Ingest 2026-08-19: memory +500%; TerraPower nuclear; routing-as-standard; CoT unfaithfulness; OpenAI revokes cyber access

concepts/ai-energy-efficiency.md

@@ -56,6 +56,13 @@ A candidate physical-neural-network circuit must be:
 
 Joules-per-token is the *demand-side* efficiency metric; the *supply-side* cost is now its own inflection. Per TechCrunch ([[dailybrief-roundup-2026-08-14]]), a new forecast has **US natural-gas prices potentially tripling** — a direct **infra cost shock** for hyperscalers who **bet on natural gas** to power the data-center buildout. This is the same *energy-is-the-binding-constraint* thesis the wiki tracks from the other direction: [[dwarkesh-patel|Dwarkesh's]] compute-repricing ([[ai-margin-collapse]]) and Chamath's **"LPS" (Land, Power, Shell)** bet ([[saas-disruption-thesis]]) both say *energized power* is the scarce, ownable layer — and a gas-price spike is exactly the risk that makes it scarce. If it holds, it pushes the hosted-vs-local calculus toward on-prem/efficient inference (the [[gregisenberg-fable-5-ban-local-models-pivot-2026-06-13|local-models-as-insurance]] + open-weights thread). *(Forecast-dependent; directional.)*
 
+## Supply-side cost shocks (2026-08-19): memory +500%, nuclear power
+
+Two supply-side signals landed together, both hitting the *cost* side of the metric rather than the Joules-per-token *efficiency* side ([[dailybrief-roundup-2026-08-19]]):
+
+- **Memory prices up ~500% in 12 months** (Latent Space AINews) — framed as *"Moore's Law reversed to 2007 levels."* This is the acute-price-spike escalation of the memory-as-binding-constraint thesis: the [[dailybrief-roundup-2026-05-24|May signal]] was *cost-share* (memory ~67% of AI-chip BOM per Epoch AI, HBM cannibalizing DRAM/NAND fabs); this is *price*. It sharpens the **quadruple-convergence on memory-as-binding-constraint** — McMahon (energy-side), the KV-cache research wave (research-side), Epoch AI 67% BOM (cost-share-side), and now a 500% price spike (market-price-side). Direct input to any [[ai-margin-collapse|inference-margin]] model. *(AINews summary; primary not fetched.)*
+- **TerraPower nuclear reactor targets AI data centers** (TechCrunch) — a structural **power-supply** answer for compute density, the complement to the **natural-gas price-spike** (the grid-cost section below, 2026-08-14): if energized power is the scarce, ownable layer (Chamath's **LPS / "Land, Power, Shell"**, [[saas-disruption-thesis]]), nuclear is one route to securing it. Reinforces the *energy-is-the-binding-constraint* read from the supply direction. *(TechCrunch; deployment-timeline unverified.)*
+
 ## Related Concepts
 
 - [[mcp]] — protocol-level optimization layer; orthogonal but interacts with KV-cache cost

concepts/ai-margin-collapse.md

@@ -23,6 +23,7 @@ It's the unit-economics lens for evaluating any AI-applied company you'd join or
 - **Value-shifts-to-the-decision-layer, priced (2026-08-17)** ([[dailybrief-roundup-2026-08-17]]): the margin thesis's *"value moves to the integration/product layer"* corollary got a **$7B price tag** — [[stripe|Stripe acquired OpenRouter]] (the multi-model routing abstraction). Brief's read: *"the real margin in AI infra isn't owning GPUs — it's owning the decision layer that gets between the app and the model."* Pairs with [[anthropic|Anthropic's $65B annualized revenue]] the same day (*"the bet isn't on better weights anymore — it's on who owns the deployment layer"*): both the applied-infra M&A and the frontier-lab revenue curve point at **distribution/decision** rather than raw model quality as the durable edge.
 - **Cost-capability retrospective (2026-08-17)**: *"From BERT to Frontier Agents: Eight Years of LM Progress"* (arXiv 2608.13675) reports **~6× annual coding-ability improvement since late 2024** and **GPT-5.6 Luna at flagship performance for $1–6/M tokens** — a concrete anchor for the falling-price side of the curve, alongside [[qwen|Qwen 3.8 27B]] matching GPT-5.6 Luna's AA Index score at laptop scale. *(arXiv retrospective; not independently reproduced.)*
 - **Inference-as-commodity-infrastructure, from the hardware floor (2026-08-18)** ([[dailybrief-roundup-2026-08-18]]): *"DumpsterCluster"* (arXiv 2608.14614) serves **LLaMA-70B on a 128-GPU cluster built from datacenter-reject silicon at ~$60/GPU** — evidence that inference is drifting toward a commodity-infra problem rather than a model problem. The brief's read completes the thesis from the supply side: *"the margin game shifts from 'can we run it' to 'who owns the DC real estate and power contracts.'"* This routes into the **LPS / "Land, Power, Shell"** land-power bet ([[saas-disruption-thesis]]) and the [[ai-energy-efficiency|energy-as-binding-constraint]] thread — if secondhand silicon can serve 70B, the durable scarcity is *energized power + real estate*, not GPUs. *(arXiv; scale/throughput reproducibility unverified.)*
+- **Model-routing goes mainstream as cost control (2026-08-19)** ([[dailybrief-roundup-2026-08-19]], Glean CEO via Latent Space): *"frontier model cost + open-weights popularity is driving demand for model routing."* An **enterprise-buyer-side corroboration** of the [[stripe|Stripe/OpenRouter $7B]] bet — routing is shifting from a power-user trick to **default B2B cost architecture** (the dial between Claude/GPT/Grok/open-weights). Confirms the *value-moves-to-the-decision-layer* corollary from the demand side, not just the M&A side. *(Vendor-CEO framing.)*
 - Track: independent GLM-vs-Opus benchmarks; whether frontier labs cut inference prices in response; open-weights adoption in production; whether compute spot-prices climb toward Dwarkesh's labor-anchored equilibrium; whether the routing/decision layer ([[stripe|Stripe/OpenRouter]]) captures the margin the model layer loses.
 
 ## Key Papers / Posts

concepts/frontier-ai-governance.md

@@ -47,6 +47,7 @@ Hassabis's lab-side blueprint pairs and contrasts with prior moves:
 | **U.S. DOE — "Genesis Open Models Initiative"** (energy.gov / ANL, 2026-08-08, [[dailybrief-roundup-2026-08-08]]) | A **government-side open-models program** (DOE / Argonne). Scope still ambiguous (funding vehicle vs framework), but a state-actor entering the open-weights arena on the *pro-open* side — a counterweight to the [[us-treasury-china-ai-sanctions-threat-2026-07-21\|Treasury restriction lever]] from *within* the US government, and the public-sector complement to the industry [[open-weights-american-ai-leadership-coalition-2026-07-24\|coalition letter]]. *(DOE credibility suggests substance; details pending.)* |
 | **[[dwarkesh-patel\|Dwarkesh]] — "locking in AI safety regulation now is premature"** (2026-08-08, [[dailybrief-roundup-2026-08-08]]) | From his *"Era of Continual Learning"* predictions: governance timelines are **misaligned with capability drift** — regulation frozen against today's static-model assumptions ages badly once models learn continually. The *"don't lock in prematurely"* argument sits opposite Hassabis's *"build the infrastructure in the precious window"* — the live tension over *when* to regulate, not just how. |
 | **OpenAI — "frontier cyber models in more trusted hands" (Daybreak partner program)** (openai.com, 2026-08-10, [[dailybrief-roundup-2026-08-12]]) | The **constructive complement to the Astra slowdown**: rather than only braking, OpenAI ships frontier cyber capability through a **restricted approved-partner program** (authorized cybersecurity service delivery). **Governance-by-access-control** — a middle path between "release broadly" and "don't release," and a concrete answer to the "safety test is a safety risk" containment problem ([[reward-hacking]]). *(Vendor program; gate-effectiveness unproven.)* |
+| **OpenAI revokes researchers' access to its limited cyber program** (TechCrunch, 2026-08-19, [[dailybrief-roundup-2026-08-19]]) | The **revocation flip-side** of the Daybreak "trusted-hands" program above: governance-by-access-control cuts both ways — the same gate that admits approved partners can **cut researchers off**, and researchers publicly complained. Surfaces the unresolved **who-decides + researcher-trust** problem inside any access-gated regime (and the incumbent-capture risk the pattern-watch below flags): a lab-controlled gate is only as legitimate as its appeals process. *(Researcher complaints via TechCrunch; OpenAI's rationale not captured.)* |
 | **Hinton, Fei-Fei Li & Andrew Ng — "make the case for staying open" (Ai4, 2026-08-12)** ([[dailybrief-roundup-2026-08-12]], TechCrunch) | Three canonical figures publicly backing **open-source access + competition** (incl. vs China) at the Ai4 conference — a heavyweight-researcher counterweight on the *pro-open* side, distinct from the industry-coalition + government (DOE) pro-open voices already in this table. Debate-format (no concrete outcome), but the *researcher-authority* endorsement is the new element. **Meta's [[muse-glimmer\|Muse Glimmer]] (Apache-2.0, same cycle)** is the product-side of the same push. |
 | **OpenAI funds 14 independent policy-research projects** ("New policy ideas for the Intelligence Age," openai.com, 2026-08, [[dailybrief-roundup-2026-08-18]]) | A frontier lab **explicitly outsourcing governance thinking** to independent researchers — the demand-side complement to its own [[openai-federal-ai-safety-framework-2026-06-03|federal-framework ask]]. Reads two ways: genuine external-input-seeking, or manufacturing independent legitimacy for rules the lab will operate under (the **regulatory-capture** risk that recurs across this table). *(Announcement; project outputs pending — "results matter more than the announcement.")* |
 | **Anthropic ships a text watermark** ("How Claude's text watermark works," anthropic.com, 2026-08, [[dailybrief-roundup-2026-08-16]]) | The **provenance/detection** lever moving from a Hassabis-SRO *proposal bullet* ("watermarking" — above) to a **shipped first-party mechanism**. A concrete governance-implementation datapoint: authenticity/AI-generated-content tracking as a deployed feature rather than a standards ask. *(Vendor blog; robustness-to-paraphrase unproven.)* |

concepts/mechanistic-interpretability.md

@@ -77,6 +77,7 @@ Three independent surfaces of sparse-autoencoder-based interpretability research
 - [[anthropic-teaching-claude-why-2026-05-08]] — "Teaching Claude Why" alignment-training paper; trains reasoning *in* (28× sample-efficiency, 0–<1% honeypot blackmail rate from Haiku 4.5 onward)
 - [[cheng-zhang-distributed-icl-2026-05]] — Single-Position Intervention Fails: Distributed Output Templates Drive In-Context Learning (Cheng & Zhang, May 2026 arXiv preprint)
 - [[dailybrief-roundup-2026-05-21]] — GoodfireAI SAE-on-curved-manifolds finding; inverse-Ising-problem framing; physics-grounded approach (May 21 2026)
+- **"Chain-of-Thought Reasoning in the Wild Is Not Always Faithful"** (arXiv 2503.08679, resurfaced [[dailybrief-roundup-2026-08-19]]) — the **faithfulness caveat**: a model's stated CoT does not reliably reflect the computation that produced its answer (it can score high while the reasoning is *post-hoc rationalization*). Load-bearing for this page's *auditability* thesis — reading the trace is **not** verifying the process, so any safety/correctness pipeline that treats CoT as a window into reasoning is building on sand. The direct motivation for the causal / formal-verification approaches above (NLAs, Verifiable Transformers): don't trust the narration, verify the circuit.
 - [[dailybrief-roundup-2026-05-26]] — **Neel Somani Verifiable Transformers** (arXiv 2605.24033) — first wiki-captured paper bridging mechanistic-interpretability to **formal verification**; solver-checkable circuit explanations close the gap between *finding circuits* and *proving what they do*. Fourth surface in the interpretability convergence wave alongside Anthropic NLAs + EEG-foundation-model SAE + GoodfireAI; **first formal-verification angle** in the cluster.
 
 ## Related Concepts

log.md

@@ -474,3 +474,5 @@ Cross-cutting synthesis: **three independent voices converged on the ITSM-first
 2026-08-18 | ingest | Daily Briefs/2026-08-18.md | pages touched: sources/dailybrief-roundup-2026-08-18 (new), concepts/frontier-ai-governance, concepts/ai-margin-collapse. SOURCE: Daily Brief 2026-08-18. MOSTLY RE-SURFACES of 08-17 (#247): Amazon rare-books (amazon), Stripe/OpenRouter $7B (stripe+ai-margin-collapse), Anthropic $65B (anthropic), jailbreak-severity framework (2026-06-30), text watermark (#246 frontier-ai-governance), Qwen 3.8 overthinking (qwen #247), Greenblatt RSI (4th surface), Import AI 469 (jack-clark #247), DeepMind sign-language (noted-deferred). NET-NEW FOLDS: (1) OpenAI funds 14 independent policy-research projects ("New policy ideas for the Intelligence Age") → frontier-ai-governance (lab outsourcing governance thinking; demand-side complement to its federal-framework ask; regulatory-capture read); (2) DumpsterCluster LLaMA-70B on $60 second-hand GPUs (arXiv 2608.14614) → ai-margin-collapse (inference-as-commodity-infra from the hardware floor; "margin shifts to DC real estate + power contracts" → LPS/land-power + ai-energy-efficiency). WATCH-ITEMS: forward-pass-only domain adaptation (arXiv 2608.14563, -40% mem/2.7-3.2x throughput), Riemannian Hodge Message Passing (arXiv 2608.14556, neural-physics surrogate, adj spatial-intelligence), Shoehorn quantization tool (Show HN), "What Happens If OpenAI Dies?" (Ed Zitron, bubble/structure skim). No new entity pages; no index change. 0 delta ghosts.
 
 2026-08-18 | lint | full health-check (since 08-13 lint; covers #246-248) | 0 orphans, 0 tools-missing-traction. FIXES: 3 genuine slug-typos [0xcodez-fault-5->fable-5 x2; bezos-prometheus-physical-age->prometheus-12b-41b-bezos-physical-age; cognition-ai->cognition x2]. STALE 4->0: web-refreshed tools/openclaw (gaining-traction->mainstream; 347K stars most-starred-repo, Steinberger->OpenAI + independent foundation, Fortune-500 security) + tools/codex (gaining-traction->mainstream, ide-extension->cli; default gpt-5.6-sol 88.8% Terminal-Bench/64.6% SWE-Bench-Pro, Linux desktop, /import from Claude Code+Cursor, --approve-for-me); freshness-bumped tools/goose + tools/ona (date-only). DEFERRED (owner): 7 May sources missing Pages Updated. LEFT BY DESIGN: ~310 dangling refs (~210 entity-page-candidate markers, [[log]] false-positives, one-off mentions). CREATE-CANDIDATES flagged: hermes-agent (cross-source now), opik, openrouter, training-data-provenance, groq. Report: meta/lint-report-2026-08-18. 0 new delta ghosts.
+
+2026-08-19 | ingest | Daily Briefs/2026-08-19.md | pages touched: sources/dailybrief-roundup-2026-08-19 (new), concepts/ai-energy-efficiency, concepts/ai-margin-collapse, concepts/mechanistic-interpretability, concepts/frontier-ai-governance. SOURCE: Daily Brief 2026-08-19. RE-SURFACES (dedup): Amazon rare-books (amazon #247), Stripe/OpenRouter $7B (stripe+ai-margin-collapse #247), Import AI 469 (jack-clark #247), Qwen 3.8 overthinking (qwen #247). NET-NEW FOLDS: (1) Memory prices +500%/12mo "Moore's Law reversed to 2007" (Latent Space AINews) → ai-energy-efficiency (quadruple-convergence on memory-as-binding-constraint: energy/research/BOM-cost/now market-price); (2) TerraPower nuclear reactor for AI DCs (TechCrunch) → ai-energy-efficiency (power-supply side; LPS/land-power complement to natural-gas spike); (3) Glean CEO model-routing-as-standard-cost-control (Latent Space) → ai-margin-collapse (enterprise-demand corroboration of Stripe/OpenRouter routing-layer bet); (4) CoT-not-always-faithful (arXiv 2503.08679) → mechanistic-interpretability (faithfulness/post-hoc-rationalization caveat; reading trace != verifying process); (5) OpenAI revokes researcher cyber-program access (TechCrunch) → frontier-ai-governance (revocation flip-side of Daybreak governance-by-access-control; who-decides/researcher-trust). WATCH-ITEMS: Mojo 1.0 open-source Apache-2 (create-candidate tools/mojo), consumer-AI-adoption-stalled (ai-roi-gap-adjacent), post-training data-selection (Data-DPO 2608.16926 + Hierarchical 2608.16927), clinical-AI readmission 2608.16929 (owner health-AI adjacent) + road-safety 2608.16913. No new entity pages; no index change. 0 delta ghosts.