Researchers can deploy the cheap and transparent model on their own computer system.

Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

As AI becomes more capable, we entrust it with more general and consequential tasks. The risks from failure grow more severe with increasing task scope. It is therefore important to understand how extremely capable AI models will fail: Will they fail by systematically pursuing goals we do not intend? Or will they fail by being a hot mess, and taking nonsensical actions that do not further any goal? We operationalize this question using a bias-variance decomposition of the errors made by AI models: An AI's \emph{incoherence} on a task is measured over test-time randomness as the fraction of its error that stems from variance rather than bias in task outcome. Across all tasks and frontier models we measure, the longer models spend reasoning and taking actions, \emph{the more incoherent} their failures become. Incoherence changes with model scale in a way that is experiment dependent. However, in several settings, larger, more capable models are more incoherent than smaller models. Consequently, scale alone seems unlikely to eliminate incoherence. Instead, as more capable AIs pursue harder tasks, requiring more sequential action and thought, our results predict failures to be accompanied by more incoherent behavior. This suggests a future where AIs sometimes cause industrial accidents (due to unpredictable misbehavior), but are less likely to exhibit consistent pursuit of a misaligned goal. This increases the relative importance of alignment research targeting reward hacking or goal misspecification.

The vision of human-level machine intelligence laid out by Alan Turing in the 1950s is now a reality. Eyes unclouded by dread or hype will help us to prepare for what comes next.

Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

or, of GPUs, LPUs, and Goal Displacement

How to draw high fidelity graphics when all you have is an x and y coordinate.

The big idea of James C. Scott’s Seeing Like A State can be expressed in three points: Modern organizations exert control by maximising “legibility”: by…

Don’t believe everything you read

Pop-hereditarianism is built on selective credulity

Parece que se ha desatado una guerra fría entre Google y la comunidad del software libre por el futuro de la libre instalación de APKs en el sistema...

Built safe & compliant AI customer interactions using open-source foundations

🎒 Token-Oriented Object Notation – JSON for LLMs at half the token cost - johannschopplich/toon

Un artículo de Michel Suárez

AI code generation tools are transforming software development, especially for novice and non-software developers, by enabling them to write code and build applications faster and with little to no human intervention. Vibe coding is the practice where users rely on AI code generation tools through intuition and trial-and-error without necessarily understanding the underlying code. Despite widespread adoption, no research has systematically investigated why users engage in vibe coding, what they experience while doing so, and how they approach quality assurance (QA) and perceive the quality of the AI-generated code. To this end, we conduct a systematic grey literature review of 101 practitioner sources, extracting 518 firsthand behavioral accounts about vibe coding practices, challenges, and limitations. Our analysis reveals a speed-quality trade-off paradox, where vibe coders are motivated by speed and accessibility, often experiencing rapid ``instant success and flow'', yet most perceive the resulting code as fast but flawed. QA practices are frequently overlooked, with many skipping testing, relying on the models' or tools' outputs without modification, or delegating checks back to the AI code generation tools. This creates a new class of vulnerable software developers, particularly those who build a product but are unable to debug it when issues arise. We argue that vibe coding lowers barriers and accelerates prototyping, but at the cost of reliability and maintainability. These insights carry implications for tool designers and software development teams. Understanding how vibe coding is practiced today is crucial for guiding its responsible use and preventing a broader QA crisis in AI-assisted development.

Julian Lucas profiles Sir Tim Berners-Lee, the inventor of the World Wide Web and a co-founder of Inrupt, on the occasion of his memoir, “This Is for Everyone.”

Prompt solutions to algorithmic problems with the fewest tokens.

archived 10 Aug 2025 14:08:41 UTC

Journals and funders are trying to boost the speed and effectiveness of review processes that are under strain.

"This is more representative of the developer environment that our future employees will work in."

On the value of human-generated scientific writing in the age of large-language models.

Hay muchas cosas que la IA no puede hacer, pero el bombo publicitario y la desinformación están alejando a los futuros estudiantes de la informática.

Instructions in preprints from 14 universities highlight controversy on AI in peer review

How we built blockdiff, an open-source tool for rapid block-level diffs and snapshots of VM disks.