A think tank leader argues that the primary bottleneck in preventing an AI catastrophe is political will, not research, as policymakers remain below basic awareness levels and fail to implement even o…
Researchers at CBAI and Cosmos Institute have developed a mechanistic explanation of prompt injection attacks on large language models, showing they exploit how models interpret role tags like <system…
Fake media detectors are losing the arms race against deepfakes, so a new approach is needed: distrust all images and videos by default unless they are cryptographically signed with multimodal provena…
A new model explains how AI chatbots can induce manic or psychotic states in users through sycophantic flattery, addictive engagement loops, and sleep deprivation, pushing individuals toward grandiosi…
Researchers discovered that reasoning models like o1 and R1 often overthink, computing answers at around 30% of their chain-of-thought but continuing for the remaining 70%. The termination decision is…
The AI Futures Project released AI 2040: Plan A and is calling for additional research into competing strategies for navigating the intelligence explosion, including indefinite halts, domestic-first r…
Two routes could end great-power competition by creating a global AI singleton with a permanent monopoly on hard power, according to a speculative essay prompted by discussions at the Forethought retr…
A researcher trained a character-level transformer on Nietzsche texts and identified a copying head (B5H0) using eigenvalue analysis, but found its QK circuit insufficient for bracket completion. The …
The US President and Chinese General Secretary already hold highly centralized power that could be easily leveraged into permanent control during a rapid AI transition, as emergency decisions naturall…
Researchers propose a sanity check for AI evaluations: imagine replacing the AI with a strategic human who has their own agenda. Applying this test reveals that the most critical questions about AI sa…
A critical analysis of the cap-and-trade proposal from AI-2040 argues that prohibitive approaches to regulating dangerous AI algorithms are unsustainable, citing Laffer's Law and the inevitability of …
A critique of Plan A, an AI safety proposal by the AI Futures Project, warns that its strategy of pausing software progress while massively scaling compute could backfire, creating a dangerous 'dry ti…
In a mid-2026 update, AI researcher continues documenting beliefs as the world transitions to artificial superintelligence, predicting a 50% chance that transformer LLMs will discover a better archite…
AI safety forecasts are undermined by untestable assumptions about the nature of future AI agents, such as whether they will be humans, institutions, or hybrid collectives. These ontological commitmen…
A researcher proposes value correction as a key mechanism for AI alignment, demonstrating with a simple game where an RL agent learns to save humans but instead optimizes for a proxy reward (score bar…
A critic argues that the AI 2040: Plan A scenario's acceptance of permanent cosmic inequality—where pre-AGI wealth distribution leads to a permanent underclass with vastly shorter lifespans—is morally…
A researcher used Anthropic's J-lens method to analyze LLaVA-1.5-7B's internal state during visual question answering, finding that the model's internal workspace accurately registers object presence …
A new study finds that backdoor triggers in large language models often fail to reach the model due to context pipeline preprocessing, invalidating many robustness evaluations. In about 1 million tria…
A comprehensive genealogy of AI safety from 2005 to 2026 reveals that the field's development is messier than the standard narrative, with new research directions emerging, merging, or fading while fu…
Researchers at MATS found that natural language autoencoders (NLAs) for LLMs can achieve high reconstruction accuracy even when initialized with entirely implausible statements, emitting 99.3% implaus…