> You/we are all susceptible to this sort of thing, and I call BS on anyone who says they check every little thing their agent does with the same level of scrutiny as they would if they were doing it manually.
Why? I do that. I give it broad permissions but I also give it very specific instructions, especially when it's about deleting resources. I work in small chunks and review before committing, and I push before starting another iteration (so that if something goes wrong, I have a good state I can easily restore).
I'm the one with the brain. The LLM can regurgitate a ton of plumbing and save days of sifting through libraries, but it'll still get something wrong because at the core it's still a probabilistic output generator. No matter how good it becomes, it still cannot judge whether it's doing something a human will immediately spot as "stupid".
Interacting with and fixing API calls automatically is something that normally works for me, but allowing the agent to run a terraform destroy is something I'd have never let it execute, I'd have been very specific about that.
Yeah, I agree. I've used it until 3 days ago, then after ~1y I got tired of taking the occasional pictures with HDR on and waiting 3-5 minutes for them to be processed and saved, while producing 3 other copies in the gallery.
Quality looked amazing, but the pre-installed phone camera gets close enough and it's instant.
> Postmarketos clearly exists and works, EVEN if LLMs were absolutely perfect for speeding up development ten folds, is there any absolute moral necessity to use them?
There's no moral necessity, but if you want to survive as a project moving forward, you'll have less and less velocity compared to projects using LLMs, so you'll eventually shrink and die as a project, because less people will contribute to a project that gets less features and bug fixes.
I don't understand why these projects have such a strong "moral" stance of "no AI ever", and instead they don't deploy LLMs to automatically review PRs based on their own guidelines, so that if the contribution is valuable, it gets through no matter if it was written by an LLM or not.
Mine was more a generic argument against the "ban all AI" stance that I've recently seen pop up more often.
At the moment, there isn't another project (that I know of) like PostmarketOS filling the same niche. If a new project were to appear, and were using LLMs, it'd likely progress faster.
Regardless, I've had success with LLMs and while I understand the maintainers' concern, if used properly they're a powerful tool to quickly iterate on huge amounts of information. They could be used to automate reviews of the spam of low-quality PRs, for instance (if they were to materialize).
But having read their policy page, their stance is more on ethical grounds, not moral: https://docs.postmarketos.org/policies-and-processes/develop... . So while I still stand by my argument in the general case, here it's not applicable, and while I see their ethical concerns, one project boycotting a tool doesn't really fix the systemic issues they mention.
In Github it needs to be explicitly configured (Settings > General > Delete head branches after merging), Gitlab is the same.
A lot of my developer colleagues don't know how git works, so they have no idea that "I merged the PR" != "I deleted the feature branch". I once had to cleanup a couple repositories that had hundreds of branches spanning back 5+ years.
Nowadays I enforce it as the default project setting.
No, Gitlab has it too (don't know about the others offerings). Gitlab integrates Advanced Search via Elasticsearch, last I checked they had plans to make zoekt available as an alternative.
As another Tesla owner, I've had this discussion with my brother one too many times. He'll insist he needs buttons, he needs Android Auto/CarPlay and whatnot. Every time I step into one of those cars I'm overwhelmed. Half of the times it doesn't connect, when it connects I get useless notifications for everything. It's not well integrated, and it'll randomly break during the trip.
I understand it has become a standard but it's not a particularly good one, and adding it "just because it's a standard" would detract from the car experience in my opinion. It's a separate device, with a separate OS, kernel, apps etc where you can install almost anything, that's supposed to take over a piece of equipment that belongs with the car and controls all its functions. I'd really rather not have that.
If the infotainment is the basic "show 2D maps and a couple settings", then Android Auto/CarPlay can serve as a viable replacement for low-end cars. But when the car costs >30k and the screen is also the central command console, no thanks. I'd rather have proper OTA updates, give feedback, and see it evolve over time for the better.
I mostly agree with the points, but I've also managed to throw AI efficiently at the problem.
We're running a self-hosted GitLab -> hosted GitHub migration at my company (which to me feels a downgrade), and without LLMs I would have spent weeks just researching syntax for how to implement the requirements I had.
I asked Claude to simply "translate these GL templates to GH actions, I want 1 flow for this, 1 flow for that, etc" and it mostly worked. Then in the repos I link the template and ask Claude to write the workflow that uses the template with the correct inputs. I think I saved maybe 3 months worth of coding and debugging workflows. Besides maybe picking slightly outdated actions (e.g. action@v4 instead of action@v6), 95% of the work was ok, and I had to tweak a couple things afterwards.
Looks like my job is ensuring stuff builds, tests and ships correctly, not learning the 100th no-design botched homegrown language that will keep changing for the next 10y until it's a different thing altogether. And because I'm one person out of two in a ~15ppl company, where time and efficiency matter, LLMs really helped out.
> Tesla deserves credit but it’s not Tesla that accelerated the world. It’s China.
China supplied the batteries, sure.
But there's definitely a pre-Tesla and post-Tesla world regarding the vehicles themselves. Tesla changed the image of EVs available to the general public by making performant and low maintenance vehicles that looked futuristic and were capable of things basically only supercars could, for a fraction of the price. And they built the DC charging infrastructure all over the world to support long-range trips, which was non-existent before Tesla. EVs before Teslas were basically niche experiments.
> EVs before Teslas were basically niche experiments.
The Nissan Leaf (which predates any Tesla production car) was pretty much an electric version of previous Nissan cars. The VW eGolf (contemporaneous with the Tesla Model S) was _literally_ an electric version of a previous VW car. The VW ID.3 and 4, which are currently leading the European market, are also pretty much like VW electric cars. In practice, 'weird' electric cars mostly failed.
Have you ever driven more than 200km at an average of 80km/h with enough turns on the highway? Perhaps after work, just to see your family once a month?
Driver fatigue is real, no matter how much coffee you take.
Lane-keep is a game changer if the UX is well done. I'm way more rested when I arrive at destination with my Model 3 compared to when I use the regular ICE with bad lane-assist UX.
EDIT: the fact that people that look at their phones will still look at their phones with lane-keep active, only makes it a little safer for them and everyone else, really.
If you're on a road trip, pull the fuck over and sleep. Your schedule isn't worth somebody else's life. If that's your commute, get a new apartment or get a new job. Endangering everybody else with drowsy driving isn't an option you should ever find tenable.
Maybe I wasn't clear. I've always stopped to take a nap when I had the inferior car, because I knew I was risking my well being and that of others. Ever since I bought my Model 3, I never had to.
As an example, after a refreshing 8h night's sleep, you put yourself at the wheel and you know you need to drive for 8h. After 2h, you'll feel the need to stretch your legs, have a bite, go to the toilet.
But if you have something like Tesla's AP, you'll still be alert and awake. Otherwise - at least for me, but I bet for many people - the constant nagging of micro-correcting the wheel, keeping the distance to the car in front, keeping the right speed, takes a huge toll on your mental resources, so much that it puts you to sleep whether you want it or not, and many people will try and push through that regardless.
You're saying that cruise and lane assist are dangerous because they're used to do bad things.
I'm saying they're very liberating from the busywork that is driving, and even people that use them as an excuse to stay on their phones are better off for it, because something more alert and precise is driving for them and as a consequence, less harm can derive from their stupidity.
I'm not saying that people that do stupid things should not be punished, and I'm quite happy that Tesla had to implement better driver monitoring. I know how things work and that I always need to monitor the system, but many people assume it's a magic button and disconnect their brain.
> The problem is knowing which lane keeping assist systems are good and which are not. Every dealer just treats it as a 'checkbox' item and implementations vary by model and year.
Yep, I agree.
I used to travel to my parent's home 300km away once a month, and changing from a 2010 no-assist car to a Tesla Model 3 with AP (not FSD) back in 2019 was a game changer. I used to drive there on a Friday evening after work, and I basically collapsed into bed when I got there. With AP I was still tired of course, but also still functioning and way more alert. In my experience Tesla's AP UX is very good: chime when engaging, chime when disengaging, you don't need to look at the screen to know the state you're in, and if you touch controls it lets you know (via chime) and deactivates.
One of the most horrible UXes for me has been on a new Hyundai i10 with the basic lane assist (and I know it's very similar on a new VW Golf that my cousin is leasing):
- there's no chimes, you're forced to look at the screen at the center of the dashboard
- said display is 100x400 (or sth similar) 16-color pixel screen in the center of the dashboard
- out of said display, you need to look in the very corner for an icon of 10x10 pixels that can be yellow, green or white (which under low backlight/high contrast conditions can be tough to decipher)
- lane-keep is on by default at every car start, and tends to butt in on twisty roads (very common where I live), so half-way through a turn you'll feel the steering wheel literally lose force-feedback, while you're still applying force, and swearing ensues
- someone thought that constantly reminding people of the speed limits was a good idea, so the car will scream incessantly at you for being 50.1 over 50
- but will happily let you change lanes and re-engage auto-steer automatically (you need to manually enable this) while doing 120km/h on the highway without any hint that it's re-engaging automatically
- the speed warning is yet another setting that you can turn off at runtime, but you can't persist properly
- auto-steer, after is manually engaged, will stay happily engaged even after you leave they highway and are at very low speeds, and will try to correct you when doing roundabouts
I think the Tesla UX is way better there, and I think regulatory bodies should start preventing things like the i10 assist to be sold to customers, because they're actively dangerous. I've literally had minor heart attacks due to the lane-keep butting in on twisty roads - I thought the front tires were slipping for some reason.
I was driving until not long ago a VW with all those systems and, while less pleasant than the current Audi, I never met any of those inconveniences you met. Yes I know sample size, but... So, lane assist had zero visuals involved - activated or not activated, and giving your wheel a soft bump in that direction (it never tried hard turns, for better or for worse). The downside is that construction sites with colorful lines would indeed confuse it, so you'd either press the disable button, or keep your hands on the wheel (wife complained about it, but for me it was acceptable). The Audi handles construction sites perfectly, so far (soon a year). Speed limits meant on both cars only a visual some place, so you could ignore it (Audi highlights on the HUD the speed limit). So all in all, I believe Hyundai f'up it big time, or something got really wrong in that car - a perfectly good reason to give it up either way.
Why? I do that. I give it broad permissions but I also give it very specific instructions, especially when it's about deleting resources. I work in small chunks and review before committing, and I push before starting another iteration (so that if something goes wrong, I have a good state I can easily restore).
I'm the one with the brain. The LLM can regurgitate a ton of plumbing and save days of sifting through libraries, but it'll still get something wrong because at the core it's still a probabilistic output generator. No matter how good it becomes, it still cannot judge whether it's doing something a human will immediately spot as "stupid".
Interacting with and fixing API calls automatically is something that normally works for me, but allowing the agent to run a terraform destroy is something I'd have never let it execute, I'd have been very specific about that.
reply