Reference: Conceptual backdrop for the MegaPrediction Training PRD. Use that PRD for implementation; this page highlights ideas that reach beyond the current POR.
MegaPrediction adds a speculative region to the MegaContext Tree so the system can plan future context, refine it, and commit when ready—now leveraging the full wLOD working-context tree surfaced by MegaAttention. Instead of bolting on standalone heads, we reuse the final hidden state from the most recent node at each wLODn to forecast tokens, block gists, and ultra-long summaries in one pass.
- Present cursor: separates committed history from speculative spans.
- MegaAttention readouts: reuse the shared LM head on the most recent wLODn hidden states (see MegaPrediction Training).
- Latent planners: operate in gist space before reifying tokens.
- Hierarchical refinement: progressively decode from LOD2 gists to LOD0 tokens.
- LensNet role: scores speculative spans to decide refinement order.
- Training: reuse ΔNLL losses, explore RL for compute/latency trade-offs.
MegaAttention-Driven Readouts
- The mixed-LOD batch fed into MegaAttention already carries wLOD0→wLOD2 nodes. MegaPrediction simply taps the latest node per level, applies the shared LM head, and interprets outputs at the appropriate time scale (token, 32-token gist, 1024-token summary).
- No extra projection heads are required; supervision flows through MegaPrediction Training by pairing each readout with ground-truth gists produced by GistNet over the teacher-forced future horizon.
- Consistency across iterative edits relies on Hierarchical KV Caching Strategy, so speculative spans never attend to stale summaries.
Beyond the PRD: ideas to incubate
The PRD focuses on wiring speculative readouts into the runtime. Once that lands, we can push further in a few directions:
- Speculative workspaces: treat the “future” partition as a collaborative scratchpad with fork/merge semantics so agents can iterate safely before committing to history.
- Richer DeGistNet decoders: experiment with diffusion-style refinement (LOD2 → LOD1 → LOD0) so we can explore multiple drafts cheaply before materializing tokens.
- Hybrid planners: blend latent planning with selective refinement—LensNet scores every speculative span, but only the high-utility regions get de-gisted.
- Policy learning: go beyond greedy editing with RL/bandit objectives (accuracy × latency × compute), turning MegaPrediction into a tunable control policy rather than a fixed loop.
- Telemetry feedback: log speculative-span outcomes (accepted vs. discarded) and reuse that signal to teach LensNet/Focus Allocator which edits are worth exploring.
Capture new concepts here first; once they graduate into deliverables, open or extend a PRD so the POR remains the single source of truth.