AI · page 2

88% of US firms use AI vs 26% in Japan, but depth beats speed, HBR authors argue

• Four researchers say US firms adopt AI widely but shallowly: McKinsey finds 88% use it somewhere, yet only 39% report any EBIT impact, and Deloitte says 30% have redesigned key processes.
• Japan lags on uptake (26%, per Yano Research) and ranks 60th for business agility in IMD's index, but goes deep where it adopts: Toyota factory workers had built 10,000 AI models by 2024.
• Their fix is a six-rung ladder. US firms should redesign a few workflows end to end and cap AI tools per role; Japanese firms need a fast lane for small AI decisions.

Other side: The figures come from different surveys, years and definitions, so the gap is rough. And the authors themselves note productivity often dips before it surges with a general-purpose technology, so today's shallow US use may be an early phase.

See the KKReckons explainer →

Summary based on Harvard Business Review (via The Economic Times). Read the full story ↗

John Thornhill: will AI's 'cognitive capitalism' spread wealth or concentrate it?

• Thornhill weighs Venkatesh Rao's argument that AI worriers are "cognitive mercantilists", treating intelligence as scarce and to be hoarded, as pre-industrial elites treated wealth.
• Rao's hope: cheap AI lets billions build and start microbusinesses, a new "industrious" revolution, with AI agents adding demand.
• The counter: AI is concentrating power, US stocks are at record value and concentration, and Pew finds 40% of Americans expect harm versus 16% benefit, though 49% use chatbots. DeepMind's Alex Imas says the missing piece is organisations built for humans and AI agents.

Other side: The analogy may flatter AI. Unlike looms, frontier models and compute are owned by a handful of firms, so abundance for users can sit alongside tighter concentration for owners.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Europe is falling behind on AI and tech, and Britain must chart its own path, says Keegan McBride

• Keegan McBride, writing with Sam Cash, says frontier AI scales with compute, capital and energy, and Europe is short of all three; Goldman Sachs puts AI-related capex at just under $8tn over five years.
• He cites DIGITALEUROPE research finding Europe behind in seven of eight critical technologies, including old strengths such as robotics and space.
• With only the US and China offering full tech ecosystems, countries must pick partners. Britain, he argues, should become a "tech-enabled state": bet on space, quantum and life sciences, stay close to the US, and rebuild government with AI.

Other side: The case leans heavily on aligning with the US. Sovereignty advocates argue that depending on American firms is itself a strategic risk, and Europe's caution on rules could look wiser if AI harms mount.

See the KKReckons explainer →

Summary based on Technology, Geopolitics, and Bureaucracy (Substack). Read the full story ↗

Martin Wolf: AI buildout at 3.6% of US GDP raises the risk of a bust

• Martin Wolf draws on a new Brookings paper by Columbia's Stijn Van Nieuwerburgh, which projects US spending on the AI buildout at 3.6% of GDP a year to 2032, a bigger sector boom than the railways of the late 1800s.
• Big Tech's spending is now outrunning its cash flow, he notes, so data-centre developers, infrastructure funds and private credit are filling the gap, much of it through loans outsiders can barely see.
• His bigger worry is stocks: US valuations sit where they have historically come before crashes, and the longer the boom and the financial engineering run, the harder the landing.

Other side: Booms that end in busts often leave useful infrastructure behind, and if AI profits arrive fast enough today's spending will look prudent. Wolf himself grants AI may still prove transformative.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

AI is speeding up hacks. Are banks in India and abroad ready?

• An ET Prime analysis notes that India's CERT-In, in an October 15 advisory, saw attackers using AI agents go from first access to impact within hours, exploiting familiar gaps such as exposed credentials and excessive permissions.
• It argues bank boards should judge how fast they can spot and shut a breach, and tightly limit what fintech and payment partners can instruct.
• In the FT, Chicago Booth's Anil Kashyap calls the Binance hack a warning shot: an AI-aided breach of a major bank is a matter of when, and lightly overseen crypto could be the weak link.

Other side: Banks are among the most watched and best-guarded firms, and AI strengthens defenders too. Years of warnings have passed without a crippling bank breach, so 'when, not if' may overstate the near-term odds.

See the KKReckons explainer →

Summary based on The Economic Times and Financial Times. Read the full story ↗

After a blue wave, data centre fans look to space

• Democrats won widely on Tuesday, taking the US House and Senate and doing well in Texas, a data-centre hot spot, with many winners campaigning against the industry, Wired reports.
• Elon Musk, who spent more than $400 million backing Republicans who all lost, posted that data centres in space now look appealing; investors, researchers and fans joined the chorus.
• Experts told Wired change in Texas may be limited: Republicans still hold the state senate, developers' permits carry strong vested rights, and counties lack zoning power over data centres.

Other side: Orbit is no quick escape: Wired notes space data centres cost far more, cannot reach hyperscale size and depend on launches. With a split Texas legislature, change may be slower than either side expects.

See the KKReckons explainer →

Summary based on Wired. Read the full story ↗

Binance says fake job interviews opened the door to its $6.3bn heist

• Binance's own investigation says the thieves behind last week's theft, which the FBI blames on the North Korea-linked group TraderTraitor, posed as recruiters and gave an engineer an interview task laced with malware.
• That foothold, plus an earlier tampered download on another engineer's laptop, let them use Binance's in-house AI tool to take over colleagues' machines and trigger the attack on 29 October.
• The exchange says it has rebuilt its wallet systems and AI tools, set aside money to cover users' losses and reopened withdrawals.

Other side: This is Binance describing its own breach, and it is keeping some security details private. Its account may stress the attackers' cunning over its own control gaps; an independent review would tell more.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

AI's borrowing bill is rising as bond buyers demand more

• The FT's editorial board notes that Alphabet, Meta and Oracle went back to bond markets this week but had to pay clearly more than a few months ago; Oracle's new bonds carry coupons above 8.5%.
• Insuring against default is dearer too: protection on Meta's debt now costs about 1.2% a year, which the board likens to junk-rated South Africa.
• Its view: the AI boom may not end with a stock crash but with a slow squeeze in credit. Oracle has even pledged to cap the new money it raises from markets until 2030.

Other side: These firms still hold strong ratings and cash-rich core businesses, and they are still raising huge sums. Repricing risk is credit markets working, not breaking.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

AI is shutting the door on new hires, not pushing workers out

• Sarah O'Connor argues the real AI jobs shock is fewer hires, not mass lay-offs. History fits: boot-sewing machines in Victorian England choked off new artisans rather than sacking existing ones.
• Today's signs match. Lay-offs stay low, entry-level software hiring is falling while demand for seniors holds, Danish firms using AI simply hire more slowly, and call centres shrink because hiring slowed.
• She wants policy aimed at the young: closer ties between schools and employers and more support for start-ups, since retraining schemes such as Sweden's often leave them out.

Other side: It is early and the data are thin. A recession could still bring abrupt cuts, and young people have often adapted fastest to new technology.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Why AI is not yet showing up in productivity data

• Zerodha Varsity revisits Robert Solow's 1987 quip that computers were visible everywhere except the productivity statistics; US trend labour productivity growth had slid from 3.8% in 1948 to 1.2%.
• History shows a lag: factories gained from electricity only after redesigning around individual motors, lifting manufacturing productivity 5.3% a year in the 1920s. Computers paid off in the late 1990s.
• AI forecasts diverge: Daron Acemoglu sees under 1% over a decade, assuming about 5% of tasks shift to AI; Goldman Sachs sees about 1.5 points a year, assuming about 25%.

Other side: The J-curve is a hypothesis, not a promise: if the gains never arrive, today's AI spending looks less like early investment and more like overbuilding.

See the KKReckons explainer →

Summary based on Zerodha Varsity. Read the full story ↗

Elliott warns of FOGO as bear-market alarm bells ring

• In an October investor letter seen by the WSJ, Paul Singer's Elliott likens today's debt-fuelled, hype-driven market to the Roaring '20s and says almost every bear-market warning is flashing.
• It lists US deficits, sticky inflation, private credit, AI hype and circular AI deals, plus gambling-like trading, and warns the mood could flip from FOMO to fear of getting obliterated.
• Elliott itself has lagged, gaining under 2% in the first nine months of 2026 against nearly 13% for the S&P 500, and is keeping capital ready for forced sellers.

Other side: Elliott's letters are known for gloom, and frothy markets can run for years; being early on a crash can cost nearly as much as missing it.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

OpenAI and Anthropic say AI disproved the exponential time hypothesis

• OpenAI and Anthropic each said this week that their models disproved the exponential time hypothesis, a 27-year-old assumption that some hard problems take exponentially more effort as they grow. Both posted the proofs on GitHub, in plain language and in the machine-checkable Lean format.
• Independent researchers have not verified them yet. OpenAI has asked the UC San Diego scientists who proposed the hypothesis to rally volunteers and has promised to fund a conference.
• Reactions are split. Linköping's Peter Jonsson says much of his upcoming conference paper no longer works, while Copenhagen's Thore Husfeldt calls it a once-in-a-generation result but says the steady run of AI results has left colleagues demoralised.

Other side: Nothing is settled until people check the proofs. Machine checking confirms the logic, but humans still have to confirm the formal statement says what the hypothesis says. And after earlier rushed AI paper dumps with errors, some academics suspect the labs time big claims to impress investors.

See the KKReckons explainer →

Summary based on Wired. Read the full story ↗

OpenAI's discounts lure customers from Anthropic before its IPO

• The Information reports that OpenAI's heavy discounts on its models and its Codex coding tool are pulling some business away from Anthropic. At HR software firm Rippling, Anthropic's share of coding work has dropped to 18% since May, with OpenAI now handling most of it.
• Replit moved most of its AI spending to OpenAI after a discount of more than 30%, a person familiar said, once Anthropic declined to extend its own discount without a bigger commitment.
• Anthropic has largely held firm on pricing and margins, and its new Opus 5.5 model has since retaken the lead on routing service OpenRouter.

Other side: Discounts that deepen losses cannot run forever, and a better model can win customers back fast, as the OpenRouter rebound suggests. A handful of startup anecdotes may also say little about large enterprise contracts.

See the KKReckons explainer →

Summary based on The Information. Read the full story ↗

AI agents widen what workers attempt, not just how fast they work

• Perplexity researchers compared the same users on its chat assistant and its autonomous agent from February to May 2026. The agent worked alone far longer per task and paused often to check in, yet users rarely stopped it.
• Counting the human follow-up, agent workflows cut task time by about 87% in their estimate, and users were less often dissatisfied.
• The bigger shift was scope: people handed agents harder, multi-step work outside their own field, and about 23% of agent tasks were ones they never tried with the assistant.

Other side: The authors work at Perplexity, studied early power users of their own product and call the magnitudes approximate. Crossing into unfamiliar fields also means more output that its user cannot properly judge.

See the KKReckons explainer →

Summary based on Harvard Business Review. Read the full story ↗

Wall Street gets choosier about funding the AI build-out

• Lenders and investors who funded AI with few questions now demand tougher terms. Neoclouds such as CoreWeave, Nebius and Nscale have effectively been shut out of the bond market, and one CoreWeave bond now yields above 15%.
• Strong borrowers feel it too: hyperscaler bonds trade at discounts, Broadcom sweetened terms for banks, OpenAI's round has fallen short of its goal, Lambda cut its target valuation and Nscale is pushing back its IPO.
• Worries include local pushback against data centres, AI profits arriving late and investors overloaded on one theme. Nebius is leaning on customer prepayments.

Other side: Money is choosier, not gone: creditworthy firms still closed giant chip financings this week. A pause that weeds out shaky projects could leave the build-out on firmer ground.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

AI's job hit so far looks like a Big Tech story, not an economy one

• A US analysis found productivity jumping in heavy AI-adopting sectors, and a Goldman Sachs note tied higher AI use to steeper job losses. John Burn-Murdoch says the picture blurs once you widen the lens.
• Outside the US, and in finance and professional services, the link largely fades. The gap between output and headcount is mostly a tech story, and it opened in 2022, when AI meant chatbots, not agents.
• His reading: rate hikes and uncertainty made firms stop hiring, and Big Tech cut deepest, long before production-ready AI.

Other side: If firms now believe AI lets them do more with fewer people, a pause that began with rate hikes could harden into lasting AI-driven cuts. Does the original cause matter once the jobs don't return?

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

New malware uses AI chatbots to profile and rank its victims

• Cisco Talos found an infostealer, dubbed Volta-Autodoxer, that grabs passwords, tokens, cookies, cards, crypto seed phrases and gaming and chat sessions, then digs through email for receipts and order history.
• It bundles the haul into a dossier and asks an AI model for an analyst-style assessment of the victim, to rank who is worth targeting first. If a model refuses, it falls back to more permissive ones.
• Talos did not track whether it has been used in the wild, and the malware works without the AI step. Clues suggest it was built for sale to many attackers.

Other side: The AI layer is an add-on to an ordinary stealer, and its first-choice model may refuse. Basic habits, like not saving passwords in browsers, still blunt the core attack.

See the KKReckons explainer →

Summary based on Wired. Read the full story ↗

AI agents get better, but businesses see barely any agent orders

• Wired reports Meta says over 10 million people in the US and Canada have activated its Muse agent, with shopping the top task. Yet most firms on recent earnings calls, from DoorDash to Hilton, saw little agent activity.
• Some are preparing anyway: Chipotle is fixing its site so agents can check out, and Meta plans a fee on agent sales. Amazon softened its stance after blocking Muse; Insurify reversed its block.
• In a separate Wired test of six agents over a day, OpenAI's Dots scored best and Muse came a close second as the easiest to set up. All still made mistakes.

Other side: Low numbers may say more about early days than demand, and one reporter's day of tests is a small sample. Price-comparing agents could also help shoppers more than they hurt sellers.

See the KKReckons explainer →

Summary based on Wired. Read the full story ↗

Why the D-Day weatherman is a warning for the AI age

• An ET column uses the film Pressure to retell how the Allies settled the D-Day date: British meteorologist James Stagg built his forecast from fresh readings and the jet stream, while American Irving Krick extrapolated from historical charts. Stagg, who admitted uncertainty, was right.
• The lesson, the writer argues, is that causes come before correlations: numbers and statistics are facts, not truth, in weather as in language, behaviour and markets.
• The worry is commercial hype that thinking and creativity can be handed to AI's pattern matching, costing truth first, then reasoning and the scientific temper.

Other side: Stagg relied on data too, so the line between reasoning and pattern-matching is blurrier than the parable suggests. The real risk may be blind trust in any forecast, human or machine.

See the KKReckons explainer →

Summary based on The Economic Times. Read the full story ↗

AI's power bill means data centres must be designed as energy assets

• Flipspaces founder Kunal Sharma notes data centres used about 415 TWh of power in 2024, and the IEA expects that to more than double to roughly 945 TWh by 2030. In the US they could drive nearly half of electricity demand growth.
• He argues energy must shape the building itself: site, rack density, cooling routes and power distribution planned together, so facilities can take denser chips without costly redesigns.
• Waste heat can become a resource: Microsoft projects in Denmark feed surplus heat into district heating networks.

Other side: Better design trims waste but cannot create electricity; generation and transmission remain the real bottleneck. Heat reuse also needs nearby homes or businesses to take it, which remote sites may lack.

See the KKReckons explainer →

Summary based on The Economic Times. Read the full story ↗

India saw 369 million cyber incidents in a year. Is it AI-ready?

• Citing the Data Security Council of India, an ET commentary says 369 million security incidents were detected in 2024, about 702 a minute, with healthcare, hospitality and BFSI hit hardest. It flags a 42% shortage of skilled cyber staff and deepfakes used to beat KYC and money-laundering checks.
• The writer calls for a national AI Act that fixes algorithmic liability, a stronger AI Safety Institute able to test models used in India, and company rules on AI use, noting a global study where 57% of employees hid their AI use at work.

Other side: Part of the jump may reflect better detection rather than more attacks. And strict rules too early could slow AI uptake in a country that already ranks low on adoption.

See the KKReckons explainer →

Summary based on The Economic Times. Read the full story ↗

From Vizag to Thane, data centres meet local pushback

• Villagers near Google's $15bn AI hub in Visakhapatnam allege land was taken back without proper consultation, and activists object to tree-clearing near the Kambalakonda sanctuary. In Thane, residents oppose Amazon's proposed 53-acre hyperscale site near homes, schools and hospitals.
• An ET commentary says India needs the compute but should earn consent: water-saving cooling and clean power, siting away from cities where possible, visible benefits for host communities, and rules set before the damage. It suggests treating big data centres as quasi-public utilities.

Other side: Sites far from cities mean higher costs and slower service for some uses, and community payouts can slide into bribes for bad projects. Utility-style rules could also deter the investment India is courting.

See the KKReckons explainer →

Summary based on The Economic Times. Read the full story ↗

Intel's record crash caps a week when the AI trade cracked

• Intel fell 29% on Friday, its worst session ever, after warning that memory-chip prices inflated by the AI boom will keep PC sales weak, the WSJ reports. The Nasdaq composite fell 2.9%, its worst day since June.
• Alphabet had its worst day on record as Google's search-ad growth slowed, partly due to AI mode, while Amazon's free cash flow plunged. With the Fed raising rates in September, heavy AI spending now reads as a burden.
• An FT Unhedged column argues cash flow is what moves these stocks now: Meta, Alphabet and Amazon were punished for spending more than expected, while Microsoft's restraint kept its shares steadier.

Other side: One bad week is not a bust. Some investors expect AI listings and a calmer Middle East to restart the rally, and steep falls in quality names have often proved buying chances.

See the KKReckons explainer →

Summary based on The Wall Street Journal and Financial Times. Read the full story ↗

Apollo lends as OpenAI-backed Thrive buys ADT for $15bn

• Thrive Holdings, spun out of Josh Kushner's Thrive Capital and backed by OpenAI, is buying home-alarm network ADT for $15bn including debt. Apollo, a past owner that made modest returns on it, is financing the deal.
• The FT's Lex column says a new breed of AI-focused owners is promising bold makeovers of old businesses, while firms like Apollo shift from buyouts to corporate lending: steadier interest, less risk and less upside.

Other side: An AI-led makeover of an alarm company is still a theory, and a buyer paying a big premium has less room for error. Lenders also lose if a heavily financed bet goes wrong.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

AI now leads a quarter of AI research at Anthropic, company data show

• A WSJ column spotlights an Anthropic chart tracking how much of its AI research is done by AI. The share where AI leads the work, with a human supervising, rose from 1% in March to 26% in August; none of it is yet fully autonomous.
• The idea behind it is recursive self-improvement: each better model helps build a better one. Lab leaders call it vital to stay ahead and also a serious risk, and Anthropic itself says it is not inevitable but could arrive before institutions are ready.
• Co-founder Jack Clark has pulled his own estimate for it forward to next year.

Other side: The numbers come from Anthropic's own AI rating its own work, on a company scale with no outside audit. Leading a task under supervision is not the same as independent invention, and a company heading for a listing has reasons to show fast progress.

See the KKReckons explainer →

Summary based on The Wall Street Journal. Read the full story ↗

Alphabet's ad slowdown and Anthropic's IPO test the AI money machine

• The FT's Lex column says Alphabet lost nearly half a trillion dollars of market value after ad sales slowed unexpectedly: AI answers leave less room for ads, and data-centre spending is outrunning cash. Lex still thinks the AI lead can be won back.
• A John Thornhill column says Anthropic's IPO pits fear of missing out against fear of messing up. The filing shows first-half revenue of $16.3bn and an $80.7bn net loss, mostly non-cash, Xinhua reports, plus large non-cancellable compute deals.

Other side: Both columns concede the upside: AI leadership has flipped quickly before, and Anthropic's revenue growth is remarkable. A sharp share drop or a pricey IPO is not proof the boom is a bust.

See the KKReckons explainer →

Summary based on Financial Times. Read the full story ↗

Anthropic says Claude drafted 35,000 antibody designs in a month

• Anthropic says it spent months working with six drug developers, including Bristol Myers Squibb and Generate Biomedicines. For Divergence Labs, Claude produced about 35,000 antibody designs for a food-allergy drug in roughly a month; in early tests, three of the 20 that dislodged IgE did better than the standard medicine.
• At Enveda, Claude checked 1,061 pages of manufacturing paperwork, caught all 11 issues human reviewers found plus seven more, and cut review time by 89%. BigHat's best antibody variants bound more than 10 times more tightly.
• In Anthropic's own muscular dystrophy programme, a candidate went from idea to promising mouse results in under 18 months; human trials could start in late 2028 if results hold.

Other side: Almost all of this is preclinical and reported by the vendor and its partners. Most drug candidates fail in human trials, so speed at the start says little yet about cures.

See the KKReckons explainer →

Summary based on Anthropic. Read the full story ↗

Snowflake CEO: AI labs risk a capacity bust, not a bank run

• Snowflake CEO Sridhar Ramaswamy argues in a thread on X that AI labs commit to compute for years while selling usage customers can cut tomorrow. In his hypothetical, $100 of capacity against $150 of expected revenue leaves $50; a 30% revenue miss cuts that to $5.
• He draws lessons from banks, commodity traders, airlines, utilities and cloud giants, and warns that clouds which invest in labs and also sell them compute can make one promise look like several revenue streams.
• Citing Anthropic's S-1, he says compute cost fell to 55% of first-half sales of $16.3bn, from 201% a year earlier.

Other side: He is not a neutral observer: Snowflake sells data and AI services. And contract-backed cash flows that lenders treat as safe are only as strong as the labs signing them.

See the KKReckons explainer →

Summary based on Sridhar Ramaswamy on X. Read the full story ↗

China's AI stocks crash 40% as US neocloud debt stress spreads

• Reuters reports China's STAR50 index of AI and chip leaders has fallen 40% in under four months, erasing about $480bn; AI stocks are 42% of China's market value, per Goldman Sachs. The Nasdaq has slipped only 6%.
• In the US the strain shows in credit. ET says CoreWeave's 2032 bonds trade below 80 cents, Musk predicts a die-off among 200-plus neoclouds, and Gartner sees $8.5tn of revenue needed by 2030 against at most $3.6tn.
• Ed Yardeni says air is leaking without a burst. An FT editorial warns hedges are getting dearer; Columbia's Neeraj Kaushal argues AI can transform the economy and still destroy capital.

Other side: The biggest spenders are among the most cash-rich firms ever, and overbuilt railways and fibre were later put to use. Gartner's gap is a model, not a measurement, so losses may stay with leveraged lenders rather than the wider economy.

See the KKReckons explainer →

Summary based on Reuters, The Economic Times and Financial Times. Read the full story ↗

Silicon Valley parents push schools to cut back on classroom tech

• Bay Area tech workers are paying for screen-light schooling, the Guardian reports: San Francisco Waldorf, two miles from Anthropic's headquarters, charges $40,000 a year for kindergarten, and many new families work in tech.
• Parents are pushing public districts too. San Mateo has stopped sending Chromebooks home with pupils up to second grade and bans high schoolers' phones; Mountain View is moving away from the i-Ready program and Palo Alto bans smart glasses.
• Experts warn poorer districts may keep the screens richer ones drop, while school-choice vouchers move public money to private schools.

Other side: Good tools can personalise practice and fill gaps where teachers are scarce, and the parents' main ask is evidence and vetting, not a blanket ban. A few wealthy districts may not speak for families who rely on school devices.

See the KKReckons explainer →

Summary based on The Guardian. Read the full story ↗