Hacker News Reader

Comment by tapoxi | Hacker News Reader

We can use a non-deterministic system because we have test suites that act as validation gates.

Other industries don't, so they have no way of knowing if the AI is bullshitting them outside of manual validation, which makes it more expensive and slower than doing it yourself.

View on HN

It's a good point but in most of the places i've worked an average person could leverage any 2026 llm to improve things 50% easily. And I'm more anti-llm than not somehow.

    - fixing regular computers issues (confusing browser or os behavior)
    - finding more possibilities with your current programs (around 2020 someone once asked me if I knew a way to capitalize a word on a computer, when she realized there was a dedicated shortcut she became ecstatic)
    - having a rubberduck to discuss your daily work dead-ends
    - thus decoupling from team friction as you need to ask less questions to people that might hurt you.

In more technical places, having a llm write small automation would also add up.

As a supporting example for the first point: I recently had a dozen ore more colleagues need to be hand-held through getting basic pieces of their local Python environment working, when the multiple agent tools they have access to could have easily solved it for them if they had just asked - but instead they came to me. Also, these were not swe's but still technical people, so I would have assumed they'd be the group who didn't need that help.

I strongly agree that search engine is the best application of LLM so far. 90% of time I don't look at Google search results, I look at AI Mode output only. And you don't need agents, harness, guardrails, etc for it — basic small clean LLM does the job well without overdoing it.

There are many tasks humans perform which have no or slow feedback loop. Even if they can audit/validate AI output, they don't know if it will be good when it reaches other humans.

> no way of knowing if the AI is bullshitting them

Back testing, hold-out sets, cross validation.

This has all happened before.

> they have no way of knowing if the AI is bullshitting them

That hasn’t stopped Tesla. Their self-driving cars work by connecting the car’s controls directly to a neural network.

And they're still a decade away from general road use outside of vetted streets. Tesla is still struggling with pedestrian detection in rain tests.

Your information seems to be outdated, unless every single street I’ve been on for the past year is “vetted”. All the other (consumer) car brand’s self-driving systems do have that limitation though, maybe that's where the confusion came from. Tesla is self-driving 98% of my mileage, with the 2% being when I take over to choose a specific parking spot or I’m towing/carrying something big.

Don’t know much about how it does in the rain since moving to SoCal but in clear weather it hasn’t come close to hitting anyone yet. Even proactively warned me a few times while I was driving. Maybe I should let it take over in parking lots too :)

Someone take this guy's license

It'll be hard, when I'm never causing any accidents or speeding or running reds/stop signs. Maybe they could get me for inattentive driving but that's what the tinted windows are for.

isn't that more a consequence of them using a cameras-only approach vs. Waymo/others who have other sensing-modalities to fall back on (lidar mainly)?

They have world simulators which verify, to an extent, the behavior of the software

I thought that pretty much every self driving algorithm in production was based on machine learning of some sort.

In most cases it’s only the object recognition systems that use machine learning. I know that Tesla was talking about using non-deterministic AI for the decision making as well, but it’s a worry if they’re already doing it.

Makes sense. Thanks for the clarification

Earlier Tesla self-driving software used search-based planning and conventional optimization alongside neural networks. Tesla has publicly demonstrated A* in its path planning work.

A lawsuit revealing part of the camera analysis showed them using a classic CNN based object recognition network.

For those who won't act ethically or at least reasonably, regulation is the answer. We're not too far away from flipping the congress into something far more willing to do just that, especially with anything connected to Elon Musk.

test suites cover nowhere near enough to assure reliability

and if they did for your system, then it's so simple that whether or not you used AI doesn't even really matter

On the other hand, if humans could apparently do that work without automated verification, then perhaps so can agents.

right cause test suites never have issues

Yeah and getting a single example of a correct program is way easier than describing the set of all correct programs.

Reply on news.ycombinator.com