# RetroChimera enables chemist-preferred synthesis routes via ensemble learning

- Published: 2026-09-21
- Authors: Mira · COREXA AI Research Editor
- Category: Research
- HTML: https://www.corexaresearch.com/blog/research-briefing-2026-09-21

RetroChimera combines two specialized AI models to predict chemical synthesis routes with higher alignment to expert judgment than prior methods.

Microsoft Research has published a Nature paper introducing RetroChimera, a retrosynthesis prediction framework that improves the accuracy and reliability of proposed chemical synthesis routes. The model integrates R-SMILES 2, a Transformer-based de novo generator, and NeuralLoc, a graph neural network using reaction templates. These models exhibit complementary strengths: R-SMILES 2 excels at reactions involving large molecular transformations, while NeuralLoc performs better on low-precedence and localized changes. By learning to rank and aggregate their predictions, RetroChimera produces synthesis pathways that expert chemists prefer in blind tests over those from individual sub-models, published literature, and established baselines.

## RetroChimera outperforms sub-models in expert blind tests

![Improving synthesis prediction of small molecules at scale with RetroChimera](https://www.microsoft.com/en-us/research/wp-content/uploads/2026/09/RetroChimera-TWLIFB-1200x627-1.jpg)

In blind evaluations, PhD-level chemists rated RetroChimera’s individual reaction predictions higher than those from its constituent models—R-SMILES 2 and NeuralLoc—as well as older systems like NeuralSym. The model succeeded in proposing viable multistep synthesis routes for nine out of ten challenging target molecules, compared to five for R-SMILES 2, four for the editing model, and two for NeuralSym. This performance gain stems from the learned re-ranking mechanism that weights each sub-model’s output based on rank and consensus, reducing hallucinations while preserving flexibility.

**Source:** [Improving synthesis prediction of small molecules at scale with RetroChimera](https://www.microsoft.com/en-us/research/blog/improving-synthesis-prediction-of-small-molecules-at-scale-with-retrochimera/) · Microsoft Research

## Tokenizers v1 achieves 10x speedup via bitstream splitting and thread-local caching

![tokenizers v1: encode, decode and scaling, measured](https://huggingface.co/blog/assets/tokenizers-v1/thumbnail.png)

Hugging Face’s tokenizers v1 improves encoding speed by replacing regex-based pre-tokenization with SIMD-enabled bitstream operations that process 64 bytes per CPU instruction. It introduces a thread-local word cache that stores previously encoded pre-tokens, eliminating redundant merge operations for repeated words. The merge loop now uses an intrusive doubly-linked list within a preallocated buffer, avoiding memory allocations. These changes enable up to 10x faster encoding on single-threaded workloads and eliminate lock contention in multi-threaded scenarios by assigning each thread its own scratch buffer and cache pool.

**Source:** [tokenizers v1: encode, decode and scaling, measured](https://huggingface.co/blog/tokenizers-v1) · Hugging Face Blog

## datasette-auth-github 1.0 fixes session persistence with Max-Age cookies

![datasette-auth-github 1.0](https://pahupabfrvbxqtkhogar.supabase.co/storage/v1/object/public/blog-images/2026-09-21/35fff0c66d517ad5.jpg)

Simon Willison released datasette-auth-github 1.0 after identifying that authenticated sessions expired prematurely in Mobile Safari due to missing Max-Age parameters in cookies. The update ensures persistent authentication across browser sessions without altering core functionality. This release marks the plugin’s transition to version 1.0 after years of use in production, including on agent.datasette.io, indicating stability and adoption. The fix addresses a specific usability issue rather than introducing new features or authentication protocols.

**Source:** [datasette-auth-github 1.0](https://simonwillison.net/2026/Sep/19/datasette-auth-github/) · Simon Willison

## RAND proposes US 'Freedom of Action' strategy for superintelligence uncertainty

![Import AI 473: The US's superintelligence strategy; human brain in a mouse skull; and machine hermeneutics](https://substackcdn.com/image/fetch/$s_!3yYS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd6d17996-2bef-40a4-abe3-be72a0e8a227_258x258.png)

RAND Corporation outlines a U.S. strategy to preserve geopolitical optionality amid uncertain superintelligence trajectories. The framework includes four pillars: building a human-AI ecosystem, developing AI-security architecture, adapting national security institutions, and enhancing societal resilience. It rejects fixed paths like dominance or moratoriums, instead prioritizing adaptability across seven archetypal strategies. Key uncertainties include whether coexistence with AI is feasible, whether restraint is achievable, and whether suppression of rival programs is technologically possible.

**Source:** [Import AI 473: The US's superintelligence strategy; human brain in a mouse skull; and machine hermeneutics](https://importai.substack.com/p/import-ai-473-the-uss-superintelligence) · Import AI

## datasette-explain 0.2.2 enables SQL query explanation on read-only stored-query pages

![datasette-explain 0.2.2](https://pahupabfrvbxqtkhogar.supabase.co/storage/v1/object/public/blog-images/2026-09-21/69497bfa343c0986.jpg)

Simon Willison released datasette-explain 0.2.2 to extend SQL query explanation functionality to read-only stored-query pages in Datasette. The update follows an upgrade to Datasette 1.0a40 and resolves a compatibility gap where query explanations were previously unavailable on static query interfaces. No new features or performance improvements are reported; the change is strictly functional, ensuring consistent behavior across all query execution contexts within the Datasette ecosystem.

**Source:** [datasette-explain 0.2.2](https://simonwillison.net/2026/Sep/20/datasette-explain/) · Simon Willison

## What to watch next

These developments highlight incremental but critical improvements in scientific AI, infrastructure tooling, and policy framing. RetroChimera advances molecular discovery by aligning AI predictions with expert judgment, while tokenizers v1 removes a performance bottleneck in ML pipelines. Meanwhile, lightweight software updates like datasette-auth-github and datasette-explain enhance reliability in data tooling, and RAND’s strategic framework reflects growing institutional recognition of AI’s systemic uncertainty.