Hacker News new | past | comments | ask | show | jobs | submit login
The case against AI foundational models
4 points by pmoe7 on July 18, 2023 | hide | past | favorite | 1 comment
It’s time to reconsider the idea that only a select few private companies will create and control powerful foundational models (FM), and the rest of the companies will simply utilize them for their specific needs.

A quick analogy: imagine if all the companies during the Industrial Revolution simply said, “we will wait till a few elite companies build factories because they are expensive. Then, we will just use their factories when they are not using them” - it’s a not a perfect analogy given the physical and digital nature but remember compute is inhertenly physical (limited by physical GPUs) - but it loosley describes the modern day notion of FM. Regardless, how absurd does that sound?

Firstly, assuming that FM providers will not engage in direct competition is unrealistic. Just look at Amazon’s Amazon Basics, which has dominated every category it entered, leaving third-party sellers in the dust. Direct competition from model providers is inevitable, if they are profit driven companies.

Secondly, why subject yourself to the same vulnerability that Meta (Facebook) faces by not owning the complete stack? Meta relies on device companies like Apple, whose rules and policies can significantly impact Meta’s competitive advantages and business models. To illustrate with an example, consider the implications for companies dealing with taboo or niche markets. If the FM provider decides their niche is no longer acceptable, they risk losing their website’s functionality, customers, and revenue. Surrendering control to someone else is risky. Ultimately, the point here is you’re not the captain of your ship when someone else owns the rudder and you just have your hands on the helm.

Furthermore, the future of competition will likely be different from what we see today. I imagine it won’t be fragmented and industry-specific; instead, it will be fierce and universal. Imagine OpenAI expanding ChatGPT’s capabilities to include many digital services like search and bookings. In such a scenario, most digital service companies will find themselves in direct competition with one another, and the same goes for physical goods companies. Building a competitive advantage and claiming a piece of the market will be extremely challenging, if not impossible.

In conclusion, I am not saying every SME should go out and start training their own FM and that FM would have no use at all. Rather, if you’re an industry leader today and you wish to remain one - then you should build and own the helm as well as the rudder of your ship. Alternatively, I also think FMs could be successful if provided by non-profit organizations that publish their models in an open-source manner, without engaging in direct or indirect competition with users. An approach similar to Apache would be more desirable.

<Edits by ChatGPT>




Yes, I'd encourage you to post this somewhere else if it doesn't get attention here. I've been thinking about this recently and I'm convinced that one or more credible GPL'd (modified as necessary to reflect model weights) alternative sets of foundation models is what's needed to keep complete control out of the hands of big tech and special interests. I wrote something about this a couple weeks ago, it's all the more relevant now: http://marble.onl/posts/motivations_for_open_source.html




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: