs human judgment, plus a check that nothing else breaks. So the reports say more discoveries mean a growing backlog.Voices of alarm and sales of protective products are arriving together. It is natural for protection to become a product. But the stalled repair work remains after you buy one. Buying does not by itself bring safety.Even after a remedy is shipped, the receiver has to apply it. The problem does not vanish on the day of release. Choosing when to apply it and checking afterward remain the receiver's job.In a pharma company, systems that handle records likely have the most checks before any update. Faster discovery could stall those steps. So I see value in naming an owner and setting priority criteria before trouble starts. So far, this was about finding weak spots and repairing them. Next comes people deliberately targeting a machine's weak spots.
CH 04 Hidden orders to a court's AI
When a machine becomes the reader, a document changes meaning. Characters people cannot see can reach the machine as commands.The core point is this. Legal professionals tried to steer a machine reader by tampering with a document. The party they wanted leniency from was a system, not a person. And a machine caught it.I think the penalty itself matters more than the amount. It shows that deceiving a machine is no mere prank.I read these moves as a search for who holds responsibility. Judgment stays with people, damage becomes a money claim, and those who create articles go to court. AI problems seem to be moving from technology to who bears the burden.What makes this trick troublesome is that it only takes having a machine read someone else's document. The concealed characters get read too. A person rereading the page may not notice.In pharma, machines increasingly summarize papers and contractor reports. Treat the content of outside documents as information, not as commands. Build that setting and a human check into your intake procedure. Document problems can be handled through procedure. But the electricity and parts that run the machines cannot be solved by procedure alone.
CH 05 Power, components and money
AI looks like something inside a screen. In fact it is an industry that needs land, electricity and parts. That fact is now surfacing.What I see here is technical competition widening into a contest for resources. An executive travels abroad to visit a parts supplier. Residents take generating equipment to court. Huge sums are being raised as well. Securing electricity and parts, more than being clever, may increasingly set the speed of development.Without enough computing parts, even the finest design will not run. The effect of a shortage may eventually reach users as usage fees or waiting time.It also matters that the dispute rests on an environmental law. A project can be stopped as a question of air, without anyone debating AI itself. Winning the technical argument may not be enough without local consent.Pharma professionals contract for a service behind a screen. Its stability can depend on consent from residents far away and on parts deals. If long-term work runs on it, naming a fallback for outages may become a procurement condition. That is how outside conditions are moving. So what moved today inside the pharma industry itself?
CH 06 In pharma practice
Pharma readers may stop first at the figures comparing doctors with a machine. The wider the gap, the more it feels like we could hand things over.The gap looks more than double, yet it comes from a constructed test. Real care involves a patient in front of you and someone answerable for the result. Several stages of checking remain between winning a test and being usable in care. Pharma people already know not to compare numbers from trials run under different conditions.On the same day, pharma investment was quiet. Plenty of reports came in, yet none deserved keeping. Automated tallies can miss things. Still, on a day of flashy numbers, no report showed money moving in pharma. Keep buzz separate from the movement of funds.So I see no reason today to rush a change in pharma decisions. What you can do first is check who can see your internal documents. I have read only the headline on this. But it deserves attention before letting a machine read widely. This work helps whether or not you adopt. If decisions need not be rushed, the next question is how solid the decision material itself is.
CH 07 Headline-only reports
The bigger the topic, the more easily it spreads unchecked. Today's reports include some whose substance we cannot confirm. Headlines are built to catch the eye, so impressions stick before content does.Huge sums and famous names stay in memory. But the items here are ones we have not traced to their substance. Some headlines break off midway. With others, we cannot tell whose story it is. This screen exists to show what we do not know as unknown. Judging from headlines alone can lead you to misread the story.A report from one outlet is not bad in itself. But until separate reporting confirms it, all we know is that a report exists. This matters when you speak inside your company. "Reported" and "found" are different words.Pharma professionals do this work daily in promotional review. They trace each claim to its basis, go to the original, and separate what can be said from what cannot. Apply the same eye to AI news, and unverified numbers stay out of internal briefings. The power to verify is something pharma work already has. It applies to today's whole story as well.
Wrap-up
What grows short as AI spreads is not intelligence but the capacity to absorb it, or so today's reports read to me. Set the allowed range, measure cost, settle the repair order, doubt outside documents, trace the basis. An existing organization can start all of this. Tomorrow I want to see whether the much-discussed tool works as promised. That answer will shape how the receiving side prepares.
- CH 01Pasquale Pillitteri「DeepSeek releases the open source prototype of its Super-agent」 pasqualepillitteri.it
- CH 01Global Times「DeepSeek unveils ‘Super-agent’ design to handle multiple tasks in parallel and improve efficiency」 globaltimes.cn
- CH 01Bloomberg.com「DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic」 bloomberg.com
- CH 01BusinessToday Malaysia「DeepSeek Expands AI Agent Ambitions With New Coding Prototype」 businesstoday.com.my
- CH 02Business Insider「Anthropic's internal AI use is now worth more than employee pay, at the prices it charges customers」 businessinsider.com
- CH 02The Motley Fool「1 Reason to Buy Microsoft Stock for 10 Years」 fool.com
- CH 02The Decoder「OpenAI is steering ChatGPT users toward more paid usage」 the-decoder.com
- CH 03Startup Fortune「AI agents are finding more software bugs than Silicon Valley can patch」 startupfortune.com
- CH 03Forbes「Elon Musk Issues Serious Hack Warning Amid Bitcoin And Crypto Price Crash Fear」 forbes.com
- CH 03The Times of India「Dell expands cybersecurity portfolio with new AI, identity and quantum-resilience capabilities」 timesofindia.indiatimes.com
- CH 04Tom's Hardware「Brazilian lawyers busted for attempting to manipulate court's AI with prompt injection hidden in filing — court's AI flags hidden text that demanded superficial legal scrutiny, lawyers fined $16,000」 tomshardware.com
- CH 04New York State Bar Association「NYSBA Releases In-Depth Guidelines To Advise Attorneys on Ethical Use of AI」 nysba.org
- CH 04The Business Journals「AI is already creating insurance claims. Is your business covered?」 bizjournals.com
- CH 04Yellow Scene Magazine「The Entire Journalism Industry Is On A Collision Course With AI」 yellowscene.com
- CH 05Open Magazine「EXPLAINED | Why Are Neighbours Taking Elon Musk’s AI Power Plant to Court?」 openthemagazine.com
- CH 05The Korea Times「OpenAI CEO's Codex keyboard gift to Samsung presidents draws attention」 koreatimes.co.kr
- CH 05EnergyNow.com「AI Backlash Hits Home in Virginia’s ‘Data Center Alley’」 energynow.com
- CH 05Wowtale「Sovereign AI Firm Cohere Announces Series E Initial Close, Targeting Up to $5.1B」 en.wowtale.net
- CH 06Yahoo Tech「Microsoft AI beats doctors in simulated diagnosis test」 tech.yahoo.com
- CH 06The Business Journals「Before you roll out Copilot: The data access problem most companies miss」 bizjournals.com
- CH 07Fortune「Anthropic’s $2T IPO could create millionaires—but it’s asking candidates what they’d do if stock hit zero」 fortune.com
- CH 07New York Post「AI giant donates nearly $1M to California Dems just days before election」 nypost.com
- CH 07The Times of India「Microsoft CEO Satya Nadella says that when Dario Amodei, Ilya Sutskever, Sam Altman, Greg Brockman and El」 timesofindia.indiatimes.com
- CH 07Startup Fortune「Yann LeCun's tiny 109M model aligns with vision better than 72B LLMs」 startupfortune.com
Articles used
- AI Daily News 2026-11-02
- AI Daily News: Article Summaries (Nov 2 Morning)
- AI Daily News: Article Summaries (Nov 1 Evening)
- AI and the Pharma Industry 2026-11-02
Today's related reports
- AI Daily News November 2, 2026
- AI & Economy News November 2, 2026
- AI & Finance News November 2, 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.