Anthropic Report: AI Is Already Writing 80% of Its Own Code, Nearing Semi-RSI

·1 views

Anthropic published a report titled 'When AI builds itself,' revealing that its Claude model now generates approximately 80% of the company's production code, with engineers merging code at eight times the rate seen between 2021 and 2025. In one notable instance from April 2026, Claude resolved over 800 bugs in a task that human engineers estimated would have taken four years to complete manually. Recursive self-improvement, or RSI, refers to a hypothetical process where an AI system rewrites its own code to become progressively more capable without human direction — a concept first theorized by mathematician I. J. Good. Experts, including the CSA research center, clarify that what is occurring today is semi-automated RSI under human oversight, not fully autonomous RSI where a system redesigns itself without any human involvement. The 2026 International AI Safety Report, authored by over a hundred independent experts, ranks loss of control through RSI among the most serious national-security-level risks currently facing the field.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Freelancer loses long-term client to AI, learns which work is truly irreplaceable

A freelancer lost a client of over a year not due to poor performance, but because the client chose to replace routine tasks with an AI tool at a lower cost. The writer reflects that the work lost was simple, repetitive, and required little creative judgment — making it an easy target for automation. The experience revealed that predictable, low-effort work carries the highest risk of being displaced by AI tools. Clients who continued working with the freelancer were those who needed complex, judgment-driven problem-solving that AI cannot reliably handle. The writer now uses this as a lens to reconsider how freelancers should position their skills in an AI-augmented market.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How barometer.today builds its animated global isobar maps using open weather data

Barometer.today, a web tool showing air pressure forecasts for cities worldwide, also features an animated global isobar map spanning one day back to four days ahead. The map is generated from ECMWF forecast data sourced via Open-Meteo, delivered as a pressure grid at 0.25-degree resolution and updated four times daily. To prevent jagged contour lines, the team applies a Gaussian blur, but uses a normalized convolution technique to stop inflated sea-level pressure values over high plateaus — such as the Himalayas and the Andes — from distorting surrounding data. Elevated terrain above 2,500 metres is masked using a digital elevation model, and isobars crossing these regions are rendered as dashed lines rather than omitted, keeping the map visually continuous while signalling extrapolated values. The underlying code is written in Python using numpy, scipy, contourpy, and shapely, and the approach is applicable to anyone drawing contour maps from gridded geospatial data.

0 comments Read more at DEV Community

ProgrammingHacker News ·

CIA Declassifies 71 President's Daily Briefs Tied to 9/11 Anniversary

The CIA has released 71 declassified President's Daily Brief documents related to the September 11 attacks. The release was made in commemoration of the 25th anniversary of 9/11. President's Daily Briefs are highly classified intelligence summaries prepared for the sitting U.S. president. The declassification provides the public with a rare look at the intelligence landscape surrounding the 2001 terrorist attacks. The documents are available through the CIA's official website.

0 comments Read more at Hacker News

ProgrammingGitHub Blog ·

GitHub Engineer Automates APAC Marketing Event Ops Using Code-Based Workflows

A GitHub engineer developed an automation system to streamline the full lifecycle of marketing events for the company's Asia-Pacific team. The approach is built on the principle that any documented, repeatable workflow can be automated. The system covers event operations from initial planning through to post-event follow-up. By treating marketing processes as code, the solution reduces manual effort and improves consistency across campaigns. The project was shared via the official GitHub Blog as a practical example of applying engineering thinking to marketing operations.

0 comments Read more at GitHub Blog