meet @mostik_ai!

what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can’t say much about while the competition is still running. and this, which I can.

everyone’s arguing about whether open models will catch up to frontier models. we think it’s the wrong question. here’s the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning?

we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure – no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched.

how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance.

we’re committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we’ve done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst

WIRED has the first external account of the company and the work: http://wired.com/story/russian-startup-mostik-ai-models-communication/… full writeup, the setup, and all the numbers: http://mostik.ai/read-more

These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words

Source: https://www.wired.com/story/russian-startup-mostik-ai-models-communication/ I recently metwith some brilliant Russian mathematicians who showed me a way forartificial intelligencemodels to communicate via something akin to machine telepathy.

The mathematicians work for astartupcalled Mostik—the Russian word for bridge. It’s a nod to the group’s approach, which allows different models to interact using the mathematical values found in theirweights—the things that determine how a prompt gets turned into an output. In practice, this means the capabilities of a larger model can be fed to a smaller model to ramp up its intelligence much more efficiently.

The startup used the approach to build a model that has rocketed to the top ofARC-AGI3, a notoriously difficult competition for AI models. (They wouldn’t tell me more because they want to win the contest.) To demonstrate the idea, however, they also created a bridge between two Chinese open-weight models: the largest version of GLM-5.2, which has 753 billion parameters; and a 4-billion-parameter version ofQwen-3.5that can run on a mobile device. The resulting hybrid system costs one-twentieth of the full GLM model, and its performance is exactly halfway between the two.

“It’s well-known in machine learning that ensembles of models perform better than individual ones,” Sasha Malysheva, Mostik’s CEO, told me over coffee.

Malysheva, who developed the approach, shared a running joke inside the company: The future of AI is similar to guessing the weight of a pig. In math circles, it’s well-known that a handful of random people can more accurately estimate a pig’s weight than an expert when their guesses are combined and averaged.

Much like communally eyeballing porcine heft, combining the outputs of several AI models often nets better results. Typically, this involves feeding the output of one model into another, which takes a good chunk of time and money. The Mostik team, however, figured out a way for AI models to talk to one another without producing text output. If it takes off, it could increase the value of open-weight models, allowing them to better compete with the closed, proprietary models offered by frontier labs likeAnthropicandOpenAI.

Malysheva says that combining lots of different models may turn out to be a better way to advance AI. “I personally do not think we will have a monolithic model [in the future] or that the capabilities of models will come from scaling,” she told me, referring to the strategy of making models larger and feeding them more data.

“If Mostik makes it possible to pair frontier models with domain-specific models—think biology, physics, and so on—many more specialized models would be trained,” says Vladimir Arustamian, the tech lead at the AI software company Lovable, who knows the Mostik team. “This team has been at it for a matter of months and already has something running that I would have guessed was years out.”

The Mostik technique means “you can approach large-model quality without the large model handling the entire loop, giving you substantial improvements with just a smaller model running alongside,” says Karl Tuyls, a former computer scientist at Google DeepMind who is familiar with the company’s tech. The method is a no-brainer for anyone tasked with running models as efficiently as possible, Tuyls says.

Stanislav Smirnov, a professor at the University of Geneva and a 2010 Fields Medalist, is Mostik’s chief scientist. He says finding common ground between two AI models is surprisingly difficult. “There seems to be no appropriate mathematical language yet,” he says. In the interim, Mostik’s approach is a way to quite literally bridge the gap.