AI: Ignore first instincts from DeepSeek, do the OPPOSITE. RTZ #617
...need MORE Global AI Infrastructure, AI Distillation,Bigger LLMs, and LESS China curbs
There is a common tendency for us all to make quick linear connections from events, wired into our primeval brains. Action—> Reaction. That is something we’ve seen all week from the DeepSeek moment this past weekend. And the reactions are still reverberating and coming, mostly in this linear A to B way.
But as Seinfeld showed us in a famous episode, at times the right answer is likely to DO THE OPPOSITE. More non-linear logic the better.
In that vein, let me highlight FOUR things that should be kept in mind, totally counter to the first instincts the DeepSeek revelations seem to be evoking in most stakeholders thus far, in this AI Tech Wave:
Going to need MORE AI Infrastructure, NOT LESS: Despite the knee-jerk selloffs in AI exposed big tech companies, especially chip companies led by Nvidia, a key takeaway from DeepSeek is that we’re going to need to accelerate our AI infrastructure needs post-DeepSeek. The open source innovations they laid out, point to faster ways to do AI training and inference, that are far more efficient in operation and price. Thus the reactions by Meta and Microsoft particularly in their earnings calls yesterday, that they’re continuing to lean into AI investments, is in the right direction. As is the OpenAI/Softbank announcement of up to $25 billion being invested in OpenAI beyond Softbank’s $15 billion commitment for their joint new Stargate Infrastructure entity with Oracle, MGX and others, is also the right ‘Opposite’ thing to do post-DeepSeek.
Distillation is a Feature, NOT A BUG: One development that got a lot of attention and hand-wringing amongst technologists and regulators post DeepSeek is that the company ‘distilled’ larger models, both closed and open, from providers like OpenAI, Meta and others. This is a popular and growing practice in recent months to use larger models to ‘train’ smaller models with more specific prompts for reinforcement learning, both supervised and non-supervised. DeepSeek came up with some core innovations here, but the underlying distillation practice was not unusual for them to do. They just did it better, and open sourced new ways to do it better. And that’s a feature, not a bug. Good for the AI industry going forward. It’ll likely be baked into more licensing agreements from the larger model companies, and may even become a new product for the LLM AI companies, via APIs and direct use. We’re likely to see what I’d call ‘micro-distillation’ that automatically does this for billions of mainstream users, as they use AI applications and services with their personal data and usage priorities. This one is a key area to do the opposite by licensing distillation as a product.
Bigger LLM AI Models are NOT A COMMODITY: One concern I heard a lot from smart investors post the DeepSeek news, is that the large AI hyperscalers, spending tens of billions to train and scale bigger AI models, may be getting commoditized. That they may not have a good, fundamental long-term business model. Here, the OPPOSITE is more true than initial linear instincts. We have not yet begun to scratch the surface of Scaling AI, especially with data beyond all the internet to date. Both the online and offline (physical) worlds offer FAR MORE, NEVER-ENDING streams of data to leverage larger language models going forward. That is why OpenAI, Anthropic, Google and other LLM AI companies have bigger, more valuable commercial opportunities than ever before. DeepSeek just accelerates those ongoing trends if applied well.
Less US/China tech AI curbs, NOT MORE: This instinct alas will likely not be curbed, given the current laser focus by DC (White House, House and Regulators) on their favorite framework of viewing AI through the keyhold/framework of a AI Chip curbing competition with China. A ‘Sputnik’ moment/AI Space Race as it were. A Super-bowl of AI, where there is presumably a singular moment of victory or defeat. A landing on the moon, even though the world then presumably loses all interest (see Apollo 13). A vivid moment that AGI is declared, and everyone assumes one side won and the other lost. Fortunately, THIS AI ‘race’ is not a ‘win-lose’ game like that. It’s just an iterative innovation, Moore’s Law/AI Scaling driven AI Tech Wave that will likely go on for decades. With no COUNTRY a winner or loser. Just Billions of users and Millions of businesses. And trillions in global GDP prosperity added due to AI, to a $110 trillion+ global economy. Where billions get to use AI in unimaginable number of useful and profitable applications and services, just like we use regular computing today. And have been for over 70 years. Not to mention Electricity for over a hundred. This time, we’re just using AI Data Centers, and ‘AI Factories’, with sufficient electrical power to produce ‘Intelligence Tokens’, both as input and output of never-ending calculations, to figure out how to the things billions of people do everyday, just far better than we did it until now. So the AI Race is a race for ALL of us. Not one country against another. It’s what would be clearer, if DeepSeek had been anything but a company in China. Imagine if DeepSeek were an Israeli company as I argue here (video chatper guide at 11 minute mark). So if the instinct to curb AI Tech trade with China was curbed itself, we would all likely be taking away the BIGGEST revelation of what DeepSeek showed the world with their Open Source work under constraints.
Those are four instincts we should try and ignore this extraordinary week of DeepSeek news.
And do the Opposite in this AI Tech Wave. Stay tuned.
(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here)