To clarify, does this mean that Anthropic employees don't understand Claude Code's code since it's level 7? I've got to believe they have staff capable of understanding the output and they would spend at least some time reviewing code for a product like this?
A friend who does game design gave me a good tip -- start with the core game loop first and only focus on that with low poly / representative shapes for game elements that you can refine in the future. Not until the core game loop is fun does it make sense to spend a minute of time on any other aspect of the game.
It definitely does not respond to flashing headlights in that manner. You’re observing its default behavior when at a 4 way stop with other vehicles not moving.
You're right I don't have inside information, but we've been interacting with them on the street for years in SF. Waymos don't wait for human subjective guidance to give them clearance to pass, as evidenced by tons of videos and IRL experience. As soon as they come to a required stop, and if a vehicle or other object's linear travel path does not intersect it, it will go. Flashing lights will not change this behavior. (Yes you're right there is a regulatory requirement to respond to safety officer guidance, but compliance is spotty as evidenced by a lot of videos of vehicles entering active crime zones, etc.)
Unlike the traffic cops directing traffic that would likely require special programming, "proceed if the other car flashes its lights at you" is completely the kind of thing that could just accidentally fall out of a neural network learning to imitate humans.
Hopefully if they ever go to Sri Lanka they get localised tuning because I was surprised to find out flashing your lights over there doesn't mean "go ahead", it means "if you don't get out of my way I will ram you"
And then there's trucks flashing an indicator to say it's safe to overtake if you're behind them. In the UK it's the nearside indicator, which makes sense: it's a bit like the truck is pulling over to let you pass. In Aotearo, it's often the off-side indicator, so you think the truck is going to pull out in front of you. I've never understood what the Aotearoa drivers are thinking there
It also doesn't make sense because "get out of my way or I will ram you" is the default state of operating a motor vehicle. Not the goal but the physical reality of it.
I think we're not interpreting the original comment in the same way.
In most places, I think, when driving on the highway, flashing your lights when behind someone means basically 'I would like to overtake you'. Same here in the UK. But that's very specific to that context. You would never see a 'go ahead' context that would mean 'get out of my way', right?
But what the original comment means is there are some countries where you'd think it was 'go ahead' but it really means 'get out of the way'. Like if you're both on a main road, and you are signaling to turn into a side road, the opposing car flashes the lights and that means you can turn. I assume the same in Serbia.
But in some places that can actually mean don't turn, I'm going first. Which I think is what the parent is describing.
You are right that I did not read it the same way, and yes, the unwritten rules are matching in Serbia. FWIW, I've mostly switched to using left-turn signal to indicate "I'd like to overtake", which I've seen done on EU highways.
Waymo has published a ton about the imitation learning they've been using since 2018. They're not imitating random cars but their drivers who are paid to drive around and follow traffic laws.
It's not enough so they use heavy reinforcement learning etc. but it's still a huge foundation to build on.
Waymo immitates humans insofar as its neural net trained on avoiding collisions after millions of miles of video footage and LIDAR data on roads shared with humans causes it to immitate humans.
It's likely manually programmed not to (incorrectly) turn the wheel to the left while stopped and waiting for an opportunity to turn. If you get rear-ended, you'll end up in the lane of oncoming traffic. It's certainly programmed to use its turn signals to indicate when it is going to turn. But after driving around thousands of cars without turn signals on but with their wheels pointed left, it "knows" to predict that they're about to turn, and might immitate humans by anticipating that action and moving to pass the stopped car on the right.
What else is an LLM supposed to do with this prompt? If you don’t want something done, why are you calling it? It’d be like calling an intern and saying you don’t want anything. Then why’d you call? The harness should allow you to deny changes, but the LLM has clearly been tuned for taking action for a request.
Ask if there is something else it could do? Ask if it should make changes to the plan? Reiterate that it's here to help with anything else? Tf you mean "what else is it suppose to do", it's supposed to do the opposite of what it did.
I think there is some behind the scenes prompting from claude code for plan vs build mode, you can even see the agent reference that in it's thought trace. Basically I think the system is saying "if in plan mode, continue planning and asking questions, when in build mode, start implementing the plan" and it looks to me(?) like the user switched from plan to build mode and then sent "no".
From our perspective it's very funny, from the agents perspective maybe very confusing.
First, that It didn't confuse what the user said with it's system prompt. The user never told the AI it's in build mode.
Second, any person would ask "then what do you want now?" or something. The AI must have been able to understand the intent behind a "No". We don't exactly forgive people that don't take "No" as "No"!
for the same reason `terraform apply` asks for confirmation before running - states can conceivably change without your knowledge between planning and execution. maybe this is less likely working with Claude by yourself but never say never... clearly, not all behavior is expected :)
Genuine questions: what do you think the request was? To build the plan? To prepare the commit? Do you never have a second thought after looking at your output, or realize you forgot something you wanted to include? Could it be that they saw "one new function", thought "boy, there should really be two... what happened?" and changed their mind?
To your original comment, it would be like calling your intern to ask them to order lunch, and them letting you know the sandwich place you asked them to order from was closed, and should they just put in an order for next Tuesday at an entirely different restaurant instead? And then that intern hearing, "no, that's not what I want" saying "well, I don't respect your 'no'" and doing it anyways.
"Do X" -> "Here are the anticipated actions (which might deviate from your explicit intent), should I implement?" -> "no, that's not actually what I want"
is a clear instruction set and a completely normal thought pattern.
My point is that I don't see any scenario in which sending a ton of input tokens (the whole past conversation) and expecting output that literally does nothing is absolutely pointless.
Like, sure, if the model were "smarter" it would probably generate something like "Okay, I won't do it.". What is the value of the response "Okay, I won't do it."? Why did you just waisted time and compute to generate it?
> Do you never have a second thought after looking at your output, or realize you forgot something you wanted to include? Could it be that they saw "one new function", thought "boy, there should really be two... what happened?" and changed their mind?
Sure, all of those are totally valid. And in each of this cases it would be better to just don't make the request at all or make the request with the correction.
> like calling your intern
LLM is not human. It can't act on it's own without you triggering it to act. Unlike the intern in your example who will be wasting time and getting frustrated if they don't receive a response from you. With model you can just abandon this "conversation" (which is really just a growing context that you send again and again with every request) forever or until you are ready to continue it. There is no situation when just adding "no" to the conversation is useful.
Yeah meat is another dimension, as is potato. So we're up to 4 dimensional breakfast latent space. I hate to think what's in the dark breakfast black hole of that 4 dimensional latent space...
I feel like there's a lot of unexplored area in the carb-soaked-in-egg category that French Toast fits into. The major analogues being chiliqiles and matzoh brie. I recently did something like french toast bites where I cubed some sourdough bread, soaked it with egg and fried it up with small pieces of bacon mixed in. But what if you did that with a glazed donut? Or a waffle?
[edit] just also why this post touched my heart - I think form is as important as ingredients whenever you're dealing with relatively few ingredients. I have a breakfast I particularly love making that's just hash browns, egg and cheese. But the trick is, you griddle the hash browns, then flip them and smash them on griddled cheese, then crack an egg on top while the cheese fries and flip the whole thing again. The result is a crispy potato pancake where one side is fried cheese and the other is embedded fried egg. The same 3 ingredients, but it can be held in hand and it's got the perfect balance in each bite.
i've done the french toast pizza and it wasn't bad. not sure if it was worth the effort. maybe there's an ideal type of pizza or combo of toppings that makes this spectacular. either way it's worth trying once just to say you did.
Also re: "I spent longer arguing with the agent and recovering the file than I would have spent writing the test myself."
In my humble experience arguing with an LLM is a waste of time, and no-one should be spending time recovering files. Just do small changes one at a time, commit when you get something working, and discard your changes and try again if it doesn't.
I don't think AI is a panacea, it's just knowing when it's the right tool for the job and when it isn't.
Anyone not using version control or a IDE that will keep previous versions for a easy jump back is just being silly. If you're going to play with a kid who has a gun, wear your plates.
Once, I told a friend that it was stupid that Claude Code didn't have native IDE integration. His answer: “You don't need an IDE with Claude Code.”
I've begun to suspect response that this technology triggers a kind of religion in some people. The technology is obviously perfect, so that any problems you might have are because of you.
I find that I vastly prefer Gemini CLI to antigravity, despite the latter being an ide. Others feel the opposite. I believe it comes down to how you are using AI. It's great they both options exist for both types of people.
I don’t think it’s “just” that easy. AI can be great at generating unit tests but it can and will also frequently silently hack said tests to make them pass rather than using them as good indicators of what the program is supposed to be doing.
> AI can be great at generating unit tests but it can and will also frequently silently hack said tests to make them pass rather than using them as good indicators of what the program is supposed to be doing.
Unit testing is my number one use case for gen AI in SWE. I just find the style / concept often slightly different than I would personally do, so I end up editing the whole thing.
But, it’s great at getting me past the unpleasant “activation energy threshold” of having a test written in the first place.
Yeah it’s not accurate at all. Not the OPs fault but the purpleair sensors are placed by users. Right now it says fidi is 9° warmer than haight. Plausible, but it could also be the only 1 sensor reporting from fidi is on a balcony near a drier vent.
reply