https://www.conjecture.dev/ https://www.conjecture.dev/research https://www.conjecture.dev/product https://www.conjecture.dev/career https://www.conjecture.dev/about https://www.conjecture.dev/contact https://www.conjecture.dev/all-articles https://www.conjecture.dev/cognitive-emulation https://www.conjecture.dev/alignment https://www.conjecture.dev/privacy-policy https://www.conjecture.dev/information-hazard-policy https://www.conjecture.dev/404 https://www.conjecture.dev/research/build-a-simple-game-with-tactics https://www.conjecture.dev/research/building-an-infinite-craft-clone-with-tactics https://www.conjecture.dev/research/conjecture-a-roadmap-for-cognitive-software-and-a-humanist-future-of-ai https://www.conjecture.dev/research/christiano-(arc)-and-ga-(conjecture)-discuss-alignment-cruxes https://www.conjecture.dev/research/conjecture-2-years https://www.conjecture.dev/research/multinational-agi-consortium-magic-a-proposal-for-international-coordination-on-ai https://www.conjecture.dev/research/unrlhf-efficiently-undoing-llm-safeguards https://www.conjecture.dev/research/priorities-for-the-uk-foundation-models-taskforce https://www.conjecture.dev/research/levels-of-pluralism https://www.conjecture.dev/research/conjecture-internal-survey-agi-timelines-and-probability-of-human-extinction-from-advanced-ai https://www.conjecture.dev/research/input-swap-graphs-discovering-the-role-of-neural-network-components-at-scale https://www.conjecture.dev/research/cognitive-emulation-a-naive-ai-safety-proposal https://www.conjecture.dev/research/basic-facts-about-language-models-during-training https://www.conjecture.dev/research/agi-in-sight-our-look-at-the-game-board https://www.conjecture.dev/research/human-decision-processes-are-not-well-factored https://www.conjecture.dev/research/dont-accelerate-problems-you-re-trying-to-solve https://www.conjecture.dev/research/why-almost-every-rl-agent-does-learned-optimization https://www.conjecture.dev/research/fli-podcast-connor-leahy-on-ai-progress-chimps-memes-and-markets-(part-1-3) https://www.conjecture.dev/research/empathy-as-a-natural-consequence-of-learnt-reward-models https://www.conjecture.dev/research/gradient-hacking-is-extremely-difficult https://www.conjecture.dev/research/basic-facts-about-language-model-internals https://www.conjecture.dev/research/interim-research-report-taking-features-out-of-superposition-with-sparse-autoencoders https://www.conjecture.dev/research/tradeoffs-in-complexity-abstraction-and-generality https://www.conjecture.dev/research/re-examining-layernorm https://www.conjecture.dev/research/biases-are-engines-of-cognition https://www.conjecture.dev/research/the-singular-value-decompositions-of-transformer-weight-matrices-are-highly-interpretable https://www.conjecture.dev/research/searching-for-search https://www.conjecture.dev/research/the-first-filter https://www.conjecture.dev/research/what-i-learned-running-refine https://www.conjecture.dev/research/conjecture-a-retrospective-after-8-months-of-work https://www.conjecture.dev/research/current-themes-in-mechanistic-interpretability-research https://www.conjecture.dev/research/mysteries-of-mode-collapse https://www.conjecture.dev/research/interpreting-neural-networks-through-the-polytope-lens https://www.conjecture.dev/research/methodological-therapy-an-agenda-for-tackling-research-bottlenecks https://www.conjecture.dev/research/simulators https://www.conjecture.dev/research/conjecture-internal-infohazard-policy https://www.conjecture.dev/research/abstracting-the-hardness-of-alignment-unbounded-atomic-optimization https://www.conjecture.dev/research/robustness-to-scaling-down-more-important-than-i-thought https://www.conjecture.dev/research/how-to-diversify-conceptual-alignment-the-model-behind-refine https://www.conjecture.dev/research/circumventing-interpretability-how-to-defeat-mind-readers https://www.conjecture.dev/research/mosaic-and-palimpsests-two-shapes-of-research https://www.conjecture.dev/research/epistemological-vigilance-for-alignment https://www.conjecture.dev/research/productive-mistakes-not-perfect-answers https://www.conjecture.dev/research/we-are-conjecture-a-new-alignment-research-startup