Hacker News Reader

Comment by AngryData | Hacker News Reader

I think the tech industry is overestimating its value. Because it can code they think it can do anything else, but unlike code a lot of other work can't have a bunch of little bugs and mistakes because you can't open up real life and edit it after the fact. Plus it lacks actual reasoning to solve novel problems.

Hiring an engineer to finetooth comb blueprints for mistakes before construction will take nearly as much time as having an engineer draft them themselves. And they will be smart enough to not do something silly like putting the electrical panel on the back of the shower wall. If you just vibecode some blueprints and start construction without the comb you could lose way more than you saved with something as simple as pouring a support pillar or building a wall 2 inches off and having to tear it out later and rebuild.

View on HN

The success of AI doesn't hinge on whether you can vibecode it all or even one particular sector really well. For example, despite several attempts to make vibecoding PCBs, it's still pretty crap. But it's really useful as a copilot, human in the loop for targeted tasks in electronics. Same for CAD work, not so good at drawing but still useful at looking at an image, understanding it, and answering specific questions.

Whether the value attached to these companies is grounded in reality is a different question.

The reality is experience isn't codified in text.

It's passed person to person over long periods of time.

And you can't train an LLM on that.

Yes. For every line of code or documentation I write there is a lengthy internal dialog that can never be scraped & used by an LLM.

It can barely read a datasheet correctly, in my experience, getting confused between different registers.

Yeah throw it at a moderately complex STM32 clock arrangement and see what comes out.

I'll be more forgiving: I found it gets confused easily when it extracts text from a PDF because datasheets tend to be written in a not so parseable way for a machine. But 1) if you tell it to take screenshots of said datasheet, it'll have greater success. 2) We're in year n<5 (depends how you count) of the age of LLMs, this will get better over time. Either on the LLM side or the thing you feed to it.

screenshot of the 300-page datasheet? (not to mention the 4000 page TRM?) no thanks.

If the datasheet is only a few pages I'll read it myself :).

Definitely. The secret will be identifying use cases where AI usage is a potential upside with limited downside, not the current blanket statements about replacing all jobs without considering lifetime ROI. There’s a lot of boring work AI can automate with minimal risk. There’s also the potential to decrease risk with AI too, including ensembles of different AIs modals and AI + human.

A human that hand-compiles 10,000 lines of C code is a very silly person. A human that works on device drivers and drops into assembly code for a dozen highly critical lines to enable real time communication can be irreplaceable. AI is a tool that can be highly useful and it's a tool with a number of large flaws that you need to acknowledge and account for. Knowing when it's worth using is a vital skill.

"There’s a lot of boring work AI can automate with minimal risk. There’s also the potential to decrease risk with AI too, including ensembles of different AIs modals and AI + human."

I think the trouble, economically speaking, is that while it will be possible from a purely technical standpoint to unbundle a job performed by a human into separate tasks, many of which can be "done" by agents, the new process will not present a cost savings overall once the entire lifecycle of the task is taken into account. The economist David Autor has written about these challenges extensively, and his theory accords with my experiences.

Conversations about the costs of inference never consider the reality that API pricing is significantly higher than the operating costs.

Nor do they ever consider that the cost of datacenter hosted inference has to crash when the bubble pops and hardware vendors can't fill orders at sky high prices created by demand anymore and the hyperscalers can't keep things running near capacity at the high demand prices.

All of which leads to the ROI math for implementing AI looking much different.

Has everybody forgotten how much money Nvidia, TSMC, and all the hyperscalers are making, today, in pure profit? The costs of inference are high because we're in a bubble.

I think many of these problems still arise if inference is effectively free in monetary terms to the end user. In many economic processes, time to getting the final and correct answer is the major driver of profitability.

I think also, just seeing non-technical friends and family interact with it, there’s a lot of massaging you have to do right now to get it to work that just goes over their heads. Until they gets pushed behind an abstraction layer I see adoption crawling at a certain point.

Reply on news.ycombinator.com