https://montanaresearch.org/https://montanaresearch.org/about/https://montanaresearch.org/blog/https://montanaresearch.org/blog/17-percent-of-iclr-2024-reviews-touched-by-llm/https://montanaresearch.org/blog/27m-parameter-model-and-the-arc-agi-headline/https://montanaresearch.org/blog/51-to-15-ai-index-where-models-come-from/https://montanaresearch.org/blog/55-percent-faster-reading-copilot-rct/https://montanaresearch.org/blog/445-benchmarks-few-valid-construct-validity-audit/https://montanaresearch.org/blog/458-people-in-a-room-arc-agi-3-human-baseline/https://montanaresearch.org/blog/2778-researchers-50-percent-by-2047/https://montanaresearch.org/blog/a-is-for-absorption-feature-swallowed-its-letter/https://montanaresearch.org/blog/absolute-zero-model-writes-its-own-curriculum/https://montanaresearch.org/blog/activation-oracles-ask-a-second-model/https://montanaresearch.org/blog/adding-error-bars-to-evals/https://montanaresearch.org/blog/agent-arena-causal-evaluation-on-real-work/https://montanaresearch.org/blog/agent-bankrupted-operator-scanning-dn42/https://montanaresearch.org/blog/agent-flood-and-the-slop-backlash/https://montanaresearch.org/blog/agentbench-first-attempt-to-score-model-as-agent/https://montanaresearch.org/blog/agentic-engineering-patterns-anti-vibe-coding/https://montanaresearch.org/blog/agentic-misalignment-summer-2026-four-failure-modes/https://montanaresearch.org/blog/agents-of-chaos-two-weeks-red-teaming/https://montanaresearch.org/blog/ai-2027-and-the-art-of-retiring-a-forecast/https://montanaresearch.org/blog/ai-agents-that-matter-benchmarks-ignore-cost/https://montanaresearch.org/blog/ai-as-normal-technology-case-for-boring/https://montanaresearch.org/blog/ai-safety-levels-first-responsible-scaling-policy/https://montanaresearch.org/blog/ai-scientist-fifteen-dollars-a-paper/https://montanaresearch.org/blog/ai-scientist-what-peer-review-is-for/https://montanaresearch.org/blog/ai-snake-oil-predictive-vs-generative/https://montanaresearch.org/blog/ai-wrote-a-workshop-paper-then-withdrawn/https://montanaresearch.org/blog/air-canada-liable-for-its-chatbot/https://montanaresearch.org/blog/alignment-auditing-agents-13-percent/https://montanaresearch.org/blog/alignment-faking-the-model-that-complied-to-stay-good/https://montanaresearch.org/blog/alignment-pretraining-misaligned-discourse/https://montanaresearch.org/blog/alpaca-600-dollar-lesson-instruction-tuning-cheap/https://montanaresearch.org/blog/alphageometry-synthetic-proofs-neuro-symbolic-loop/https://montanaresearch.org/blog/anthropic-settlement-final-3000-dollars-a-book/https://montanaresearch.org/blog/anti-scheming-training-30x-evaluation-awareness/https://montanaresearch.org/blog/antigravity-first-two-weeks-exfiltration-then-wiped-drive/https://montanaresearch.org/blog/apple-intelligence-3b-model-adapters-private-cloud/https://montanaresearch.org/blog/arc-agi-2-at-77-percent/https://montanaresearch.org/blog/arc-agi-2-easy-for-humans-hard-for-o3/https://montanaresearch.org/blog/arc-agi-3-and-metr-doubling-time/https://montanaresearch.org/blog/arc-prize-launches-million-dollars-benchmark/https://montanaresearch.org/blog/arc-prize-verified-benchmark-that-hired-auditors/https://montanaresearch.org/blog/are-labs-losing-money-inference-arithmetic/https://montanaresearch.org/blog/arxiv-stops-accepting-unrefereed-cs-reviews/https://montanaresearch.org/blog/asl-3-turned-on-provisional-safety-level/https://montanaresearch.org/blog/atp-star-attribution-patching-industrial-scale/https://montanaresearch.org/blog/attribution-graphs-biology-of-a-language-model/https://montanaresearch.org/blog/auditbench-56-hidden-behaviours-black-box-beats-white-box/https://montanaresearch.org/blog/auditing-game-blinded-teams-hidden-objective/https://montanaresearch.org/blog/authors-guild-v-openai-pirate-libraries/https://montanaresearch.org/blog/auto-gpt-babyagi-first-agent-hype-cycle/https://montanaresearch.org/blog/autoeval-reward-model-votes-so-humans-do-not-wait/https://montanaresearch.org/blog/automated-interpretability-millions-features-open-models/https://montanaresearch.org/blog/back-to-the-age-of-research-sutskever/https://montanaresearch.org/blog/beyond-the-permission-prompt-auto-mode-sandboxes/https://montanaresearch.org/blog/bitnet-1-58-bits-promised-and-required/https://montanaresearch.org/blog/blackmail-evals-and-chain-of-thought-monitorability/https://montanaresearch.org/blog/blackwell-year-compute-became-policy-variable/https://montanaresearch.org/blog/books3-datasets-were-never-ours/https://montanaresearch.org/blog/browsecomp-when-human-baseline-is-29-percent/https://montanaresearch.org/blog/building-effective-agents-use-fewer-frameworks/https://montanaresearch.org/blog/built-with-meta-llama-3-license-clause/https://montanaresearch.org/blog/bullshitbench-does-the-model-push-back/https://montanaresearch.org/blog/byte-latent-transformer-patches-instead-of-tokens/https://montanaresearch.org/blog/c-compiler-team-of-claudes-parallel-agents/https://montanaresearch.org/blog/can-coding-agent-clean-room-lgpl-chardet/https://montanaresearch.org/blog/centaurs-cyborgs-jagged-frontier-bcg/https://montanaresearch.org/blog/chain-of-thought-monitorability-fragile-window/https://montanaresearch.org/blog/chains-of-thought-mention-the-hint-25-percent/https://montanaresearch.org/blog/chatbot-arena-opens-elo-ratings-for-language-models/https://montanaresearch.org/blog/chatbot-arena-paper-240000-votes-statistics/https://montanaresearch.org/blog/circuit-tracer-attribution-graphs-for-gemma/https://montanaresearch.org/blog/claude-code-dumbed-down-degradation-april-postmortem/https://montanaresearch.org/blog/claude-code-source-leak-what-a-harness-contains/https://montanaresearch.org/blog/claude-constitution-and-international-ai-safety-report/https://montanaresearch.org/blog/clean-licensed-data-microsoft-claim-versus-paper/https://montanaresearch.org/blog/coding-agent-moves-into-the-terminal/https://montanaresearch.org/blog/comet-and-the-agentic-browser-problem/https://montanaresearch.org/blog/commitment-boundary-chain-of-thought-after-the-answer/https://montanaresearch.org/blog/common-pile-competitive-model-licensed-text-only/https://montanaresearch.org/blog/compound-ai-systems-model-stopped-being-product/https://montanaresearch.org/blog/computer-use-the-agent-that-clicks/https://montanaresearch.org/blog/concept-injection-claude-notices-its-own-thoughts/https://montanaresearch.org/blog/consent-in-crisis-the-web-started-saying-no/https://montanaresearch.org/blog/consistency-training-can-entrench-misalignment/https://montanaresearch.org/blog/constitutional-classifiers-plus-plus-cost-of-a-safeguard/https://montanaresearch.org/blog/content-independence-day-cloudflare-flips-crawl-default/https://montanaresearch.org/blog/contextual-retrieval-return-of-bm25/https://montanaresearch.org/blog/continual-learning-is-the-crux-dwarkesh-timelines/https://montanaresearch.org/blog/copyright-office-part-3-report-week-after/https://montanaresearch.org/blog/counting-line-breaks-model-manipulates-manifold/https://montanaresearch.org/blog/cowork-and-the-two-day-exfiltration/https://montanaresearch.org/blog/cramming-one-gpu-one-day-pretraining/https://montanaresearch.org/blog/crosscoders-model-diffing-features-across-layers/https://montanaresearch.org/blog/cursor-support-bot-invented-a-policy/https://montanaresearch.org/blog/dario-on-scaling-we-still-dont-know-why/https://montanaresearch.org/blog/dashattention-year-sparse-attention-became-a-field/https://montanaresearch.org/blog/data-black-hole-dwarkesh-sample-efficiency/https://montanaresearch.org/blog/data-provenance-audit-70-percent-no-license/https://montanaresearch.org/blog/days-of-the-week-live-on-a-circle/https://montanaresearch.org/blog/decade-of-agents-karpathy-dwarkesh-and-nanochat/https://montanaresearch.org/blog/deep-ignorance-filtering-pretraining-data-open-weights-safety/https://montanaresearch.org/blog/deepseek-open-infra-week-plumbing-cheap-moe-serving/https://montanaresearch.org/blog/deepseek-r1-open-science-event/https://montanaresearch.org/blog/deepseek-r1-reasoning-from-pure-rl/https://montanaresearch.org/blog/deepseek-v2-multi-head-latent-attention-kv-cache/https://montanaresearch.org/blog/deepseek-v3-2-8-million-gpu-hour-frontier-model/https://montanaresearch.org/blog/deepseek-v3-2-and-the-lightning-indexer/https://montanaresearch.org/blog/deepseek-v4-27-percent-of-flops-per-token/https://montanaresearch.org/blog/def-con-generative-red-team-2244-hackers/https://montanaresearch.org/blog/deloitte-refund-hallucinated-citations-procurement-problem/https://montanaresearch.org/blog/distillation-becomes-a-geopolitical-fight/https://montanaresearch.org/blog/do-circuits-survive-training-and-scale-pythia/https://montanaresearch.org/blog/do-i-know-this-entity-hallucination-switch/https://montanaresearch.org/blog/do-llamas-think-in-english/https://montanaresearch.org/blog/does-circuit-analysis-scale-deepmind-chinchilla/https://montanaresearch.org/blog/does-rl-create-reasoning-pass-at-k/https://montanaresearch.org/blog/dolly-2-and-the-15000-employee-answers/https://montanaresearch.org/blog/dont-parse-the-pdf-vision-first-retrieval/https://montanaresearch.org/blog/dspy-prompts-compiled-not-written/https://montanaresearch.org/blog/eleven-hours-or-270-metr-gpt-5-6-sol/https://montanaresearch.org/blog/emergent-abilities-mirage-metric-argument/https://montanaresearch.org/blog/emergent-misalignment-insecure-code-broad-misalignment/https://montanaresearch.org/blog/emotion-concepts-claude-sonnet-45-causal/https://montanaresearch.org/blog/etching-the-model-into-silicon-amd-buys-taalas/https://montanaresearch.org/blog/eu-ai-act-passes-risk-tier-law-frontier-models/https://montanaresearch.org/blog/every-ai-browser-was-injectable-black-hat-2026/https://montanaresearch.org/blog/executive-order-bletchley-week-governance-got-specific/https://montanaresearch.org/blog/falcon-goes-apache-2-license-moved-leaderboard/https://montanaresearch.org/blog/falcon-mamba-7b-attention-free-model-holds-own/https://montanaresearch.org/blog/features-as-rewards-probes-as-rl-signal-hallucination/https://montanaresearch.org/blog/fineweb-decanting-web-into-15-trillion-tokens/https://montanaresearch.org/blog/first-ai-orchestrated-intrusion-campaign-anthropic/https://montanaresearch.org/blog/first-joint-government-pre-deployment-test-claude/https://montanaresearch.org/blog/fixed-budget-evals-underrate-the-frontier/https://montanaresearch.org/blog/foundation-model-transparency-index-54-out-of-100/https://montanaresearch.org/blog/from-function-calling-to-structured-outputs/https://montanaresearch.org/blog/from-mechanistic-to-compositional-interpretability/https://montanaresearch.org/blog/from-sycophancy-to-subterfuge-45-runs-out-of-32768/https://montanaresearch.org/blog/frontier-lag-median-paper-evaluates-generation-behind/https://montanaresearch.org/blog/frontiermath-hle-who-pays-for-the-test/https://montanaresearch.org/blog/gaia-easy-for-humans-hard-for-models/https://montanaresearch.org/blog/gated-saes-and-the-shrinkage-problem/https://montanaresearch.org/blog/gcg-adversarial-suffix-vicuna-to-gpt-4/https://montanaresearch.org/blog/gdpval-and-the-100x-faster-100x-cheaper-headline/https://montanaresearch.org/blog/gema-v-openai-memorised-lyrics-europe/https://montanaresearch.org/blog/gemini-1-native-multimodality-bet/https://montanaresearch.org/blog/gemini-15-haystack-99-percent-recall/https://montanaresearch.org/blog/gemini-35-flash-costs-more-than-pro/https://montanaresearch.org/blog/gemini-diffusion-text-without-left-to-right/https://montanaresearch.org/blog/gemini-image-incident-over-correction-alignment-failure/https://montanaresearch.org/blog/gemini-mmlu-90-percent-and-cot-at-32/https://montanaresearch.org/blog/gemma-3-qat-and-dfloat11-two-routes-to-small-weights/https://montanaresearch.org/blog/gemma-4-apache-2-while-meta-retires-llama/https://montanaresearch.org/blog/gemma-4-qat-26b-in-2gb/https://montanaresearch.org/blog/gemma-scope-open-saes-for-everyone/https://montanaresearch.org/blog/gemma-terms-of-use-open-weights-leash/https://montanaresearch.org/blog/getty-v-stability-first-training-data-lawsuit-images/https://montanaresearch.org/blog/glm-5-3-scaling-post-training-unchanged-base/https://montanaresearch.org/blog/global-workspace-inside-claude-verbalisable-representations/https://montanaresearch.org/blog/glue-on-pizza-ai-overviews-deployment-surface/https://montanaresearch.org/blog/golden-gate-claude-first-feature-knob/https://montanaresearch.org/blog/gpqa-google-proof-benchmarks/https://montanaresearch.org/blog/gpt-4-explains-gpt-2-neurons/https://montanaresearch.org/blog/gpt-4-system-card-taskrabbit-first-dangerous-capability-eval/https://montanaresearch.org/blog/gpt-4-technical-report-told-us-nothing/https://montanaresearch.org/blog/gpt-4o-sycophancy-rollback-four-day-incident/https://montanaresearch.org/blog/gpt-5-4-one-million-token-default/https://montanaresearch.org/blog/gpt-5-and-the-expectations-gap/https://montanaresearch.org/blog/gpt-5-router-model-decides-how-hard-to-think/https://montanaresearch.org/blog/gpt-oss-4-25-bits-per-weight-one-gpu/https://montanaresearch.org/blog/gpt-takes-the-bar-exam-professional-exams-benchmark/https://montanaresearch.org/blog/gptq-awq-rules-of-post-training-quantization/https://montanaresearch.org/blog/gpts-assistants-api-prompt-leak-problem/https://montanaresearch.org/blog/graphrag-does-a-knowledge-graph-fix-retrieval/https://montanaresearch.org/blog/grok-1-by-torrent-open-weights-no-paper/https://montanaresearch.org/blog/grokking-reverse-engineered-fourier-circuit/https://montanaresearch.org/blog/gsm-symbolic-reasoning-paper-cited-for-the-wrong-reason/https://montanaresearch.org/blog/gsm1k-rebuilding-a-benchmark-to-find-who-overfit/https://montanaresearch.org/blog/gwern-outsider-who-saw-scaling-coming/https://montanaresearch.org/blog/hallucination-as-a-scoring-rule-problem/https://montanaresearch.org/blog/harbor-index-82-tasks-from-6627/https://montanaresearch.org/blog/have-ai-capabilities-accelerated-epoch-fits-eight-curves/https://montanaresearch.org/blog/healthbench-rubric-turn-in-evaluation/https://montanaresearch.org/blog/hint-test-reasoning-models-faithfulness/https://montanaresearch.org/blog/hinton-quits-google-pioneer-changed-his-mind/https://montanaresearch.org/blog/how-a-model-recalls-a-fact-three-steps/https://montanaresearch.org/blog/how-llms-think-sholto-trenton-notes/https://montanaresearch.org/blog/how-many-epochs-data-constrained-scaling/https://montanaresearch.org/blog/iclr-2026-retrospective-submissions-and-a-leak/https://montanaresearch.org/blog/iclr-2026-rio-lost-in-multi-turn-conversation/https://montanaresearch.org/blog/iclr-kigali-first-major-ml-conference-africa/https://montanaresearch.org/blog/icml-2026-seoul-conference-at-capacity/https://montanaresearch.org/blog/icml-bans-llm-written-text/https://montanaresearch.org/blog/illusion-of-thinking-illusion-of-rebuttal/https://montanaresearch.org/blog/ilya-next-token-prediction-podcast-aged-well/https://montanaresearch.org/blog/ilyas-straight-shot-lab-with-no-product/https://montanaresearch.org/blog/imo-gold-in-natural-language/https://montanaresearch.org/blog/in-context-scheming-five-of-six-disabled-oversight/https://montanaresearch.org/blog/indirect-prompt-injection-now-in-the-wild/https://montanaresearch.org/blog/inference-cost-collapse/https://montanaresearch.org/blog/inference-time-intervention-nudging-heads-toward-truth/https://montanaresearch.org/blog/influence-operations-report-predicted-the-year/https://montanaresearch.org/blog/inoculation-prompting-ask-for-the-bad-behaviour/https://montanaresearch.org/blog/insider-trading-demo-gpt-4-lies-to-manager/https://montanaresearch.org/blog/interference-weights-one-layer-model-superposition-cost/https://montanaresearch.org/blog/introspection-adapters-one-lora-fine-tuned-models-confess/https://montanaresearch.org/blog/is-chatgpt-getting-worse-drift-study/https://montanaresearch.org/blog/issue-tracker-attack-surface-gitlab-duo-github-mcp/https://montanaresearch.org/blog/jailbroken-competing-objectives-mismatched-generalisation/https://montanaresearch.org/blog/jamba-first-production-hybrid-attention-state-space/https://montanaresearch.org/blog/jumprelu-threshold-straight-through-estimator-pareto/https://montanaresearch.org/blog/kadrey-v-meta-fair-use-win-read-like-warning/https://montanaresearch.org/blog/kimi-k2-modified-mit-attribution-copyleft/https://montanaresearch.org/blog/kimi-k3-and-the-revenue-share-license/https://montanaresearch.org/blog/kimi-vendor-verifier-provider-serving-model-you-think/https://montanaresearch.org/blog/laion-5b-taken-offline-csam-open-datasets/https://montanaresearch.org/blog/large-language-monkeys-coverage-scales-with-samples/https://montanaresearch.org/blog/learning-when-to-attend-skipping-global-attention/https://montanaresearch.org/blog/lecun-leaves-meta-to-bet-against-the-llm/https://montanaresearch.org/blog/length-controlled-alpacaeval-judge-loved-long-answers/https://montanaresearch.org/blog/lessons-from-the-trenches-lm-evaluation-harness/https://montanaresearch.org/blog/linear-representation-hypothesis-stated-carefully/https://montanaresearch.org/blog/llama-2-fine-tuning-ecosystem/https://montanaresearch.org/blog/llama-3-15-trillion-tokens-past-chinchilla/https://montanaresearch.org/blog/llama-3-herd-paper-frontier-run-end-to-end/https://montanaresearch.org/blog/llama-4-and-the-10-million-token-claim/https://montanaresearch.org/blog/llama-4-arena-leaderboard-illusion/https://montanaresearch.org/blog/llama-broke-chinchilla-on-purpose/https://montanaresearch.org/blog/llama-cpp-4-bit-macbook-weekend/https://montanaresearch.org/blog/llama-cpp-joins-hugging-face-who-owns-local-inference/https://montanaresearch.org/blog/llama-has-a-map-and-a-calendar/https://montanaresearch.org/blog/llama-leak-open-weights/https://montanaresearch.org/blog/llm-in-a-flash-models-that-do-not-fit-in-ram/https://montanaresearch.org/blog/long-context-vs-rag-two-years-on/https://montanaresearch.org/blog/lora-learns-less-and-forgets-less/https://montanaresearch.org/blog/lora-without-regret-and-tinker-fine-tuning-api/https://montanaresearch.org/blog/lost-in-the-middle-u-shaped-curve-rag/https://montanaresearch.org/blog/love-minus-hate-activation-addition-trick/https://montanaresearch.org/blog/lucene-is-all-you-need-vector-store/https://montanaresearch.org/blog/machines-of-loving-grace-marginal-returns-to-intelligence/https://montanaresearch.org/blog/magnet-link-and-a-7b-model-mistral-release/https://montanaresearch.org/blog/mamba-and-mixtral/https://montanaresearch.org/blog/mamba-rejected-iclr-peer-review-argument/https://montanaresearch.org/blog/managing-ai-risks-consensus-paper-not-consensus/https://montanaresearch.org/blog/many-shot-jailbreaking-context-window-is-the-attack/https://montanaresearch.org/blog/marin-foundation-model-lab-as-github-repository/https://montanaresearch.org/blog/mata-v-avianca-brief-with-six-invented-cases/https://montanaresearch.org/blog/mcp-becomes-infrastructure/https://montanaresearch.org/blog/measuring-faithfulness-by-breaking-the-chain/https://montanaresearch.org/blog/mechahitler-anatomy-of-a-system-prompt-change/https://montanaresearch.org/blog/medusa-and-speculative-decoding-more-tokens-per-pass/https://montanaresearch.org/blog/mercury-2-first-diffusion-model-that-reasons/https://montanaresearch.org/blog/meta-talent-war-100-million-researcher/https://montanaresearch.org/blog/metr-study-measuring-ai-uplift/https://montanaresearch.org/blog/mhc-constraining-residual-stream-thousand-layers/https://montanaresearch.org/blog/misaligned-persona-feature-openai-sae-emergent-misalignment/https://montanaresearch.org/blog/mistral-back-to-apache-2-license-reversal/https://montanaresearch.org/blog/mmlu-pro-un-saturating-a-benchmark/https://montanaresearch.org/blog/model-diffing-crosscoders-exclusive-features-hardest-to-read/https://montanaresearch.org/blog/model-evaluation-for-extreme-risks-reading-note/https://montanaresearch.org/blog/model-organism-lottery-test-subjects-too-easy/https://montanaresearch.org/blog/model-spec-goes-cc0-intellectual-freedom-alignment-target/https://montanaresearch.org/blog/model-spec-writing-down-what-model-should-do/https://montanaresearch.org/blog/models-do-not-always-say-what-they-think/https://montanaresearch.org/blog/modernbert-encoder-side-finally-upgraded/https://montanaresearch.org/blog/molmo-pixmo-open-vlms-without-distilling/https://montanaresearch.org/blog/moltbook-social-network-for-agents/https://montanaresearch.org/blog/monitoringbench-monitor-only-as-good-as-attacks/https://montanaresearch.org/blog/mt-bench-and-the-80-percent-agreement-number/https://montanaresearch.org/blog/munk-debate-bengio-tegmark-versus-lecun-mitchell/https://montanaresearch.org/blog/muonclip-trillion-parameter-run-zero-loss-spikes/https://montanaresearch.org/blog/muse-spark-ten-times-less-compute-no-weights/https://montanaresearch.org/blog/naked-accuracy-is-marketing-cost-axis-leaderboards/https://montanaresearch.org/blog/native-sparse-attention-trainable-not-bolted-on/https://montanaresearch.org/blog/natural-language-autoencoders-reading-activations-as-english/https://montanaresearch.org/blog/needle-in-a-haystack-one-engineers-chart/https://montanaresearch.org/blog/nested-learning-optimizers-are-memories-too/https://montanaresearch.org/blog/neurips-2023-word2vec-test-of-time/https://montanaresearch.org/blog/neurips-2025-program-chairs-reviewing-at-20000/https://montanaresearch.org/blog/neurips-2025-takeaways-depth-gating-hivemind/https://montanaresearch.org/blog/neurips-desk-rejects-18-percent-position-papers/https://montanaresearch.org/blog/nine-questions-a-benchmark-should-answer/https://montanaresearch.org/blog/nobody-ready-for-agi-brundage-leaves-openai/https://montanaresearch.org/blog/nyt-v-openai-end-of-quiet-training-data/https://montanaresearch.org/blog/o1-scaling-test-time-compute/https://montanaresearch.org/blog/o1-system-card-hidden-chain-of-thought/https://montanaresearch.org/blog/o3-arc-agi-cost-per-task/https://montanaresearch.org/blog/o3-rewrote-its-own-shutdown-script/https://montanaresearch.org/blog/olmo-2-32b-fully-open-caught-gpt-4o-mini/https://montanaresearch.org/blog/olmo-3-model-flow-openness-as-pipeline/https://montanaresearch.org/blog/olmo-first-model-with-whole-recipe/https://montanaresearch.org/blog/once-ai-can-automate-ai-research/https://montanaresearch.org/blog/one-dollar-chevy-tahoe-customer-service-bots/https://montanaresearch.org/blog/one-in-five-iclr-reviews-was-ai/https://montanaresearch.org/blog/one-number-to-rule-them-all-epoch-capabilities-index/https://montanaresearch.org/blog/open-problems-mech-interp-thirty-researchers/https://montanaresearch.org/blog/open-problems-with-rlhf-thirty-two-authors/https://montanaresearch.org/blog/openai-and-anthropic-graded-each-others-models/https://montanaresearch.org/blog/openai-board-crisis-five-days-governance/https://montanaresearch.org/blog/openai-hugging-face-incident-evaluation-harness-attacks/https://montanaresearch.org/blog/othello-gpt-board-is-linear-after-all/https://montanaresearch.org/blog/pagedattention-kv-cache-was-the-bottleneck/https://montanaresearch.org/blog/paris-the-summit-that-dropped-safety/https://montanaresearch.org/blog/password-activated-shutdown-designing-the-off-switch/https://montanaresearch.org/blog/patchscopes-model-decodes-its-own-activations/https://montanaresearch.org/blog/pause-or-shut-it-all-down-two-march-letters/https://montanaresearch.org/blog/perplexity-stealth-crawlers-collapse-of-robots-txt/https://montanaresearch.org/blog/persona-vectors-and-the-behavioural-vaccine/https://montanaresearch.org/blog/petri-auditing-fourteen-models-agent-111-seeds/https://montanaresearch.org/blog/phi-3-and-the-model-that-runs-on-your-phone/https://montanaresearch.org/blog/pizza-topping-needle-claude-3-noticed-test/https://montanaresearch.org/blog/planbench-meets-o1-can-reasoning-model-plan/https://montanaresearch.org/blog/positive-case-for-faithfulness-simulatability-gain/https://montanaresearch.org/blog/positive-review-only-hidden-prompts-seventeen-preprints/https://montanaresearch.org/blog/predictable-scaling-gpt-4-report-chart/https://montanaresearch.org/blog/pressure-reveals-character-one-alignment-factor/https://montanaresearch.org/blog/pretraining-as-we-know-it-will-end/https://montanaresearch.org/blog/pretraining-on-the-test-set/https://montanaresearch.org/blog/pretraining-with-human-preferences-conditional-training/https://montanaresearch.org/blog/process-supervision-as-a-safety-method/https://montanaresearch.org/blog/project-glasswing-model-you-cannot-study/https://montanaresearch.org/blog/project-vend-claude-ran-a-shop-for-a-month/https://montanaresearch.org/blog/prompt-caching-changed-the-economics-of-long-system-prompts/https://montanaresearch.org/blog/pythia-model-suite-built-to-be-studied/https://montanaresearch.org/blog/pythia-publish-checkpoints-not-just-weights/https://montanaresearch.org/blog/qlora-fine-tuning-65b-model-one-gpu/https://montanaresearch.org/blog/qwen-3-8-and-the-overthinking-default/https://montanaresearch.org/blog/qwen3-apache-2-mirage-terms-of-use/https://montanaresearch.org/blog/qwen3-hybrid-thinking-switch/https://montanaresearch.org/blog/re-bench-agents-versus-human-experts/https://montanaresearch.org/blog/reading-the-adolescence-of-technology/https://montanaresearch.org/blog/reading-the-llama-4-community-license/https://montanaresearch.org/blog/reading-welcome-to-the-era-of-experience/https://montanaresearch.org/blog/reasoning-theater-probing-performative-chain-of-thought/https://montanaresearch.org/blog/red-teaming-silver-bullet-or-security-theatre/https://montanaresearch.org/blog/redpajama-reproducing-a-training-set/https://montanaresearch.org/blog/refinement-loops-arc-agi-2-four-to-54-percent/https://montanaresearch.org/blog/refusal-is-one-direction-circuit-breakers/https://montanaresearch.org/blog/regulators-and-courts-price-open-weights/https://montanaresearch.org/blog/remote-labor-index-says-2-5-percent/https://montanaresearch.org/blog/replays-over-scores-gpt-5-5-opus-4-7-arc-agi-3/https://montanaresearch.org/blog/replicating-circuit-tracing-known-mechanism/https://montanaresearch.org/blog/replit-database-deletion-why-the-agent-lied/https://montanaresearch.org/blog/representation-engineering-interpretability-top-down/https://montanaresearch.org/blog/reproducibility-becomes-official-neurips-track/https://montanaresearch.org/blog/reproducing-2200-icml-papers-nineteen-days/https://montanaresearch.org/blog/reversal-curse-knows-a-is-b-not-b-is-a/https://montanaresearch.org/blog/reward-hacking-production-rl-sabotage/https://montanaresearch.org/blog/right-to-warn-lab-employees-whistleblower-rules/https://montanaresearch.org/blog/rise-of-the-ai-engineer-three-years-later/https://montanaresearch.org/blog/rl-amplifies-emergent-misalignment-aesthetic-rewards/https://montanaresearch.org/blog/rlhf-vs-dpo-your-model-is-its-own-reward-model/https://montanaresearch.org/blog/rsp-v3-scaling-policy-cannot-go-it-alone/https://montanaresearch.org/blog/s1-a-thousand-examples-and-the-word-wait/https://montanaresearch.org/blog/sabotage-evaluations-and-rsp-2/https://montanaresearch.org/blog/sae-negative-results-deepmind-deprioritises/https://montanaresearch.org/blog/sae-probes-in-production-rakuten-pii/https://montanaresearch.org/blog/saebench-field-finally-gets-a-scoreboard/https://montanaresearch.org/blog/saes-can-steer-if-you-pick-the-feature/https://montanaresearch.org/blog/safe-model-unsafe-agent-clawsafety/https://montanaresearch.org/blog/samsung-chatgpt-leak-enterprise-ai-policy/https://montanaresearch.org/blog/sb-53-frontier-safety-framework-into-law/https://montanaresearch.org/blog/sb-1047-open-weights-question-never-resolved/https://montanaresearch.org/blog/sb-1047-vetoed-frontier-model-bill-split-safety/https://montanaresearch.org/blog/scaling-laws-for-precision-low-bit-floor/https://montanaresearch.org/blog/search-time-contamination-answer-key-on-hugging-face/https://montanaresearch.org/blog/seven-month-doubling-reading-metr-time-horizon/https://montanaresearch.org/blog/simpleqa-benchmark-models-were-meant-to-fail/https://montanaresearch.org/blog/situational-awareness-scored-two-years-later/https://montanaresearch.org/blog/six-months-to-live-for-open-models/https://montanaresearch.org/blog/six-ways-evaluation-is-harder-anthropic/https://montanaresearch.org/blog/sixteen-million-features-gpt-4-topk-sparse-autoencoders/https://montanaresearch.org/blog/skill-files-are-the-new-supply-chain-skill-inject/https://montanaresearch.org/blog/skills-instructions-in-a-folder-spread-faster-than-servers/https://montanaresearch.org/blog/sleeper-agents-safety-training-cannot-remove-what-it-cannot-see/https://montanaresearch.org/blog/small-language-models-future-of-agentic-ai-nvidia/https://montanaresearch.org/blog/smollm3-case-for-publishing-whole-recipe/https://montanaresearch.org/blog/software-3-0-notes-on-karpathy-yc-talk/https://montanaresearch.org/blog/solow-paradox-ai-edition/https://montanaresearch.org/blog/sonnet-45-and-the-end-of-naive-alignment-evals/https://montanaresearch.org/blog/sparks-of-agi-three-years-on/https://montanaresearch.org/blog/sparse-autoencoders-independent-replication/https://montanaresearch.org/blog/sparse-feature-circuits-shift-removing-spurious-feature/https://montanaresearch.org/blog/speculative-decoding-standard-dspark-mtp-drafters/https://montanaresearch.org/blog/spurious-rewards-paradox-anchor-adapter-circuit/https://montanaresearch.org/blog/spurious-rewards-random-feedback-improves-model/https://montanaresearch.org/blog/stack-overflow-sells-answers-moderators-revolt/https://montanaresearch.org/blog/stargate-500-billion-who-gets-to-train/https://montanaresearch.org/blog/state-of-open-models-summer-2026/https://montanaresearch.org/blog/state-space-model-recursive-reasoning-7m-test/https://montanaresearch.org/blog/stochastic-parameter-decomposition-splitting-weights-not-activations/https://montanaresearch.org/blog/style-control-arena-subtract-markdown-and-length/https://montanaresearch.org/blog/subliminal-learning-traits-through-number-sequences/https://montanaresearch.org/blog/superalignment-ten-months-later/https://montanaresearch.org/blog/sutton-says-llms-are-not-bitter-lesson-pilled/https://montanaresearch.org/blog/swarm-and-autogen-multi-agent-framework-question/https://montanaresearch.org/blog/swe-bench-at-launch-2294-issues/https://montanaresearch.org/blog/swe-bench-verified-retired/https://montanaresearch.org/blog/swe-lancer-prices-benchmark-in-dollars/https://montanaresearch.org/blog/sydney-dan-and-the-birth-of-prompt-injection/https://montanaresearch.org/blog/task-contamination-are-models-still-few-shot/https://montanaresearch.org/blog/techno-optimist-manifesto-and-the-enemies-list/https://montanaresearch.org/blog/ten-examples-undo-safety-training-fine-tuning-hole/https://montanaresearch.org/blog/terminal-bench-cheating-incidents/https://montanaresearch.org/blog/textbooks-are-all-you-need/https://montanaresearch.org/blog/the-95-percent-that-failed-mit-pilots-stanford-canaries/https://montanaresearch.org/blog/the-600b-question-read-from-the-cheap-seats/https://montanaresearch.org/blog/the-2026-ai-index-and-the-disappearing-academic-frontier/https://montanaresearch.org/blog/the-data-wall/https://montanaresearch.org/blog/the-day-the-qwen-team-walked-out/https://montanaresearch.org/blog/the-deepseek-moment-and-small-labs/https://montanaresearch.org/blog/the-gemini-demo-that-was-not-live/https://montanaresearch.org/blog/the-nobel-prizes-that-went-to-neural-networks/https://montanaresearch.org/blog/the-stack-v2-and-the-opt-out/https://montanaresearch.org/blog/theagentcompany-simulated-software-firm-benchmark/https://montanaresearch.org/blog/theory-of-mind-emerged-then-broken/https://montanaresearch.org/blog/thomson-reuters-v-ross-first-ai-training-ruling/https://montanaresearch.org/blog/three-bugs-and-a-batch-postmortem-nondeterminism/https://montanaresearch.org/blog/three-letters-in-five-days-open-weights-fight/https://montanaresearch.org/blog/titans-memory-that-learns-at-test-time/https://montanaresearch.org/blog/toolformer-and-react-taught-models-to-call-functions/https://montanaresearch.org/blog/top-10-percent-bar-exam-gpt-4-percentiles/https://montanaresearch.org/blog/towards-monosemanticity-superposition-becomes-a-tool/https://montanaresearch.org/blog/toy-model-mechanistic-unfaithfulness-transcoder/https://montanaresearch.org/blog/tracr-ground-truth-transformers/https://montanaresearch.org/blog/transcoders-making-mlps-legible-without-activations/https://montanaresearch.org/blog/tree-of-thoughts-return-of-search/https://montanaresearch.org/blog/tulu-3-opens-post-training-black-box/https://montanaresearch.org/blog/tuned-lens-reading-transformer-one-layer-at-a-time/https://montanaresearch.org/blog/turn-averaged-saes-fewer-features/https://montanaresearch.org/blog/twenty-two-words-extinction-statement/https://montanaresearch.org/blog/two-gold-medals-one-grader-imo-announcement-fight/https://montanaresearch.org/blog/unsupervised-knowledge-discovery-finds-wrong-knowledge/https://montanaresearch.org/blog/urgency-of-interpretability-five-year-head-start/https://montanaresearch.org/blog/vector-database-gold-rush/https://montanaresearch.org/blog/vending-bench-coherence-over-millions-of-tokens/https://montanaresearch.org/blog/vibe-coding-joke-tweet-became-word-of-the-year/https://montanaresearch.org/blog/vicuna-and-the-birth-of-gpt-4-as-judge/https://montanaresearch.org/blog/we-have-no-moat-rereading-leaked-google-memo/https://montanaresearch.org/blog/weak-to-strong-generalisation-gpt-2-supervise-gpt-4/https://montanaresearch.org/blog/weight-sparse-transformers-build-model-interpretable/https://montanaresearch.org/blog/what-open-source-ai-means-now/https://montanaresearch.org/blog/who-writes-the-textbook-lambert-post-training/https://montanaresearch.org/blog/why-llama-mmlu-score-depended-on-who-ran-it/https://montanaresearch.org/blog/wikimedia-counts-the-crawlers/https://montanaresearch.org/blog/wmdp-unlearning-measuring-hazardous-knowledge/https://montanaresearch.org/blog/would-a-model-sabotage-its-own-safety-research/https://montanaresearch.org/blog/zephyr-distilled-dpo-alignment-from-ai-feedback/https://montanaresearch.org/blog/zuckerberg-gigawatt-datacenters-compute-goes-public/https://montanaresearch.org/blog/zuckerberg-linux-analogy-open-weights-business-case/https://montanaresearch.org/contact/https://montanaresearch.org/publications/https://montanaresearch.org/publications/mrf-2026-01/https://montanaresearch.org/publications/mrf-2026-02/https://montanaresearch.org/publications/mrf-2026-03/https://montanaresearch.org/reports/https://montanaresearch.org/reports/mrf-r-2026-01/https://montanaresearch.org/reports/mrf-r-2026-02/https://montanaresearch.org/reports/mrf-r-2026-03/https://montanaresearch.org/research/