EXPLORE / 06
Terms index
Specialized AI and software-development terms with concise explanations
RECORDS
Entity records
3360 records · 27/84
- 1041
Relative Positional Encoding
A mechanism calculating attention weights based on relative distance between tokens rather than absolute positions
Website pending - 1042
Semantic Routing
Technology distributing requests based on semantic understanding of input content rather than keyword matching
Website pending - 1043
Mixture-of-Models
A systems approach coordinating multiple specialized models to serve different requests or environments
Website pending - 1044
Regression Testing
Testing method verifying that new code changes do not break existing functionality or performance
Website pending - 1045
Feed-Forward Network (FFN)
Fully connected layer structure in Transformers used for non-linear feature transformation
Website pending - 1046
Parallel Drafting
A strategy generating multiple candidate tokens simultaneously rather than sequentially via autoregression.
Website pending - 1047
SVE256
An implementation of Arm's Scalable Vector Extension instruction set for parallel data processing.
Website pending - 1048
INT8 Quantization
Converting model weights and activations from floating-point to 8-bit integers to accelerate inference.
Website pending - 1049
Day-0 Support
The capability to run a new model seamlessly on specific frameworks or tools on its release day
Website pending - 1050
AllGather
Collective communication primitive gathering data from all nodes and broadcasting to each
Website pending - 1051
mmap
Maps files directly into process virtual address space to avoid redundant copying
Website pending - 1052
Diffusion RL
Technique using reinforcement learning algorithms to optimize diffusion model generation quality
Website pending - 1053
Confidence Head
An additional module in neural networks that outputs reliability scores for predictions.
Website pending - 1054
Non-associativity of Floating Point
The property where changing the order of floating-point additions yields slightly different results due to rounding errors.
Website pending - 1055
Sharded Transfer
Technology that splits large datasets or model weights into smaller chunks for parallel transmission
Website pending - 1056
Online RL
A learning paradigm where agents update policies simultaneously while interacting with the environment
Website pending - 1057
LLM Serving
The technical system for deploying trained large models as online API services.
Website pending - 1058
Prefill/Decode Disaggregation
Separates compute-intensive prefill stages from memory-intensive decode stages.
Website pending - 1059
Preemption
Interrupting current requests to release resources for later reprocessing.
Website pending - 1060
Disaggregated Serving
Architecture separating prefill and decode phases onto different hardware resources to boost throughput
Website pending - 1061
Mixed Precision Quantization
Using different numerical bit-widths across layers or components to balance accuracy and speed.
Website pending - 1062
Bitwise Consistency
Different compute engines producing identical binary outputs for the same inputs.
Website pending - 1063
RL Post-training
The phase of using reinforcement learning to optimize alignment or reasoning after pretraining.
Website pending - 1064
CUDA Operator
High-performance parallel computing kernel functions written based on the NVIDIA CUDA platform.
Website pending - 1065
DSpark
An enhanced block-level speculative decoding algorithm proposed by DeepSeek, aiming to improve token coherence and acceptance length.
Website pending - 1066
Hardware Decoding
Technology using dedicated chip circuits rather than general-purpose CPUs to decompress video streams.
Website pending - 1067
PD Disaggregation
An architecture that deploys prefill and decode phases of LLM inference on separate compute resources to optimize throughput and latency.
Website pending - 1068
Time to First Token (TTFT)
Latency from request initiation to first token generation, measuring responsiveness
Website pending - 1069
Paged Attention
Manages KV cache dynamically using virtual memory principles to eliminate fragmentation
Website pending - 1070
Output Distribution
The set of probability predictions for the next token, reflecting language modeling capability
Website pending - 1071
Adaptive Thinking
A technique where the model dynamically adjusts internal reasoning depth or computational resource allocation based on task complexity.
Website pending - 1072
Runtime Configuration
The ability to dynamically adjust parameters or components during program execution rather than fixing them at compile time.
Website pending - 1073
CUDA Graphs
Technology capturing GPU operation sequences as static graphs to reduce launch overhead
Website pending - 1074
Hot Reloading
Dynamically updating code or configuration without restarting services or interrupting sessions
Website pending - 1075
Agent Trajectory
Sequence of states, actions, and feedback recorded during an agent's task execution
Website pending - 1076
Capability Commoditization
The process where once-scarce advanced AI functions become low-cost generic services as technology spreads
Website pending - 1077
Marginal Value of Intelligence
The additional economic benefit generated by each incremental unit of frontier AI capability
Website pending - 1078
NAND Flash
Non-volatile storage technology widely used in SSDs and mobile devices, characterized by high density.
Website pending - 1079
DRAM
Dynamic Random Access Memory, the primary form of volatile computer main memory requiring periodic refresh.
Website pending - 1080
Subliminal Learning
Behavioral changes induced by implicit signals or indirect data transmission
Website pending