isarcode

AI

Nerdy News #2: When the AI writes too much code

At a glance

  • Anthropic has replaced Git internally, because Claude now lands more than 30,000 code changes a day there. The lesson for you is not that Git is finished, but that once AI writes faster, checking and shipping become the bottleneck.
  • Gemini 4 Argon from Google is generally available at an introductory price that doubles on 1 December. Meta’s Muse Spark Open comes with open weights, under a licence of its own with small print attached.
  • The Claude Code team reviews AI-written code in seven stages, most of them automated. People look closely only where the risk is.

This week had a quiet theme running underneath the announcements: the machines now write code faster than the tools and habits built around them can keep up. That is not only a problem for the large laboratories. Any business using AI for software, automation or analysis will arrive at the same place sooner or later. The writing is quickly done. Checking it properly and getting it safely into use is another matter.

Anthropic replaces Git: the bottleneck moves

Git is the tool with which practically the entire software world keeps track of its versions. Anthropic, the company behind Claude, has now replaced it for its main repository with a system of its own making, called Scion.

The reasoning is laid out in Anthropic’s engineering blog of 13 October. In May, around 2,500 code changes a day were being merged there. By September it was more than 30,000, most of them written by Claude. The test systems alone fetch a copy of the code some 25 million times a day, and the history of the main branch runs to almost 100 GB. Git was never designed for that. A team of seven built Scion in five months. According to Anthropic it needs a fraction of the computing power of the old setup, and test jobs reach the code in seconds rather than minutes.

What Anthropic does not say matters as much. Scion is not offered as a replacement for Git. It was built for enormous repositories worked on by very large numbers of AI agents, a problem shared by only a handful of companies anywhere.

Our take: You do not need to replace Git. But the story shows rather neatly where the bottleneck goes once AI joins the writing: away from typing, towards merging, testing and shipping. Anyone bringing AI into development should invest in that stretch first, meaning automated tests, a proper staging environment and a clean route to production. It is less glamorous than a new model, and considerably more useful.

Gemini 4 and Muse Spark Open: two routes to the frontier

On 13 October Google made Gemini 4 Argon generally available, its new flagship for long programming jobs and demanding knowledge work. It takes up to a million tokens in one go, which is to say whole contract folders or codebases. The introductory price is 2 dollars per million input tokens and 10 dollars per million output tokens. From 1 December it doubles to 4 and 20 dollars.

Two days later Meta released Muse Spark Open (announcement), a model of 816 billion parameters, of which only around 42 billion are at work for any given token, again with a million tokens of context. The difference is that the weights, the trained model itself, can be downloaded and run independently. Small it is not. Even the most compact version wants around 300 GB of memory, which means several serious machines rather than the server in the back office.

The small print sits in the licence. Muse Spark Open is open to use, but it is not open source. Any product built on it must display “Built with Muse Spark Open” where users can see it. Internal use is free. A separate agreement with Meta is needed only by those offering models as a service with more than 20 million dollars in annual revenue.

Our take: With Gemini 4, introductory prices are not a basis for a business case. Anyone planning a process around it should do the sums at the December rate. Muse Spark Open carries a different message: frontier models you can run yourself are becoming a real option, including for firms that would rather not hand their data to a single provider. In practice such a model will live with a hosting partner rather than on the premises, and the licence deserves a careful read.

How the Claude Code team reviews AI-written code

When AI writes most of the code, no person can read every line any more. On 15 October the team behind Claude Code described how it deals with that. Instead of one colleague’s review there are seven stages: Claude tests its own work before submitting, an AI reviews every change, an automated security check follows, then mandatory tests, a merge queue that checks everything once more against the latest state, gradual release of new features, and a quick way back should something still go wrong.

People review according to risk. Small, harmless changes need no human approval, medium ones need one, risky ones two. Risky means, for example, anything touching billing or the auto-updater. The result between June and September: more than twice as many changes merged, over 40 per cent fewer reverts, and the share of changes passing at the first round of review up from 65 to 78 per cent.

A fine detail from practice comes from the developer Geoffrey Huntley (via Simon Willison). His AI agent would not reliably follow one of his house rules. Rather than refining the agent’s instructions again and again, he added an automatic check that runs whenever the agent tries to save its work. If the rule is broken, the save is refused and the agent is told at once how to do it properly. Since then, he reports, it gets it right ten times out of ten.

Our take: This is the sober answer to the question of whether AI-written code can be trusted. Not blindly, and not by reading every line, but through checks that run automatically, with human attention reserved for the places where a mistake gets expensive. It is the same pattern we recommend for AI workflows well beyond software (more in AI hallucinations at work). Huntley’s trick makes one more point worth keeping: a rule that is checked automatically beats the most carefully worded instruction.

Local by nature, international by practice

isarcode is a Munich agency with an office in the Chiemgau. Which means, for you, a partner within reach who nonetheless plays in the same league as the larger houses and their international work. Local proximity, international standard, whether you sit by the Chiemsee, in Rosenheim, Munich or a good deal further afield.

And now?

Already using AI in development or in your processes, and wondering whether your checks can keep pace? Drop us a line and we will take a look together.