Around 1726, Montesquieu wrote, “Le mieux est le mortel ennemi du bien.” Translated to English it reads “The best is the mortal enemy of the good’
Simplified to “Perfect is the enemy of good,” in modern times, it warns against chasing goals which are so lofty as to prevent more practical accomplishments that solve for “good enough.”
I really like Richard P. Gabriel’s variation on this for software development, “Worse is better,” which suggests that fewer capabilities usually means higher quality achieved at lower cost. This contrasted with the earlier Stanford/MIT software development approach, “The Right Thing,” that stressed the need to support all reasonable use cases in simple, correct, and consistent ways.
Today, the Chinese frontier large language models used in modern “AI” cost millions to tens of millions of dollars while the frontier models in the US cost hundreds of millions to develop. This, of course, influences the prices these companies much (eventually) charge to support continued development and a return for investors.
According to the Institute of Electrical and Electronics Engineers (IEEE) the latest benchmarks comparing the leading Chinese models with the leading US models show the Chinese models lagging only 1.7% behind the US ones, the gap closing from 9.26% in the last year.
I believe that the fast followers with slightly worse models and far better prices are going to be the ultimate winners. During this bubble period, the leading US makers have all the cash they need to subsidize end user prices, but that cannot last. At some point, likely not far in the future, the bill will come due; if OpenAI, Meta, or Google have not found ways to lock in users, they will quickly be surpassed by more affordable model makers with “good enough” tech.
There is no “perfect” here to be had. The claims the big US model makers have been making about LLMs becoming self aware and taking over all of human kind’s work burdens were always bullshit.
Bullshit sells in Silicon Valley. Only about 1/3rd of 1% of startups here, including ones with millions or even billions of venture capital dollars invested, ever achieve significant success. Nevertheless, the money spigot is locked in the “on” position, because even just one of those startups taking off, achieving widespread success, can provide the VCs with a lucrative return. See Google and Facebook, Netflix, Uber, Spotify, etc., but consider that for each of those, there were hundreds of failures.
This generation of AI likely will produce significant value, as there are some very solid use cases emerging, such as language translation, code generation, image recognition, materials science, and sentiment analysis. But these practical use cases will soon not require frontier models to achieve their primary goal, user satisfaction. Language translation, for example, has been quite good for most popular languages for years, already working quite well for most users and eeking out a 1% improvement will not grow that audience meaningfully.
The fast follower models, the far cheaper, worse is better, good enough models whose makers deliver something that’s minimally sufficient to real customers, will surpass the ridiculously expensive frontier models from companies investing billions in pursuit of now diminishing returns, hoping to achieve a fantasy.
LLM AIs are not going to “wake up” and they’re not going to displace the entire human work force. That’s all nonsense. Further, the top models in are no longer seeing the drastic improvements with each newer and more costly release.
LLMs are reaching a plateau where ingesting even more data and buying even more GPUs to crunch that data is hardly moving the needle. In that light, the billions that continue to pour into these attempts is mostly wasted. Those who pivot early to building practical and user friendly features on top of far cheaper models will take the lead.
Microsoft has already signaled such an approach. They have been de-emphasizing the frontier models for a while, to focus on smaller, more targeted models that support the most practical use cases yet to emerge. They’re pulling back on new data centers and letting OpenAI and others chase perfection while they begin to tackle good enough. In this case, as with most popular software over the last 30 years or so, worse will be better.
Comments