I think any piece of work that was billed as T&E (Time and Expenses) or on an hourly basis, is now going to be radically affected. That being said, Software Engineering (for example) can be "solved" by AI - potentially before the end of the decade, as per Anthropic's Sholto Douglas and Trenton Bricken.
But once you enter into specialised areas, I do think AI will be more of an augmenting agent rather than a replacement, potentially looking at huge savings in terms of time and money.
The Mint report that RSJ quotes from gets the patent data wrong by ascribing all patent counts to a single year. The original WIPO report actually states that the numbers are for a 9 year period.
Inventors based in China were responsible for more than 38,000 patent families between 2014 and 2023, based on the inventor addresses published on patents…. With around 6,300 patent families between 2014 and 2023, the US is the second most important research location for GenAI patenting. The Asian countries Republic of Korea, Japan and India are other key research locations for GenAI, all ranking in the top 5 countries worldwide (third, fourth and fifth respectively).
A key part that these numbers miss is made clear in Fig. 20 of the same report, showing that 40K patent families were filed in China, while only a little over 10K filed in the US. It likely is a result of Chinese research institutions and firms filing primarily in China, not confident of obtaining returns on their patent filings elsewhere.
This essentially corroborates Pranay’s take “China's large number of patents reflects its institutional emphasis on patent filings as a metric rather than necessarily breakthrough innovation.”
Industrial use
RSJ and Pranay take different tracks to determining the impact of GenAI on Indian industry – RSJ goes for field research – “sitting through real examples” and having conversations with folks. His reporting suggests that GenAI penetration is already here in many business organisations, including core sectors, and that organisations are able to bypass traditional IT vendors in identifying use cases for and deploying GenAI effectively. Against, this background, he posits “The bottom line is that the Indian IT and BPM services sectors are in for a rude shock over the next three years”
Pranay’s addendum relies on the distinction between jobs and tasks. While he’s right that none of the GenAI tools is plug-and-play (in the sense traditional software is), I think RSJ’s field evidence proves that what’s already available is sufficient to get the task done. The “coordination, negotiation, and other duties” that Pranay refers to can actually reduce as the number of persons involved in getting a project done drops as many tasks are taken up by GenAI. In my limited experience and conversations with folks in Indian finance/fintech industry, the story is closer to how RSJ tells it.
Indian companies
I find Pranay’s take confusing and possibly wrong. It is unclear what “enterprise AI” means, or what he means by “small language models” (there’s “smaller” LLMs like Gemini Flash with a much lower cost of inference but a drop in quality from a parent model, and then there’s small language models, which are very different and not general-purpose.
Economic logic suggests cost-efficient LLMs will find a greater range of applications. The pace of innovation over the last 2-3 years likely means that on performance scale (both inference cost and inference quality), open-source models will continue to lag. Also, while academic publishing continues (and Pranay is right regarding the fundamental breakthroughs), as GenAI technology moves out of research labs and gains real-world applications, walls will start to go up, through trade secrets (a lot of this technology is hard to reverse engineer) and proprietary tech stacks (chips, data centres, datasets, etc.). Advantage is to big tech here, not to Indian IT players.
While distilled (smaller size) models for sectors/contexts are likely, it seems farfetched to me that “here Indian IT companies have a significant opportunity to convert their services prowess into AI products”.
I'm a bit sceptical about some of RSJ's arguments. IT services is a very competitive industry - the largest player has <10% share of the global market. It is unlikely that all companies have successfully colluded in resisting mass adoption of AI tools and changes to the billing model. The obvious incentive for any mid-size challenger is to package some of these tools into a neat solution, offer chunky price cuts, and snatch revenues from large incumbents. If this isn't happening, it is likely because these tools are over hyped to some extent (very powerful in specific tasks but not across tech workflows) or that we're still early in the journey.
Totally agree with the risks to some white collar jobs (caveat - there could be offsets which are unknown).
Very insightful read! I lean closer to Pranay's viewpoint on AI.
Dan Davies has an interesting essay viewing Generative AI similar to an offshore center in Chennai. The main point being that it is not easy to outsource business functions to India and manage.
However, AI could cause a painful recession and has adversely impacted the job market for US college graduates. AI is also leading to productization of services. It will be faster to figure out how to downsize and get more productive with AI and take longer to figure out new opportunities for the surplus. That will likely cause a recession.
I tend to disagree with Pranay/Arvind. Reflecting on my own work while I want to believe my work cannot be automated as it involves coordination/negotiation with other humans, the fact is if we remove those roles there is not much lost in organisations. Coordination becomes a big deal as humans inherently have different judgments or assessment of risk for same situation. This will stop mattering as we automate most of our workflows or perhaps create new workflows in future. I suspect a lot change is coming our way
I lean closer to @Pranay's argument on implications from AI diffusion in India. But I'm surprised both of you focus entirely on the automation/substitution effects only. There's a whole world of possibilities on the augmentation side where innovation not in the model layer but with data & applications can make a bigger difference. BPO's that transition from offering labor arbitrage to new capabilities(a-la-GCC's) will still flourish. Biotech will experience a new tailwind. All the UPI data we now have can be used to create new busineeses assuming this data is made available with safeguards. Lots of upside if the policy focus shifts from building new models to finding other leverage areas where we have comparative advantage.
I think any piece of work that was billed as T&E (Time and Expenses) or on an hourly basis, is now going to be radically affected. That being said, Software Engineering (for example) can be "solved" by AI - potentially before the end of the decade, as per Anthropic's Sholto Douglas and Trenton Bricken.
But once you enter into specialised areas, I do think AI will be more of an augmenting agent rather than a replacement, potentially looking at huge savings in terms of time and money.
Patents
The Mint report that RSJ quotes from gets the patent data wrong by ascribing all patent counts to a single year. The original WIPO report actually states that the numbers are for a 9 year period.
Inventors based in China were responsible for more than 38,000 patent families between 2014 and 2023, based on the inventor addresses published on patents…. With around 6,300 patent families between 2014 and 2023, the US is the second most important research location for GenAI patenting. The Asian countries Republic of Korea, Japan and India are other key research locations for GenAI, all ranking in the top 5 countries worldwide (third, fourth and fifth respectively).
A key part that these numbers miss is made clear in Fig. 20 of the same report, showing that 40K patent families were filed in China, while only a little over 10K filed in the US. It likely is a result of Chinese research institutions and firms filing primarily in China, not confident of obtaining returns on their patent filings elsewhere.
This essentially corroborates Pranay’s take “China's large number of patents reflects its institutional emphasis on patent filings as a metric rather than necessarily breakthrough innovation.”
Industrial use
RSJ and Pranay take different tracks to determining the impact of GenAI on Indian industry – RSJ goes for field research – “sitting through real examples” and having conversations with folks. His reporting suggests that GenAI penetration is already here in many business organisations, including core sectors, and that organisations are able to bypass traditional IT vendors in identifying use cases for and deploying GenAI effectively. Against, this background, he posits “The bottom line is that the Indian IT and BPM services sectors are in for a rude shock over the next three years”
Pranay’s addendum relies on the distinction between jobs and tasks. While he’s right that none of the GenAI tools is plug-and-play (in the sense traditional software is), I think RSJ’s field evidence proves that what’s already available is sufficient to get the task done. The “coordination, negotiation, and other duties” that Pranay refers to can actually reduce as the number of persons involved in getting a project done drops as many tasks are taken up by GenAI. In my limited experience and conversations with folks in Indian finance/fintech industry, the story is closer to how RSJ tells it.
Indian companies
I find Pranay’s take confusing and possibly wrong. It is unclear what “enterprise AI” means, or what he means by “small language models” (there’s “smaller” LLMs like Gemini Flash with a much lower cost of inference but a drop in quality from a parent model, and then there’s small language models, which are very different and not general-purpose.
Economic logic suggests cost-efficient LLMs will find a greater range of applications. The pace of innovation over the last 2-3 years likely means that on performance scale (both inference cost and inference quality), open-source models will continue to lag. Also, while academic publishing continues (and Pranay is right regarding the fundamental breakthroughs), as GenAI technology moves out of research labs and gains real-world applications, walls will start to go up, through trade secrets (a lot of this technology is hard to reverse engineer) and proprietary tech stacks (chips, data centres, datasets, etc.). Advantage is to big tech here, not to Indian IT players.
While distilled (smaller size) models for sectors/contexts are likely, it seems farfetched to me that “here Indian IT companies have a significant opportunity to convert their services prowess into AI products”.
I'm a bit sceptical about some of RSJ's arguments. IT services is a very competitive industry - the largest player has <10% share of the global market. It is unlikely that all companies have successfully colluded in resisting mass adoption of AI tools and changes to the billing model. The obvious incentive for any mid-size challenger is to package some of these tools into a neat solution, offer chunky price cuts, and snatch revenues from large incumbents. If this isn't happening, it is likely because these tools are over hyped to some extent (very powerful in specific tasks but not across tech workflows) or that we're still early in the journey.
Totally agree with the risks to some white collar jobs (caveat - there could be offsets which are unknown).
Very insightful read! I lean closer to Pranay's viewpoint on AI.
Dan Davies has an interesting essay viewing Generative AI similar to an offshore center in Chennai. The main point being that it is not easy to outsource business functions to India and manage.
Using AI is like managing an intern. Most people are not good managers especially if it involves giving negative feedback. https://substack.com/@backofmind/p-163203692
However, AI could cause a painful recession and has adversely impacted the job market for US college graduates. AI is also leading to productization of services. It will be faster to figure out how to downsize and get more productive with AI and take longer to figure out new opportunities for the surplus. That will likely cause a recession.
I tend to disagree with Pranay/Arvind. Reflecting on my own work while I want to believe my work cannot be automated as it involves coordination/negotiation with other humans, the fact is if we remove those roles there is not much lost in organisations. Coordination becomes a big deal as humans inherently have different judgments or assessment of risk for same situation. This will stop mattering as we automate most of our workflows or perhaps create new workflows in future. I suspect a lot change is coming our way
I lean closer to @Pranay's argument on implications from AI diffusion in India. But I'm surprised both of you focus entirely on the automation/substitution effects only. There's a whole world of possibilities on the augmentation side where innovation not in the model layer but with data & applications can make a bigger difference. BPO's that transition from offering labor arbitrage to new capabilities(a-la-GCC's) will still flourish. Biotech will experience a new tailwind. All the UPI data we now have can be used to create new busineeses assuming this data is made available with safeguards. Lots of upside if the policy focus shifts from building new models to finding other leverage areas where we have comparative advantage.