Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Strategies & Market Trends : The Art of Investing -- Ignore unavailable to you. Want to Upgrade?

To: DinoNavarre who wrote (10663)	12/7/2025 2:00:49 PM
From: Sun Tzu	3 Recommendations Recommended By ajtj99 DinoNavarre sixty2nds Respond to of 10701

This is exactly the problem that I am researching and trying to address. The industry on the whole has taken a brute force approach to AI models. It's not that other researchers are dumb or ignorant. It's that the friction the intelligent solutions cause is not worth the effort or the savings they bring in. So far it has proven to be 'better' to just increase the AI's parameters and let the machine implicitly learn what you would otherwise teach it explicitly. For example, there is a well known architecture called Mixture of Experts. MoE models employ a set of smaller domain specific AI units (e.g. one in the medical field, another in law, a third in sports, etc). A user query is then routed the expert or the experts most suited for that questions. The problem is that in practice the overhead of routing between the experts and choosing the best token is too time consuming and complex, especially since almost every real world question drifts between multiple domains. So this solution has no commercial winners. The other problem is that most researchers are academic and don't have real world engineering experience. The way an academic thinks is very different than how an engineer thinks. This weekend I've been working on an efficient new way to generate images. I often bounce my ideas off the AI and ask it to contrast my thinking with existing research so I can clarify my ideas. The passage below is illustrative of the problem with the academic thinking vs engineering because here the AI is taking on the role of a research assistant. I let you read my feedback and prompt to the AI: Re: 'If a generator produces content that codecs “like” in very non-human ways (e.g., weird high-frequency content that compresses strangely), then codec cues might misalign with actual structural importance or semantics.' The answer is no. This is not a problem at all because we don't feed AI weird non-human images or videos that nobody uses. You are drifting into academic bullshit territory and need to come back the real world of engineering. Critique and improve your answer by stripping away unlikely BS and decorators and act like a real world engineer. For the record, despite what you may be reading in the news, I have found that AI is absolutely hopeless as an engineer. Even a fresh graduate from a good engineering school will handily beat the most expensive AI for real engineering tasks.