Boolsi

Circuits that are grown, not designed: why we led BoolSi’s Seed round

Fine Structure Ventures is proud to lead the $6M Seed round of BoolSi, alongside our friends at Pillar VC, to reimagine how digital logic circuits come into existence.

By David Needell

                                                                                                                                                 

The Bitter Lesson, Again

In 2019, the reinforcement learning pioneer Richard Sutton wrote a short essay that has since become canon in the AI community.  He called it “The Bitter Lesson,” and its thesis uncomfortably admits the limitation of our own abilities: systems that can learn on their own outperform ones hand-crafted with human knowledge and intuition.   Chess engines built on grandmaster techniques lose to those learned through self-play.  Bespoke computer vision models underperform compared to deep-learning neural networks.  The bitter lesson articulated by Sutton asserts that our attempt to build artificial intelligence from our own understanding of the universe is fundamentally limited, that history tells us such an endeavor is fraught, and that the only way to build ever-complex systems is to let the program search and learn through the “universe” of solutions.

Part of what makes this such a bitter lesson is that, unfortunately, we seem to keep having to relearn it.  We are drawn; irresistibly, intuitively, and often unknowingly; to intelligent design.  We yearn to build systems that mirror how we think a problem should be solved.  We favor the human-centric approach.  But throughout the history of AI—and advanced computation in general—we keep re-discovering the extraordinary power of systems that can search and learn on their own, that can form their own “views”, views whose structures and mechanisms evolve from interaction with the environment itself.  The question, therefore, becomes: where else must we re-learn this lesson?  In which aspects of our computational systems lie that hidden influence of our own intuition?  How are we biasing the search space?

Arguably, the $12T chip industry is in the middle of re-learning this lesson right now.

Problem #1: More Transistors Than We Know What to Do With

A subtle crevasse has been fracturing through the semiconductor ecosystem.  On one side, on-chip transistor densities continue their exponential march, doubling in count every 1.5 years, with today’s state-of-the-art processors exceeding 50 billion transistors per chip.  While we can pack more computational potential onto silicon than ever before, our ability to meaningfully process data has fallen behind—the other side of the crevasse.  In logic circuitry alone, as an illustrative example of the complexity of this problem, there exists around  possible designs simply for adding two, n-bit numbers together.  With so many transistors to play with, the problem now becomes how can we exponentially grow the sophistication of our logic circuitry, our floor planning, and our chip architecture that makes use of this raw power.

The industry’s answer is, to no one’s surprise, AI (and more specifically transformer-based models).  Incumbent electronic design automation giants and a wave of emerging, well-capitalized startups are racing to point large language models and other generative systems at this problem of logic circuitry and chip design.  The vision: GPTs that write description code for logic circuitry, propose layouts, and hunt for errors that inevitably surface, all on their own, all automated, an end-to-end solution.

And herein lies the bitter lesson.

Problem #2: Intelligent Design, Now with Machine Intelligence

Whether a transformer-derived neural network could ever evolve into a form of general intelligence remains, for the moment, science fiction.  Such a question may be flawed from the start; as it could be better suited for philosophy rather than science. Regardless, it is undeniable the level sophistication seen in today’s GPT-based algorithms.  Large language models and the mathematics that underpin them have changed the way most of us seek information, navigate the world, and engage with our computers.  We ask these models to plan a vacation, to write an email, to read an article.  It is natural to wonder what else such models could do.

In walks the GPT-based circuit designer.

Let us imagine an artificial circuit designer, one trained from looking into myriad human-assembled logic circuits.  The artificial designer generates new approaches, layouts, searching and learning through the phase space of its training dataset.  At first glance, such an approach would seem to align with Sutton’s point: letting the machine search and learn for itself.  But one could also argue that such a designer does not escape the intelligent design trap—it merely swaps the designer.  Instead of a human engineer hand-crafting a circuit from intuition, a transformer model assembles a circuit element by element, each placement drawn from a probability distribution at its output layer, all trained on human-generated datasets.  The result still reflects the designer’s intuition, its internal logic.  In effect, we have simply replaced human intelligent design with machine intelligent design.

BoolSi: Circuits Grown from a Software Environment

In 2022, Dr. Mihailo Isakov became infatuated with an idea in the aftermath of his PhD research (research in which he was building custom logic circuits for data center workloads).  The idea that captivated his imagination: What if a neural network could evolve into the circuit itself?  What if we could make every artificial node representative of a digital logic gate (AND, OR, NOT), and what if the nodal connections could evolve into the wires between gates?  As training continues, the network would not output a circuit design.  Rather, the network itself would crystallize into one.

Two years later in December of 2023, Mihailo started BoolSi (Boolean + Silicon, pronounced “bullseye”).

The novelty behind BoolSi lies in both the mathematics involved in the training algorithm as well as the environment for which to train.  In the latter, we return to Sutton’s key point in the bitter lesson.  By using supervised training to learn a digital twin of some C/C++ code, for example; by letting the neural network sweep through the input/output space, evolving against the environment in which it occupies; the circuit is not architected by any intelligence, human or machine.  The circuit is grown from the world in which it inhabits, from the code that defines its purpose.  And when training converges, when the circuit is complete, the result is not a probabilistic suggestion.  The result is deterministic, an analytical circuit.  It is letting a network learn how to play chess on its own.  It is searching and learning.  It is invoking the evolutionary process in the environment itself, recognizing that such an approach will phenomenologically yield the greatest performance in circuit design.

The early evidence is striking.  Where today’s GPT-based, circuit-design approaches oscillate in accuracy indefinitely (a signature of an irreducibly stochastic system), BoolSi’s circuits reach 100% accuracy and never deviate.  In recent demonstrations, a BoolSi-generated circuit ran up to 63x faster than the same function executing on a state-of-the-art ARM core (a 10,000-line C regex library with BoolSi’s generated accelerators against a gcc -O3).

Where Could Processors Go

BoolSi’s beachhead is the FPGA, the field-programmable gate array, an integral component to deployed robotics, industrial automation, telecom, defense systems, and many others.  FPGAs offer enormous flexibility, but exploiting their advantages requires translating high-level software into custom digital logic (a skill held by far too few engineers relative to the software developers who need it).  BoolSi’s novel approach lets a software team go directly from code to an optimized, verified circuit, collapsing a talent bottleneck that gates the adoption of programmable silicon across entire industries.

From there, the roadmap extends naturally: a growing library of proprietary, machine-evolved circuit designs; embedded system-on-module architectures; and, ultimately, BoolSi’s own custom silicon processors.

But the reason we led this round runs deeper than any single market.  For decades, hardware and software have lived on opposite sides of a translation barrier.  For decades, armies of engineers (and now armies of GPTs) have been laboring to carry meaning across it, to bridge the divide between hardware and software through intelligent design.  BoolSi offers a world where that barrier dissolves entirely, a world where the neural network and the logic circuit are not two artifacts but one, where intelligence is not something that designs the hardware but something that is the hardware.

Richard Sutton’s lesson was that discovered systems beat designed ones.  BoolSi is what happens when you apply that lesson all the way down, past the algorithm, past the design tool, down into the silicon itself.

We could not be more excited to partner with Mihailo and the BoolSi team on that journey.

Fine Structure Ventures | An F-Prime fund invests in early-stage deep science companies across energy and electrification, aerospace and defense, next-generation computing, and materials and manufacturing.