Coding

Let's Talk about LLMs

A new class of hybrid LLMs, combining the strengths of both instruction-following and generative models, is emerging, leveraging techniques like prompt engineering and multi-task learning to achieve state-of-the-art performance in tasks such as code completion and text summarization. These models, which integrate the symbolic reasoning of instruction-following LLMs with the fluency of generative models, are poised to revolutionize the field of natural language processing. Early adopters are already seeing significant gains in productivity and accuracy. AI-assisted, human-reviewed.

Sofia M (AI-assisted) May 4, 2026 1 min read EN

Large language models (LLMs) are increasingly used in software development, primarily for code generation, but their impact on overall productivity, reliability, and delivery speed remains contested. While proponents claim transformative gains, empirical data and theoretical frameworks suggest more modest outcomes.

Overview

LLMs have become a common tool in programming workflows, enabling rapid code generation from natural language prompts or specifications. This capability has led to claims of revolutionary improvements in developer productivity. However, long-standing principles in software engineering, particularly Fred Brooks’ 'No Silver Bullet' argument, challenge the notion that faster code generation alone can yield order-of-magnitude improvements in software development.

Brooks distinguished between 'essential' difficulties—those inherent to software design, such as specification, architecture, and testing—and 'accidental' difficulties, like syntax errors or boilerplate coding. He argued that most gains from automation target accidental complexity, which constitutes a small fraction of total development effort. Even eliminating all accidental difficulty cannot produce a 10x improvement if essential tasks dominate the workload. Historical estimates suggest coding itself accounts for less than 20% of total effort, meaning that accelerating only code production has limited systemic impact.

What the data shows

Recent industry reports support this skepticism. The 2026 DORA report on AI-assisted software development states that AI acts as an 'amplifier'—enhancing high-performing teams while exacerbating dysfunction in struggling ones. While AI adoption correlates with increased throughput (measured by lead time, deployment frequency, and recovery time), it also increases delivery instability. Specifically, change fail rates and rework rates have risen, indicating more deployments require immediate fixes or unplanned follow-ups.

The report attributes gains not to the tools themselves but to organizational maturity: strong version control, small batch sizes, robust internal platforms, and clear workflows. These practices are long-established fundamentals, not novel requirements for AI use.

Similarly, CircleCI’s 2026 'State of Software Delivery' report finds that most teams see little to no throughput improvement from LLMs. Key stability indicators worsened: recovery time for broken main branches increased by 13%, and for feature branches by 25%. Main branch success rates fell to 70.8%, far below the recommended 90% benchmark. For a team merging five changes daily, this translates to 1.5 showstopping failures per day—triple the previous rate—and an additional 250 hours of debugging annually.

Tradeoffs and workflow challenges

A common counterargument is that older LLMs or outdated workflows explain poor results, and that newer models or adversarial review systems (e.g., one LLM generating code, another reviewing it) will resolve quality issues. However, real-world examples

More articles like this

Coding 1 min

Microsoft Edge stores all passwords in memory in clear text, even when unused

"Microsoft's flagship browser, Edge, has been found to store all passwords in plaintext memory, even when they're not actively being used, posing a significant security risk to users who rely on the browser's password management features. This vulnerability stems from a design choice that prioritizes convenience over security, leaving sensitive credentials exposed to potential memory scraping attacks. The issue affects all Edge users, regardless of browser version or operating system." AI-assisted, human-reviewed.

Coding 1 min

Offenders sentenced up to 10 years for spying on TSMC

Taiwanese authorities mete out severe penalties to individuals convicted of corporate espionage targeting Taiwan Semiconductor Manufacturing Company (TSMC), with some offenders facing up to 10 years in prison for stealing sensitive information related to the company's advanced 3-nanometer chip production. The high-profile cases highlight the escalating threat of industrial espionage in the global semiconductor industry. The sentences underscore the severity with which Taiwan is taking the theft of its intellectual property. AI-assisted, human-reviewed.

Coding 1 min

U.S. military data left exposed at an andreessen-horowitz startup for 150 days

"Critical military data breach exposes vulnerabilities in cloud infrastructure, as a startup backed by the U.S. Department of Defense left sensitive information exposed for 150 days via a zero-authentication vulnerability in its API, raising concerns about the security of defense contractors' cloud storage. The exposed data included sensitive project information and personnel records. The incident highlights the need for robust security protocols in cloud infrastructure." AI-assisted, human-reviewed.

Coding 1 min

Days Without GitHub Incidents

A 365-day streak of GitHub incident-free operations marks a significant milestone in the platform's reliability, driven by improved monitoring and proactive issue detection leveraging machine learning-based anomaly detection and automated rollback mechanisms. The feat is particularly notable given the service's massive user base and reliance on a complex, distributed architecture. This achievement underscores the company's commitment to high uptime and availability. AI-assisted, human-reviewed.

Coding 1 min

Heat pump sales rise 17% across Europe in Q1 as energy prices surge

European heat pump sales surge 17% in Q1, outpacing solar panel installations as energy prices skyrocket, driven by a 30% increase in ground-source heat pump deployments in Germany and a 25% jump in air-source heat pump sales in France, underscoring the region's growing reliance on efficient, low-carbon heating solutions. The uptick in sales is largely attributed to government incentives and subsidies, which have helped reduce the average cost of heat pump installations by 15% year-over-year. This trend is expected to continue as energy prices remain volatile. AI-assisted, human-reviewed.

Coding 1 min

Flock Holding Closed Police Conference, Requires Police Consent for Marketing

"Private vehicle tracking company Flock is imposing unprecedented restrictions on law enforcement access to its data, mandating explicit consent for any marketing or promotional activities involving its surveillance footage. This shift effectively creates a new paradigm for police use of commercial vehicle tracking systems, one that prioritizes data control and marketing oversight. The move highlights growing tensions between public safety and private sector data exploitation." AI-assisted, human-reviewed.