Free Script · CRISPR's One-Dose Proof · July 24, 2026
The Week in One Number
89%. That's the reduction in monthly hereditary angioedema attacks requiring on-demand treatment (weeks 5–28), from Intellia's additional Phase 3 HAELO data for lonvoguran ziclumeran, published in NEJM (June 13, 2026). A single IV dose of LNP-delivered CRISPR-Cas9 knocks out KLKB1 in the liver, permanently lowering plasma kallikrein. That figure marks in vivo gene editing crossing from promise into durable, one-time therapeutic effect.
Three Moves
AMD ships MI455X into volume. At Advancing AI 2026 (July 23), AMD moved the MI400 family to volume production — 432GB HBM4 at 19.6 TB/s per GPU, a 145% bandwidth jump over MI350's 288GB HBM3e. The Helios rack packs 72 GPUs and 2.9 EFLOPS FP4. So-what: memory bandwidth, not FLOPS, now gates inference throughput.
Quantinuum removes distillation overhead. Quantinuum and academic partners reported (July 17, 2026, Nature) the first universal topological gate set on a 54-qubit non-Abelian S3 state on the H2 processor. The mechanism sidesteps magic-state distillation, which often dominates physical-qubit count and runtime — a structural cut in fault-tolerance cost.
Frontier token pricing collapses. BenchLM's Token Price Index (updated July 2, 2026) reports frontier average output pricing down ~94.5% since March 2023 (index 100 → 5.5), with DeepSeek V3 at $1.10/1M output tokens. So-what: reasoning "thinking tokens" quietly inflate effective cost per useful output, muddying the headline drop.
One Failure Worth Studying
OpenAI's Frontier Evals team announced (Feb 23, 2026) it would stop reporting SWE-bench Verified. The reason: contamination had made the 500-issue score partly a measure of memorization rather than skill. The tell was Claude Opus 4.8 falling ~14% under restricted-internet conditions, exposing retrieval-based score inflation. The models aren't worse — but any benchmark with discoverable ground-truth on the open web decays into a memorization test. Retrieval closes the gap that reasoning was supposed to.
Scoreboard Snapshot
| Segment | System | Number | Status |
| AI accelerator | AMD MI455X | 432GB HBM4 / 19.6 TB/s | Volume production |
| AI accelerator | NVIDIA Rubin | 3.6 EFLOPS dense FP4 | Full production; partners H2 2026 |
| In vivo editing | Intellia lonvo-z | 89% attack reduction | Phase 3 (NEJM) |
| Sequencing | Element VITARI | $100 whole genome | Introduced Feb 2026 |
| Quantum | Quantinuum H2 | 54-qubit topological gate set | Published (Nature) |
| Eval ceiling | HLE (Nature 2026) | Gemini 3 Pro 37.5% / human ~90% | Leaderboard current |
The Upgrade Line
Thursday's paid issue ran a Deep Read on the MI455X-versus-Rubin memory-bandwidth economics and what 432GB HBM4 does to inference batching, plus a Tool Box on auditing benchmark contamination in your own eval pipeline. Independent technical analysis — not investment advice or engineering counsel.
---
You're reading the free edition of Systemic Script — the architecture of complex technology, decoded for senior engineers.
Script Pro subscribers receive the full Thursday issue — the Deep Read with mechanism, cost math, and sourced figures, the complete Scoreboard tracker, and the monthly Architecture Teardown — for $19/month or $180/year.