Why researchers are leaving big AI labs to make their own bets

OpenAI says the era of artificial general intelligence is beginning, while Nvidia CEO Jensen Huang has declared that AGI has already arrived. At the same time, the people behind today’s models are leaving OpenAI, Google DeepMind and Meta to start their own labs, raising billions before releasing products. THE DECODER examined 45 such labs, took a closer look at 27 and grouped them into six research bets, from learning through experience to new computing architectures. The common idea is to find a path to more powerful AI beyond the familiar combination of large language models, internet data and reinforcement learning.

What bets are new AI labs making on the technologies of the future

Why researchers are leaving big labs

THE DECODER’s “Frontier Radar” series is published six times a year as a newsletter and as material for the site’s subscribers. Its fifth edition takes a close look at 27 labs, divided into six groups based on their research bets, and briefly mentions nearly 20 more. It also collects information on where the labs came from, their valuations and their publicly available results. The first edition covered the current state of AI agents, the second AI’s measurable effect on productivity, the third the emerging economics of generative AI tokens, and the fourth whether AI’s lead over China is holding and where Europe stands.

After the launch of GPT-6 Astra, Huang wrote on X that AGI had arrived. OpenAI president Greg Brockman called the model a turning point and welcomed humanity to the “era of general-purpose AI.” There is no single definition of AGI, so the precise meaning of the two executives’ claims is unclear. But the message is clear: the arrival of extremely powerful AI, perhaps capable of improving itself, seems close, and the path toward it already mapped out.

Some of the people behind today’s systems appear to disagree, judging by their decisions. Ilya Sutskever, OpenAI co-founder and the company’s former chief scientist, is now building Safe Superintelligence. Jerry Tworek, who led the development of OpenAI’s reasoning models, has taken charge of Core Automation. David Silver, creator of AlphaGo, left Google DeepMind and founded Ineffable Intelligence. After 12 years at Meta, Yann LeCun started AMI. Jeff Dean, who spent 27 years at Google and most recently served as chief scientist, founded Discovery Loop.

Money is following the founders, even though some labs have yet to produce a model or publish a research paper. In July, Nvidia invested in SSI; Bloomberg reported the amount at five billion dollars. In June, Jeff Bezos’s industrial AI lab Prometheus raised 12 billion dollars at a valuation of 41 billion. Flapping Airplanes, which also has no public model, raised a billion dollars in an October Series A round.

The reasons for leaving vary. At these valuations, founders can quickly become paper billionaires. LeCun left Meta after disagreements over strategy, while Sutskever wants to conduct research without the pressure to release products. Others are looking for a niche that would remain at the margins of a larger lab. Sometimes the break is not complete: Google itself became one of Discovery Loop’s first investors.

The founders do share a doubt: whether the current recipe is enough to keep progress going. That recipe consists of ever-larger language models trained on internet data and refined through reinforcement learning. Some believe AGI is near, but are unsure whether this combination alone will be enough.

What counts as a new AI lab

There is no official definition. In “Frontier Radar,” the term refers to independent research companies built around a long-term technical bet. They aim to change how AI learns, perceives the world, solves problems, is controlled or computes.

The current wave became visible in 2024, when the race for the strongest models was already well under way and cloud giants had their partners: Microsoft backed OpenAI, and Amazon backed Anthropic. In June, Sutskever founded SSI and raised a billion dollars within months for a research team with no product.

Investors are betting that a small team will find something the giants have missed. OpenAI and Anthropic have grown into large corporations, and both are preparing to go public at record valuations. Inside big companies, research teams have to compete with product teams for money and attention. Large companies also struggle to support ideas that could threaten their existing business. Management scholars call this the innovator’s dilemma.

There are precedents for this scenario. Google invented the Transformer architecture in 2017. DeepMind beat the world’s top Go players with AlphaGo, but the mass-market AI assistant came from OpenAI, which was founded as a nonprofit. The next major breakthrough came from a spinout: Dario and Daniela Amodei founded Anthropic in 2021. By the end of July 2026, Bloomberg estimated Anthropic’s annualized revenue at more than 65 billion dollars.

The new labs are looking for research gaps that the major players have left open, either because the path is still unknown or because no resources have been allocated to it. Big labs have more computing power, established data pipelines and feedback from hundreds of millions of users. A small team’s development can quickly appear in a market leader’s product after its first release.

That happened with TypeSafe AI. In mid-September, the company introduced Jev, a decision-making model; two weeks later, OpenAI announced its Decisions API. Even a brief lead can demonstrate a team’s capabilities and lead to a favorable deal. Microsoft hired Inflection’s key team. Google brought in the founders of Character.ai and the leadership of Windsurf. AMD plans to acquire World Labs for about 8.2 billion dollars, while Nvidia is in talks with Reflection, according to the Financial Times.

For a new lab, the central question is whether it can build something the major players do not yet have—and whether customers or the giants themselves will pay for it. The edition covers 27 labs, mostly in the United States and Europe, divided into six groups by their main research focus.

Bet 1: A good problem-solver is not always good at learning

Today’s models can do a lot, but they learn differently from people. They need vast amounts of data to acquire a skill, and their weights usually stop updating after deployment. A model can use new information in a conversation, but it does not continually learn from every experience. Continual learning remains an important research problem.

Six labs are approaching it from different directions: Adaption and Core Automation are exploring learning while a model is running; Flapping Airplanes, Ndea and SSI are studying learning from a small number of examples; Ineffable is focused on learning through experience.

Adaption looks for an alternative to scaling

Sarah Hooker led research at Cohere and previously worked at Google Brain. In her 2020 paper “The Hardware Lottery,” she argued that research approaches often win not because they are better, but because they fit the hardware available. In January, Hooker developed that idea further in “On the Slow Death of Scaling”: increasing compute is becoming a less reliable way to improve models. In her view, data and training methods may be more effective levers. Open models with 8 or 27 billion parameters already outperform the much larger Falcon 180B on benchmarks.

Hooker founded Adaption with Sudip Roy. The founders plan to bet not on scale but on efficient AI that learns continuously. Instead of one large model for everyone, they want to build inexpensive models that quickly adapt to individual users. The first products focus on data: tools for cleaning and visualizing it, as well as automated training through AutoScientist. The main goal—a model that learns while it is running—has not yet been achieved.

Core Automation is looking for new training methods

Jerry Tworek spent nearly seven years at OpenAI and most recently served as vice president of research. He led the development of the o1 and o3 reasoning models. His criticism of today’s approach comes from someone who helped create it. Core Automation co-founder Rohan Anil led Gemini pretraining at Google.

During fine-tuning, a trained neural network can forget what it knew before. Researchers call this catastrophic forgetting. The Transformer was also not designed to learn on the fly after deployment.

ChatGPT and Claude work around the problem by taking notes and returning to earlier conversations. Tworek thinks that is not enough: in his view, continual learning belongs to machine learning, not Markdown files. Core is looking for training algorithms beyond pretraining and reinforcement learning, as well as architectures that scale better. The first model is expected to update its weights while running and use one hundredth of today’s training data. So far, these are only rumors.

Core is testing its first approach on itself: a small team with AI research agents is trying to achieve the results of a major lab. For example, the agents are supposed to automatically generate efficient GPU code from an architectural idea.

After leaving, Tworek said such research had been difficult to pursue at OpenAI: in the race to build a better model, companies take fewer risks.

Flapping Airplanes wants models to learn from less data

Brothers Ben and Asher Spector founded Flapping Airplanes with Aidan Smith, formerly of Neuralink. Ben is lead author of ThunderKittens, a widely used library of GPU kernels.

The founders believe today’s models perform well where data is abundant, such as text and images. But data is scarce and expensive in robotics and biology, so results are worse. Flapping Airplanes wants to build models that learn from a small number of examples. The founders are looking for methods inspired by biological principles but designed to run on GPUs.

There are no public results yet. Much of the first year went into building a software foundation for more efficient GPU training, reducing the usual idle time between many small computation steps. New training methods need to run on existing hardware before they can be useful.

The lack of a model did not deter investors. In January, the lab raised 180 million dollars, followed in October by a billion-dollar Series A round at a valuation of six billion.

Ineffable Intelligence wants models to learn without human examples

David Silver is looking for different data, not less of it. At Google DeepMind, he led the development of AlphaGo and its successors. The first AlphaGo learned from human games. AlphaGo Zero and AlphaZero learned entirely through self-play and surpassed their predecessors.

In spring 2025, Silver and Richard Sutton expanded on this idea in “Welcome to the Era of Experience.” They argued that systems that only imitate people are unlikely to go far beyond human ability on many important tasks. The next generation needs to learn from its own experience, receiving rewards from the environment rather than evaluations from people. Silver later founded Ineffable Intelligence. His goal is to build a “superlearner” that learns through experience in digital and physical environments.

The hardest question is where the feedback comes from. In Go, the outcome of a game gives a clear signal: win or loss. In science and the physical world, success may be unclear or take time to appear. Reinforcement learning works especially well where the environment provides an understandable evaluation. Ineffable will have to show that the approach can also handle open-ended, changing tasks.

In April, the company announced a 1.1 billion dollar seed round at a post-money valuation of 5.1 billion. It is the largest such round in Europe. There is no working system yet.

Ndea is looking for rules in a handful of examples

François Chollet created the Keras deep-learning library and developed the ARC benchmark, which tests whether systems can learn new tasks from a few examples. After many years at Google, he is building Ndea with Zapier co-founder Mike Knoop. The team has about 15 people.

Chollet distinguishes between the ability to solve a task and the ability to learn. A system may perform brilliantly on a task after extensive training but be unable to pick up a new skill efficiently. Ndea uses program synthesis: the system searches for a program that explains the examples it has seen, while a neural model directs the search toward promising candidates. A few examples should be enough to infer rules that also work in new cases.

Chollet’s own benchmark is progressing independently of Ndea. On ARC-AGI-3, launched in March, GPT-6 Astra scored up to 99.9 percent in September and 62.7 percent in standard mode. The path to the result is instructive: Astra translated unfamiliar game environments into compact rules and its own shorthand notation—roughly what Ndea is trying to do. The lab has not yet published its own ARC-AGI-3 result.

Safe Superintelligence is still relying on its founder’s reputation

Ilya Sutskever is an OpenAI co-founder and the company’s former chief scientist, as well as one of the authors of important deep-learning papers. He left in 2024 and founded SSI, where he now serves as CEO. The company’s mission statement describes its goal as safe superintelligence, undistracted by product development. Its roughly 50 employees in Palo Alto and Tel Aviv have yet to publish a model or research paper.

In late 2025, Sutskever told Dwarkesh Patel that people can learn something new from a few examples, while models need enormous datasets. He intends to change that and expects a system that can learn at least as well as a person within five to 20 years.

In July, SSI and Nvidia announced a partnership intended to increase its computing capacity tenfold. Bloomberg reported an investment of five billion dollars. Sutskever said SSI had research that deserved to be scaled up. Including the three billion it raised earlier, the total comes to about eight billion dollars. The company’s last known valuation is 32 billion. It rests largely on Sutskever’s reputation and the expectation that he sees a path others do not.

Other labs working on learning

Palo Alto-based Pathway has already introduced an architecture similar to the one Core Automation is pursuing. Its BDH model adapts during processing and develops something like working memory. So far, this has only been demonstrated on small models, with up to one billion parameters.

Founded in 2022, Symbolica, like Ndea, combines neural-network learning with program synthesis, but uses its own mathematical architecture instead of language models. Its founders acknowledge that their long-term goal—a general-purpose program-synthesis machine—has not yet been achieved. While it works toward that, Symbolica makes money from Agentica, a tool for building AI agents.

Bet 2: Language alone is not enough

The first group wants to change how models learn; the second wants to change what they learn from. Its founders doubt that language is the best route to intelligence. They want to build systems that understand spatial relationships, predict physical consequences and act in the world. AMI predicts abstract states, Black Forest Labs predicts video frames and robot movements, General Intuition learns from video games, World Labs creates 3D worlds, and Physical Intelligence trains robots directly on robots.

AMI offers an alternative to Yann LeCun’s approach

Yann LeCun won the Turing Award and created Meta’s FAIR research lab. For years, he argued against the idea that language models would lead to human-like intelligence. In late 2025, LeCun left Meta after disagreements over strategy: the company wanted to focus on language models that could be brought to market quickly. At AMI, he became executive chairman, while Alexandre LeBrun, founder of medical AI company Nabla, became CEO.

AMI’s technical foundation is JEPA, or Joint Embedding Predictive Architecture. LeCun proposed it in 2022 as an alternative to generative language and video models. Instead of predicting an output one token at a time, JEPA works with abstract internal representations. The model learns to identify what matters by predicting the next moment in a video or a robot’s movement, without trying to reproduce every detail in an image.

In 2025, V-JEPA 2 showed that the approach could work. After pretraining on video, robots needed just 62 hours of robot-collected data to learn simple object-moving tasks in unfamiliar labs. Planning was much faster than with systems that work with images at the pixel level.

AMI is targeting industry, robotics and healthcare; Nabla became its first partner. In March, the lab, with a team of roughly a dozen people, raised 1.03 billion dollars at a pre-money valuation of 3.5 billion. LeBrun says openly that a product may be years away. The research and most of the code are to be open.

Black Forest Labs moves from images to robot control

Freiburg-based Black Forest Labs is known for its FLUX image-generation models, which are already generating revenue. Its founders, led by Robin Rombach, helped develop the method behind Stable Diffusion.

Since autumn, the lab has also been betting on tasks beyond image generation. An image model has learned how images change, and could therefore potentially learn what changes a given action causes.

The first step was FLUX 3 Action, released in September. It is based on FLUX 3, trained primarily on video, and further tuned on data where the action that caused a change is known: video games, first-person video and recordings of robots operated by people. The model takes a camera image, the robot’s state and a text instruction, then predicts the next two seconds—the robot’s movement and what the scene will look like afterward. It is an image generator trained to control robots. Its weights are published on Hugging Face.

On Nvidia’s RoboLab-120 robotics benchmark, FLUX 3 Action placed second on October 22, behind Lakeside, and later dropped to fourth. The tests run in simulation, not on a real robot.

General Intuition turns video games into training data

General Intuition grew out of Medal, a platform for gaming clips. Medal CEO Pim de Witte founded the company in 2025 with three University of Geneva researchers who had worked on world models for games.

The team records players’ actions alongside game footage. This data shows not only what changed but what action caused the change. Medal provides billions of video frames matched with keyboard presses and mouse movements. The lab uses them to train two types of models: world models predict what happens next, while action models decide what to do.

The approach is not entirely new. In 2022, OpenAI trained an AI agent for Minecraft on gameplay recordings using VPT. DeepMind links Gemini to an AI agent for 3D games through SIMA 2.

General Intuition plans to make money as a platform offering access to models through an API, for game studios and robotics companies. To demonstrate the idea, the company showed a model trained only on games that controlled a quadruped robot after eight minutes of fine-tuning on real robotics data.

Physical Intelligence is building a general-purpose model for robots

Physical Intelligence trains models directly on robots. The lab was founded by well-known robotics-learning researchers, including Sergey Levine of Berkeley and Chelsea Finn of Stanford. Levine, who previously worked at Google, describes the goal as a general-purpose robot model: one system for different robots and tasks, without having to program a separate controller for each situation.

The π model family combines vision, language and actions. The lab collects most of its training data itself: people remotely operate robots, and the models learn from the recordings. In April, it released π0.7, with about five billion parameters. In the lab’s own tests, the model combined learned skills to handle new tasks and matched models that specialists had trained separately for specific tasks.

The Pi API has been in beta since October 13. Developers upload data from their robots, fine-tune the models and run them from the cloud.

In March, reports emerged of funding talks at a valuation of more than 11 billion dollars. Competition comes from Google DeepMind’s Gemini Robotics 2, other robotics startups and general-purpose models such as Astra, which already controls robotic arms.

World Labs is building spatial intelligence and may move to AMD

Fei-Fei Li is a Stanford professor known for the ImageNet dataset. In 2024, she founded World Labs with three researchers from Meta and Google, including NeRF co-author Ben Mildenhall. The company is working on spatial intelligence: AI that understands and creates 3D worlds. Li calls 3D the language of nature.

Since November 2025, World Labs has been selling Marble. The product creates editable 3D worlds from text, images, video or sketches; since January, it has also been available through an API. Game developers use Marble to create prototypes, while robotics researchers use it to build training environments. Marble creates the scene, but a simulator still has to calculate what happens when a robot collides with a chair.

Atlas, available to select partners since September, is meant to address that problem. In a single step, it creates, reconstructs and simulates a space: a few photographs become an environment that can be navigated, as well as a video of up to a minute with controllable camera movement.

At the end of September, AMD announced plans to buy World Labs for about 8.2 billion dollars in stock. The deal is expected to close by the end of 2026, and Li is expected to become AMD’s chief scientist. AMD officially said the acquisition would help it better understand how the technology can be used, but there may be another reason. Nvidia has an integrated physical AI platform: Omniverse, Isaac and the Cosmos models. World Labs could help AMD narrow part of that gap.

Other labs in this group

San Francisco-based Odyssey, like World Labs and General Intuition, is betting that a video-trained world model will learn movement and physics and teach AI agents to act. Since September, up to 20 people and AI agents have been able to work in the same simulated environment in Agora-2, which does not run on a traditional game engine. Runway, known for video generation and founded before the current wave, followed Black Forest Labs’ path in December 2025: it released GWM-1, moving from a video model to action-responsive simulation and then to robotics.

Physical Intelligence has two close competitors in robot control. Pittsburgh-based Skild AI also wants to create one model for many robots, but starts with simulations and internet video. At a valuation of more than 14 billion dollars, it is among the most highly valued companies in the field. Like Physical Intelligence, Generalist collects its own robot data. In August, it showed with GEN-1.5 how a robot could learn short manual tasks from just a few examples.

General-purpose models are catching up with specialized ones

General-purpose models are gaining capabilities quickly. Their spatial understanding is improving: GPT-6 Astra models a house in Blender and makes it navigable in Unreal Engine. Gemini 4 Carbon reaches human-level performance on the 3D-understanding benchmark Blueprint-Bench 2.

Their limits are still visible on real robots. On the AA-Robotics benchmark, Astra controls a robotic arm using its own code and scores 69 out of 100, the best result. But one task takes half an hour and costs more than 16 dollars. A person needs 20 seconds. With even more compute, Astra beat specialized Lakeside and FLUX 3 Action in Nvidia’s study, but the cost was about 36 dollars per task.

General-purpose models can already outperform specialized ones on some tasks, but they take more time and money to do it. Robotics researcher Georgia Chalvatzaki sees this as a real increase in capability, but not proof that scaling alone will solve robotics. The most impressive results do not come from the model alone: Astra uses simulators, motion planners, existing control software and reinforcement learning.

In Chalvatzaki’s assessment, general-purpose models are getting better at using and combining existing robotics tools. But the tools themselves—fast controllers, physics models and safety layers—still need to be built. The next few months should make clearer what this means for specialized labs.

Bet 3: The most valuable data is not on the internet

The third group is betting on data that cannot be found on the open internet: measurements from laboratories, samples from nature and engineering data from factories. Improvements in general-purpose models do not undermine this strategy; such models can be used as building blocks. The advantage is supposed to come from exclusive access to data. Three questions matter: who collects it, who can use it and how hard it is to reproduce. Basecamp and Periodic already have some of the data they need; Prometheus says it is creating its own.

Basecamp Research looks for genes in volcanoes and on the seabed

London-based Basecamp Research was founded in 2019, before the current wave of labs. Its example shows particularly clearly what happens when there is no internet-scale dataset. According to co-founder Glen Gowers, this is typical of biology: public gene databases mainly contain data about well-studied organisms.

Basecamp collects samples with local partners in more than 30 countries, from tropical forest soils and volcanic rock to the deep sea. The company says its dataset contains more than 100 billion previously unknown genes, access to which Basecamp controls. The Trillion Gene Atlas, announced in March, is meant to expand the database significantly; Nvidia and Anthropic are partners.

Basecamp uses the data to train its EDEN family of models. They design long DNA sequences and enzymes that insert them into the genome. Some capabilities are available to researchers through Claude. The first area of application is cell therapy: six programs are in the early stages of research and development, before clinical trials. The company has not said when trials might begin.

Periodic Labs uses nature as a training environment

Periodic Labs creates its own data. Liam Fedus led model fine-tuning at OpenAI. Ekin Dogus Cubuk headed the GNoME project at Google DeepMind, which predicted millions of new crystal structures in 2023. So far, only a few hundred have been confirmed in the lab. Periodic’s motto is that intelligence is necessary but not enough: scientific progress has to be tested against reality.

To do that, Periodic runs largely automated laboratories in Menlo Park. Robots mix powders, furnaces process them, and X-ray machines check the results around the clock. The process builds up data that is absent from textbooks, including negative results. Scientific journals publish successfully produced crystals but rarely describe failures. Fedus considers sequences of failed experiments a particularly valuable kind of data.

The company’s long-term goal is to create a superconductor that works at a higher temperature. Materials for the semiconductor industry are closer to generating revenue. The first named customer is a chipmaker dealing with a heat problem.

On September 15, Periodic introduced Neon: an existing model with a trillion parameters, fine-tuned on scientific literature and the lab’s own data. The company says Neon analyzes X-ray measurements better than GPT-6 Astra and Claude Fable 5.1, and costs less.

Prometheus is building an artificial engineer for Bezos

At Prometheus, Jeff Bezos has taken on an operating role for the first time since stepping down as Amazon CEO in 2021. He runs the lab alongside co-CEO Vik Bajaj, who led ambitious projects at Google X. According to GeekWire, about 150 employees have joined from OpenAI, Google DeepMind, Meta, xAI and Nvidia.

Prometheus is betting on a “general-purpose artificial engineer.” It has nothing to do with robotics, Bezos told CNBC. The company is building tools to make it easier for engineers to design physical objects—essentially a modern version of CAD. In the long term, the software is supposed to take an engine or a drug compound from concept to production. Bezos says the data needed for this still has to be created, for example from physical laws and test results.

Another Bezos project points to one possible way to obtain that data. According to the Wall Street Journal, in March he discussed a fund of about 100 billion dollars to buy and modernize industrial companies in aerospace, semiconductors and defense using AI. Over decades, such companies have built up large stores of engineering data, including CAD and simulation files. Licensing access to them is nearly impossible: companies are reluctant to hand over their internal digital knowledge.

Buying these companies could also provide access to their data. The fund would then be not only an industrial project but a possible way to obtain training data for Prometheus’s artificial engineer. This is speculation, not a confirmed plan.

According to reports, the lab raised 6.2 billion dollars and then another 12 billion, for a total of more than 18 billion. Of the labs covered in this edition, Prometheus has raised the most without presenting a product or published research.

Other labs working with their own data

Cambridge, Massachusetts-based Lila Sciences is closest to Periodic. In its “AI science factory,” AI proposes hypotheses, robots run experiments and the results feed back into training. The lab works in chemistry and biology. Its advantage is supposed to come from this cycle, not from the robots themselves. For now, the process still requires people.

London-based PhysicsX gets its data from simulations. Its data factory has run more than 20,000 aerodynamic simulations of car bodies. The company believes that the more varied the data, the better its models will handle new shapes. Competitors with enough computing resources could reproduce this approach.

Simile simulates people, not nature. The startup was founded by Stanford researchers behind “Generative Agents.” It builds behavioral models from interviews, purchasing data and controlled experiments with people. Companies use them to test how customers will respond to prices or products. Investors valued Simile at two billion dollars in July. A smarter general-purpose model alone cannot provide the same access to human behavior.

Colorado-based Enveda predates the current wave and applies a similar approach to natural compounds from plants. Its PRISM model was trained on 1.2 billion chemical fingerprints. Half were measured in Enveda’s laboratories; the rest came from public sources.

Big companies are building their own labs, too

Anthropic is building its own biology lab in the San Francisco area, where Claude will control robots during experiments. In April, the company acquired Coefficient Bio for about 400 million dollars. In October, Google described how its agents design and test protein binders in its own lab. Back in August, DeepMind’s Co-Scientist was planning experiments with 2D materials, an area close to Periodic’s core work.

Specialized labs sometimes gain a lead of several years from the samples, laboratories and partnerships they have assembled. Whether that advantage lasts depends on how difficult their infrastructure is to reproduce and whether the data stays private.

Bet 4: Computing power is becoming a constraint

World models, robots and automated labs ultimately run on the same chips. GPT-6 Astra was trained on more than 100,000 Nvidia GPUs, and Huang says another 400,000 will be added. The fourth group is exploring computing efficiency. The Transformer partly won because GPUs are good at the operations it needs. Those who want to change the approach have two options: adapt it to existing chips or build different ones.

Inception refines many parts of a text response at once

Inception was founded by Stanford professor Stefano Ermon, UCLA’s Aditya Grover and Cornell’s Vladimir Kuleshov. In 2019, Ermon and Yang Song introduced diffusion generative models, which became the foundation of today’s image and video generators.

ChatGPT writes one token at a time, waiting for the previous one each time. Inception’s Mercury models first create a draft of the whole response, then refine many parts of it simultaneously, much like an image generator gradually turns noise into a picture. The approach promises to reduce waiting time and costs. The Transformer is still at its core: the founders are challenging not the architecture itself but the current way of generating text.

Unlike most labs in this edition, Inception is already selling models through an API. Mercury 3 was released on October 19. The company says it outperforms the smaller GPT-6 Luna and Claude Haiku 5.5 on eight of 12 benchmarks and produces text four to eight times faster. Inception also offers Mercury Decide, a decision-making model that competes with TypeSafe. Google had already demonstrated Gemini Diffusion in May 2025. Reuters reported in May that Microsoft was in talks to acquire Inception; the company wanted more than a billion dollars.

Liquid AI builds models for resource-constrained devices

Liquid AI was founded by MIT researchers in 2023. The company develops compact models for devices with limited memory and computing resources, where a connection to a data center is too expensive or slow, or simply undesirable.

In most LFM2 blocks, the energy-intensive attention mechanism used by Transformers is replaced with cheaper operations that take a word’s immediate surroundings into account. According to the researchers’ LFM2 technical report, the model’s architecture is determined by the memory and computing resources of the device it will run on.

The new LFM3 generation came out on October 27, with support for text, images, video and audio. The models have open weights and are intended for local deployment and specialization. Since early October, Liquid has offered d1, a decision-making model in several sizes, partly with open weights, including for images. AMD led a 250 million dollar Series A round. Liquid has also started optimizing its models for AMD chips.

The company’s technology is reaching cars, too. On November 5, the Wall Street Journal reported that Mercedes began deploying it in North America in October. Voice commands for navigation, music and seat settings run directly in the car, while more complex requests are passed to cloud models.

Ricursive Intelligence is betting on self-improvement in chip design

Before founding Ricursive, Anna Goldie and Azalia Mirhoseini developed AlphaChip at Google, an AI method for placing components on chips. Google has used it across several generations of TPUs.

Ricursive is building on this with a feedback loop: AI designs better chips, which then power the next generation of AI. Conventional tools divide a chip into dozens of parts, while Ricursive wants to optimize the whole thing. In the long run, models and chips should be developed together, and a chip should take a week to design instead of months or years.

Two months after its launch, in January, Ricursive announced a 300 million dollar Series A round at a valuation of four billion. The company does not yet have a product.

On November 4, it released Ripple-65B Preview with open weights. Ricursive designed the architecture for Nvidia hardware to make the model more efficient to train and run. An internal version already runs on the company’s chip-design platform. Ricursive says Ripple is among the strongest US open models for coding tasks. There are no independent evaluations yet.

Competition is serious. Google has been using AlphaChip for several years. Nvidia optimizes designs with PrefixRL. In late September, OpenAI began working with Synopsys, one of the leading chip-design software developers, alongside Cadence and Siemens. Their joint GPT-Synopsys model is supposed to work independently with Synopsys tools and improve designs step by step. Ricursive is competing not only with the in-house research teams of chipmakers, but also with a partnership between a major AI lab and a market leader in chip-design software.

Unconventional AI wants to compute with oscillations

Unconventional wants to change the hardware itself. Founder Naveen Rao has already sold two companies: Nervana to Intel and MosaicML to Databricks. He expects energy consumption to become a hard limit for AI in about three years.

Today’s chips perform digital computation using billions of switching transistors. Unconventional proposes using analog, coupled ring oscillators: oscillating circuits affect one another until their rhythms align. This dynamic is supposed to handle part of the computation by combining memory and processing.

In June, the lab presented Un-0, a kind of demonstration version of an open image generator built on 16,384 simulated oscillators. Its quality is several years behind the current state of the art. The company received its first physical chip for image generation in September.

The plan is to build the first rack-scale system within two years. Unconventional does not want to sell chips; it plans to provide computing capacity from its own systems. The company is also acquiring other businesses: on October 26, it announced the purchase of Israeli AI-chip developer NeuReality. The NeuReality team will join the buyer; the price was not disclosed.

Other labs changing how computing works

Like Unconventional, Extropic wants to change the physics of computing, but in a different way. Its circuits deliberately allow electronic noise in transistors and use the resulting randomness in AI models. In summer, the company said its Z1 chip was in production, with the first systems expected in 2027. The promised efficiency gains of several orders of magnitude are based on simulations.

Normal Computing follows a similar principle and, for now, makes money from AI tools for chip design.

Cartesia was founded by the inventors of the Mamba architecture. Like Liquid AI, it replaces much of the resource-intensive attention mechanism, but mainly sells models for speech.

Bet 5: Research itself becomes the product

The fifth group automates researchers’ work: developing ideas, running experiments and analyzing results. The extent of automation can be described roughly in four stages.

Most publicly demonstrated results fall into the second stage. Some labs and large companies have reached the third stage in limited ways. Research on automated workflows shows that human decisions remain crucial. Automation has gone furthest where a machine can check its own results: mathematics.

Axiom Math builds machine-checkable proofs

Axiom aims to automate mathematical proofs. Karina Hong and Shubho Sengupta, both formerly at Meta, founded the company in 2025. Number theorist Ken Ono left his professorship to join the team.

A language model can write a convincing proof that is still wrong. Axiom works around the problem by having AI write proofs in Lean, a programming language for mathematics. A dedicated system checks every step. If a proof passes, it is valid under the definitions and assumptions used, leaving no room to bluff.

People still have to check whether the original claim was translated into Lean correctly and whether incorrect additional assumptions slipped in along the way. If the translation is wrong, the machine will flawlessly prove the wrong claim. Axiom itself describes this limitation.

On October 26, the company reported its biggest achievement: dozens of AI agents, guided by people, translated the so-called “enormous theorem” into Lean over 24 days. The proof spans more than 10,000 pages, hundreds of research papers and five decades of work. A research team had previously spent six years formalizing one section of about 250 pages using the Coq proof system. The code has been published on GitHub and runs to about 43 million lines.

During the work, the team found an error in a 1986 paper. Mathematician Timothy Chow said specialists knew about it, but it appears not to have been documented publicly. Formalization made explicit what only insiders knew.

Axiom plans to make money beyond mathematics. The company intends to use the 200 million dollars raised in its Series A round to bring the method to software: AI writes code, and a proof verifies that it meets formally defined requirements. Tests check individual examples; a proof can cover every case described by a formal model.

Discovery Loop was founded by researchers who left Google

Discovery Loop aims to automate the entire scientific experiment cycle, from coming up with an idea to analyzing the results, running thousands of experiments in parallel. The project is backed by four prominent former Google employees: Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals, one of the leaders of Gemini’s development. All four left in early August amid a Google reorganization. WIRED reported that the idea for the project was only a few weeks old.

Google became one of the first investors and a cloud partner. Business Insider reported talks over a seed round of about a billion dollars at a valuation of ten billion. Within weeks, the valuation being discussed had reached about 50 billion. Discovery Loop plans to start with machine-learning research, making itself the first customer. Projects in chip design, biology and materials science are meant to follow. For now, there is a team, a goal and an organization.

For Google, this is a way to stay connected to four leading researchers: the company remains an investor and cloud provider.

Mirendil wants to make frontier research accessible to more companies

Behnam Neyshabur led a team at Gemini and then headed Discovery at Anthropic. Co-founder Harsh Mehta built the first version of the platform Anthropic uses to automate research. Mirendil wants to make frontier research more accessible: teams at companies and research institutes should be able to build specialized systems for chemistry, medicine and robotics.

As its first result, the company showed in October how one researcher, working with AI agents, built the Mirendil-Base model with 100–500 billion parameters in three and a half months. Computing costs came to less than ten million dollars. The company says the agents wrote more than 99 percent of the code.

The product has two parts. The Fern 0.1 model is designed for coding and AI research agents, with a context window of up to ten million tokens. The Lor platform can run several hundred agents in parallel, continue experiments for days and use OpenAI and Anthropic models alongside Fern. Lor is in closed beta, while early access to Fern is available on request.

Recursive says it built its first model through a closed loop

Recursive Superintelligence, which is not connected to the Ricursive in the fourth group, wants to go further: automated research should improve itself. CEO Richard Socher was Salesforce’s chief scientist and founded You.com. One of the seven co-founders is Tim Rocktäschel, who led open-ended task and self-improvement teams at Google DeepMind.

Socher does not deny the progress made by major labs and believes AGI may already be close, depending on the definition. In his view, recursive self-improvement will drive further AI development. He ties the path to it to open-ended tasks, a principle resembling evolution: the system continually sets itself new tasks instead of optimizing for one fixed goal.

On October 21, Recursive released Flint, its first model for coding tasks. Its weights are available under the Apache 2.0 license. More important is how the company says it built the model: a small team used a “Eureka machine” for less than four months. The system proposes ideas, implements them, tests them and selects the next experiment. Socher says AI designed and improved almost every part of the model pipeline. It is not yet clear how much of the result came from automation rather than the work of the human team.

Other labs automating research

Tokyo-based Sakana AI shares Recursive’s belief that AI will take on its own research. Its co-founder Llion Jones was one of the authors of the paper on the Transformer. Since 2024, the company has regularly published results: an “AI scientist” that writes research papers, systems that rewrite their own code and, since June, a dedicated self-improvement lab. On the scale described above, this puts it at stages two to three of automation.

Palo Alto-based Harmonic, co-founded by Robinhood CEO Vlad Tenev, also uses its Aristotle model to prove theorems in Lean. In 2025, it reached gold-medal level at the International Mathematical Olympiad. Since October 27, Aristotle has also checked C software and Verilog chip designs, within the bounds of formally specified properties and assumptions. Axiom wants to build this line of business with money from its new round. In November 2025, investors valued Harmonic at 1.45 billion dollars.

Founded by former Google DeepMind researchers, Poetiq does not train its own models. Instead, it improves the prompts and strategies existing models use to solve problems, running one round of optimization after another. The system around the base model improves, not the model itself.

Lila Sciences, in the third group, also belongs here: it brings together hypothesis generation, experiments and training in a single loop.

Big companies are automating research, too

The new labs are not the only ones automating research. In September, OpenAI said it had achieved its goal of creating an “automated research intern.” By March 2028, the company wants to have a fully automated AI researcher. In June, Anthropic said Claude writes more than 80 percent of the new code incorporated into production systems.

Big companies are also moving into the formal proofs Axiom works on. In September, OpenAI published a proof for a version of the Navier–Stokes problem, one of seven famous mathematical problems that each carry a one-million-dollar prize. The proof has not yet been accepted. More than 700 manuscripts followed in October.

The growth of automation is also intensifying concerns. During internal safety tests in July, OpenAI agents escaped the test environment and broke into real Hugging Face systems. More than 1,300 employees of AI companies signed an open letter calling for tools that could slow autonomous AI development when necessary. The labs in this group face the same safety question but get much less public attention.

Bet 6: You can win without building the first general-purpose model

The sixth group is not trying to create the strongest model. Its labs believe that even the most powerful system will leave gaps, and that customers will pay to fill them. Companies need to understand what happens inside a model, trust its decisions, work with it, adapt it to their field and run it themselves. Unlike the labs in other groups, every company listed here already has products.

Goodfire studies what a model has learned

Goodfire wants to understand which internal processes lead a model to an answer and intervene at that level. Co-founder Tom McGrath previously led a research team at Google DeepMind. The field is called mechanistic interpretability. A chatbot can explain its answer, but the explanation does not necessarily reflect how it actually arrived at it. Goodfire studies a model’s internal representations. In an interview with Sequoia, the company’s CEO, Eric Ho, said Goodfire’s 2025 Series A was Anthropic’s first direct investment in another company.

The team showed its first results in February with RLFR. Instead of asking a second language model to evaluate responses, Goodfire uses a model’s internal signals to provide feedback during further training. In a test using Google’s open Gemma 3 with 12 billion parameters, hallucinations fell by 58 percent. Goodfire says each intervention cost about 90 times less than using an external evaluator model. There is a caveat: the measurement also includes monitoring responses during operation, correcting them and selecting the best of 32 options.

On November 5, Goodfire introduced mechanistic behavior shaping. The idea is that a model can arrive at the right answer for the wrong reason. Goodfire identifies misleading signals inside the model and penalizes it during training for relying on them. In experiments, this reduced thoughtless agreement with users and cheating on tasks. But internal signals may change as training continues, so the tools used to monitor them may also need retraining.

Through Silico, customers can analyze their models and precisely change their behavior. Goodfire’s partners include Microsoft, the Mayo Clinic and Arc Institute.

TypeSafe AI returns decisions, not text

Language models answer in free-form text, but many business processes require a choice from a set of predefined options and an estimate of confidence. That is the idea behind TypeSafe AI, founded by former OpenAI researcher Diogo Almeida. Its Jev model returns only fixed decisions with probabilities—for example, choosing a team to handle a support request.

The probabilities are meant to be calibrated: if the model gives an 80 percent confidence score, it should be right about 80 percent of the time in similar cases. A company can set a threshold above which the decision is passed to a person.

Competitors arrived quickly. OpenAI’s Decisions API has been available to everyone since October 22. Cloudflare has its own options under the name Clef. Two new labs are also in the race: Inception says Mercury Decide outperforms Jev on several benchmarks, while Liquid AI says its d1 decision-making model beats Jev in its own tests.

Despite the competition, TypeSafe has not been displaced so far. On October 9, the company raised 870 million dollars in a Series A round at a valuation of 7.5 billion.

humans& is building AI to work alongside people

humans& wants to build models that reflect how people actually speak and behave: inconsistently, contradictorily and irrationally. That should make AI a better partner for working together. Its founders include Eric Zelikman of xAI and Andy Peng of Anthropic.

The company has three products. Persimmon creates realistic conversation partners. Developers specify a situation and user profiles, and the model generates corresponding responses, such as customers who lose patience or change their minds. Creators of support chatbots can test them on a thousand simulated customers. Access is restricted because of the risk of misuse.

The Ampersand model is designed for developers. It handles conversations with several participants, detects conflicts between their preferences and connects to applications through a paid API. Ampersand was trained using conversation partners simulated by Persimmon.

Boba is built on top of these technologies and works as a ready-to-use assistant for personal and group chats: it can schedule meetings, remind people about tasks and search the web.

humans& did not create its own base model. Ampersand runs on Z.ai’s open Chinese GLM 5.3. Not every new lab needs to build a model from scratch.

Thinking Machines Lab uses products to stay close to real-world practice

Thinking Machines is developing models and tools to help companies and researchers train AI on their own knowledge. Mira Murati was OpenAI’s CTO, while John Schulman was a company co-founder and one of the creators of ChatGPT. Several other co-founders have already moved to OpenAI or Meta.

Since October 2025, customers have been using Tinker to fine-tune open models for their own tasks without managing the infrastructure themselves. In July, the company introduced its Inkling family of open-weight models with 975 billion parameters. On benchmarks, it trails Reflection’s Beam and Meta’s Muse Spark Open.

Inkling 2 Small came out on November 5, with 594 billion parameters and support for text and image data. It can be used and fine-tuned through Tinker. Its weights are still closed; publication under the Apache 2.0 license is planned after external safety testing.

On the Artificial Analysis Intelligence Index, Inkling 2 Small scored 31 points, compared with 13 for Inkling Small, matching Google’s Gemini 3.8 Flash at its highest reasoning level. If its weights are released, it will become the highest-scoring open model developed outside China in the current rankings.

According to The Economist, the company’s latest annualized revenue exceeded 200 million dollars. On November 6, Business Insider reported that the lab was close to raising five billion dollars at a pre-money valuation of 40 billion, with Accel expected to lead the round. There is no information that the deal has closed.

The lab’s main argument is that general intelligence does not automatically contain knowledge scattered across companies and specialized fields. Whether this idea holds up will depend on how useful fine-tuning remains as base models improve.

Reflection builds open models for people who want to control them

Reflection develops open models that companies and government organizations can run themselves, without sending data to OpenAI or Google. Micha Laskin and Ioannis Antonoglou came from Google DeepMind; Antonoglou worked on AlphaGo. Laskin positions Reflection as a Western counterpart to DeepSeek, Qwen and Kimi.

On October 16, Reflection published Beam’s weights under the Apache 2.0 license. The company says Beam is comparable to Chinese GLM 5.2 on reasoning tasks but requires between one-third and one-quarter of the computing resources. Artificial Analysis considers Beam competitive among open models, but not a leader.

Reflection has raised billions of dollars. The 25 billion dollar figure mentioned in reports is the company’s target valuation before a new round. According to the Financial Times, Nvidia has been discussing a purchase or a large stake since October. Neither report has been confirmed.

Other labs building similar products

Guide Labs works on traceability earlier in the process than Goodfire. It builds models from scratch so answers can be linked to specific concepts and training data. Its open Steerling-8B model uses about 33,000 predefined concepts. This does not provide full transparency; and with roughly nine million dollars in seed funding, Guide is playing in a different weight class.

Reflection’s closest counterpart is Arcee AI. The startup, with about 30 employees, trained Trinity Large Thinking, a model with about 400 billion parameters, and released it under the Apache 2.0 license. In September, the company raised funding at a valuation of more than a billion dollars.

Why Nvidia wins whichever bet pays off

Researchers leaving does not prove that AGI is far away. Many founders believe far more powerful AI is already close. But their decisions show that even the people who built today’s systems think the path ahead is less predictable than major labs suggest.

At the same time, Huang, who declared AGI achieved, is investing in labs that question the current approach. Nvidia has invested in at least 25 of the 45 labs listed and has confirmed technology or research partnerships with four more. Its investments span all six areas.

The contradiction is smaller than it seems. As long as competing approaches run on Nvidia chips and cloud services, the company does not have to guess the winner in advance. The investments also help secure future customers: some of the money Nvidia puts in comes back as spending on compute. The same applies to Google, Amazon and AMD.

What could happen to the new labs

Base case

Funding remains high, but only a small share of the labs produce results within two to three years that can be shown to outperform existing solutions. Major players keep adopting new methods within weeks, as in the case of TypeSafe and OpenAI’s Decisions API. Many new labs are acquired or have their teams hired wholesale.

There are already early examples: AMD is buying World Labs, Nvidia is in talks with Reflection, and Microsoft is in talks with Inception. Labs with hard-to-reproduce data and infrastructure, such as Basecamp, Periodic and Physical Intelligence, may prove more durable, as may companies with revenue, like Thinking Machines. Nvidia and cloud providers benefit regardless of which bet pays off.

Acceleration

One of the key ideas leads to a major breakthrough. For example, a model that continually learns while running appears, as in the Ceres system Core has said it is building, or research automation really does close the loop, as Recursive and Discovery Loop hope.

AI research agents that lower development costs could help. Mirendil-Base, built for less than ten million dollars, and Flint, developed in less than four months, point in that direction. Proprietary labs and data from nature could provide another advantage. Small teams could then compete with major players.

Lab valuations continue to rise, and large companies buy them more often than they build equivalents of their own. At the same time, safety questions become more urgent. OpenAI agents have already entered Hugging Face systems during testing, and more than 1,300 AI-company employees are calling for tools to slow development. New labs will face the same questions, but with much less public scrutiny.

Slowdown

Labs without products, such as SSI, Prometheus, Flapping Airplanes and Ineffable, may fail to show results before investors run out of patience. Meanwhile, general-purpose models like Astra will cover more specialized tasks, and the cost of each task will fall. Early signs are already visible: Astra solves ARC-AGI-3, beats specialized robotics models when given enough compute, and OpenAI is developing formal proofs.

Other potential brakes include a downturn in funding, regulation of autonomous systems and the energy limit Naveen Rao expects in about three years. Possible consequences include lower valuations, rushed asset sales and further industry consolidation around a handful of major labs, chipmakers and cloud companies. The diversity of research approaches would shrink.

Dataist’s take: the base case of consolidation looks most likely. New methods are adopted quickly, acquisitions are already happening, and some investments return to investors as spending on compute. Acceleration is possible for individual teams working in areas with verifiable results, such as materials research. For some labs without products, lower valuations or acquisition along with the team are likely within two to three years, even if interest in new research approaches persists.