Comment by cadamsdotcom | Hacker News Reader

Comment by cadamsdotcom | Hacker News Reader

The value of knowledge is plummeting. It can, will, and should be subsumed by LLMs. Consider London’s black cabs. Drivers pass a gruelling exam to prove they know “everything” about one part of London. But, Google Maps puts that knowledge and more on every rideshare driver’s dashboard, for free, no advance study needed.

Map knowledge doesn’t make you able to drive. Driving happens in real time - steering, obeying road rules, avoiding accidents.

In engineering, the real time part is when you use mental models you’ve painstakingly developed - how the hardware works so you can debug it, what syntax is valid in your programming languages, what APIs exist in your stack. But, those depend on knowledge - so mental models can be self-taught with AI as tutor in your own time. The LLMs supply the knowledge, you integrate it via study - or a side project.

An organization cannot clone itself a team, it’s true. They need engineers now, not in 6-12 months. But - a motivated individual can use AI to make themselves employable as a full stack engineer faster now than at any point in history.

How cool is that!

One of the big issues with LLMs for e.g. systems software is that there are broad domains where training data effectively does not exist. Consequently, you can’t learn much from them of value. It is the blind leading the blind.

The lack of training material for LLMs is of course a lack of training material for people too. Some areas of software have a long history of relying almost entirely on an oral tradition to pass down knowledge. This has some advantages but it doesn’t scale and it makes it basically “dark knowledge” for LLMs or people without access to those that know it. If you want to get into an area like this, you often need to find a way to spend a lot of time with people that already know it.

>This has some advantages but it doesn’t scale and it makes it basically “dark knowledge” for LLMs or people without access to those that know it.

We call this "tribal knowledge" in games. I despise it. It's done for a few reasons:

- NDAs make public knowledge a landmine. And every game studio makes you sign NDAs. Even at the interview stage.

- Churn. No one gets time to really develop expertise as they work on a project for 2-3 years and then layoffs come. Only a relative few become experts, and probably not because of the studio itself

- lack of incentives. Games aren't very connected with acedemia to begin with, despite relying so much on cutting edge tech. So the best resources for sharing such techniques is shafted. This is slowly getting better as more tech conferences talk about games tech, but it's a pretty slow tricke unless you come from one of the largest studios and specifically come in for R&D.

>If you want to get into an area like this, you often need to find a way to spend a lot of time with people that already know it.

All too true. Open source development is one bastion for this, but that's overall why I keep trying to stay in this domain. You literally can't get the knowledge elsewhere. And it's knowledge that directly leads to better looking, more optimal, and less buggy games overall.

I work in high-end data infrastructure, not games, but almost identical incentives and dynamics are at play. The state-of-the-art research isn’t coming out of academia for the most part. The R&D being done in private industry is slathered in NDAs that only slowly leaks out a decade or more after it was put into production. Many people don’t stay with it long enough to really master it.

There are some elegant and sophisticated techniques related to database kernels that have been passed around for a decade or two over beers that still don’t have a single reference in literature that I can find. The original researchers probably stayed quiet because it was under strict NDA but also likely retired years ago. No one writes it down because it sort of feels wrong to claim second-hand knowledge of unknown origin as your own, or to even lead people to assume as much. You show people, and they think it is amazing, but when they ask you for references or sources you have no idea where it came from.

There is a real gap here and it is getting worse.

From a purely humanity-improving perspective this lack of dissemination is quite sad.

> No one writes it down because it sort of feels wrong to claim second-hand knowledge of unknown origin as your own, or to even lead people to assume as much.

Wouldn't it be possible to publish informally (say in a blog post) while fully disclaiming first-authorship or invention?

>It can, will, and should be subsumed by LLMs.

okay. So what to do in the meantie while an LLM can no in fact help you much with kernel programming? People love talking about the future but people also have current business needs.

> a motivated individual can use AI to make themselves employable as a full stack engineer faster now than at any point in history.

I'd love to hear one case of that happening. I'll even take someone ramping up from zero to freelance work as long as they have a good portfolio.

Maybe have a chat to these folks in the article?

Seems they are plucking self-starters directly from universities. You could either use that funnel or build your own with the same idea.

Sounds like a good plan. University recruitment is how you invest in society's future. I don't know how large this company is, so it's hard to determine if they can afford a training program.

Reply on news.ycombinator.com