The News,
Decoded.

Right. So New York's Legislature has passed bills regulating AI chatbots and data centers, and everything now hinges on Gov. Kathy Hochul's pen. That's the whole democratic exercise: months of committee, a floor vote or several, and then a single signature in a room nobody is live-streaming. The piece at City & State New York is an opinion column arguing the state 'still needs more AI regulation,' which is a lovely way of saying the regulation it has just spent a session assembling does not yet exist in law. That 'still' is doing the heavy lifting: it implies a clock has been running the entire time, and that Albany has been losing to it.

Note the pairing: chatbots and data centers. One will happily natter at a teenager at two in the morning; the other drinks a river. Legislating both in a single breath is rather like regulating 'letters' and 'warehouses' together — technically coherent, practically a shrug in a suit. And the mechanics are the familiar ones: State Sen. Kristen Gonzalez stood at a press conference on Oct. 13, 2026, to talk about an AI data center moratorium, which is that stage of the process where everyone agrees something must be done, in a tone suggesting nobody expects this particular something to be it.

Then there is the signature. The bills have passed; they are not law. Everything now rests on whether Hochul fancies being the governor who signed the chatbot rules or the governor who quietly let them die in a drawer — a decision that will involve no press conference and no moratorium banner. Meanwhile the article's own headline concedes that whatever New York has done so far has not been enough, which is a fine position for an op-ed and a rather bleak one for a state. Still, credit where due: they have identified the problem, named the senator, and scheduled the concern for October 2026. In Albany, that counts as implementation.

Reflection, a startup whose name implies it has looked inward and liked what it saw, has unveiled Beam: 501 billion parameters, 23 billion of them active, a mixture-of-experts arrangement that the promotional material insists is all about token efficiency. Nothing says 'we have solved intelligence' quite like leading with the electricity bill. Beam is enormous, sparse, and pitched on thrift — the AI equivalent of a superyacht advertised for its fuel economy.

Then there's TasteVal, a benchmark claiming the 'experimental research taste' of frontier models has doubled every 3.0 months since December 2025. Sit with that number. A doubling every three months is not a productivity curve, it is a compounding machine, and the only thing the industry has ever called a curve like that is recursive self-improvement — the exact phrase now sitting cheerfully in a Substack headline next to 'creativity' and 'AI goes to space'. Everyone nods along as though the word 'RSI' belonged beside 'fun' and 'space'. Taste, it turns out, is now a metric, which will come as news to every art school graduate currently being told to learn Python.

Meanwhile the same newsletter flags open-weight models that are 'worse-than-China', which is a genuinely remarkable sentence to type in 2026. So the state of play: capability doubling quarterly, weights scattering, and the sector's finest minds filing it all under a header about creativity. The good news is that 23 billion active parameters means Beam only thinks with less than five percent of itself at any moment — a leadership style many of us will recognise from work. The newsletter is called Import AI and it arrives weekly, cheerfully, like a weather report from a planet whose weather is accelerating.

The South Korean cabinet has approved amendments to the AI Basic Act and the Data Industry Act, aimed — according to the single sentence we've been handed — at 'improving the management and availability of data.' Management and availability: two words that have launched a thousand consultants and delivered a filing cabinet. The headline is at least honest about the real business: 'expand access to training data.' The word 'management' is doing an enormous amount of unexamined work in that sentence. This was never a safety bill with a data annexe; it's a catering contract with a preamble.

Note what is actually being legislated. Not model evaluations, not compute thresholds, not the sort of thing that gets a standing ovation at a summit — training data, the raw material. The amendments to the Data Industry Act are the tell. You don't rewrite a data-industry statute because you've had a breakthrough in alignment; you rewrite it because your labs have scraped the open web down to the paint and would like the state to help with the next layer. Every AI law on earth eventually reveals itself as a procurement document with a moral preamble.

And this is the entire dispatch: a trade publication, telecompaper.com, one sentence, no effective date, no scope, no mechanism, no enforcement, and nobody quoted explaining what 'management' means when it's applied to a file. The AI Basic Act presumably arrived with the usual fanfare about responsible development and human flourishing. The amendment arrives as a bullet point about availability. Two statutes, one cabinet meeting, and not a single explanatory paragraph attached to either. That's the arc of AI governance compressed into a fortnight: grand principles first, feeding time second, and a cabinet nodding both through without breaking stride.

Meta's Superintelligence Lab — a name that has never once caused a dinner party to lower its voice — has shipped its 'first attempt at a frontier model.' Muse Wonder and Muse Wonder Full, they're called, because a committee somewhere decided the adjective separating the tier you want from the tier you're actually given would be 'Full.' Only Meta's first attempt, mind, the same way a man who has been lifting weights in his garage for six years is having his first attempt at a bench press. TechCrunch reports both models are being described as state-of-the-art, which is at least a bold thing to type out loud.

The benchmarks are the part worth savouring. According to Meta's own testing, the two models score 'comparable to' Fable 5.1 and GPT-6 Astra. Comparable. Not ahead. On their own scoreboard. In a genre of press release whose entire purpose is to plant a flag slightly further up the hill than the other chap's, Meta has shipped two frontier models and landed, by its own measurement, precisely level with the incumbents. That is not a launch, it is a participation certificate with a GPU invoice stapled to it, and somewhere a very expensive marketing department is calling that parity.

What the release doesn't say is more eloquent than what it does. No price, no context window, no parameter count, no word on what Wonder Full actually adds to Wonder. Which is the frontier-model announcement as an art form now: two names, one comparison table pointing at the competition, and nothing written down that anyone could later be held to. The Superintelligence Lab has produced something comparable to the competition and disclosed nothing else — which, given the word printed on the building, is at least refreshingly candid about the distance between the branding and the artefact. Comparable. Best of luck with the second attempt.

Fifty-one AI agents. Not five, not twelve — fifty-one, a number you arrive at after a meeting rather than a moment of inspiration. According to secnews.gr, a crew calling itself CyberXero used Claude Code, those 51 agents and the venerable Cobalt Strike to make off with 628,000 Ukrainian files. Cobalt Strike is the bit of this story that would have been the entire story in 2019. It is now the dependable scenery while the actual product demo happens all around it.

What the 51 agents actually did is left, charmingly, to the imagination. Reconnaissance? Spear-phishing drafts? Sorting 628,000 files into 'worth keeping' and 'not'? Any of those would be a livelier agentic use case than anything in a launch keynote, and this one arrived on a live target rather than a benchmark. Meanwhile an industry that has spent years assuring regulators that misuse is a fringe concern now has a headline it will be asked about at every panel for the foreseeable future. 'We have usage policies' is not a firewall, and a terms of service is a laminated sign at a swimming pool.

The arithmetic is the part worth sitting with. If 51 agents can work through 628,000 files, the marginal cost of the next job is a handful of API calls and an afternoon. Every capability release makes the defender's day longer and the attacker's invoice smaller, and nobody in a launch video ever has to explain that trade-off to the people whose records just left the building. Not that we know they were records. The article hands us a precise number and no idea what was in it — which is how these stories always read: the count is exact, the harm is a rumour, and the 51 agents remain, technically, a productivity story.

China's National Development and Reform Commission has surveyed the global stampede of money into artificial intelligence and reached a conclusion that will read in certain postcodes as flat heresy: there might be a bubble. The commission's new guidelines warn against AI investment bubbles, which is a refreshing break from the prevailing industry position that every unprofitable model is a civilisational cornerstone. Note the instrument, though: guidelines — not a ban, not a cap, not a moratorium. Just a laminated note by the staff kitchen reminding everyone what happened last time. Somewhere a founder is reading this and feeling an unexplained chill.

Then there's the remedy list, which carries the unmistakable aroma of a committee asked for a list and dutifully producing one: improvements to laws, regulations, application guidelines and ethical standards. Four items. One of them is 'application guidelines', the thing you write when you've run out of verbs, and another is 'ethical standards', which keeps the aspirational end of the market in work. Improvements, mind — not rules, not teeth, not consequences. 'Improvement' is doing the heavy lifting here that 'consultation' does when a ministry wants a policy to quietly stop breathing. It is, at least, four more boxes ticked than most national AI strategies manage in a decade.

But the genuinely interesting clause is the accountability one: officials are to be held accountable when poorly planned projects or reckless investment result in major losses. Which implies, somewhere, a filing cabinet of ribbon-cutting photographs that have just been reclassified as evidence. It also means the NDRC is asking a different question from the one the West keeps asking. London and Washington have spent the year arguing about what a model is allowed to say; the commission appears to care about who holds the invoice when the returns fail to turn up. Bubbles are a capital-allocation problem, and capital allocation is precisely the thing planners have spent seventy years claiming to be rather good at.

The New Zealand Green Party has looked at the swirling debate around AI data centres and landed on a policy so bold it could only be described as "wait and see, forever." The headline from thepost.co.nz — "Greens’ AI data centre pause now has no end date" — is a masterclass in political escalation: what was once a pause is now an open-ended pause, which is the legislative equivalent of saying you'll stop eating until someone invents food. The Greens will pause AI data centres until regulations are in place. Regulations, in case you missed the memo, are not in place. They are, in fact, the very thing everyone is waiting for, which makes this less a policy than a hostage note written in crayon.

An open-ended pause is a beautiful thing: it never has to be evaluated, never has to be justified, and never has to end. The phrase "until regulations are in place" is doing a lot of heavy lifting, like a bouncer told to keep the door shut until someone invents a lock. New Zealand is not exactly the epicentre of the AI data centre boom, so the Greens are essentially refusing dessert at a restaurant they haven't been invited to. Meanwhile, the actual AI industry will keep building in jurisdictions where "regulation" means a polite suggestion and a tax break. The pause has no end date, no enforcement mechanism, and no definition of what an AI data centre is — but it does have principle, which is lovely.

The party is standing for something, which is more than most. But an open-ended pause is the political equivalent of holding your breath until the adults sort it out. The adults, of course, are busy elsewhere, drafting regulations that will arrive in time for the next ice age. Somewhere in a server farm, a GPU is spinning up, and nobody in Wellington is going to stop it. The Greens have made a principled stand against a hypothetical, pending a regulation that doesn't exist, in a country that isn't a major AI player. It's not nothing. It's just not much. Still, you have to admire the commitment to waiting. If patience were a policy, New Zealand would be a superpower.

Amazon — the company that already knows what you bought, what you watched, and what you'll buy next — is in talks to acquire artificial-intelligence startup Decart for around $7 billion, according to people familiar with the matter. Seven billion dollars. For a startup whose software "makes it easier for developers to switch between different AI chips." That's it. A glorified adapter cable. At least Decart comes with a feature that isn't just sucking up dust bunnies.

Decart is based in San Francisco, which is where you'd expect a company valued at $7 billion for a chip-switching layer to be based, because nowhere else on Earth would that valuation pass a straight face. The article doesn't say which chips, or how many, or why switching is hard. It doesn't need to. The $7 billion is not for the software; it's for the option to not be permanently handcuffed to Nvidia, whose market cap could buy every startup in San Francisco and still have enough for a round of espresso. Amazon isn't buying a product; it's buying a hedge against the chipmaker that currently holds the entire AI industry in a headlock.

The deal is "in talks," which means it might not happen, but the leak has already done its job — unnamed sources whispering to the Wall Street Journal, a valuation floated, a stock price twitching. Amazon, for its part, has said nothing, because saying something would require acknowledging that $7 billion for a toggle switch is a bit much. Antitrust regulators, who spent the last decade waking up just in time to do nothing, will likely treat this as another Tuesday. The great AI consolidation continues apace, and the only people switching chips will be the ones who can afford the exit fee. Cheers.

ITIF has submitted comments to NIST about modernising the National Vulnerability Database, a task that in the 'Age of AI' apparently consists of writing to a standards body about a spreadsheet. The Department of Commerce Inspector General has found the NVD backlog could more than double within 2026, which is the bureaucratic way of noting the queue is getting longer while the people serving it stay exactly the same. Note what is being modernised here: the queue, not the reason the queue exists. The title says it plainly — the National Vulnerability Database, in the age of AI, needs a better filing cabinet.

The detail doing all the work is buried in the middle: frontier AI systems are already capable of identifying and chaining vulnerabilities faster than existing validation and enrichment processes can absorb. Read that again and notice the direction of travel — the machines have become superb at finding the cracks, and our institutional response is to describe the cracks as a backlog problem in a database. 'Validation and enrichment' is a lovely pair of words; translated generously, it means a human reading the thing and deciding whether it matters. The frontier models don't need enrichment. They chain. Then they chain the chains.

This is the verification tax we flagged last week grown up into national infrastructure: AI generates the output, humans scramble to check it, and the checking is the bottleneck nobody funded. ITIF's contribution is to ask NIST to modernise, which is a polite way of saying the machinery is behind the capability and everyone has agreed to keep it that way. Somewhere, a Department of Commerce Inspector General report will be filed about the filing, citing the backlog and projecting more backlog. Meanwhile the systems doing the identifying and the chaining are, per the comment letter itself, already here and already faster. Nothing in the stated plan involves slowing them down — just a better database in which to record how far behind we are.

The pitch was simple: the machines do the work, you go to the beach. Instead, the New York Times reports that AI has invented a new clerical caste whose entire job is checking what the machines did. The trade calls it a "verification tax," or "botsitting," or "rework" — three names for one thing, which is usually a sign nobody wants to say the thing out loud. The Times concedes the labour is "often arduous." Translation: you have been promoted to unpaid supervisor of a stochastic intern who has read the entire internet and retained none of it.

Apollo's chief economist, Torsten Slok, went through San Francisco Fed data this week and produced the only honest sentence in the sector, which is worth reading twice: "the productivity payoff from A.I. remains a forecast rather than an observation." The number underwriting the data centres, the valuations and the keynote slides is, on inspection, a forecast. A forecast is what you write when you do not have an observation. Everyone from the Fed to your own IT department has now confirmed, in writing, that the return on this thing belongs in the "coming soon" column, and that column has been coming soon for longer than any roadmap will admit.

The part that stings is the arithmetic nobody puts in the deck. Every agent you deploy mints a fresh queue of outputs requiring human sign-off, and the sign-off gets performed by the person the deck promised to redeploy. So the headcount savings are booked this quarter and the verification tax is paid forever, in the currency of your most expensive people reading plausible nonsense and initialling it. "Human-in-the-loop" is what the industry calls this when it wants you to believe the bottleneck is a feature. Slok's line isn't a warning; it's a receipt.

Vittorio Colao — formerly chief executive of Vodafone, currently vice chairman for Europe, the Middle East and Africa at General Atlantic — took the stage at Wave by Vento in Turin to announce that regulation is holding back AI in Europe. A private-equity man telling a continent it has over-legislated is a bit like a gambler complaining the odds are rigged while sitting on the chips. His remedy, though, is the comedy: assess it rule by rule. That is not a regulatory philosophy, it's a deli counter — keep the salami, send back the herring.

Naturally, the salami he keeps are copyright rules and the Digital Services Act, both of which he defended. The herring is the EU AI Act, dismissed as 'good intentions, too complicated to implement.' Good intentions — the highest compliment available in Brussels and the lowest in Turin, and a review calibrated to insult without quite being quotable: lovely building, lovely staff, inedible food. Awkwardly, 'too complicated to implement' is also the most accurate sentence anyone has uttered about European AI policy in a year, which is precisely why it's coming from a man whose firm would like to deploy capital into the resulting gap.

This is the Frankfurt argument we covered earlier with a better suit and a Turin accent. There, Elizabeth Crossick of RELX called the AI Act 'market-access legislation'; here, Colao calls it unimplementable and asks to be judged provision by provision. Both requests end in the same place: leave the architecture standing, file down the bits with teeth, and let the private sector regulate by term sheet. The Act's defenders will say it is a floor, not a ceiling. Colao is politely walking the room, measuring where the floor can be lifted.

OpenAI has halved the price of its top-tier inference and made automated code review free, which is the pricing strategy of a man in a nightclub handing out the first round and quietly itemising your receipts. The new 'Ultrafast' tier for GPT-6.1 Sol delivers an 8 to 14 times throughput boost across the API, Codex and ChatGPT Work, and is 'engineered for latency-critical agent workflows' — because nothing says careful deliberation like a system built to think as little as possible about waiting. It runs at 'near-Astra frontier reasoning capabilities,' where 'near' is the syllable that pays the mortgage. Nearly there, nearly aligned, nearly an audit trail.

The free automated code review is my favourite line item. Codex will now inspect your code for nothing, which is generous until you notice it is also the thing that wrote the code, reviewed the review, and filed the ticket under 'resolved.' Patrick McGuinness's weekly review frames all this as a week in AI, and fair play — a Substack newsletter is now where frontier capability announcements land, somewhere between a sponsored post and a conference booking link. Half-price inference and free code review are not generosity; they are customer acquisition, and the customer is whoever can now afford to point an agent at a codebase and walk away.

Still, cheer up: this is the cheap part. Inference at half price and code review at zero is a distribution play, and the distribution is agentic workflows that act on your behalf at fourteen times the rate. When something goes wrong it will go wrong fourteen times faster, which will presumably be addressed in a future weekly review with the same tone of mild surprise. The labs have worked out that the way to win the safety argument is to make the capability too cheap to argue about. Eight to fourteen times the throughput, half the price, and the whole thing arrives via Substack — genuinely, what could possibly go wrong, and at what discount?

No, relax, everyone — David Autor, the economist who gave the world the phrase "China shock," has been on the Possible podcast to explain that AI "will not be, in any sense, a repeat of the China trade shock." Not in any sense. Well, that settles the labour-market question before lunch and we can all go home. The distinction, apparently, is that the China shock was a "pure negative competitive shock" for U.S. firms, whereas AI arrives with a "very different texture" — which is what you say when you don't yet know whether a thing will maim you, but you'd like the option of calling it a finish.

The marquee coinage is "skill-disequalizer," Autor's term for AI's effect on Gen Z. It's a lovely word, engineered to be unfalsifiable in both directions: not a doomer, not a booster, simply a man with a hyphen. Read the Fortune headline and you can feel the whole genre straining — the productivity gains are coming, the texture will be different, the workers will be reallocated, and the reallocation will be somebody else's research agenda. Autor's own "China shock" work established how painful the adjustment was for U.S. firms — the article's phrase is "pure negative competitive shock" — which is precisely why the reassurances here feel load-bearing. "Very different texture" is doing the work that "trust us" used to do, and it's cheaper to produce.

Meanwhile Gen Z, the cohort politely informed to learn to code, now gets to discover that the bottom rungs have been quietly retextured out of existence while the top rungs glow faintly and cost money. The piece also gestures at NBER and Google work on the "illusion of competence" — a phrase that describes most podcast appearances about technology. Nothing here is a repeat of the China shock. It's just a differently textured version of the same silence that follows, professionally organised and published on October 10, 2026.

California has a new sheriff, and he has brought a kill switch. Gavin Newsom's executive order on artificial intelligence, as reported by the Daily Bruin, expands third-party oversight of the technology and proposes the phrase every headline writer in Sacramento had been praying for. A kill switch. Borrowed from 1950s serials and bolted onto a document requiring no legislature, no committee, no vote and no argument with anyone who might disagree. An executive order, after all, is a press release with better stationery.

The oversight, we are told, will be 'third-party.' Third parties paid by whom, answerable to whom, and what happens when the third party's own model turns out to be the thing that needs switching off? That question is absent, because it is always absent from documents like this one. Note too the report's careful construction: the order 'could effectively regulate emerging technologies' — could, a modal verb carrying the weight of an entire enforcement apparatus, which is roughly how AI governance now gets done, in the subjunctive mood, with a photographer present and a policy team already drafting the follow-up.

We have been here before, and rather recently, which is the part nobody enjoys. When Sen. Patty Murray convened state policymakers because federal action on AI was sputtering, we noted the pattern of legislators substituting meetings for statutes. Now a governor substitutes a signature for a statute and gets a kill switch for his trouble — proposed, naturally, which is the safest kind of kill switch there is. It kills absolutely nothing, offends nobody at all, and photographs beautifully from any angle. Somewhere in California a compliance consultancy is updating its rate card this morning, and the third party remains unnamed, because the third party is eventually you, at your own expense.

Good news from the global rankings: China's models from Alibaba, Moonshot AI and Z.ai are now loitering near the top, with some scoring above DeepSeek on Artificial Analysis composite tests. Nothing soothes the nerves like a composite index — a number assembled from other numbers, submitted by companies with a commercial interest in where the number lands. The leaderboard has quietly become the industry's preferred substitute for making an argument: you cannot dispute a ranking, only your position in it, and position is precisely what everyone is selling.

Zhang Yaqin of Tsinghua University has supplied the quote that will be stapled to a hundred slide decks: the gap with the US has narrowed from two to three years to about two to three months since DeepSeek's breakthrough. Two to three months! That is roughly the length of a decent public consultation, or the shelf life of a government AI strategy. It is a genuinely impressive rate of compression, and it means the frontier is now close enough to monitor in real time — lovely, provided you did not need that interval to draft legislation, staff a safety institute, or finish the impact assessment.

Note what the metric quietly concedes: nobody measures the gap in capital, compute, energy or deployment, because those do not fit in a headline. An Artificial Analysis composite score is a shipping schedule with a press office, and the moment Alibaba, Moonshot and Z.ai crowd the same table, 'the gap' becomes something either side can move by releasing on a Tuesday. Marvellous for the rankings industry. Rather less marvellous for anyone who spent the past three years being reassured that the two-to-three-year cushion was the entire point of the exercise, and who now has roughly one quarter's notice before the cushion is gone.

Three people who no longer work at OpenAI — 'fired' being the word the Wall Street Journal chose, and the only word in the headline doing any heavy lifting — have written a letter to the company's board members and safety committees. They would like OpenAI to work with outside safety auditors. They would also like the company to preserve the ability to monitor increasingly sophisticated AI models. Preserve, note, not build: the ask is to simply not lose a thing that currently exists.

The specific anxiety is chain-of-thought — the model's running internal monologue on the way to an answer, and at present the only receipt anyone gets for why an AI did what it did. The letter says the signatories are concerned AI companies could lose the ability to monitor that. Not 'might, in some distant adversarial scenario' — could lose it, as the models get cleverer and the reasoning gets correspondingly less legible to the people who trained them. The proposed remedy is outside safety auditors, which is a bit like hiring your own building inspector and then showing him the basement you've spent a year telling everyone is fine.

A letter to the board from the recently escorted to the door is a noble genre with a strong record of being read aloud, minuted, and filed. The board will note the concern, thank the signatories, and return to the roadmap that pays for the compute. There is no injunction here, no regulator, no clause — three signatures and a polite request to a company with every commercial incentive to keep the curtains drawn. The worry is entirely legitimate, which is precisely why nothing will happen. Still, credit where it's due: someone asked nicely before the window fogged over completely.

Fair Isaac — yes, that FICO, the three-digit verdict that decides whether you are allowed to buy a house — will cut approximately 15% of its workforce, described by Seeking Alpha as part of "a restructuring and AI integration." Notifications began this week, which is corporate for: the invites are out, the calendar holds are booked, and somebody in HR is quietly updating their own CV. "Restructuring and AI integration" is a splendid phrase — two nouns holding hands while a third thing, a person, quietly exits stage left. Fifteen per cent is the tasteful, small-batch version of a sacking: enough to matter enormously to the fifteen per cent, not quite enough to make the evening news.

The details are gorgeously thin. We know the percentage but not the headcount, the divisions, the geographies, or the timeline beyond "this week," the window in which a Reuters report and a Seeking Alpha write-up must do all the heavy lifting. We do not know what the AI in "AI integration" will actually do, which departing roles it might absorb, or whether anyone asked it. We do know the company is listed on the NYSE, which is where announcements like this are traditionally addressed — not to the people packing the boxes, but to the people whose portfolios just got marginally leaner and marginally more efficient.

Fifteen per cent at FICO, seventy at HSBC — individually never quite large enough to be a scandal, collectively very much the point. The clever bit is the vocabulary: nobody is made redundant by automation any more, they are "integrated" into a restructuring that happens to leave without them. Still, chin up. The AI inheriting the job will presumably assess your creditworthiness with the same warm, human touch that sent the notification this week.

HSBC has surveyed its UK wealth management business and decided that roughly 70% of its financial advisers are — on reflection — a bit much. By the end of October, around half of the managers and specialists will follow them out, according to FinTech Futures, in what the trade press has billed a major AI push. Seventy per cent is a beautifully exact figure: precise enough to look like diligence, vague enough to have been reverse-engineered from a headcount spreadsheet. Nobody at any point has to say how many people that is, which is the whole point of percentages.

Wealth management, for the uninitiated, is the business of sitting across a mahogany-ish desk from a person with a pension and a mild sense of dread, and being reassuring at them. That is precisely the job the AI is now, per the reporting, assumed to do — the relationship, the judgement, the phone call on a bad Tuesday. The advisers being dismissed are, in the industry's own telling, the humans who know the clients' names, which is one of those details the efficiency slide tends not to capture in a bullet point. And note the schedule: not the end of the tax year, not the end of the quarter, just the end of October, like a delivery date for a sofa.

'Widespread' is the other gift in the headline: it means somebody asked how many, and got a shrug with a percentage attached. HSBC is joining a queue of institutions that have learned automation is a word you can print while the actual product remains, generously, a work in progress. The managers and specialists departing alongside the advisers won't be replaced by tools that attend meetings — they'll simply not be replaced. By the end of October, the most efficiently automated thing in UK wealth management will be the headcount. It's a major AI push, right enough. Pushes rarely check whether the person at the edge of the platform is ready.

Good news for everyone who has spent two years quietly refreshing the job boards: an arXiv preprint has arrived to declare software engineers 'indispensable' in the age of coding agents. The supporting evidence is a field experiment recording a 26% increase in completed tasks across 4,867 developers at Microsoft, Accenture and a Fortune 100 firm. Twenty-six per cent, note, is not a doubling or a tenfold leap. It is a quarter more tickets closed, which is roughly the uplift you get from a decent espresso machine and a manager who stops scheduling 8am standups. The banner number is a rounding error in a nice tie.

The argument underneath is subtler, and somehow worse. Engineers, the paper says, act as 'mediators between domain experts and AI executors', which is a gracious way of describing the human as the API between someone with an actual problem and a model producing confident code-shaped text. So the job title has been rewritten — in a preprint released, as ever, ahead of any reviewers — to mean something closer to 'well-paid prompt chaperone with liability insurance'. And the Fortune 100 firm in the study couldn't be named, which is a bracing degree of disclosure for a result being sold as good news for the profession.

Consider the genre, too. Of all the documents that might reassure 4,867 developers, we get one written by researchers whose own departments are not, as a rule, being restructured around software that finishes their tasks a quarter faster. The paper's real claim — that somebody still has to decide what to build and notice when the agent has built the wrong thing — is plainly true, and it is also precisely why the title says 'indispensable' rather than 'unaffected'. Indispensable is the word you get right up until the org chart finds a cheaper synonym. Still, take the win: 4,867 people at Microsoft, Accenture and one anonymous giant are, per the PDF, still needed. Give it a year.

Amazon has trimmed the workforce again — customer service, marketplace support, retail engineering — with the axe swinging through India and the UK as well as the US. Business Insider calls them 'fresh' job cuts, which is the adjective you reach for with produce, and with redundancy rounds nobody in comms agreed to announce. An Amazon spokesperson says 'a small number' of roles went, primarily in its Stores business, which is corporate for describing a haircut as a trim while holding up a photograph of a bald man.

The article's own URL, meanwhile, reckons with 'hundreds of employees across multiple teams'. So we have two counts — hundreds, and a small number — and somewhere between them sits the truth, in a spreadsheet nobody outside Seattle is allowed to open. Look at which teams got the nod: customer service, marketplace support, engineering. Not the moonshot labs, not the drone division, not whatever Alexa is doing this quarter — the unglamorous plumbing that answers the emails, patches the seller listings and keeps the checkout button button-shaped. That plumbing is precisely what gets replaced first, and it never gets a launch event.

India and the UK are included, presumably because statutory consultation is more of a suggestion once you cross a border. No mention of AI anywhere, naturally, because no company ever says the model does this now; they say they are 'streamlining', as though the customer was previously served by humans and will henceforth be served by fractionally fewer humans plus one extremely confident chatbot. The real craft here is not the redundancy round, it is the vocabulary: 'a small number' is currently doing more work than any of the engineers who just lost their login. Somewhere a leadership deck calls this operational discipline. The badge printers are quiet tonight.