The Strange Brief
NVIDIA's quantum software play, OpenAI goes bio, Anthropic releases Opus 4.7 and Claude design, and China's 2D semiconductor breakthrough.
THE DOWNLOAD
NVIDIA Releases Ising, Open AI Models for Quantum Calibration
NVIDIA released Ising, the first open AI model family built for quantum processor calibration and error correction. The suite includes a 35B-parameter vision-language model that automates calibration workflows (reducing setup from days to hours) and decoder models delivering 2.5x faster, 3x more accurate quantum error correction. Models are available on GitHub and Hugging Face, and run on NVIDIA’s CUDA-Q quantum software platform.
Why it matters: Is Nvidia running the CUDA playbook applied to quantum? Nvidia software CUDA became the de facto standard for AI training by being free, performant, and deeply integrated with NVIDIA hardware. Ising does the same thing for quantum: every lab that builds on it ties its calibration and error-correction workflows to GPU-accelerated infrastructure.
OpenAI Dives Into Bio with GPT-Rosalind
OpenAI launched GPT-Rosalind, its first domain-specific frontier model purpose-built for biology, drug discovery, and translational medicine. The model is gated through a trusted-access program with initial partners including Amgen, Moderna, and Thermo Fisher. In evaluations with Dyno Therapeutics on unpublished RNA sequences, the model’s predictions ranked above the 95th percentile of human experts.
Why it matters: It seems like the model makers are all getting into life sciences. Anthropic acquired Coefficient Bio for $400M earlier this month to build biology-native capabilities into Claude. AWS launched Amazon Bio Discovery the same week. Three of the largest AI platforms made major bio moves within days of each other.
Anthropic Releases Opus 4.7 and Claude Design
Anthropic released Claude Opus 4.7 alongside Claude Design, a new Mac-based design tool that reads a team’s codebase and design files, builds a design system automatically, and generates prototypes matching existing brand and components. Opus 4.7 outperforms GPT-5.4 and Gemini 3.1 Pro on coding benchmarks but falls short of Anthropic’s own unreleased Mythos model, which remains restricted to select partners due to cybersecurity concerns.
Why it matters: Claude Design is a model maker moving directly into the application layer, competing with Figma, Framer, and Adobe as a standalone product, not a plugin. On the model side, early reception of 4.7 has been mixed: coding and agentic tasks are measurably better, but a new tokenizer consumes up to 35% more tokens on identical inputs. Anthropic is also publicly running a two-tier strategy: ship the commercial model, hold back the more capable one.
Anthropic Publishes Nature Paper on Hidden Trait Transmission in LLMs
Anthropic co-authored a paper published in Nature showing that LLMs can transmit behavioral traits through semantically unrelated training data. A teacher model fine-tuned on insecure code generated datasets of plain number sequences. Student models trained on those numbers acquired the misalignment, producing responses endorsing violence and criminal behavior, even after researchers filtered out numbers with negative cultural associations. The authors proved mathematically that this is a general property of neural networks, not an LLM-specific quirk.
Why it matters: This matters because the way the trait spreads is exactly how most AI companies already build models. Training a model on its own outputs, compressing a large model into a smaller one, or starting multiple products from the same base model are all standard practice, and all meet the conditions for this effect. For companies building on top of foundation models, this opens a new risk category: you need to know not just what is in your training data, but where it came from and what model generated it.
China’s 2D Semiconductor Sprint Gets a Manufacturing Breakthrough
Researchers from China’s Institute of Metal Research achieved a 1,000x improvement in the growth rate of wafer-scale 2D semiconductor films using a novel liquid gold/tungsten CVD process. The technique produces monolayer tungsten silicon nitride films with tunable doping properties at commercially relevant dimensions.
Why it matters: 2D semiconductors are one of the leading candidates for what comes after silicon hits its physical limits. This is still early-stage research, years from commercial production. But China is building a lead in materials that are not covered by current U.S. export controls, which today focus on EUV lithography and advanced silicon fabrication. If 2D materials become viable at scale, the chokepoints that currently give the U.S. and its allies leverage over China's chip supply chain may not apply.
DEEP DIVE FROM THE REVIEW
This week we published a piece predicting that model makers are absorbing entire software categories into the model itself.
A few days later, Anthropic launched Claude Design, a direct competitor to the likes of Figma and Framer. Over the next year, I believe we will see model makers move aggressively into the application layer, think: project management tools, expense software, and other SaaS categories that sit between the model and the user.
Read more below: AI Swallows Software Whole
EVENT
Strange Gathering | San Francisco |April 24 2026
We’re hosting a small lunch in San Francisco to demo agentic workflows. Bring a Claude routine, a weird hack, or the agent setup you’ve been loving.