Vol. I, No. 16
Covering 17 July - 30 July 2023
Monday, July 31, 2023
Late edition · A Relentless publication
All the fortnight that mattered, in technology and in the world, read next to what we were building at the time.
ISRAEL

The Knesset passes the first piece of the judicial overhaul, and half the country calls it the end of something

After 29 weeks of the largest protests in the nation's history, the government stripped the Supreme Court of one of its powers. The other half of the country is not going home.

Israel's parliament passed the first law of Benjamin Netanyahu's judicial overhaul on July 24, voting 64 to 0 after the opposition walked out, to strip the Supreme Court of the power to strike down government decisions it finds "unreasonable." The reasonableness standard is a technical-sounding thing; the fight over it is not. For 29 straight weeks, since the plan was announced in January, hundreds of thousands of Israelis have filled the streets every week in the largest sustained protest movement the country has ever seen, arguing that a government with no written constitution and a single legislative chamber is dismantling the only real check on its power. Military reservists, including hundreds of air-force pilots, said they would stop reporting for duty. Business leaders, former security chiefs and the country's own president warned of a slide toward autocracy or civil rupture.

The government, the most right-wing and religious in the country's history, says it is restoring democracy by returning power from unelected judges to elected politicians. Its opponents say that in a country with no other brake, the court is the only thing standing between a temporary majority and permanent power. Both understand that this first law is the opening move, not the whole game, with further bills to change how judges are chosen still to come. The Supreme Court will now hear challenges to a law that limits the Supreme Court, a constitutional knot with no clean way out. A society that spent 75 years arguing about its neighbours is now tearing at the question of what it is to itself.

WEST AFRICA

Soldiers seize power in Niger, and the last Western foothold in the Sahel wobbles

Soldiers of the presidential guard detained Niger's democratically elected president, Mohamed Bazoum, on July 26 and announced a military takeover, the seventh coup in West and Central Africa in three years and the most damaging to the West. Niger was the last reliable Western partner in the Sahel, host to French and American troops and drones fighting a jihadist insurgency, and a country the United States and Europe had held up as a democratic exception in a region sliding toward the juntas and the Russian mercenaries of Wagner. Crowds in the capital waved Russian flags. The regional bloc ECOWAS threatened force to reverse the coup; the junta closed the borders and dared it. Another piece of the map tilted away from the West, toward Moscow, without a shot fired by anyone who meant to lose it.

TECHNOLOGY

Meta gives away a powerful AI model, and the White House gets promises

Meta released Llama 2 on July 18, a large language model roughly competitive with the previous generation of the best commercial systems, and did something the others have refused to do: it made the model free to download and use commercially, in partnership with Microsoft, for almost anyone. It is the strongest open model yet released by a major company, and it hands every business the option of running capable AI on its own infrastructure, under its own control, instead of renting it by the token from OpenAI. Three days later, on July 21, the White House announced that seven leading AI companies, including OpenAI, Google, Microsoft, Meta and Anthropic, had made voluntary commitments to safety testing, watermarking and independent scrutiny. Voluntary is the operative word; the commitments have no force but the companies' word, and arrive as the substitute for legislation Congress has not written.

IN BRIEF

Barbenheimer; X kills the bird; Trump's cases multiply; the fever

Two films opened on the same July weekend and became a cultural event: Barbie, a candy-coloured comedy, and Oppenheimer, a three-hour drama about the atomic bomb, drew crowds who saw both back to back and gave the cinema its best weekend in years, more than a billion dollars between them within weeks. Elon Musk renamed Twitter "X" on July 23, tearing the bird logo off a 17-year-old brand worth billions in recognition and replacing it with a letter. Trump's legal jeopardy deepened: new charges landed in the documents case, and his lawyers were told he was a target of the federal January 6 investigation, with an indictment expected imminently. And July was confirmed as the hottest month in the history of measurement, as wildfires forced tens of thousands to flee the Greek islands.

The Column

The model you can hold in your hand

Meta gave away a capable AI model this fortnight, free to download, run and own. The gift quietly reframes the whole question of who controls this technology, and it is not the question everyone is arguing about.

Amid the judicial crisis in Israel and the coup in Niger, a technology story this fortnight will matter longer than either: Meta released Llama 2, a genuinely capable AI model, and made it free to download and use commercially, for almost anyone. That sounds like a narrow industry event. It is not. It is the clearest move yet in a fork that will shape the next decade of this technology, the fork between AI you rent and AI you own, and I want to explain why the second option changes everything even though the first is where all the attention is.

Here is the fork. On one path, the powerful AI models live behind the doors of a few companies, OpenAI, Google, and you reach them by renting access, by the token, sending your requests to their servers and getting answers back, never holding the thing itself. On the other path, the model is a file you can download and run on your own machines, under your own control, seeing exactly what it does, keeping your data on your own premises, modifying it to your needs, dependent on no one. Until this fortnight the second path led only to weaker models. Llama 2 is the first time a major company has handed out something powerful enough to matter, free, to be held in your own hand. That is a different kind of event from another chatbot launch.

Why does it matter which path wins? Because ownership and dependence are the whole game. A business that runs its AI on a model it controls keeps its data, sets its own rules, cannot have the service changed or priced up or switched off under it, and can inspect what the thing actually does. A business that rents its AI from a giant has convenience and the frontier's cutting edge, and also a dependency it does not control, on a company that owns its access, holds its data, and can change the terms whenever it likes. This is the oldest question in enterprise technology, the one this paper keeps finding under everything, dressed in new clothes: do you control your tools, or do they control you. Meta, for its own competitive reasons, just made "control your own" a real option at the frontier for the first time, and in doing so handed every business, and every country, a choice that the closed labs would much rather they did not have. The argument everyone is having is whether open models are dangerous. The question that will actually matter is whether, when the dust settles, the most important technology of the age is something you own or something you are permitted to rent.

Field Notes
A Relentless build, told plainly

Building a tool to hide the real data, so the people testing the system never see it

The front page has the White House asking AI companies to promise they will handle data responsibly. Here is a smaller, concrete version of that promise, one I actually built: a way to make sure that the copies of a company's system are full of realistic data that belongs to no real person. No client is named.

Here is a problem most people outside enterprise software have never had to think about, and everyone inside it loses sleep over. A company runs its real system, full of real customers: their names, their addresses, their financial details, their private information. To improve that system, developers and testers need copies of it to work in, because you do not test changes on live customer data any more than a surgeon practises on a patient. So the company makes copies, sandboxes, and by default those copies contain the real data, faithfully reproduced. Which means every developer, every tester, every contractor with access to a sandbox is now looking at real people's private information, in an environment with a fraction of the security of the real one, for no reason anyone would defend if you asked them out loud.

I built a tool to solve this, and it is on the Salesforce marketplace. What it does is simple to describe and fiddly to do right: it goes through a copy of a system and masks or clears the sensitive information, so that the copy keeps the shape and realism a developer needs to work, real-looking names, real-looking numbers, the right formats and relationships, while belonging to no actual person. The tester sees a database that behaves exactly like the real one and exposes nobody. The private data stays in the one place it is properly protected, and the ten copies floating around the organisation stop being ten copies of everyone's secrets.

The reason I want this in the same issue as the White House AI commitments is that it is the concrete, unglamorous version of the thing those commitments gesture at. Everyone agrees, in the abstract, that companies should handle personal data responsibly. Almost nobody, in the specific, has looked at how many copies of their customers' private information are sitting in test environments that were never built to hold it. Responsible data handling is not a promise you make at a podium. It is a hundred small, boring engineering decisions, and one of them is this: before you copy the system so your developers can work, strip out the real people. The regulation everyone is waiting for will, eventually, require some version of this. The companies that were doing it already, because an engineer built the tool and made it routine, will barely notice. The ones treating data protection as a press release will discover how many copies of everyone they had been keeping, usually on the day one of those copies leaks.

The Ledger
AI
Meta's Llama 2 makes a genuinely capable model free to run and own, the strongest push yet for open AI; the White House extracts voluntary safety promises with no teeth. Control of the technology keeps splitting between the closed leaders and the open field.
Data centers & power
July was the hottest month ever recorded, straining the same grids being asked to power the AI build-out. The energy question under the technology story grows harder to ignore each month.
Rates
The Fed raised a quarter point on July 26, to 5.25-5.50 percent, the highest in 22 years, and left the door open to more or none. Most now believe it is the last rise of the cycle, but Powell will not say so.
Real estate
Mortgage rate 6.81 percent (July 27). The freeze holds: prices sticky because nobody sells, sales low because nobody can afford to buy at 7 percent.
India tech
Llama 2 matters especially for India's services industry: an open model a firm can run on its own infrastructure, for its own clients, without renting from a US lab, is the kind of thing Bengaluru can build a practice on. Data control becomes a selling point.
What we called wrong
Our running "Data centers & power" watch has repeatedly said no one is pricing the electricity. The record-heat summer suggests we should broaden it: it is not only whether the grid can power the data centres, but whether it can power the data centres and cool the cities in the same July. We will widen the line.
The Back Page

Barbenheimer

Nobody planned it. Two films could hardly have been more opposite: a plastic-pink comedy about a doll discovering the real world, and a sombre three-hour epic about the man who built the atomic bomb, released on the same Friday, July 21, by rival studios who surely each wished the other had picked a different weekend. Instead the internet decided they were a double feature, coined a portmanteau, and turned seeing both in one day, ideally the bomb first and the doll as a chaser, into a global event. People dressed up. They queued. Cinemas that had spent three years wondering whether the habit of going to the movies had died sold out both, and took more than two billion dollars between the pair. There is a lesson in it about counter-programming and about the stubborn human wish to be in a dark room with strangers watching something big, but mostly it was just a good time, arriving in a heavy year, and the culture grabbed it with both hands. Sometimes the mushroom cloud and the dream house open on the same day, and you go, and it is wonderful.