A look at “liquid” neural nets, which change their underlying algorithms based on observed inputs, making them more flexible than standard ML neural networks
“Liquid” neural nets, based on a worm's nervous system, can transform their underlying algorithms on the fly, giving them unprecedented speed and adaptability. Tweets: @quantamagazine , @mit_csail , @quantamagazine , @quantamagazine , @ramin_m_h , @glenngabe , and @quantamagazine Tweets: @quantamagazine : So-called liquid neural networks could help autonomous vehicles navigate complex environments with animal-like efficiency. https://www.quantamagazine.org/ ... @mit_csail : MIT scientists created a new class of speedy “liquid” neural networks that can change underlying algorithms on the fly, sometimes outperforming convolutional neural nets: https://www.quantamagazine.org/ ... Image v/@QuantaMagazine https://twitter.com/... @quantamagazine : Researchers have built an artificial neural network made out of equations that mirror the cognition of real-life roundworms. https://www.quantamagazine.org/ ... @quantamagazine : Artificial intelligence researchers have built a new class of “liquid” neural networks that process inputs in a nonlinear way. In certain scenarios, these networks handily outperform conventional neural networks. Steve Nadis reports: https://www.quantamagazine.org/ ... Ramin Hasani / @ramin_m_h : Thank you @QuantaMagazine & Steve Nadis for featuring such a great article about our work on “liquid neural networks” https://www.quantamagazine.org/ ... It describes the journey @mlech26l, Daniela Rus & I took @MIT @MIT_CSAIL to develop liquid networks and what's next Glenn Gabe / @glenngabe : Wait until Google's ML systems for Search start using liquid neural nets. Or are they already? :) -> Liquid neural nets, based on a worm's nervous system, can transform their underlying algorithms on the fly, giving them unprecedented speed & adaptability https://www.quantamagazine.org/ ... https://twitter.com/... @quantamagazine : The roundworm 𝘊. 𝘦𝘭𝘦𝘨 𝘢𝘯𝘴 is one of the only creatures with a fully mapped-out nervous system. This profile has gifted computer scientists with a new source of inspiration for artificial neural networks. https://www.quantamagazine.org/ ...
Context & Ripple Effects
This explainer lands mid-search for alternatives to the dominant deep-learning recipe. Earlier coverage traced how transformers revolutionized language modeling and spread into vision, while neurosymbolic AI represented one hybrid attempt to move past pure deep learning. MIT's Ramin Hasani and Daniela Rus offer a different route: models built on the wiring of a worm's 302-neuron nervous system that rewrite their underlying equations based on what they observe.
What makes the piece worth tracking is where the arc goes next — the same lab lineage surfaces in Liquid AI's launch of non-transformer LFM models claiming state-of-the-art performance at every scale, turning this research curiosity into a commercial challenge to transformer orthodoxy.
First-order effects
- MIT's group gains a high-profile validation point for worm-derived architectures, with the coverage explicitly pitching them at autonomous-vehicle navigation where adapting behavior to shifting inputs matters most.
- Researchers working on adaptive and brain-inspired models — including the Meta-lab EEG cognition work in related coverage — get fresh evidence that biological wiring diagrams can yield practical ML designs.
Second-order effects
- Transformer-first labs now face a named architectural challenger rather than an academic curiosity, since the spinoff path shows the idea can ship as competitive production models rather than stay in papers.
- Compute-constrained deployments such as vehicles, robots, and edge devices become the natural battleground for these faster, more adaptable models, pressuring incumbents whose stacks assume large fixed-weight transformers.
Third-order effects
- If non-transformer designs keep posting competitive results, foundation-model supply diversifies beyond a single architecture, weakening the field's working assumption that scaling one blueprint is the only road to capability.
- The MIT-to-spinoff pipeline points toward a structural pattern in which neuroscience findings feed startup formation, tightening the feedback loop between studying brains — as the EEG cognition research does — and building commercial models.
The trend: Neural-network research is diversifying beyond the transformer toward adaptive, biologically-inspired architectures, carried from university labs into the market through spinoffs like Liquid AI.