LeCun founded Meta’s Fundamental AI Research lab, known as FAIR, in 2013 and has served as the company’s chief AI scientist ever since. He is one of three researchers who won the 2018 Turing Award for pioneering work on deep learning and convolutional neural networks. After leaving Meta, LeCun will remain a professor at New York University, where he has taught since 2003.
LeCun has previously argued that large language models like Llama that Zuckerberg has put at the center of his strategy are useful, but they will never be able to reason and plan like humans, increasingly appearing to contradict his boss’s grandiose AI vision for developing “superintelligence.”
For example, in May 2024, when an OpenAI researcher discussed the need to control ultra-intelligent AI, LeCun responded on X by writing that before urgently figuring out how to control AI systems much smarter than humans, researchers need to have the beginning of a hint of a design for a system smarter than a house cat.
Mark Zuckerberg once believed the “metaverse” was the future and renamed his company because of it. Credit: Facebook
Within FAIR, LeCun has instead focused on developing world models that can truly plan and reason. Over the past year, though, Meta’s AI research groups have seen growing tension and mass layoffs as Zuckerberg has shifted the company’s AI strategy away from long-term research and toward the rapid deployment of commercial products.
Over the summer, Zuckerberg hired Alexandr Wang to lead a new superintelligence team at Meta, paying $14.3 billion to hire the 28-year-old founder of data-labeling startup Scale AI and acquire a 49 percent interest in his company. LeCun, who had previously reported to Chief Product Officer Chris Cox, now reports to Wang, which seems like a sharp rebuke of LeCun’s approach to AI.
Zuckerberg also personally handpicked an exclusive team called TBD Lab to accelerate the development of the next iteration of large language models, luring staff from rivals such as OpenAI and Google with astonishingly large $100 to $250 million pay packages. As a result, Zuckerberg has come under growing pressure from Wall Street to show that his multibillion-dollar investment in becoming an AI leader will pay off and boost revenue. But if it turns out like his previous pivot to the metaverse, Zuckerberg’s latest bet could prove equally expensive and unfruitful.
Lots of startups use Google’s productivity suite, known as Workspace, to handle email, documents, and other back-office matters. Relatedly, lots of business-minded webapps use Google’s OAuth, i.e. “Sign in with Google.” It’s a low-friction feedback loop—up until the startup fails, the domain goes up for sale, and somebody forgot to close down all the Google stuff.
Dylan Ayrey, of Truffle Security Co., suggests in a report that this problem is more serious than anyone, especially Google, is acknowledging. Many startups make the critical mistake of not properly closing their accounts—on both Google and other web-based apps—before letting their domains expire.
Given the number of people working for tech startups (6 million), the failure rate of said startups (90 percent), their usage of Google Workspaces (50 percent, all by Ayrey’s numbers), and the speed at which startups tend to fall apart, there are a lot of Google-auth-connected domains up for sale at any time. That would not be an inherent problem, except that, as Ayrey shows, buying a domain with a still-active Google account can let you re-activate the Google accounts for former employees.
With admin access to those accounts, you can get into many of the services they used Google’s OAuth to log into, like Slack, ChatGPT, Zoom, and HR systems. Ayrey writes that he bought a defunct startup domain and got access to each of those through Google account sign-ins. He ended up with tax documents, job interview details, and direct messages, among other sensitive materials.
You have to close up shop, not just abandon it
Reached for comment, a Google spokesperson provided a statement:
We appreciate Dylan Ayrey’s help identifying the risks stemming from customers forgetting to delete third-party SaaS services as part of turning down their operation. As a best practice, we recommend customers properly close out domains following these instructions to make this type of issue impossible. Additionally, we encourage third-party apps to follow best-practices by using the unique account identifiers (sub) to mitigate this risk.
Notably, Ayrey’s methods were not able to access data stored inside each re-activated Google account, but on third-party platforms. While Ayrey’s test cases and data largely concern startups, any domain that used Google Workspace accounts to authenticate with third-party services and failed to delete their Google account to remove its domain link before selling the domain could be vulnerable.
Enlarge/ A single logical qubit is built from a large collection of hardware qubits.
One of the more striking things about quantum computing is that the field, despite not having proven itself especially useful, has already spawned a collection of startups that are focused on building something other than qubits. It might be easy to dismiss this as opportunism—trying to cash in on the hype surrounding quantum computing. But it can be useful to look at the things these startups are targeting, because they can be an indication of hard problems in quantum computing that haven’t yet been solved by any one of the big companies involved in that space—companies like Amazon, Google, IBM, or Intel.
In the case of a UK-based company called Riverlane, the unsolved piece that is being addressed is the huge amount of classical computations that are going to be necessary to make the quantum hardware work. Specifically, it’s targeting the huge amount of data processing that will be needed for a key part of quantum error correction: recognizing when an error has occurred.
Error detection vs. the data
All qubits are fragile, tending to lose their state during operations, or simply over time. No matter what the technology—cold atoms, superconducting transmons, whatever—these error rates put a hard limit on the amount of computation that can be done before an error is inevitable. That rules out doing almost every useful computation operating directly on existing hardware qubits.
The generally accepted solution to this is to work with what are called logical qubits. These involve linking multiple hardware qubits together and spreading the quantum information among them. Additional hardware qubits are linked in so that they can be measured to monitor errors affecting the data, allowing them to be corrected. It can take dozens of hardware qubits to make a single logical qubit, meaning even the largest existing systems can only support about 50 robust logical qubits.
Riverlane’s founder and CEO, Steve Brierley, told Ars that error correction doesn’t only stress the qubit hardware; it stresses the classical portion of the system as well. Each of the measurements of the qubits used for monitoring the system needs to be processed to detect and interpret any errors. We’ll need roughly 100 logical qubits to do some of the simplest interesting calculations, meaning monitoring thousands of hardware qubits. Doing more sophisticated calculations may mean thousands of logical qubits.
That error-correction data (termed syndrome data in the field) needs to be read between each operation, which makes for a lot of data. “At scale, we’re talking a hundred terabytes per second,” said Brierley. “At a million physical qubits, we’ll be processing about a hundred terabytes per second, which is Netflix global streaming.”
It also has to be processed in real time, otherwise computations will get held up waiting for error correction to happen. To avoid that, errors must be detected in real time. For transmon-based qubits, syndrome data is generated roughly every microsecond, so real time means completing the processing of the data—possibly Terabytes of it—with a frequency of around a Megahertz. And Riverlane was founded to provide hardware that’s capable of handling it.
Handling the data
The system the company has developed is described in a paper that it has posted on the arXiv. It’s designed to handle syndrome data after other hardware has already converted the analog signals into digital form. This allows Riverlane’s hardware to sit outside any low-temperature hardware that’s needed for some forms of physical qubits.
That data is run through an algorithm the paper terms a “Collision Clustering decoder,” which handles the error detection. To demonstrate its effectiveness, they implement it based on a typical Field Programmable Gate Array from Xilinx, where it occupies only about 5 percent of the chip but can handle a logical qubit built from nearly 900 hardware qubits (simulated, in this case).
The company also demonstrated a custom chip that handled an even larger logical qubit, while only occupying a tiny fraction of a square millimeter and consuming just 8 milliwatts of power.
Both of these versions are highly specialized; they simply feed the error information for other parts of the system to act on. So, it is a highly focused solution. But it’s also quite flexible in that it works with various error-correction codes. Critically, it also integrates with systems designed to control a qubit based on very different physics, including cold atoms, trapped ions, and transmons.
“I think early on it was a bit of a puzzle,” Brierley said. “You’ve got all these different types of physics; how are we going to do this?” It turned out not to be a major challenge. “One of our engineers was in Oxford working with the superconducting qubits, and in the afternoon he was working with the iron trap qubits. He came back to Cambridge and he was all excited. He was like, ‘They’re using the same control electronics.'” It turns out that, regardless of the physics involved in controlling the qubits, everybody had borrowed the same hardware from a different field (Brierley said it was a Xilinx radiofrequency system-on-a-chip built for 5G base stationed prototyping.) That makes it relatively easy to integrate Riverlane’s custom hardware with a variety of systems.
