Token Labhttps://tokenlab.fyi/en/AI learning articles and research tutorialsIndustry AI briefing: monitoring actions and specifying behavior | Token Labhttps://tokenlab.fyi/en/company-frontier/2026-10-30/https://tokenlab.fyi/en/company-frontier/2026-10-30/Fri, 30 Oct 2026 00:00:00 GMTSelf-improving agents must preserve authorization | Token Labhttps://tokenlab.fyi/en/frontier/authorization-after-learning/https://tokenlab.fyi/en/frontier/authorization-after-learning/Fri, 30 Oct 2026 00:00:00 GMTIndustry AI briefing: restricted models, enterprise actions and local speech | Token Labhttps://tokenlab.fyi/en/company-frontier/2026-10-28/https://tokenlab.fyi/en/company-frontier/2026-10-28/Wed, 28 Oct 2026 00:00:00 GMTSelf-improving agents: reward gains are not capability gains | Token Labhttps://tokenlab.fyi/en/frontier/adaptive-agent-evaluation/https://tokenlab.fyi/en/frontier/adaptive-agent-evaluation/Mon, 26 Oct 2026 00:00:00 GMTRobot recovery: when retrying helps, and when it hides failure | Token Labhttps://tokenlab.fyi/en/frontier/robot-recovery/https://tokenlab.fyi/en/frontier/robot-recovery/Mon, 26 Oct 2026 00:00:00 GMTAdam: persistent moments and bias correction | Token Labhttps://tokenlab.fyi/en/interview-code/adam/https://tokenlab.fyi/en/interview-code/adam/Mon, 26 Oct 2026 00:00:00 GMTAttention and Multi-Head Attention: write the shapes first | Token Labhttps://tokenlab.fyi/en/interview-code/attention/https://tokenlab.fyi/en/interview-code/attention/Mon, 26 Oct 2026 00:00:00 GMTBeam Search: global pruning, EOS and sequence scores | Token Labhttps://tokenlab.fyi/en/interview-code/beam-search/https://tokenlab.fyi/en/interview-code/beam-search/Mon, 26 Oct 2026 00:00:00 GMTK-Means: assignment, centroid updates and empty clusters | Token Labhttps://tokenlab.fyi/en/interview-code/kmeans/https://tokenlab.fyi/en/interview-code/kmeans/Mon, 26 Oct 2026 00:00:00 GMTSoftmax: stable probabilities and gradients | Token Labhttps://tokenlab.fyi/en/interview-code/softmax/https://tokenlab.fyi/en/interview-code/softmax/Mon, 26 Oct 2026 00:00:00 GMTTool-using assistants: validate, execute and trace | Token Labhttps://tokenlab.fyi/en/applications/bounded-tools/https://tokenlab.fyi/en/applications/bounded-tools/Sun, 25 Oct 2026 00:00:00 GMTFile Q&A: build an evidence-first document assistant | Token Labhttps://tokenlab.fyi/en/applications/document-evidence/https://tokenlab.fyi/en/applications/document-evidence/Sun, 25 Oct 2026 00:00:00 GMTDocument extraction: JSON shape is not factual correctness | Token Labhttps://tokenlab.fyi/en/applications/validated-extraction/https://tokenlab.fyi/en/applications/validated-extraction/Sun, 25 Oct 2026 00:00:00 GMTFrom LLM Calls to Agent Harnesses | Token Labhttps://tokenlab.fyi/en/frontier/agent-harness-engineering/https://tokenlab.fyi/en/frontier/agent-harness-engineering/Sun, 25 Oct 2026 00:00:00 GMTDiffusion language models: masking and iterative decoding | Token Labhttps://tokenlab.fyi/en/frontier/diffusion-language/https://tokenlab.fyi/en/frontier/diffusion-language/Sun, 25 Oct 2026 00:00:00 GMTSelective state spaces: Mamba and sequence memory | Token Labhttps://tokenlab.fyi/en/frontier/selective-state-spaces/https://tokenlab.fyi/en/frontier/selective-state-spaces/Sun, 25 Oct 2026 00:00:00 GMTTest-time compute: search, verification and budgets | Token Labhttps://tokenlab.fyi/en/frontier/test-time-compute/https://tokenlab.fyi/en/frontier/test-time-compute/Sun, 25 Oct 2026 00:00:00 GMTAdam: understand the update, not just the defaults | Token Labhttps://tokenlab.fyi/en/papers/adam/https://tokenlab.fyi/en/papers/adam/Sun, 25 Oct 2026 00:00:00 GMTAdamW: where you apply weight decay matters | Token Labhttps://tokenlab.fyi/en/papers/adamw/https://tokenlab.fyi/en/papers/adamw/Sun, 25 Oct 2026 00:00:00 GMTCalibration: being right is different from knowing how often | Token Labhttps://tokenlab.fyi/en/papers/confidence-calibration/https://tokenlab.fyi/en/papers/confidence-calibration/Sun, 25 Oct 2026 00:00:00 GMTCPC: learn by identifying the right future | Token Labhttps://tokenlab.fyi/en/papers/contrastive-predictive-coding/https://tokenlab.fyi/en/papers/contrastive-predictive-coding/Sun, 25 Oct 2026 00:00:00 GMTDeep RL that Matters: a score is a distribution | Token Labhttps://tokenlab.fyi/en/papers/deep-rl-reproducibility/https://tokenlab.fyi/en/papers/deep-rl-reproducibility/Sun, 25 Oct 2026 00:00:00 GMTDeepSeek-R1: rewards, reasoning and what the evidence establishes | Token Labhttps://tokenlab.fyi/en/papers/deepseek-r1/https://tokenlab.fyi/en/papers/deepseek-r1/Sun, 25 Oct 2026 00:00:00 GMTDPO: preference learning through relative log probabilities | Token Labhttps://tokenlab.fyi/en/papers/dpo/https://tokenlab.fyi/en/papers/dpo/Sun, 25 Oct 2026 00:00:00 GMTFlashAttention: optimize data movement before counting FLOPs | Token Labhttps://tokenlab.fyi/en/papers/flashattention/https://tokenlab.fyi/en/papers/flashattention/Sun, 25 Oct 2026 00:00:00 GMTLLM-as-a-judge: evaluate the evaluator | Token Labhttps://tokenlab.fyi/en/papers/llm-as-a-judge/https://tokenlab.fyi/en/papers/llm-as-a-judge/Sun, 25 Oct 2026 00:00:00 GMTLoRA: constrain the update, not the whole model | Token Labhttps://tokenlab.fyi/en/papers/lora/https://tokenlab.fyi/en/papers/lora/Sun, 25 Oct 2026 00:00:00 GMTPPO: inspect the clipped objective one sample at a time | Token Labhttps://tokenlab.fyi/en/papers/ppo/https://tokenlab.fyi/en/papers/ppo/Sun, 25 Oct 2026 00:00:00 GMTQLoRA: a memory budget is more than weight precision | Token Labhttps://tokenlab.fyi/en/papers/qlora/https://tokenlab.fyi/en/papers/qlora/Sun, 25 Oct 2026 00:00:00 GMTReAct: close the loop between actions and observations | Token Labhttps://tokenlab.fyi/en/papers/react/https://tokenlab.fyi/en/papers/react/Sun, 25 Oct 2026 00:00:00 GMT