Redshift > Blog

Introduction

Welcome! This post outlines a preliminary vision for a Bell Labs-style "idea collider", focused on solving the world's most challenging problems and bringing innovations that benefit all. This sounds ambitious, and it is. But I'll try to argue that Bell Labs offers a precedent that can plausibly be adapted from a 19th century state monopoly to a 21st century startup.

Three ways

First, let's ask the question: what is best way to solve real-world problems in science and technology? Broadly speaking, there are two approaches. The first is academic. It occurs in universities, and is funded by grants to (often tenured or tenure-track) principal investigators. These investigators may be part of larger collaborations, consortia, or research initiatives. But since hiring and funding tend to reward individual publication metrics, Goodhart's Curse tells us that these will be maximized. The focus on individual publication record leads to siloing and an obvious bias against work which is practical, messy, or heuristic, thereby precludes advances on concrete problems in the real world.

The second approach is private enterprise. In principle, this is better placed to solve messy problems, since a for-profit operation can command more capital, has strong incentives against siloing, and must yield practical outcomes if it hopes to survive. But here Goodhart's Curse cuts even deeper, since the pressure to generate revenue limits research horizons and timelines. Solutions may well be profitable, but are likely to be suboptimal, both from the research perspective and that of distributive fairness.

Is there a way to enjoy the benefits of both? Industrial-academic partnerships are one option, but since these involve two parties operating under separate constraints, it is some ways the worst of both worlds rather than the best. (Naturally, there are exceptions.) More promising, perhaps, is the industrial R&D lab. These vary widely in scope, operation, and profit motive; we'll focus on the "blue sky" efforts which lie in the tail of the innovation distribution. The goal here is usually major competitive advantage from an unexpected breakthrough, but because these breakthroughs are rare, blue sky research is risky and usually only attempted by larger companies which can absorb the possible losses. Modern examples range from the Skunkworks projects of the aerospace sector to the moonshot factories of Big Tech.

The Bell Labs recipe

However innovative these modern exemplars, history shows a clear outlier: Bell Labs. After Alexander Graham Bell invented the telephone in the 1870s, he formed a company to develop a continental telephone system for the United States. Constructing the national infrastructure for a new technology, from scratch, naturally required a research division to solve the many questions that arose. By the mid-20th century, this research division—Bell Labs—had became the largest and most productive industrial research facility in the world.

Let's reel off a few of its successes: the transistor, the laser, C, UNIX, photovoltaic cells, quality control, information theory, modern cryptography, error-correcting codes, support vector machines, charged couple devices, radio astronomy, electron diffraction, optical fibres, and the Fast Fourier Transform, to name a few. No other R&D lab in history boasts so many Nobels, Turing awards, or IEEE prizes. No other lab touches so many branches of fundamental and applied science. It didn't invent the US telephone system; it invented the 20th century.

In the '00s, Bell began to crumble—undermined by its own inventions—and Bell Labs crumbled with it. We'll ask why it failed below, but first: why was it was so successful for so long? It's easy to point to the monopoly, or the telephone system itself, but I think it was due to a number of factors:

Domain. The lab was embedded in a problem-rich domain of application.
Mandate. The lab was required to effectively solve the problems in this domain.
Freedom. Researchers had freedom to pursue and initiate research projects connected to the domain.
Stability. Researchers did not have to apply for funding, and were encouraged to think long-term.
Interaction. The environment was deliberately collaborative and interdisciplinary.
Separation. Research was strictly separated from development/product, ensuring total focus.
Feedback. Results from the domain, including development/product, fed back into research.
Talent. The lab hired researchers of high academic calibre who were a good cultural fit.
Structure. Managers and executives collaboratively enable all of the above.

You can read more about why I selected these factors here.

Monopoly isn't enough

A state monopoly with a broad infrastructural mandate helped realize each of the factors above, but wasn't strictly necessary. It seems clear that, at least in principle, Big Tech could do it today. But while subsidiaries like Google X work on "blue sky", high-impact projects, they are cartoonishly divorced from Google's operational concerns and, it seems, designed by committee in advance. They are neither adequately coupled nor adequately free.

So why don't we see Bell Labs everywhere in Big Tech? The answer is probably that, by sharing its innovations, Bell Labs put itself out of business, and Big Tech very much wants to stay in business. DeepMind is a case in point. Formed in the UK in 2010, the company did groundbreaking work in AI and reinforcement learning, particularly in the context of games, culminating in AlphaGo's defeat of Lee Sedol in 2015. DeepMind was acquired by Google the year before, and despite its bids for autonomy, was recently merged with Google Brain to form Google DeepMind. Its focus has now almost entirely shifted away from the open, heterogenous AI research that made it famous, to closed R&D for large language models like Gemini. This reflects Google's frantic struggle to catch up in the language model arms race.

DeepMind could be given a long leash when times were good, but once executives started to panic, blue sky became a luxury the company couldn't afford. This clearly shows the value of a monopoly for industrial research: it stops the market from spilling into the laboratory. But if a giant like Google—which, according to the US Department of Justice, does have a monopoly—can't manage it, how can anyone else?

Everything in its right place

There is a subtle difference between Bell Labs and DeepMind which suggests that corporate governance and organizational structure are more important than monopoly per se. DeepMind was acquired as an unprotected asset that could later be mined for talent; Bell Labs was operationally essential from the beginning and so operationally protected until the end. In terms of our criteria, DeepMind fails the requirement of mandate, and Google that of structure.

Another cautionary tale is OpenAI. With its track-record of "repeatable innovation" and cross-disciplinary insight, it may be the closest modern analogue of Bell Labs. The company started out as a non-profit entity whose goal was to research AI, safely, independently, and openly, for the good of all humanity. I would guess that it initially ticked our boxes. But it has seemingly become a victim of its own success, transitioning, after the release of ChatGPT, to a closed, capped-profit entity enmeshed with Microsoft. This might have left the research culture intact for all I know, but the ongoing exodus of senior figures suggests otherwise.

Time will tell whether these changes are in fact necessary for the company to achieve its mission. The point is that, either way, the initial structure could not bear the weight of ChatGPT's unexpected commercial success. At Bell, the research had a commercial outlet from the beginning; there was no need to choose between open, non-profit research and closed, for-profit product, because the two already existed in distinct places.

The virtuous cycle

The history of Bell, Google and OpenAI suggest that, for repeatable innovation and long-term success, the industrial lab ("Research") should be paired from the outset with a commercial wing ("Product"). Product supplies capital and contact with the problem domain, and receives the mature outcomes of Research in return. This arrangement can only work on the culturally- and institutionally-enforced understanding that Product does not interfere with Research.

To get the ball rolling, there needs to be something for Product to commercialize, and Research to understand, analogous to the telephone system at Bell or the transformer at OpenAI. I propose three projects:

StateCraft. A high-level functional language for quantum programming.
Invisible Ink. A tool for watermarking large language models at both training and output stages.
Paχ. A partially centralized cryptocurrency.

Each project couples a research mandate to a product; in turn, the projects are in overlapping domains and lend themselves to cross-disciplinary sequels. The goal of Redshift is to implement these initial projects and see if our expectations pan out. If so, we hone the methodology, scale up, and hope to become a Bell Labs-style center for repeatable innovation and creative applied research. And by honing culture, structure, and corporate governance, I think we can avoid the issues that plague blue sky innovation in Big Tech.

I'll end with a draft mission statement:

Our mission at Redshift is to solve the world's most challenging problems, with a spirit of creative freedom, technical excellence, and interdisciplinary play, making innovations that benefit all.

You can read more about the statement here. But in it, I've attempted to distil my take on what made Bell Labs successful, and what a modern recreation ought to do with its success.

So ends Part I! In the next installment, we'll dive deeper into:

Funding models for sustainable innovation

Strategies for attracting and retaining talent

Fostering collaborative teamwork and management

Navigating ethical considerations and regulatory compliance

Defining success metrics that avoid Goodhart's Curse