DeepSeek has US AI firms talking about Jevons paradox and invigoration
No one saw this coming. DeepSeek has the Silicon Valley worried. The fact that they took their most advanced AI model.
No one saw this coming. DeepSeek has the Silicon Valley worried. The fact that they took their most advanced AI model, packaged it as an app, a web tool and an API for developers and gave it away for everyone to use for free, is just the proverbial cherry on a very real cake. No subscriptions for consumers, and much less cost intensive for businesses. Satya Nadella’s writing about Jevons paradox (it is a concept of economics, where efficiency causes the cost of a resource to drop, thereby increasing consumption) in a late night post on X. OpenAI’s Sam Altman insists such competition will be “invigorating”. A collective brave face is one thing, but the hits simply keep coming. Nvidia shed almost $600 billion in market cap in a single day of trading (January 27 on Nasdaq), making it the single biggest loss in a day in U.S. stock trading history. Nvidia’s 16.97% slide in the day’s trading may get everyone’s attention, but I’d like to point to Broadcom (-17.40%), ARM (-10.19%), AMD (-6.37%) and Intel (-2.59%) also took hit after hit. And on the matter of the hits that keep coming, DeepSeek has now released the Janus-Pro-7B imaging model, competing with OpenAI's DALL-E 3 and Stability AI's Stable Diffusion. Chinese AI, almost overnight, has the typical American tech order, scurrying to find a response.
DeepSeek claims to have spent around $5.5 million to train its V3 model, a considerably frugal approach to delivering the same results, that took the likes of Google, OpenAI, Meta and others, hundreds of millions of dollars in investments to achieve. According to research by Epoch.AI, Google and OpenAI spent roughly between $70 million and $100 million in 2023 to train the Gemini 1.0 Ultra and GPT-4 frontier models respectively. This cost goes up every year. Or at least that’s what the estimate was, till DeepSeek came into the picture. Those perceptions, optics and assumptions that AI needs massive infrastructure, have been shattered comprehensively.
There’s frugality of hardware too. “I was trained on a combination of Nvidia A100 and H100 GPUs,” the DeepSeek chatbot tells us. It doesn’t share an exact number, and this is specific to the R1 model. The question is, how did a Chinese tech company get access to number of Nvidia GPUs, amidst the trade limitations? DeepSeek CEO Liang Wenfeng is a billionaire, who runs a hedge fund and is funding DeepSeek that reportedly hired top talent from other Chinese tech companies including ByteDance and Tencent.
Let me summarise why what DeepSeek has done, worries every other AI company.
- This may well be the tipping point of AI economics. It’s easy to see why: DeepSeek R1’s API costs just $0.55 per million input tokens and $2.19 per million output tokens. In comparison, OpenAI’s API usually costs around $15 per million input and $60 per million output tokens.
- The methodology has changed too. Much like OpenAI’s o1 model, the R1 too uses reinforced learning, or RL. This means, models learn through trial and error and self-improve through algorithmic rewards, something that develops reasoning capabilities. Models learn by receiving feedback based on their interactions.
- With R1, DeepSeek realigned the traditional approach to AI models. Traditional generative and contextual AI uses 32-bit floating points (a floating point is a way to encode large and small numbers). DeepSeek’s approach uses a 8-bit floating point, without compromising accuracy. In fact, it is better than GPT-4 and Claude in many tasks. The result, as much as 75% lesser memory needed to run AI.
- Then there is the multi-token system that read entire phrases and set of words at one, instead of in sequence and one by one. That means AI will be able to respond twice as fast.
Turns out, everything we were told about the power and tech intensive nature of AI, was incorrect. Massive data centers. Big data sets. Lots of money to be pumped in. A circular economy, if you may, and money in everyone’s pockets. Or at least the Chinese researchers found a way that no one in Silicon Valley thought of. That seems unlikely.
INTENTIONS
Late last month, the Telecom Regulatory Authority of India (TRAI) made it mandatory for telecom operators to offer voice and SMS only plans. The idea is good, envisioned as a cost effective recharge for those who still use feature phones, those who simply want to keep a second number active, and purely easier on the pocket. What happened over the past couple of weeks, as Reliance Jio, Bharti Airtel and Vi released their “voice and SMS only plans” made an absolute mockery of that intent.
First, the companies announced a set of recharge packs, a few of which basically were a reconfiguration of existing recharge options, with the 4G/5G data element removed. The result — Jio’s ₹458 and ₹1,958 plans, Airtel’s ₹499 and ₹1,959 recharge packs and Vi’s ₹1,460 pack. Of course there was criticism, because even if we are to factor in validity periods for each of these recharge packs, they still aren’t exactly “affordable”.
Then some more changes happened. The result now — Jio’s ₹448 (84 days) and ₹1,748 (336 days) recharge options, Airtel’s ₹469 (84 days) and ₹1,849 (365 days) recharge plans and Vi sticking to its guns (and whatever few users remain) with the ₹470 (84 days) and ₹1849 (364 days) packs. The intent is clear, to have a user locked into the network for the duration of the validity, and that mission only truly works well if it is a long-duration recharge. The need of the hour may be 28-day or 30-day recharge packs for voice and SMS, for those who’d prefer affordability over the convenience of a longer validity recharge.
Much complication over a simple guideline, mostly borne out from an intent to rake in the moolah.
EVOLVE
What are the three key things that Xiaomi’s tablet aspirations would focus on for the rest of the year, I asked Anuj Sharma, CMO at Xiaomi India, as we sat down for a conversation. Interconnectivity with HyperOS playing a major role, improved productivity such as tackling basics including offline spreadsheet work (think about it, Google Sheets isn’t always the answer) and larger screen tablets. “I guess the question from my perspective would be, can we keep the same weight and mobility but put in a larger display,” he says.
Android tablets have evolved rapidly, in the past couple of years. Undoubtedly more so, in the previous year. Xiaomi, Samsung and OnePlus have played a major role in that transition. CyberMedia Research (CMR)’s Tablet PC India Market Report Review for Q3 2024, pegs India’s tablet shipment trajectory at an impressive 46% year on year growth. For Sharma, the reason is that “the overall expectations from consumers in terms of tablets has changed. It is no longer just a large screen consumption device. It’s now a creativity device and it’s now a product productivity device.”
In the past few months, Xiaomi has strengthened its tablet portfolio. The Redmi Pad Pro and the Xiaomi Pad 7 that follows an impressive Xiaomi Pad 6. Things are looking up for the future of Android tablets, and Xiaomi’s role in the overall ecosystem is gains even more importance.
