SNI: WEEK 44
Welcome to all the AI news you won't regret reading – across tech, biopharma, medtech, advanced manufacturing and insurance. At the end of a week in which the end of the world was foretold, priced and promptly tossed aside.
tl;dr: It's the end of the world as we know it. And we feel fine.
Reading Anthropic's IPO filing might finally convince you that AI will be humanity's ultimate cataclysm. And yet – most of us feel fine.
It appears the world still thinks Claude's maker is worth its estimated $2tn. Despite its 90-page risk register.
But those of a nervous disposition might want to look away now.
The filing reminds us that advanced AI could pose 'catastrophic or existential risks to humanity'.
Claude could meaningfully help weapons development.
Its safeguards have been circumvented before and Anthropic has run with known gaps in its defences.
The US state designated the company a supply-chain risk, while proposed laws contemplate nationalisation or shutdown.
And many of the risks could daisy-chain together. For instance, an adverse copyright ruling or a compute shortage could weaken revenue. Slower growth (which Anthropic expects) could then breach loan agreements and trigger defaults across its debts.
We could go on. And on. And on some more (see the footnote below).
And given all that, you could be forgiven for thinking that Anthropic's investors want much the same as the rioters on Europe's streets – to tear down reality as we know it. And hang the potential consequences.
Yet. But. And however:
The prospect of the end of the world has failed to spook most of us this Halloween. Which raises the trillion-dollar question: why, exactly?
Perhaps because spasms of fear now evaporate as quickly as they emerge. Days after PHP's flaws had the web reaching for the off-switch, the predicted wasteland from hacking Armageddon was nowhere to be seen. And such extreme anti-climaxes clearly make us less sensitive to Anthropic's S-1 reading like a doomsday prophecy.
But is there a deeper reason? If the probability of AI killing us is just 10%, it remains rather unlikely. And financial markets shrug because the end itself can't be hedged. But the upside can be grasped – and owning a piece of the technology might be a useful hedge against whatever else happens. Which is why, for a few days at least, the markets stopped focusing on P. Doom and started pricing in P. Boom again.
And then there's the fact that the more useful something feels, the safer it seems. Anthropic's prospectus suggests AI could help cure cancer. And the weekly news cycle tells us there's plenty of other AI happenings to feel fine about:
The ECB's chief economist says AI is now a significant contributor to euro-area growth.
AI calling has closed roughly a third of the pay gap for deaf couriers in China.
A fed-up driver's AI has logged more than 28,000 litter reports.
An AI radio host is backing immigrants and winning over listeners.
Microsoft's Azure grew 48%, a sign, The Money says, that AI spending is paying off.
You can now even hug a plushie that hums like a data centre. It sold out. And Google's AI satellite sent its first words home: a haiku.
Which was just the cherry atop Google's cake. Gemini 4 Carbon put the search behemoth back on top of the AI leaderboard. And its Genie 4 model turns a prompt into an explorable world with sound, in new and remarkable ways. Which rather overshadowed GPT-6.1 Terra being smarter than its predecessor at less than a third of the cost. Although many are not overlooking Europe's latest model. Mistral has opened up Le Chonk, a trillion-parameter model for cyber defence.
But The Power, US-style at least, still fears much of such progress and needs answers. Elizabeth Warren demanded them, the FTC says Anthropic's executives must testify, and House Democrats asked five labs to report safety incidents.
Yet should they be focusing more on the everyday? Two youth campaigners argue that the obsession with disaster ignores the more prosaic and immediate harms.
The World is clearly not listening. Anthropic's S-1 warns that the whole of the global insurance industry may lack enough capacity to cover an AI-related catastrophe. While The Money is fretting that the commercial advantage investors are expecting to pay up to $2tn for might only last a month.
And markets were indeed volatile. Alphabet lost about $500bn in a day, Meta slid as it asked investors for up to $95bn and even Microsoft gave back its early gains. Then on Friday, tech stocks shed more than $1tn. Blackstone, which anchors Anthropic's $63bn chip financing, now expects it to get a high junk rating after the IPO. No wonder capital is getting harder to find.
But whether that strikes terror into our hearts or not, there are many other predictions of what else may be ending. At least 'as we know it'.
Science is a prime candidate. OpenAI says its AI refuted a 30-year-old assumption in computing while a maths bust-up has fed fears AI is undermining how science works. And – just in case encryption goes a bit Pete Tong – the FCA is already trialling a new tool. With not one but 50 large institutions. Timely, then, that Lex Fridman asked a panel to consider whether 'mathematics as we know it is done completely'.
Physics might be on the 'done' list too. Alex Wissner-Gross reckons his AI-assisted Chain Theory 'rebuilds fundamental physics'. The Commentariat raised an eyebrow and reserved judgement. Economist Samuel Hammond called it 'either groundbreaking or
And after the slightly damp PHP squib, we might feel that rumours of the open web's demise are greatly exaggerated. Yet they persist.
Noah Smith wondered whether AI means everyone has to shell out big bucks for private security. But who will fund those big bucks? Google's AI search is claimed to have upended online publishing already. And agents going shopping on our credit cards could upend the rest. Insurify already lets them shop for cover.
After all, who needs humans when Chrome's AI autofill can now read your Gmail? There is symmetry then in Delaware wanting to let AI agents run companies day to day.
But to what extent is any of this real and present? DoorDash has only had several hundred orders from AI agents. Against about 3 million deliveries a day.
Seeing being believing, though, is almost certainly over as we know it, even in Ireland. And globally:
Blue Planet III deliberately used cloudier footage to show it isn't AI.
Prime Video's first fully AI-generated show has been panned.
Footballer Cristian Volpato has licensed his 'synthetic twin'.
Hasbro plans to have the late Don Rickles' voice cloned to play Mr Potato Head.
People with visible differences are objecting to AI tools that erase them.
Expertise, as we know it, is also moving house. Anthropic bought Mithril Medicine, a company with one person and a vast Claude habit. Not that it was the one that found four severe bugs in software that runs particle accelerators. But what might AI find as Tesla begins early robot production on a line it says can make a million a year? Let's hope the Seoul hospital planning humanoid assistants in its operating theatres is hack-proof.
And finally, we cannot ignore another potential end of the world: is AI adding to the environmental existential risk we already knew about?
Some of the money Anthropic raises will be spent on fossil fuels. And a new analysis reckons power for AI data centres will undo 15 years of cuts. Closer to home, SEAI data shows, data centres are also eating up Ireland's renewable gains.
But while campaigners accuse our EU presidency of weakening AI data-protection safeguards, perhaps the world isn't going to end?
Exhibit A: Trump says the term 'artificial intelligence' will soon be banned in favour of 'super intelligence'. He clearly hasn't read the First Amendment.
Which brings us to this week's ultimate take-away: Many things persist in the face of almost any change. They might not end up as we know them. But they also keep going in so many other ways. And those ways are usually fine.
On that not-so-super-intelligent but somewhat-philosophical bombshell, here's everything else worth reading this week.
Biopharma:
AI drug developers go shopping: Bristol Myers Squibb leads Pathos's $237m round, and Formation Bio launches a $600m buying fund.
Early responses for a 'too hard' cancer target: two of five mesothelioma patients responded to Insilico's AI-designed treatment.
Three biotechs become one AI drug hunter: Flagship folds Cellarity, Etiome and Empress into Etcelera.
The hedge against AI is healthcare: Raymond James calls the sector, up roughly 23% since late April, an 'anti-AI trade'.
Roche asks the folding models to spot a binder: the best model picks the right antibody from 50 just 29% of the time, a preprint finds.
Medtech:
The specialist says the chatbots get doses wrong: Doximity finds a serious or potentially lethal error in about 19.5 of every 1,000 answers.
The patient's agent books the appointment: Zocdoc opens booking to registered AI agents acting with the patient's consent.
The AI scribe rolls out across A&E: a pilot of TORTUS's ambient voice tool cut documentation time by 61%.
The AI's label comes off the health record: America's health-IT regulator drops a requirement to display an algorithm's training data to clinicians.
The radiographers still see more fractures: standalone AI caught 75.6% in a live emergency department, against their 91.6%.
Advanced manufacturing:
A €1.1bn European fab for AI data centres' power chips: Infineon plans what it calls the first high-volume 300mm gallium-nitride fab, in Austria.
Humanoids start making humanoids: six robots assemble actuators for the next generation at Schaeffler's new plant.
A $20bn price for robot brains, before revenue: Physical Intelligence is in talks, as Bedrock's self-driving excavators win a $4.1bn valuation.
A robot that learns from one demonstration: Machenta raises $40m, the largest seed yet for a UK robotics start-up.
Washington finds it can't make enough robots: Trump opts for trade talks over tariffs on robot and machinery imports.
Insurance:
Quoted or declined before an underwriter reads the email?: Guidewire agrees to buy London's hyperexponential for $1.7bn, plus up to $400m more.
Four in five customer contacts, handled by AI: Ping An's AI service agents logged about 1.48 billion interactions in nine months.
Snap the renewal, skip the form: Allianz Ireland's AI prefills a motor quote from a photo, 85% faster than phoning, and home cover can now be quoted inside ChatGPT.
Crop cover priced by the field: Aon and SAGE's AI will price corn and soybean growers individually in 15 states for 2027.
The boardroom sees more AI than the managers below it: 73% of insurance C-suite respondents report AI at scale, against 51% of senior managers, BlackRock finds.
But what set podcast tongues a-wagging?
AGI keeps getting closer. Yet we've barely used what's here already.
Exponential View notes that Demis Hassabis keeps shortening his AGI estimate: five to ten years at Davos in January, then three to five, now two to three in a Wall Street Journal profile. Azeem Azhar calls the Google DeepMind co-founder 'one of the field's more measured voices', and asks whether we can test and apply AI's discoveries fast enough.
Meanwhile, Professor Ethan Mollick told Alex Kantrowitz on Big Technology: 'We don't need better models for complete transformation of a lot of how we work in the next 10 years.'
He calls the unused capability 'the overhang': stop progress today and filling it would take five to ten years. The constraint, he says, is organisational: after 30 years of leaning on software vendors and consultants, firms have 'very little muscle memory left'. Those that do rebuild their teams can go far. One company he spoke to cut a $200m mainframe data-migration plan to $20m. And its CTO now wonders about $2m.
An off switch only works if we can manage without it.
RAND's Matt Chessen, Joel Predd and Ben Boudreaux have mapped seven strategies to be applied to superintelligence, from dominance to a moratorium. The evidence, they told Nathan Labenz on Cognitive Revolution, can't yet justify any of them.
But one thing that is clear – threatening to shut down a dangerous model is credible only if we can live without it. Which is why Boudreaux wants 'human backstops' – systems that can still be run by hand, paper copies of health records. Is this realistic? Well, he admits, he often fails at it himself.
Hand over the keys one sensible step at a time.
Nathaniel Whittemore reported back on his first days with Claude Duet, Anthropic's new work-focused agent, on the AI Daily Brief. It asked for his Slack, Gmail and Drive straight away and offered a stack of projects to take on unprompted. Some proved overeager: it warned him four times of a double-booking that wasn't one. But when an email arrived of the kind he'd usually star and leave for later, Duet drafted a reply he could edit or authorise from the chat. 'A couple of times I've just let it send.'
AI can close an enterprise deal in weeks. But its emails still give it away.
Rob Snyder on Lenny's Podcast said that he saw an enterprise deal close in about two weeks. 'When I was growing up, enterprise deals took 18 months.' Where the need is pressing, buyers see an AI pitch, realise it can do the job and buy almost at once.
What Snyder has yet to see work is AI-personalised outreach. It usually tells people something they already know about themselves. And it's instantly dismissed because it looks like Claude wrote it.
FOOTNOTE:
Here are some more of the risks in Anthropic's S-1 filing:
Anthropic depends on chips made by TSMC, so a natural disaster in Taiwan or tensions with China could disrupt its supply.
Anthropic is being sued over its training data. A separate antitrust suit alleges the labs agreed to slow down.
Anthropic spent about $4m on personal security for Dario and Daniela Amodei, and says its success is 'highly dependent' on their continued service.
Over 80% of third-quarter revenue came from models less than six months old. Nearly half of first-half revenue arrived through cloud providers' marketplaces, and those providers are also Anthropic's suppliers, investors and rivals.
The founders' single share carries 50.1% of the vote and the trust elects a board majority, so public shareholders may not be able to affect the outcome of any vote.
Thank you for reading Second Nature Intelligence. Come back next week for all the AI news you won't regret knowing.
