Convergence
UPDATED · AUG 06 2026Convergence points where a result on one layer of the stack changes what is true on another. The Atlas is a chain from a token price to a task price, so the connections that matter are the ones that move a number across a boundary — not thematic resemblances. **This map is seeded, not complete.** Every connection below is grounded in a source read in full and written up as a close read in `raw/`; the citation is named in each row. Connections that would require asserting what a concept page says have been left out rather than guessed. Add to it from close reads, not from plausibility — the previous version of this file was a table of confident-sounding pairings that survived a topic rename as dead links.
6
Domain pairs
6
Connections
12
Live cross-links
Harnesses ↔ Inference economics
| Harness concept | Cost concept | Connection |
|---|---|---|
| Harness Effect Measurement | Cost per Task | The harness decides how many tokens a finished task costs, including retries and runs that fail — so a token price is not a task price. Weng (2026) reports Lin et al. separating two capabilities usually conflated: harness-updating capability is flat from Qwen2-32B to Opus 4.6, while harness benefit is non-monotonic, with middle-tier models gaining most. If that holds, the harness is not a uniform multiplier on model quality, and "buy a better model" and "build a better harness" are not substitutes at every tier. Source: `harnesses/raw/harness-engineering-for-self-improvement.md`. |
Harnesses ↔ Serving
| Harness concept | Serving concept | Connection |
|---|---|---|
| Agent Memory Architectures | Conditional Memory | Memory is a context problem, not a weights problem — which makes it the harness's job and the serving layer's cost. O'Neill (2026): facts written to weights are question-keyed and do not survive later writes — after twenty sequential writes, bare-statement facts retain 1% accuracy against 46% for broad-study data, and 70% of wrong answers contain the most recently written fact. The model is being redirected, not forgetting. A forgotten fact re-supplied in the prompt recovers to 77–80%. Source: `harnesses/raw/can-a-language-model-learn-facts-continually-in-its-weights.md`. |
Serving ↔ Inference economics
| Serving concept | Cost concept | Connection |
|---|---|---|
| Prefill/Decode Disaggregation | Cost Measurement Problem | Every published cost-per-token is a point someone chose on the latency/throughput curve. Larger batches raise aggregate TPS by keeping the GPU busy but increase queueing time, TTFT and per-user TPOT; small batches invert it. Without the batch size and the SLO, two vendors' cost figures are not comparable. Goodput — requests per second meeting defined SLOs — is the metric that stops a cost number flattering a system that is fast and unusable. Source: `inference-economics/raw/modular-llm-inference-handbook.md`. |
Datacenters ↔ Inference economics
| Datacenter concept | Cost concept | Connection |
|---|---|---|
| Rack Power Density | Energy Cost of Tokens | The margin sits in the gap between selling electricity and selling compute. Somani's implied-price anchors: Anthropic's lease from TeraWulf — $19B, 400 MW, 20 years from H2 2027 — implies ~$271/MWh with no GPUs (building and cooling included), against a SpaceX→Reflection arrangement implying ~$5000/MWh for power that comes with GPUs ready to go. ⚠️ Both are the author's own derivations, one labelled "napkin math" — use for order of magnitude, never as market prints. Source: `datacenters/raw/power-2026-electricity-pricing-primer.md`. |
Energy ↔ Datacenters
| Energy concept | Datacenter concept | Connection |
|---|---|---|
| Power: The Binding Constraint | Hyperscaler Capex Cycle | Capex can be committed far faster than power can be interconnected, and the queue is the gate. The interconnection queue exists because the ISO must model each generator's grid impact before it connects. Behind-the-meter generation can "simplify or avoid the generator interconnection process" — but buys that with siting, cooling and the batteries needed to balance generation against consumption. Demand context: datacenters are already ~5% of US power consumption (DOE) and demand is doubling roughly every two years (Goldman Sachs). Source: `datacenters/raw/power-2026-electricity-pricing-primer.md`. |
Evals ↔ Harnesses
| Evals concept | Harness concept | Connection |
|---|---|---|
| Agent Evaluation Benchmarks | Harness Effect Measurement | A benchmark score is not attributable to a model unless the harness is disclosed, and the ranking can invert with the time budget. On RE-Bench, human experts scored non-zero in 82% of 8-hour attempts and 24% matched or beat strong reference solutions; agents scored 4× higher than humans at a 2-hour budget, but humans had better returns to longer budgets and exceeded agents at 8 and 32 hours. Related caution: Meta-Harness's reported gains came from a search initialised from two already-strong harnesses, which is not evidence that search finds strong harnesses from nothing. Source: `harnesses/raw/harness-engineering-for-self-improvement.md`. |
Method
Convergence points are where a breakthrough in one domain directly enables progress in another. Linked concepts open their Knowledge Base pages; the rest are tracked but not yet written up. Back to the Atlas →