Arnab Biswas

Father of two daughters

[Arnab Biswas @arnabbiswas1]

Here is a thread capturing all my talks on various ML/MLOps topics

⬇️

[Pranay Kotasthane @pranaykotas]

Dare I say that our book We the Citizens makes for a perfect gift on Children's Day (for age group 10+). See the thread below for some panels from the book.

[Arnab Biswas @arnabbiswas1]

@pranaykotas @thescribblebee @anupammanur @karthik_venk Got the book last December. Incidentally I have started reading it since yesterday evening. Told my 14 year olds to read it as well.

Very well thought through and well written. Specially at a time like this we need many more books like this.

Thanks for everything🙏

[Arvind Narayanan @random_walker]

I’m excited to announce I’ve started a YouTube channel. I plan to publish videos regularly explaining my views on AI and its present and future impacts.

My first video asks: What happens if there’s an AI crash?
https://t.co/7zC9o4GrG9

This is my first foray into video (beyond my usual academic talks) so I would love to receive feedback on what you would like to see. Thank you!

Intro: What if there’s a bubble?

AI companies have invested over a trillion dollars into data centers. They're promising that it'll be worth it because AI will transform the economy.

But so far, that's not happening. There doesn't seem to be any measurable uptick in the GDP growth rate despite all the investment. This has led to a chorus of voices saying AI is a bubble, and it's easy to see why. Even Mark Zuckerberg has been candid that meta might be spending too much money simply because of FOMO.

Zuckerberg: I think that there's a meaningful chance that a lot of the companies are overbuilding now, but the downside of being behind is that you're out of position for the most important technology for the next 10-15 years.

Suppose there is indeed an AI bubble and it bursts. I'm not saying that it will happen, but it could. Let's imagine what the scenario will look like. What will the consequences be?

Will an AI crash look like the dot com crash?

This wouldn't be the first cycle of AI hype to be corrected. This has happened twice in the past. These periods have been called AI winters. During these periods, research funding dried up and the press became negative. Back then, of course, AI was just a research field. Now it's a big industry, so the consequences of a bubble will surely be bigger.

Maybe it will be more like the dot com crash. Many companies went out of business. Enterprises paused their digitization plants. The whole of e-commerce was set back by many years. Entire business models like grocery delivery online were poisoned and no one would fund them for well over a decade.

We might expect something similar to happen if there is an AI crash. Products like ChatGPT will no longer be offered because they aren't profitable and the venture money is no longer rolling in. Enterprises will pause their AI strategies and deployments. Hype will turn sour and research will be cut back.

Perhaps most significantly countless people will lose their jobs, just like the dot com crash, which led to between half a million and a billion job losses in the tech sector alone and cascading effects in many other sectors, and a 50% increase in the unemployment rate over the next few years,

Maybe. Or maybe the dot com bubble is not a good model for an AI bubble. So how can we think clearly about this?

Unit economics is a helpful concept

Now, the AI industry is an outlier in terms of the scale of investment, but the fact that AI companies aren't profitable is not at all unusual.

Early on in the lifecycle of any new industry sector, most startups won't be profitable. So the key question is this: Is it because they're doing a lot of research and building out infrastructure, in which case the costs will come down as the companies mature? Or is it because the basic cost of providing their products or services is so high, in which case the costs won't come down?

For example, the Juicero was a juice machine that famously failed after being introduced at a ridiculous price. It was destined to flop because the machine was extremely over-engineered. The parts alone cost hundreds of dollars, so there is no way the company could have dropped it to a reasonable price point without taking a loss on every unit. In the same way, there have been subscription services that lost money on every single customer, and so were widely predicted to fail.

Now let's look at generative AI. These companies have three big kinds of costs, talent training and inference. First, talent. AI companies are paying obscene amounts of money to higher top engineers.

Next, training. Training models can cost hundreds of millions of dollars in terms of hardware and electricity, especially ones you factor in the computational needs of all the research that needs to happen in order to push the state-of-the-art and develop new model capabilities. And companies are training new and bigger models all the time.

And finally, inference. Inference means running the models so that people can actually use it. Now, we know that companies are spending a lot on data centers, but we don't know how that breaks down into training and inference.

Inference costs are low and falling

But we do know what it costs for a chatbot to respond to a single query or to output a given amount of text. And it's remarkably little. You can generate thousands and thousands of pages of text for just $1, and that cost has decreased a hundred fold in the last couple of years because engineers have been able to speed up these models by making them smaller and more efficient.

Now, there is a meme that using ChatGPT is extremely energy intensive, but this is just not true when it comes to the regular use of chatbots, though it is somewhat true when it comes to generating images or if you use the feature where it goes off on its own and does research for a while.

So knowing all this, suppose there is a market correction or crash. AI companies will decrease or maybe even eliminate their expenditure on all that expensive research and training of new models as well as the ridiculous amounts of money that they're paying researchers.

But even in this scenario, they will be able to offer chatbots like ChatGPT, because the inference costs, that is the cost of actually operating these models, are dirt cheap. They might have to limit access to a few features like image generation that are particularly expensive.

OpenAI has over 5,000 employees today. But actually running a chatbot only requires a handful of engineers. In fact, there are chatbot companies with very few employees. So if you cut out the research, operating a chatbot can actually be extremely profitable because people are willing to pay $20 a month, which translates to $240 a year for a subscription. That's a lot more than ad-based apps typically make per user per year.

Even if companies like OpenAI were to go out of business, smaller AI companies would step in to take their place. There are many openly available AI models which might not be as good as the leading ones, but whose quality is good enough for everyday users.

It’s the same story beyond chatbots

Now, of course, chatbots are not the only kind of generative AI products, but in most other cases I've looked at, it's the same story.

For example, AI agents for assisting with coding or software engineering are a lot more expensive to run than chatbots, but at the same time, the value that they bring is also much greater. If it makes a software developer, let's say, 20% more productive. What that means is that it brings tens of thousands of dollars per year of value to a software company for every single developer in that company that uses such a product.

Of course, not every developer benefits from AI assistance. It depends on the type of project, the developer's preferences and other factors. But for many people, including me, it's hard to even imagine going back to the way that programming was done before AI. Once you've made the adjustment of learning how to delegate the tedious and boring and error prone parts of software engineering to AI — admittedly, this adjustment takes a while — but having done that, it feels like trying to go back to punch cards instead of keyboards.

Here's another AI application that I looked into. One of the most computationally intensive and expensive types of generative AI is video generation. The Wall Street Journal used AI to create a moderately high quality YouTube video as part of learning about the process of using AI to create such videos. And this is what they reported: the total cost would've been around a thousand dollars for Google and runway's tools.

Now, that number seems like a lot, but it's much less expensive than the cost of producing a video with comparable quality in the traditional way.

An AI crash will look nothing like the dot com crash

So based on all this, my view is that an AI crash will look nothing like the dot com crash. It's true that in both cases we see outrageous valuations of companies and expenditures as well. But the dot com bubble was based entirely on the expectation of future profits, and those profits never materialized because customers just weren't interested. Whereas in the case of AI, it's true that there is a lot of hype, but that hype is layered on top of a technology that's already bringing lots of real value to lots of people. It's being used by hundreds of millions of people every day, and a growing number of them are paying $20 a month or even $200 a month. All of that I think will continue.

I'm making a subtle point here, so let me be clear. I'm not saying that there won't be a crash. I'm saying that even if there is a crash, its effect will be on the research that's going into AI and the development of new models. The use of existing models and products will keep going strong.

Of course, if there is a crash, the way that the media talks about and hypes up AI might change, and I think that's probably for the better.

And now for the big question, what will a potential AI crash do to jobs? Once again, I think the AI situation is very different from the internet bubble. The dot com crash was so harmful because in the run up to it, many internet companies were massively overstaffed, especially considering their lack of a business model. But in AI, despite all the hype the tech sector has actually been contracting over the last few years, which is seen as a period of correction for the hiring that happened during the pandemic.

And as for the rest of the economy outside of tech, AI is seen as a reason to cut jobs rather than to hire. So if there is a crash, some of these AI engineers may no longer receive enormous paychecks, but they'll have no trouble finding other jobs in tech and companies not having AI as a readily believable excuse to cut jobs will probably be a good thing for workers on balance.

The long view

Even though AI is different from the dot com bubble, there are a lot of similarities between AI adoption and internet adoption. Over a couple of decades, the internet gradually became the medium through which all knowledge work happens. Those of us who were around in the nineties used to log onto the internet, do stuff quickly because it was so expensive and log back out. But now we're online all the time, and the hard part is to switch off our devices.

In my book, AI Snake Oil with Sayash Kapoor, we predict that the same thing will happen with AI, for better or for worse. It will just be there all the time in the background. But getting to that point will take time. In a paper called AI As Normal Technology, we talk about the gradual decades-long process by which workers and companies will have to adapt to AI and in turn adapt AI to us. The key thing is that most of this work will happen outside the AI companies, which means it'll be relatively unaffected by what happens in the AI companies, including whether there is a crash.

So in short, I think an AI winter is unlikely, and whether we like it or not, AI is a technology that's here to stay in our lives. Regardless of what happens in the market, we need to get used to it and start figuring out what it means for each of us and for society as a whole. Let me know what you think in the comments. I'm Arvind Narayanan. Thank you for watching.

[Karthik Muralidharan @karthik_econ]

India now spends ₹2 trillion/year (0.6% of GDP) on cash transfers to 130M+ women—yet we know little about their effects. In a new paper (https://t.co/vbLgba0Wm0), we present findings from the first large-scale RCT of maternal cash transfers in India. Key results in🧵(1/10)

[Paul Graham @paulg]

"She says 'We'll bake bread tomorrow,' even when there is no flour. She says 'The windows are broken because we wanted more air,' even though that is not why the windows are broken. She says 'The baby is sleeping,' even though the baby is gone."

[Arnab Biswas @arnabbiswas1]

@radhikasathe @mukhoty I will let Sharanyaa know. :-)

Do you have any other recommendation?

[Pratik Sinha @free_thinker]

I am able to use ocrmypdf (https://t.co/NtiMhMBsqR) to add a text layer to the BIHAR SIR files that have been converted by ECI from digital to unreadable-by-machine in the age of digital. This solution can be run free of cost on your computer, takes a while to convert if you have an old computer, but it allows decent amount of searching.

The command I used was:

parallel --tag -j 2 ocrmypdf -l hin+eng --optimize 3 --deskew --rotate-pages --clean '{}' 'output/{}' ::: *.pdf

You'll have to install both english and hindi language files for tesseract.

[Arnab Biswas @arnabbiswas1]

@roocode Waiting for GPT-5 through Azure.

[Ihab Hassan @IhabHassane]

The Washington Post has published the names of all 18,500 Palestinian children killed by the Israeli army in Gaza.

[Arnab Biswas @arnabbiswas1]

@eugeneyan This love for OOP is most visible in Clude generated code. o3, o4-mini most often generate simple functions instead of multiple classes. However claude-4 specifically solved few problems for me which o* couldn't and Gemini **thought** it has solved after few iterations.

[TheLiverDoc™ @theliverdoc]

We recently published our independent analysis of Homeopathic practitioner commonly prescribed and popular over-the-counter 134 Homeopathy products marketed and sold as "medicine."

The Placebo Project: An observational study and comprehensive analysis of 134 commonly prescribed homeopathic remedies in India uncovers potential for hepatotoxicity: https://t.co/ACmo4bkMwY

Here is our plain language and visual abstract summary of what these Homeopathy products are and what we found in them.

1/3 - Classical (Diluted) Homeopathy Products

[Arnab Biswas @arnabbiswas1]

@abhicantdraw Thank you!

[Arnab Biswas @arnabbiswas1]

@abhicantdraw Sonnet version?

[Will McGugan @willmcgugan]

The Toad is out of the bag! 🛍🐸

Announcing Toad - a universal UI for agentic coding in the terminal

https://t.co/ZR8QreXbOg

[Arnab Biswas @arnabbiswas1]

@abhicantdraw 🙂

[Arnab Biswas @arnabbiswas1]

@abhicantdraw So...do you recommend it? :-)

[Ernest Ryu @ErnestRyu]

10. My career as a mathematician certainly isn't threatened by AI; in fact, I hope to leverage AI to accelerate my work. However, I'm unsure whether "mathematician" will remain a career path for my son’s generation. (10/10)

[Arnab Biswas @arnabbiswas1]

@A_K_Mandhan @grok Create such a chart for China and US, but in terms of percentage?

[Arnab Biswas @arnabbiswas1]

@MParakhin Thanks a lot for sharing your use case.

Will try o3-pro again.