#93 Thinking About Data for AI
Data for AI - An Excerpt; A Curious Case of Policy Convergence
Today, Bharath Reddy offers a sneak peek into the framework being used in an upcoming discussion document on Data and AI. Rijesh Panicker points out an emerging similarity in different countries’ regulations regarding the export of quantum computers and allied technologies.
But before that, we encourage you to participate in our latest survey, ‘The China Challenge Survey’.
This is the first iteration of an intended periodic Survey to assess public opinion on India-China relations. The Survey is prepared by the Takshashila Institution, an independent centre for research and education in public policy.
The survey should take you no more than 5 minutes to complete. Please complete it only once. Responses will be accepted until Friday, 2 August 2024, 11:59 PM IST.
Please note that any and all information provided through the Survey will be completely confidential. Please also note that by submitting your responses, you are authorising Takshashila Institution to use the information provided for research, analyses and/ or educational purposes.
In case of any queries, please reach out to us via contact@takshashila.org.in or use the comments space below the text of this edition.
Cyberpolitik: Data for AI - An Excerpt
— Bharath Reddy
**This is an excerpt from an upcoming paper on Data for AI - What should India do? You can find our latest research here.**
Access to data can be an entry barrier to developing or deploying AI systems. In business, data is regarded as a moat in some conditions, referring to the company’s ability to maintain a competitive advantage over its rivals. Martin Casado and Peter Lauten of Andreeson Horowitz write about the conditions under which data serves as a moat. Andrew Ng also echoes similar views in his AI news publication, The Batch. These conditions are summarised below:
System performance improves significantly with additional data. In applications such as text-to-speech or transliteration technologies, performance often plateaus after a point. In such applications, including additional data does not lead to significant improvements in performance. On the other hand, large language models have shown significant improvements when trained on increasingly larger datasets. Web searches with a long tail of unique queries have also shown that more data and context can lead to better results.
Access to proprietary data can be an entry barrier when sources are rare or not easily available. This barrier may arise due to the industry structure, such as TransUnion's CIBIL credit score. Additionally, exclusive access can be granted through government tenders or when compliance standards are exceptionally stringent.
Data with high variability over time can create entry barriers. In applications like social media, a continuous flow of fresh data is essential to remain relevant. The freshness of this data can depend on real-world events, news cycles, and changing attitudes, among other factors.
Data might not be an effective entry barrier in cases where applications can be built with smaller datasets. In their issue brief on “Small Data’s Big AI Potential”, Chahal et al. showcase how, in some applications, approaches such as reinforcement learning, Bayesian methods, transfer learning, data labelling, and synthetic data, which do not require massive datasets, can deliver good results.
It is also important to note that access to data works as an effective moat, often in situations when network effects and switching costs are high. The above conditions distinguish when data itself contributes to being a moat.
In Takshashila’s discussion document on AI governance, we explored AI governance from the perspective of a supply chain comprised of data, compute, models, and applications. Knowing that data is a crucial input to unlocking many applications for AI, it is vital to consider the entry barriers and market failures in the data part of the supply chain.
A framework to guide this thinking could be to think of a 2x2 along the national and commercial interests of datasets. This could inform the appropriate policy actions in the buckets of production, financing, or regulation.
We would love to hear your thoughts on the initial ideas presented here. Please write to us if you have any comments.
India is set to host the Quad Leaders' Summit in 2024. Subscribe to Takshashila's Quad Bulletin, a fortnightly newsletter that tracks the Quad's activities through the Indo-Pacific.
Your weekly dose of All Things China, focusing on Chinese discourses on defence, foreign policy, tech, and India, awaits you in the Eye on China newsletter!
The Takshashila Geospatial Bulletin is a monthly dispatch of Geospatial insights for India’s strategic affairs. Subscribe now!
Technomachy: A Curious Case of Policy Convergence
— Rijesh Panicker
In recent months, Canada, France, the Netherlands, the U.K., and Spain have implemented a spate of eerily similar regulations regarding the export of quantum computers and allied technologies.
The regulations impose export controls on quantum computers along two dimensions: qubit size and error rates. All quantum computers above 34 physical qubits will need an export permit. In addition, error rate thresholds have been specified, starting at 10−4 (1 error in 10000 calculations) and going up to 6*10−3 (6 in 1000) for larger qubit size (1100-2000 physical qubits). Canada’s report on its export control list (ECL) shows that export controls apply to all the major types of qubit technologies, such as semiconducting qubits, ion traps, and semiconductor and photonic qubits.
The regulations also control the export of the allied control and measurement components that form an essential part of any quantum computer. In addition, they impose controls on cryoCMOS integrated circuits designed to operate at temperatures at or below 4.5K−268.65C.
Canada (see report above), France, and Germany have indicated that these limits have been decided based on negotiations under the Wassenaar Arrangement, although Canada itself seems to have made amendments under Group 5 of its ECL, which is not covered by the arrangement.
The Wassenaar Arrangement is a voluntary export control regime with 42 members who exchange information on transfers of conventional weapons and dual-use goods and technologies. The aim is to promote greater responsibility amongst members in exporting weapons and critical technologies and prevent exports to states of concern for members. Signatories are expected to disclose voluntary information about their exports concerning two control lists: a munitions list covering conventional weapons and a dual-use goods and technologies list. Negotiations under the Wassenaar agreement are confidential amongst member states, so unfortunately, there is no answer to the intriguing question: why set the limit at 34 qubits?
Over the last decade, quantum computing has gone from being a theoretical possibility to one already in an intermediate stage of realisation. In the last two years, we have seen significant advances from IBM with its 1000 qubit chip and Google’s quantum computing breakthrough. Other nations like China have consistently reported larger and larger quantum computers in recent years, so it’s unclear why the export control limits are set so low. One possible answer is that 34 qubits are believed to be the smallest point at which quantum computers can no longer be simulated classically, even by supercomputers. Thus, this limit also restricts the ability to do classical simulations above a certain complexity using some hybrid classical-quantum computer.
Given that China (a leader in quantum technology research) and Israel (quantum controls and instrumentation) are not a part of the Wassenaar Arrangement, these export control laws may only have a marginal impact for now, even in terms of the limited goal of increasing transparency. In the long run, this is the first step to creating a multilateral system of coordinated control for quantum computing and allied control and instrumentation technologies.
The Wassenaar Arrangement does not (as far as I can tell) preclude member controls from imposing export controls on other members, so from an Indian perspective, an understanding of what technologies could fall under export controls could prove useful in deciding where we really need to be atmanirbhar and where we can depend on our “allies”.
If you like the newsletter, you will love to read our in-depth research and analysis at https://takshashila.org.in/high-tech-geopolitics.
Also, we are hiring! If you are passionate about working on emerging areas of contention at the intersection of technology and international relations, check out the Staff Research Analyst position with Takshashila’s High-Tech Geopolitics programme here.
What We're Reading (or Listening to)
[Takshashila Issue Brief] Looking Ahead to China’s 2024 Third Plenum [authored by the Indo-Pacific Studies Programme]
[Opinion] New Telecom Act risks normalising dangerous culture of unaccountable state intrusion, by Satya S Sahu
[Opinion] From Russian Ladas to Chinese BYDs: Central Asia’s Changing Priorities, by Rakshith Shetty
[Takshashila Discussion Document] Space Reforms in India: A Job Half Done, by Ashwin Prasad