What Comes Next - Predictions & the Horizon Map
Everything in this chapter is 🟥 speculation, including the authors’ own.
Plain words and small, concrete numbers. - Step up or down anytime with keyboard shortcuts bracket keys or the buttons above.
The authors made five predictions. This chapter presents them, extends them into a horizon map from now to 2035, and - because this entire chapter is speculation - every prediction arrives with what would falsify it.
Benchmark + harness released
Standardized evaluation suite with author-blind tasks, multiple seeds, and baseline controls.
Replications on other LLMs
Independent verification of weight-reading and editing on open-weight architectures (LLaMA-3.1, Mistral, Gemma).
Weight-diff patches shared like PRs
Small surgical weight-delta files (~1MB) distributed as signed patches against canonical model checkpoints.
Zero-data unlearning trials
Removing hazardous or copyrighted capabilities by projecting out targeted directional vectors without data or training.
LLM as circuit auditor
Automated mechanistic interpretability: LLMs scan parameter matrices to localize backdoors, induction heads, and policy bugs.
One-shot initialisation for small nets
Pre-computing functional policy and transformer weights from natural language specifications to bypass cold-start training.
Evaluation-free auditing
Certifying model safety invariants and compliance from static weight tensors alone, without dynamic test runs.
Cross-architecture grafting
Transplanting functional computational circuits between disparate architectures (e.g. from a Dense Transformer to an MoE or RNN).
WSP agents in training pipelines
Autonomous supervisor agents inspecting checkpoints and executing surgical weight modifications during training runs.
Models program their successors
Frontier models authoring the architectures, initial weight distributions, and surgical patches for next-generation systems.
Weights as reviewable source code
Neural network weights managed with git-like version control, unit tests, linters, and human-readable semantic diffs.
Zero-data unlearning trials
Technical Implementation Mechanism
Closed-form projection W' = (I - uuᵀ)W on identified task subspaces, analogous to the Exp-8 math surgery.
Empirical Falsifier (What Would Prove It Wrong)
Safety evaluations reveal high collateral capability regression across unrelated benchmark domains.
The Authors’ Five Predictions (🟥 Speculation)
- Auditing without evaluationSpotting backdoors or hidden behaviours from the weights alone, without running the model on a benchmark suite.
- Dramatically cheaper optimizationOne-shot initialisation: an LLM writes a starting point so good that training cost collapses.
- Zero-data capability steering & unlearningAdding, steering or removing behaviours by editing weights directly - no data collection, no training loop.
- Recursive self-improvementModels editing their own or their successors’ weights, closing the loop between designer and artifact.
- True architectural mergingGrafting a circuit between different architectures - not just averaging checkpoints of the same skeleton.
Empirical Grounding: What Falsifies The Map
- 1. Mid-2027 Benchmark Deadline:If no independent third-party author-blind benchmark ships by mid-2027, the entire ecosystem band stalls.
- 2. Open Model Replication Failure:If weight-reading fails on LLaMA-3.1 or Mistral under identical prompting, universal generalization is disproven.
- 3. Direction-Ablation Baseline Equivalence:If traditional activation steering matches every surgical fix result, WSP’s unique value narrows strictly to Write.
Horizon bands catalogued: Now → 6 months · 6 → 24 months · 2 → 5 years · 5+ years
Timeline: Year 2028. Generalization Active: Universal weight-reading allows all research tracks to unfold. (🟥 Speculation - all milestones carry testable empirical falsifiers).
L2 view: Near-term items (benchmarks, replications, signed patches) are viable today. Far-future items (evaluation-free auditing, models programming successors) hinge on cross-model generalisation.
🟥 SPECULATION - Based on MIT CSAIL predictions and extended horizon mapping. Positions are qualitative and paired with empirical falsifiers.
- Auditing without evaluation - Spotting backdoors or hidden behaviours from the weights alone, without running the model on a benchmark suite.
- Dramatically cheaper optimization - One-shot initialisation: an LLM writes a starting point so good that training cost collapses.
- Zero-data capability steering & unlearning - Adding, steering or removing behaviours by editing weights directly - no data collection, no training loop.
- Recursive self-improvement - Models editing their own or their successors’ weights, closing the loop between designer and artifact.
- True architectural merging - Grafting a circuit between different architectures - not just averaging checkpoints of the same skeleton.
The authors’ five predictions (🟥). ① Auditing without evaluation - e.g. spotting backdoors from weights alone. ② Dramatically cheaper optimisation - one-shot initialisation. ③ Zero-data capability steering and unlearning. ④ Recursive self-improvement - models editing successors. ⑤ “True architectural merging” - grafting a circuit between different architectures. All untested in the post.
The near horizon (now–6 months, still 🟥 but checkable): the promised benchmark + harness (“currently being built”); replications on other LLMs; shrinking context and tooling limits; the first shared weight-diff patches. The middle horizon (6–24 months): signed patch ecosystems - “git for weights”; zero-data unlearning trials; LLM-assisted interpretability as a research tool; one-shot initialisation for small specialist nets and edge devices. The far horizon (2–5 years): evaluation-free auditing if reading generalises; cross-architecture grafting; WSP agents inside training pipelines; models programming successors. The very far edge (5+ years): weights treated as reviewable source code - humans reviewing, versioning, formally verifying “compiled” behaviours.
Notice the load-bearing assumption separating near from far: does reading generalise beyond one frontier model and beyond author-designed tasks? Every far-future item quietly assumes yes. The futures slider (M13) lets you flip that switch and watch the far side dim.
Foundational Refreshers (1)