Spotlight
Spotlight is a neural network architecture proposed by Percepta that fully decouples a model's intelligence module from its memory. Through an arbitrarily sparse addressable memory mechanism, it enables memory to grow without bound while maintaining constant access cost, allowing models to acquire new capabilities without retraining.
Timeline (last 90 days)
Siri's cross-app message retrieval relies on on-device natural language processing and the Spotlight indexing mechanism, with message content indexed locally to be searchable.
Apple's first-party apps (Mail, Notes, Calendar, etc.) are deeply integrated with Spotlight, and their data can be indexed by Siri.
Google apps such as Gmail do not provide per-email content indexing to iOS Spotlight Search, preventing Siri from directly retrieving their content
Percepta claims that Spotlight is the first architecture to achieve unbounded memory growth without increasing access cost.
AI startup Percepta has publicly unveiled a new architecture called Spotlight, whose core claim is the separation of a model's intelligence from its memory, allowing knowledge and skills to grow continuously without changing the model weights.
Spotlight replaces the attention mechanism with a memory mechanism that escapes the trade-off between memory capacity and access cost.
In Spotlight, each token can read from and write to an unbounded memory; the model learns to index individual memory units, with each token touching only a very small portion of them at a time.
MoE models always activate a fixed number of experts from a fixed set, whereas Spotlight is arbitrarily sparse—no matter how large the memory grows, the number of units touched remains constant.
Spotlight splits the system into an intelligence module and memory: the intelligence module is fixed in size with weights that do not change as memory grows, while the memory stores knowledge, procedures, and working state.
In Spotlight, memory is writable—the model itself decides token by token what to load and when to overwrite—and memory can store not only facts but also skills.
12 more timeline events
All Facts (20)
Siri's cross-app message retrieval relies on on-device natural language processing and the Spotlight indexing mechanism, with message content indexed locally to be searchable.
50%UnverifiedApple's first-party apps (Mail, Notes, Calendar, etc.) are deeply integrated with Spotlight, and their data can be indexed by Siri.
50%UnverifiedGoogle apps such as Gmail do not provide per-email content indexing to iOS Spotlight Search, preventing Siri from directly retrieving their content
50%UnverifiedPercepta claims that Spotlight is the first architecture to achieve unbounded memory growth without increasing access cost.
50%UnverifiedSpotlight replaces the attention mechanism with a memory mechanism that escapes the trade-off between memory capacity and access cost.
50%UnverifiedIn Spotlight, each token can read from and write to an unbounded memory; the model learns to index individual memory units, with each token touching only a very small portion of them at a time.
50%UnverifiedMoE models always activate a fixed number of experts from a fixed set, whereas Spotlight is arbitrarily sparse—no matter how large the memory grows, the number of units touched remains constant.
50%UnverifiedSpotlight splits the system into an intelligence module and memory: the intelligence module is fixed in size with weights that do not change as memory grows, while the memory stores knowledge, procedures, and working state.
50%UnverifiedIn Spotlight, memory is writable—the model itself decides token by token what to load and when to overwrite—and memory can store not only facts but also skills.
50%UnverifiedPercepta asserts that models can acquire new capabilities without retraining, with capability expansion shifted to the growable memory.
50%UnverifiedSpotlight's memory approach is architecturally close in philosophy to early explorations of Neural Turing Machines (NTM) and Memory-Augmented Neural Networks (MANN).
50%UnverifiedAll current information about Spotlight comes from a single-source description on Percepta's own blog, with no independent benchmarks, peer review, or reproducible experimental data to support it.
50%UnverifiedmacOS's Spotlight and iOS's Photos smart albums are fundamentally based on the index-mapping principle, where search and classification results do not reorganize the original files.
50%UnverifiedRaycast is a macOS launcher similar to Spotlight but with a plugin ecosystem; its built-in clipboard history feature comes at virtually no additional cost for Raycast Pro subscribers.
50%UnverifiedSnap launched Spotlight, a short-video feed feature similar to TikTok, in 2020, and in its early days established a creator fund of $1 million per day.
50%UnverifiedBeOS's BFS file system has a built-in database-style metadata indexing feature using a B+ tree index structure, predating macOS Spotlight and Windows desktop search by nearly a decade.
50%UnverifiedEinfall captures inspiration through multiple entry points including the menu bar, share menu, Shortcuts, widgets, Spotlight search, and Siri voice.
50%UnverifiedEinfall supports capturing ideas on Mac, iPhone, and iPad via multiple methods including the menu bar, share menu, Shortcuts, widgets, Spotlight search, and Siri voice.
50%UnverifiedRaycast于2020年出现，作为第三方替代品将Spotlight范式从搜索扩展到执行，支持通过插件系统运行脚本、管理窗口、查询API等
50%UnverifiedRaycast于2020年出现，作为Spotlight的第三方替代品，将全局快捷入口范式从搜索扩展到执行
50%