The Shape of Things

The Shape of Things
3 min read

AI systems are no longer limited to responding to individual requests: they can now be assigned a job, sometimes equivalent to several hours of human work, and deliver it in minutes. In an analysis published in his newsletter, Ethan Mollick describes a rapid progression in these capabilities, which, he argues, risks increasing instability in markets, employment, and politics.

Since the arrival of ChatGPT, working with AI primarily took the form of successive exchanges, which the author calls ‘co-intelligence.’ Since late 2025, agents like Claude Code, OpenAI’s Codex, or OpenClaw have allowed for entrusting an entire task to an AI. The author now speaks of an era of ‘managing’ AIs rather than working alongside them.

A Rapid and Steady Progression

To illustrate this growth, the author revisits his ‘Otter Test, ‘ which asks models to generate an image of an ‘otter on a plane using wifi.’ Between 2022, the year ChatGPT launched, and 2025, the progress is, in his view, rapid and remarkable. He also submitted a prompt to a video model from ByteDance, TikTok’s parent company, still unavailable in the United States: the result is, according to the author, nearly perfect, save for one pronunciation error.

Benchmarks, those standardized tests that measure model performance, follow the same curve. The METR Long Tasks graph measures the amount of human work an AI can accomplish autonomously with a certain level of reliability. This graph has been criticized, and METR itself has flagged potential issues, but most capability measures show, according to the author, a similar profile. On the Google-Proof Q&A test, graduate students using Google score 34% outside their field and about 70% within it, while the best AIs reach 94%. On GDPval, the latest AIs match or exceed top-performing humans in 82% of cases, according to industry experts.

These advances remain uneven. The author notes that AI remains ‘inconsistent, ‘ capable of performing certain tasks at a high level while failing at others. Companies, for their part, are still slowly adopting the tool. Some, however, are experimenting with new structures: the security company StrongDM announced it has designed a ‘Software Factory’, where AI agents write, test, and deliver software without human intervention on the code. The two posted rules are: ‘code must not be written by humans’ and ‘code must not be reviewed by humans.’ Each engineer must spend on AI tokens, the compute units billed by models, an amount equivalent to their salary, or at least $1,000 per day.

A Turbulent Week

According to the author, the week of February 22-27 illustrates what the coming period could become. On February 22, the financial research firm Citrini Research published a fictional scenario on how AI adoption could destroy several established companies by 2028. The text, deemed partly implausible, nonetheless triggered notable movements on Wall Street. On February 26, the financial services company Block announced 40% job cuts, attributing them to AI, an explanation the New York Times considers likely exaggerated. On February 27, a public conflict erupted between the Pentagon and Anthropic over who should set the rules for the government’s use of Claude.

Toward AI Self-Improvement

The author also mentions ‘recursive self-improvement, ‘ or RSI, the idea that AIs are increasingly used to build better AIs. At Davos in January, Anthropic’s Dario Amodei explained that models proficient in code and AI research could be used to create the next generation. He noted that engineers at his company are almost writing less code themselves. In February, OpenAI introduced its latest Codex model as ‘our first model that played a decisive role in its own creation.’ Google DeepMind’s Demis Hassabis indicated during the same panel that all major labs are working on this.

The author acknowledges that it’s unknown how far this loop can go, and that technical, computational, or data limits could emerge. For him, RSI is no longer science fiction but ‘an explicit part of every major AI company’s roadmap.’

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x