IBM Research on Tuesday released MAMMAL, a multi-modal AI model that achieved state-of-the-art performance on 9 out of 11 biological benchmarks, outperforming Google’s AlphaFold 3 on several key tasks including antibody-antigen binding prediction. The breakthrough comes as Miami startup Subquadratic emerged from stealth claiming a 1,000x efficiency gain with its SubQ model architecture, though researchers are demanding independent verification of the extraordinary claims.
IBM’s MAMMAL Dominates Biological AI Benchmarks
IBM’s MAMMAL (Multi-modal AI Model for Molecular Analysis and Learning) combines protein, molecule, and gene data in a unified framework that excels at biological interaction tasks. According to the Nature paper, the model achieved superior performance across nine critical areas including drug-target interaction prediction, ligand binding affinity, and gene expression forecasting.
The model’s most significant victory came in antibody-antigen binding prediction, where it substantially outperformed AlphaFold 3. This capability is crucial for vaccine development and immunotherapy research, as it predicts how antibodies will interact with disease targets. MAMMAL also excelled at multi-modal biological reasoning, combining cellular data with molecular structures to predict biological outcomes.
Unlike AlphaFold 3, which focuses primarily on protein structure prediction, MAMMAL addresses functional prediction — determining what proteins actually do rather than just their shape. The model also demonstrated strong cross-domain generalization, applying knowledge across different biological systems without task-specific retraining.
Subquadratic’s Extraordinary Efficiency Claims Draw Skepticism
Miami-based Subquadratic raised $29 million in seed funding at a $500 million valuation, claiming its SubQ 1M-Preview model achieves nearly 1,000x reduction in attention compute compared to frontier models. The company says its architecture grows compute linearly with context length rather than quadratically — a mathematical breakthrough that would fundamentally change AI scaling economics.
According to VentureBeat, Subquadratic launched three products into private beta: an API with full 12-million-token context access, SubQ Code for command-line development, and SubQ Search. The funding round included Tinder co-founder Justin Mateen, former SoftBank Vision Fund partner Javier Villamizar, and early investors in Anthropic and OpenAI.
However, the AI research community has responded with significant skepticism. Multiple researchers questioned why the company would limit access through an early-access program if the model truly costs less than 5% of Claude Opus to serve. Critics described the benchmarks as “suspiciously perfect” cherry-picked results, drawing uncomfortable comparisons to past technology fraud cases.
Autonomous AI Achieves Scientific Discovery Breakthrough
Researchers demonstrated the first end-to-end autonomous scientific discovery by an AI system on a real physical platform. The Qiushi Discovery Engine, detailed in a preprint on arXiv, autonomously identified and experimentally validated optical bilinear interaction — a previously unreported physical mechanism analogous to Transformer attention operations.
The system consumed 145.9 million tokens across 3,242 LLM calls and 1,242 tool interactions during its investigation. It successfully reproduced published transmission-matrix experiments and converted abstract coherence-order theory into measurable observables. The AI-discovered optical mechanism suggests potential pathways for high-speed, energy-efficient optical hardware for pairwise computation.
This represents the first demonstration of an AI agent autonomously proposing and validating novel physics through real-world experimentation, marking a significant milestone for research-level autonomous systems.
Timer-XL Advances Long-Context Time Series Forecasting
Tsinghua University’s THUML lab released Timer-XL, a decoder-only Transformer foundation model designed for long-context time series forecasting. The model handles variable input and output lengths in a single architecture, eliminating the need for separate models for different forecasting horizons.
Timer-XL introduces TimeAttention, a specialized attention mechanism optimized for temporal data patterns. The model supports non-stationary univariate series, complex multivariate dynamics, and covariate-informed contexts with exogenous variables. Unlike previous models that required multiple versions for different sequence lengths, Timer-XL uses unified architecture across all forecasting scenarios.
The research team behind Timer-XL previously developed milestone models including iTransformer, TimesNet, and the original Timer model. Timer-XL can be trained from scratch or fine-tuned on domain-specific datasets for improved performance.
What This Means
These developments highlight AI’s expanding capabilities across scientific domains, from biological research to autonomous discovery. IBM’s MAMMAL demonstrates that multi-modal approaches can outperform specialized models like AlphaFold 3 on interaction-focused tasks, suggesting the future lies in unified architectures rather than narrow specialists.
Subquadratic’s claims, if validated, would represent the most significant efficiency breakthrough in AI since the Transformer architecture. However, the research community’s skepticism reflects hard-learned lessons about extraordinary claims requiring extraordinary evidence. Independent verification will determine whether this represents genuine innovation or another cautionary tale.
The autonomous discovery achievement by Qiushi Engine suggests AI systems are approaching genuine research capability, moving beyond assistance to independent scientific contribution. This could accelerate discovery timelines across physics, chemistry, and materials science.
FAQ
How does MAMMAL differ from AlphaFold 3?
MAMMAL focuses on biological interactions and functional prediction while AlphaFold 3 specializes in protein structure prediction. MAMMAL combines multiple data types (proteins, molecules, genes) and excels at tasks like drug-target binding and antibody-antigen interactions where AlphaFold 3 shows limitations.
What makes Subquadratic’s efficiency claims so significant?
If validated, a 1,000x reduction in attention compute would solve the quadratic scaling problem that limits context length in current AI models. This would enable massive context windows at practical costs, fundamentally changing what AI systems can process and remember.
What did the autonomous AI actually discover?
The Qiushi Engine identified optical bilinear interaction, a physical mechanism that processes light in ways similar to how Transformer attention mechanisms process information. This could lead to optical hardware that performs AI computations using light instead of electricity, potentially offering speed and efficiency advantages.
