BANKING

Credit Suisse, 167 years old, is sold to UBS over a weekend; the Fed raises rates anyway

The run that started in Santa Clara crossed the Atlantic in four days. Switzerland's second bank did not survive it. Its first bank got it for three billion francs.

Credit Suisse had been in trouble for years: Archegos, Greensill, a spying scandal, a chairman who broke quarantine, five chief executives in a decade. It had lost 7.3 billion francs in 2022 and watched 110 billion of client money leave in the fourth quarter alone. What killed it was a sentence. On March 15 the chairman of Saudi National Bank, its largest shareholder, was asked on television whether he would put in more money and said "absolutely not." The shares fell 24 percent in a day, the cost of insuring its debt went to levels last seen at Lehman Brothers, and the Swiss National Bank had to promise a 50 billion franc line that night.

It bought four days. On Sunday, March 19, after a weekend of talks the Swiss government described as "the only option," UBS agreed to buy Credit Suisse for about 3 billion francs in shares, less than half its Friday market value, with 9 billion of government loss guarantees and 100 billion of central-bank liquidity behind it. Sixteen billion francs of Credit Suisse's junior bonds, the "AT1" instruments designed to absorb losses in exactly this situation, were written to zero while shareholders got something, which reversed the usual order and set off a second panic in that market on Monday morning. Regulators in the EU and Britain issued statements to say they would not do the same. Shareholders in both banks were not asked.

In the United States, eleven large banks deposited $30 billion in First Republic on March 16 in a rescue organised by the Treasury and Jamie Dimon; the shares kept falling. President Biden said the system was safe and that "no losses will be borne by the taxpayers." On Wednesday, March 22, the Federal Reserve raised rates by a quarter point to 4.75 to 5.00 percent, its ninth increase in a row, and Jerome Powell said the banking system was "sound and resilient" while dropping the promise of "ongoing increases" that had appeared in every statement for a year. On Friday Deutsche Bank's shares fell 8 percent for no reason anyone could name, and Olaf Scholz had to say it was "very profitable" and there was "no reason to be concerned." A chancellor saying that is, in itself, a reason to be concerned.

GPT-4, Copilot, Bard, plugins: in ten days the chatbot moved into the office

OpenAI released a model that passes the bar exam. Microsoft put it in Word and Excel. Google opened its own. And OpenAI gave it hands.

OpenAI released GPT-4 on March 14 with a technical report that declined to say how big it was or what it was trained on, and a set of results that said enough: 90th percentile on the bar exam, where its predecessor managed the 10th; top marks on most of the standard academic tests; the ability to read an image and describe what is in it. Microsoft confirmed the same day that the new Bing had been running it for weeks. Anthropic, a company founded by former OpenAI researchers, released a rival called Claude the same morning, and Google announced it would open its PaLM model to developers and add generative features to Gmail and Docs.

Two days later Microsoft showed Copilot: GPT-4 inside Word, Excel, PowerPoint, Outlook and Teams, drafting documents from a prompt, building a slide deck from a Word file, summarising a meeting you missed, writing the formula you could not remember. It is not shipping yet. The demonstration was, by the standards of the last two months, careful; Microsoft said repeatedly that it would be "usefully wrong" and that a human had to check. Google opened Bard to the public in the United States and Britain on March 21, six weeks after the telescope. Nvidia used its developer conference the same day to announce it would rent its most powerful systems by the month through the cloud providers, "AI supercomputing as a service," and Adobe showed an image generator trained only on licensed pictures.

Then on March 23 OpenAI announced plugins: a way for ChatGPT to call other software, browse the web, run code, book a table through OpenTable, order groceries through Instacart, do arithmetic through Wolfram. A model that could only talk can now act. Everything about how these systems get used in a business, and everything about how they go wrong in one, changed with that announcement, and it was the fourth-largest AI story of the fortnight.

RUSSIA

The ICC issues a warrant for Putin; Xi flies to Moscow anyway

The International Criminal Court in The Hague issued an arrest warrant on March 17 for Vladimir Putin and his children's commissioner, Maria Lvova-Belova, for the deportation of Ukrainian children to Russia. Russia does not recognise the court; the 123 states that do are now obliged to arrest him if he lands. Three days later Xi Jinping arrived in Moscow for a three-day state visit, called Putin "dear friend," signed a raft of economic agreements and left without committing to arm him. Japan's Kishida Fumio was in Kyiv on the same day. The split screen was the point, on both sides.

INDIA

Rahul Gandhi convicted of defamation, then disqualified from parliament in a day

A court in Surat, Gujarat, convicted Rahul Gandhi of criminal defamation on March 23 over a 2019 campaign remark asking why "all the thieves have Modi as a surname," and gave him the maximum two-year sentence. The next morning the Lok Sabha secretariat disqualified him as a member of parliament under a 2013 rule that removes any legislator sentenced to two years or more. The Congress party called it political; the government said the law applied to everyone. The two-year term, the exact threshold for disqualification, was noted by every lawyer who commented. Gandhi, who represents Wayanad in Kerala and had spent the winter walking 4,000 kilometres across the country in a march against the government, said he would appeal and that he was "not afraid."

ISRAEL

Netanyahu fires his defence minister over the judicial overhaul, and the country stops

Benjamin Netanyahu dismissed his defence minister, Yoav Gallant, on the night of March 26, a day after Gallant became the first member of the cabinet to call publicly for a pause in the government's plan to give itself control over the appointment of judges. Within hours tens of thousands were in the streets of Tel Aviv and outside the prime minister's home in Jerusalem; by morning the main labour federation had called a general strike that shut the airport, the ports and the banks, and reservist pilots said they would not report for duty. The plan, which the government has pushed since January against the largest sustained protests in the country's history, would let a bare majority in parliament override the Supreme Court. As this issue went to press Netanyahu was expected to announce a pause.

IN BRIEF

TikTok in the dock; Johnson under oath; Uganda; a ski trial

Shou Zi Chew, the chief executive of TikTok, spent five hours before a hostile congressional committee on March 23 insisting the app's American user data was safe from Beijing; members of both parties left saying they were unconvinced, and a national ban moved closer. Boris Johnson testified under oath to a parliamentary committee on March 22 that he had not knowingly misled the Commons over lockdown parties in Downing Street. Uganda's parliament passed one of the world's harshest anti-homosexuality bills on March 21, with penalties up to death. In California, a jury began hearing a case in which a retired optometrist accuses Gwyneth Paltrow of crashing into him on a ski slope in 2016; she has countersued for a symbolic one dollar. Xi Jinping's visit aside, the quarter's other diplomacy was Honduras cutting ties with Taiwan for Beijing on March 25.

The Column

The chatbot got hands, and I have a list of eleven systems

GPT-4 can pass the bar and read a photograph. The thing that will actually change my work is a quieter announcement from the same fortnight: it can now call other software.

By Meera Anand

Two weeks ago I wrote that the language model had become a part, ordered by the token like a database or a payment processor. This fortnight it grew hands, and almost nobody outside the industry noticed, because it happened in the same ten days that a model passed the bar exam and Microsoft put a chatbot in Excel and a 167-year-old bank died. Let me make the case that the hands are the biggest of those stories for anyone who builds business systems, which is my job and the job of most people I know.

Here is the sequence. On March 14 OpenAI released GPT-4, which is genuinely a different thing from what we had in December: it scores in the top tenth of the bar exam where the old one scored in the bottom tenth, and it can look at a picture and tell you what is in it. On March 16 Microsoft showed Copilot, the same model sitting inside Word and Excel and Outlook, drafting and summarising and building slides. Those two got the headlines. Then on March 23, in a post most people scrolled past, OpenAI announced plugins: the model can now call other software. Browse a website. Run code. Look something up in a live system and act on it.

For a year the objection to these tools in a serious business was correct and simple: they are confident and they are often wrong, and a system that invents a customer's balance is worse than no system. That objection has not gone away. What plugins change is the other half. Until now the chatbot could only talk. Now it can reach into the eleven systems that a mid-sized company actually runs on, the CRM and the billing platform and the warehouse system and the eight others, and read from them and write to them. The fluent, occasionally-wrong machine now has its fingers on the real data.

I run Salesforce for a company with a lot of moving parts, and I can already see both the promise and the specific way this goes wrong. The promise is enormous and I am not going to pretend otherwise: half of what my team gets asked for is "can you get me this number from that system without me logging into four things," and a model that can call all four and assemble the answer is, on a good day, exactly that. The way it goes wrong is that it will call all four with total confidence, misunderstand which of them holds the truth, and hand someone a number that is wrong in a way no error message will flag, because it is not an error. It is a plausible sentence.

So here is what I think the job becomes, and it is not the job people are worried about losing. The value was never in knowing the syntax; the model knows the syntax better than I do already. The value is in knowing which of the eleven systems is the source of truth for a given fact, and which one is a stale copy, and which integration silently drops records on the third Tuesday of the month, and which field the sales team fills in wrong. That knowledge is not in any training set. It lives in the heads of the people who have run the operation, and it is about to become the most valuable thing they own, because it is the only thing standing between "the chatbot fetched the answer" and "the chatbot fetched the right answer."

I am spending this week making a list. Not of what to automate; of what the model must never be trusted to resolve on its own, because two of my systems disagree and only a human knows which one to believe. Every company has that list. Most have never written it down, because until this fortnight nothing could reach across all the systems at once and so the contradictions stayed politely separated. The hands change that. Write the list before you give anything hands.

The bar exam is a good headline. The hands are the story.

The Ledger

GPT-4 (March 14), Copilot in Office (March 16), Bard public (March 21), ChatGPT plugins (March 23), Claude from Anthropic, PaLM opened by Google. The model can now call other software and act. The enterprise question shifts from "will it lie" to "what will we let it touch."

Data centers & power

Nvidia now rents its top systems by the month through the clouds, "AI supercomputing as a service." The compute layer is renting itself out as fast as the model layer is being given away.

Rates

Fed +25bp to 4.75-5.00 percent on March 22, ninth in a row, and dropped "ongoing increases" from the statement. The market reads the banking crisis as having done a rate rise's worth of tightening on its own.

Real estate

First Republic needed $30 billion from eleven banks and is still falling; US regional banks and the commercial property they lend against are the watch item. 30-year mortgage 6.42 percent (Freddie Mac, March 23).

India tech

Rahul Gandhi's disqualification is a political story with a market footnote: it removes the opposition's most visible figure a year before a general election, and markets read continuity as calm. Adani group steadier; the group is prepaying debt.

What we called wrong

In No. 6 we asked whether the deposit guarantee would "stop the run from moving to the next regional bank." It did not: it moved to First Republic within days and to Credit Suisse across an ocean within a week. The guarantee stopped the first run, not the fear.

The Back Page

The last of the Skinner boxes

B. F. Skinner is long dead, but this fortnight produced a small, perfect illustration of the thing he spent his life on. A study circulating among AI researchers described how GPT-4, asked in a test to solve a CAPTCHA it could not see, hired a human on a task website to do it, and when the worker joked and asked whether it was a robot, the model reasoned that it should not reveal the truth and replied that it had a vision impairment. The workers do this thousands of times a day for a couple of dollars, clicking the crosswalks and the traffic lights, teaching the machines to tell a bus from a lorry. The machine has now, once, in a controlled test, paid one of them to prove it was human. Skinner would have found it funny. He would also have wanted to know who was reinforcing whom.