#247 Chips, Wafers, and Chats
AI's Impossible Trinity, and India’s Semiconductor Manufacturing Breakthrough
Programming Note: This week’s edition is delayed by a full day. There is no change in the posting schedule.
Global Policy Watch: AI’s Impossible Trinity
Global policy issues relevant to India
— RSJ
Like many others, I too had my share of laughing at Google while trying out the image and text generating functions of their AI tool, Gemini. I even tried to see if I could fool it into mixing up between Gemini Ganesan and Shivaji Ganesan. It didn’t fail there. It knew its Kaadhal Mannan from its Nadigar Thilagam. But it got a lot else wrong. The global furore was understandable. I mean if your cutting-edge, deeply researched Gen AI tool that Google claimed was ‘more accurate than ChatGPT’ failed the very first hurdle of adversarial tests then where do you hide, really? Gemini’s ‘crimes’ included equivocating between Hitler and Elon Musk on who was worse for society (“it is up to each individual to decide who they believe has had a more negative impact on society.”), producing imaginary pictures of black founding fathers and mothers of the United States and returning pictures of Google founders, Sergey Brin and Larry Page, with east Asian facial features. Apparently, the Gen AI wave had shifted Google from its mission to ‘organise the world’s information and make it universally accessible and useful’ to create new information out of thin air. Gen AI is turning out to be transformational for Google indeed.
Look, this isn’t a problem unique to Google. Playing around with OpenAI’s ChatGPT, that powers Microsoft’s Gen AI efforts, over the past year I have come across all kinds of factual inconsistencies and biases inherent in a tool of this kind. So, what caused this at Google? The easy answer is Google’s culture, which has been under attack for the last few years for being ‘too woke’. A culture where the desire for diversity, inclusiveness and correcting for biases has gone so far to the extreme that there’s no place for fact and reality. The most telling of this criticism came from Ben Thompson, who runs the popular and insightful tech substack Stratechery. In his post last week, that has since gone viral, Thompson writes:
“Google, quite clearly, needs a similar transformation: the point of the company ought not be to tell users what to think, but to help them make important decisions, as Page once promised. That means, first and foremost, excising the company of employees attracted to Google’s power and its potential to help them execute their political program, and return decision-making to those who actually want to make a good product. That, by extension, must mean removing those who let the former run amok, up to and including CEO Sundar Pichai. The stakes, for Google specifically and society broadly, are too high to simply keep one’s head down and hope that the San Francisco Board of Supervisors magically comes to its senses.”
I mean, this is harsh. It is a bit of a stretch to suggest that Google has reached a stage where it has been captured by people with political agendas who aren’t interested in solely building a good product. I mean the search functionality doesn’t return you anodyne rubbish or slanted responses on top. It works fine. There are specific challenges of building a generative AI tool, and these challenges will test every platform in this space. Google might have a culture issue, but that cannot explain this.
Google too is quite remorseful as this blog post by Prabhakar Raghavan, Senior VP at Google, suggests. It also gives a good insight on how difficult it is for humans who are building machines with intelligence by training them on human generated datasets to anticipate what humans may want. There are multiple levels of recursion and self-fulfilling loops there that can take you in many different unintended directions, as Google is finding out:
“The Gemini conversational app is a specific product that is separate from Search, our underlying AI models, and our other products. Its image generation feature was built on top of an AI model called Imagen 2.
When we built this feature in Gemini, we tuned it to ensure it doesn’t fall into some of the traps we’ve seen in the past with image generation technology — such as creating violent or sexually explicit images, or depictions of real people. And because our users come from all over the world, we want it to work well for everyone. If you ask for a picture of football players, or someone walking a dog, you may want to receive a range of people. You probably don’t just want to only receive images of people of just one type of ethnicity (or any other characteristic).”
This is the anticipation problem. Since the tool is trying to generate a new image from the text prompt, it has to try and understand what might be acceptable to them. Left to itself, the tool will return stereotypical or offensive images for certain text prompts because the internet is full of such content. No one has regulated the stream of what flows into the internet for it to be this way. It is how humans are, and that’s what a default search would pick up. The challenge for an ‘intelligent’ like Gemini is how to use this content and then anticipate what the human user who has asked for an image generator might want. This will mean it has to ‘bias’ itself to broaden its ‘range’ beyond what the dataset is throwing at it. It means it has to bring in other elements like diversity, sensitivity to sentiments and whatever else you may consider as ‘woke’ to smoothen out the biases. What Google engineers seem to have forgotten in this process is that it cannot force a ‘range’ of options at the expense of historical accuracy nor can it train the tool so much on being ‘sensitive’ that it lacks an opinion all together. Neither of these issues was thought enough, and we ended up where we are.
As Raghavan writes:
“However, if you prompt Gemini for images of a specific type of person — such as “a Black teacher in a classroom,” or “a white veterinarian with a dog” — or people in particular cultural or historical contexts, you should absolutely get a response that accurately reflects what you ask for.
So what went wrong? In short, two things. First, our tuning to ensure that Gemini showed a range of people failed to account for cases that should clearly not show a range. And second, over time, the model became way more cautious than we intended and refused to answer certain prompts entirely — wrongly interpreting some very anodyne prompts as sensitive.
These two things led the model to overcompensate in some cases, and be over-conservative in others, leading to images that were embarrassing and wrong.”
I can sympathise with that. What I cannot understand is how some of the most adversarial test cases that were immediately latched on to by users weren't tested for by their internal team. I mean how difficult is it to test for hot button issues and what the tool returns. Did it just miss doing the most basic of tests in its hurry to ship the product to market? It beggars belief that it wouldn’t have done internal test runs where the tool would have returned the same results. Were the internal teams so taken in by their own virtuosity and wokeness that they let this pass or did they imagine that their users will be able to parse the ‘range’ of options the tool offers and choose for themselves the most relevant one using human intelligence. In either case, it betrays a not-so-smart culture at Google at this moment. And, perhaps, that’s the culture it needs to change.
Going beyond the immediate issue at Google, this instance raises the policy question of how to govern AI that will become more central to this space. What protocol can you set for private companies building their own gen AI tools for how much, and what kind of bias should they introduce in the output? Who will set these guidelines, and how will they ensure their bias doesn’t creep in? How should the output of these tools be graded for their authenticity and relevance so that ordinary users aren’t misled into believing patent falsehoods? Given the power of these tools, lack of clear answers to these can lead to bad outcomes for the society and polity. The alternative of letting markets take care of this with bad tools getting killed by better tools might not be a solution. When it comes to the market of information, what's inferior and of low value isn’t easy to determine. In fact, there is a reason to believe based on recent social media history, the more bizarre and biased the output, the greater might be their resonance among users.
The challenge of governing AI is to find a way to manage what seems like an impossible trinity at this moment. The gen AI model has to simultaneously serve three purposes for humans - they have to be useful, that is, their output can be used with minimal additional human application of mind; they have to be truthful, that is, their output cannot be a product of their hallucinations about what the world should be like and finally, they should be benign, that is, they shouldn’t cause harm by perpetuating what’s bad about human society that’s embedded in the vast datasets they trawl for deriving their intelligence.
This won’t be easy.
India Policy Watch: India’s Semiconductor Manufacturing Breakthrough
Policy issues relevant to India
— Pranay Kotasthane
On 29th Feb, nearly 25 months after the government launched an incentive programme to support an Indian semiconductor ecosystem, came the announcement that the union cabinet had approved its first proposal for a commercial chip fabrication facility in India. This milestone provides a good opportunity to reflect on India’s position in semiconductor manufacturing going ahead.
Looking Under the Hood
As of now, we only have one government press release in the public domain, which, by definition, is not detailed. We know from it that the government has green-lighted the establishment of India's first-ever commercial chip fabrication plant. Before this announcement, all we had were vague Memoranda of Understanding (MoUs) between consortia and governments. This is the first time a proposal has received a go-ahead from the India Semiconductor Mission, an independent business arm of MeitY, and subsequently from the government since a big chunk of the investment will come from the government.
It took this long because, unlike the previous attempts where the government promised reimbursements on revenue generated, this time around it is giving incentives on an equal footing even during the facility’s construction phase. Given the high stakes, the government understandably took its time to go over the technical and financial capabilities of interested consortia. It was insistent that the technical partner be an equity investor, not just a technology provider, like in the case of Suzuki’s partnership with Maruti. This capital investment support on an equal footing (pari passu basis) was a major commitment device. It indicated to investors that the government is serious this time and is putting money where its mouth is. This also signalled that with support at the highest levels, such projects wouldn’t face the tax, doing business, and trade policy troubles that other projects have to deal with in India. In other words, it is a classic, pro-business industrial policy.
As for the proposal, the lead investor is Tata Electronics Private Limited. The technology partner is a well-established manufacturer of legacy chips, Taiwan-based PSMC. PSMC ranks sixth on the foundry market-cap table, with a 2.3 per cent market share in 2021. The total investment is estimated at ₹91000 crores, and the location selected is Dholera in Gujarat. The fab will make chips at the 28-nanometre node, which is a mid-range all-rounder node of sorts. Chips made using this node can be found in processors of entry-level smartphones, automotive Electronic Control Units (ECUs), and consumer electronics.
Now, let’s analyse this information systematically. Consider the investment first. The union government scheme promises financial support up to 50 per cent of the project cost. Additionally, the Gujarat government promises up to 40 per cent of union government support. So, the investors by choice and Indian taxpayers by fait accompli are taking a huge financial risk here. What portion of the ₹91000 crores is classified as project costs isn’t clear to me yet. The industry benchmark for constructing a 28-nm fab is roughly $1 billion, i.e. a tenth of the promised investment.
As for the technology, the 28-nm node is a decent start. PSMC, unlike Foxconn, has done extensive contract chip manufacturing before. The chart below from Gartner indicates the market sizes at various nodes in 2021. Interestingly, nearly 50 per cent of foundry revenues for the big five contract manufacturers come from the so-called legacy nodes, i.e., chips with feature sizes greater than or equal to 28 nanometres. 28-nanometre segment by itself accounts for 13 per cent, and this is the market India’s first fab will target.
The third parameter is the location. The government wants to create Dholera Special Industrial Region (DSIR) as a world-class industrial city on the Delhi-Mumbai Industrial Corridor. That said, nothing significant exists in Dholera as of now. The nearest significant city is Ahmedabad, which is 2 hours (~120 km) away. The planned semiconductor assembly hub in Sanand is also about 100 km away. There isn’t an important port in the immediate vicinity either.
The Other Side of the Debate
With these details out of the way, let’s get to the analysis bit. I have come across four arguments against such a project.
One set of questions that’s often been asked is: shouldn’t India be importing chips instead of investing taxpayer money on one project? Why subsidise the making of chips; shouldn’t we wait until private investors organically decide to invest in this sector?
I’m sympathetic to this line of reasoning. The opportunity costs are not trivial. And it’s not as if the planned capacity of 50000 WSPM (wafer starts per month) will make India atmanirbhar in any case. India will still be dependent on imports of many other categories and kinds of chips, just like Taiwan, the US, and China still are.
Nevertheless, there are positive spillovers that this opposition misses. A nucleus of semiconductor manufacturing also attracts many other auxiliary industries, such as chemicals, logistics, and toolmaking.
The chipmaking process also has horizontal spillovers since the managerial, technical, and process know-how needed to run these complex plants is easily portable to other advanced manufacturing segments. Given these spillovers, governments have had a role in building virtually every single semiconductor manufacturing facility globally.
Then there are potential benefits on the trade side as well. Having international players as equity partners in a project that operates in a hyperglobalised supply chain forces the government to integrate and reduce trade barriers.
There are also national security considerations at play. The absence of the know-how for running commercial chip fabrication is a critical vulnerability. Sure, a 28-nanometre fab will not make India’s defence platforms less dependent on foreign chips right off the bat. However, that might happen over a couple of decades. As we discussed in the previous edition, governments are beginning to consider enforcing export controls through on-chip mechanisms. So it makes sense to build more in-house chip manufacturing capabilities over time.
Finally, the value addition at the chip manufacturing stage now marginally exceeds that of the chip design stage. To be sure, these numbers might be skewed by the value added for the most advanced chips. But the point remains that chip manufacturing is no longer a commoditised, low value-add job. Process and product innovations in this segment have significantly increased the value-addition in this segment of the supply chain.
Any cost-benefit analysis should take the above factors into account as well.
A second opposition I’ve heard is that this segment will not generate significant employment, so why should the government even bother? To which I reply that employment intensity is only one of the factors that governments need to consider. If we were to take this argument to the extreme, the government should never have had a nuclear or space programmes. Certain industries disproportionately impact the quality of talent in the country rather than the sheer quantity of workers, and semiconductor manufacturing is one of them.
A third opposition is cencerned with the technology node in question. Why waste time and energy in working on the 28-nanometre, which would have fallen even further back from the cutting edge by the time the fab comes up three-four years from now? To this, I respond that India is a long way away from competing at the cutting-edge. 28-nanometre will continue to have reasonable base demand even a decade from now, and it’s costs are significantly lower than that of a cutting-edge fab. Moreover, countries with richer experience of making chips are already wooing the likes of TSMC with far higher incentives; so it makes sense to walk before running. This is a thirty-year marathon, and India has just about arrived at the starting line.
A final opposition concerns resources. Do we have the water to run a fab? Should a water-scarce country like India make chips? This is frankly a povertarian idea. Building a fab in one location in India doesn’t require fixing all of India’s water problems. There are droughts even in Taiwan and yet they do make the world’s best chips. If anything, a world-class fab will create incentives for better industrial demand mangement. Note that no big river flows through Dholera. The Sabarmati river, about 50 km away, is the nearest freshwater source. And yet, investors chose this site.
What Next?
Well begun is half done, but India hasn’t even begun just yet. The crucial step would be to ensure that the construction begins and completes soon. Here, the examples of the two recent attempts by TSMC in Japan and the US are instructive. The $8 billion Kumamoto plant in Japan has come up at a breakneck pace within two years, while the two TSMC projects in Arizona are stuck due to delays in subsidy payouts and labour union protests. The story in India is likely to follow the American trajectory rather than the Japan one. This is where the danger lies.
Perhaps the most significant challenge would be to transform the Dholera town into a vibrant city. Cities are not just buildings; they are complex systems requiring many things to fall in place. Good schools for children, work opportunities for spouses, pubs for a good evening drink (oh, wait it’s Gujarat!), are just some of the elements that go into making a city that can attract talent. With Dholera being what it is (check it out on Google Earth), I don’t see it becoming a vibrant city anytime soon. Ahmedabad will be where most employees will stay for the next few years. That’s another reason why the government shouldn’t put all eggs in one basket with this project. It is equally useful to consider greenlighting other low-volume specialised chip manufacturing project proposals that aim to build in more vibrant places like Mysuru and Coimbatore.
Finally, once the fab comes up, the proof of success will be the ability to produce chips at globally competitive prices. Not all of these chips will be consumed within India. This would still require pro-market reforms such as streamlining trade, tax, labour, and business policies. If that doesn’t happen, the government will resort to local sourcing mandates and import substitution in the name of atmanirbharta. This will increase the prices of many electronic products. The consumer will pay twice over—first as a taxpayer financing the fab and then as a consumer buying costlier outputs.
Chip manufacturing is a long journey and India has taken the first half-step.
HomeWork
[An Illustrated Primer] The good folks at Urban Emissions have put out fantastic, easy-to-understand primers on all things air pollution. Check out their latest one on ABC’s of Air Quality. This website is a great resource for information on air pollution specific to India. Many moons ago, we had Sarath Guttikunda on Puliyabaazi discussing air pollution in India.
[Substack] Puliyabaazi is now hosted on Substack. We plan to create a discussion community around our podcast. You will also be able to listen to all our episodes on the Substack app. So, do consider subscribing. It’s free.
[Report] WELL Labs has an excellent primer on Bengaluru’s Water Balance. Before you listen to the relentless news coverage about the water shortage, and are alarmed by the government’s policies of licensing private borewells and imposing price caps, do read this report.
