Who is Babak Hassibi?
The race in AI is usually told as a story of things getting bigger, and Babak Hassibi is betting his company on the opposite move, making models small enough to live on the phone in your pocket. He is a professor at the California Institute of Technology and the co-founder and chief executive of PrismML, a startup working on extreme compression of large language models. A respected academic in information theory, signal processing, and control, he brought decades of theoretical work on how to represent information efficiently to the problem of shrinking AI models without breaking them. PrismML emerged from stealth in March 2026 with a family of one-bit language models, an approach that compresses a model so aggressively that its weights are reduced almost to single bits, which in turn lets capable models run on ordinary consumer devices. By the middle of 2026 the company was reported to be in talks with Apple about running larger models directly on hardware, a discussion that placed Hassibi academic work at the centre of one of the most valuable questions in consumer technology, namely how much intelligence can be pushed onto a device that fits in a hand.
What does Babak Hassibi think about AI?
Hassibi approaches AI as an information-theoretic and engineering problem, asking how much a model can be compressed while preserving the behaviour that makes it useful. His work rests on the conviction that the future of practical AI is on-device, private, and efficient, and that the path there runs through better representations rather than through ever-larger clusters of accelerators. He treats efficiency not as a compromise that sacrifices quality but as the enabling condition for putting capable models directly into people hands, since a model that runs on a phone needs no data center, no constant connection, and no handing of private data to a remote server. That framing connects his decades of academic study of how information can be encoded and transmitted with a very current commercial goal, and it puts him among the researchers who believe the most important gains left in AI are in making it smaller and cheaper to run rather than only larger.
What is Babak Hassibi’s role in the AI race?
Hassibi role is at the frontier of edge and on-device AI. While much of the industry races to build and serve enormous cloud models, PrismML pushes in the opposite direction, trying to make strong models small enough to run on personal hardware without a network connection. If that effort succeeds, it changes where AI runs and who controls it, moving computation and data off remote servers and onto the devices people already own. His academic standing gives the one-bit approach a seriousness it might not otherwise command, because compression claims are easy to make and hard to deliver, and a proposal coming from a Caltech information theorist carries a different weight than one from an anonymous startup. That credibility is part of why the possibility of running large models on a phone moved quickly from a research curiosity to a topic of serious commercial discussion.
Where does Babak Hassibi work?
He is a professor at Caltech and the co-founder and chief executive of PrismML, where research is co-led by figures including Sahin Lale and Omead Pooladzandi. The combination of an academic post and a company is central to how the work proceeds, since the theoretical foundations that make extreme compression possible come out of exactly the kind of long-horizon research a university supports, while the company provides the means to turn that theory into products aimed at real devices. Hassibi leads the technical direction of PrismML while keeping his academic base, which keeps the company close to the mathematics its approach depends on.
What are Babak Hassibi’s key projects?
His central project is PrismML family of one-bit large language models, aimed at commercially viable, on-device inference on phones and other consumer hardware. This builds on a long academic career of contributions to information theory, estimation, and signal processing, fields that study precisely how information can be represented and recovered under tight constraints, which is the theoretical heart of what PrismML is trying to do in practice. The company work is to take those principles and make them hold up in a shipping model, so that a heavily compressed system still answers usefully rather than degrading into noise, and much of the challenge lies in that gap between what is possible in theory and what survives contact with a real product.
What has Babak Hassibi written about AI?
Hassibi written output is primarily academic, a large body of peer-reviewed research in information theory, control, and machine learning accumulated over a long career. His public voice on AI comes through that scholarship and through PrismML technical communications rather than through opinion writing or public commentary on the direction of the industry. He is a researcher first, and the clearest statement of his thinking is in the technical results, which is fitting for a founder whose company rests on a mathematical claim about how far models can be compressed.
Does Babak Hassibi think humanity will survive AI?
Hassibi has not made public statements about existential risk or humanity long-term survival. His focus is technical and mathematical, centred on compression and on-device inference, and it would misrepresent him to attribute a position on long-term AI catastrophe to him based on the record. What can be said is narrower and better grounded. His work points toward keeping AI capabilities private and local, on the user own hardware, which is a stance about control and privacy rather than a forecast about how the broader story of AI and humanity resolves.
Is Babak Hassibi a transhumanist?
There is no public evidence that Hassibi identifies as a transhumanist. He is best understood as an academic engineer working on the efficiency and compression of models, not as an advocate of human enhancement or a merger of people and machines. His career sits firmly in the mathematics of information and signals, and any attempt to attach a transhumanist philosophy to him would go well beyond his published work and public statements, which stay within the technical questions of how to make models smaller and more efficient.