nearbuy

We absolutely want politicians to listen to lobbyists in a way that is proportional to some parameter besides the number of people they represent. The should listen to who is right, not which lobby represents the most members. Self-interested groups lobby for bad legislation all the time. When the harm to society is diffuse, they don't always have a large lobby opposing them.

...they did.

You're misremembering. The first known use of AGI was in 1997, but that was a single, mostly unknown use in one paper. It wasn't until at least a decade later that the term started entering mainstream use after being independently reinvented in the 2000s. AGI just wasn't a term in the 90s.

It wouldn't cost $10 for their lifetime use if they used cheap models instead. Mistral Small 3 24B is probably substantially better than their Qwen 3 0.6B trained model and would give you about 500,000–2,000,000 English to Bash translations. If they're a heavy user, they might use $0.30 in total, ever.

They also used Astra for the coding, which they can't get on a $10 subscription. And then there's the actual training cost.

Given that they're just using it to avoid the googling for bash command syntax, I'm not sure they'll save in the end against the 140k training examples they generated.

No it's not. 1-2 seconds is what we find in practice for real drivers responding to unexpected hazards.

Your close following distance example doesn't show anything. A normal braking event on the highway doesn't require a fast reaction. If you're driving 100 km/h and you're following 1.5 seconds behind the car in front of you (about half the recommended following distance) and they brake to 80 km/h, you have about 9 seconds to slow down or switch lanes. That's plenty of time.

The risky scenario is if the car in front of you has to do a full, hard emergency stop and you're following too closely. That's rare, and collisions are common when it happens.

If you know the car in front of you is going to brake because you can see the traffic ahead slowing down, that's not fast reaction time. That's just you seeing cars slow down and reacting at a normal speed. An AI has just as much time to react to that as you do.

This is well studied by researchers, and 1-2s is the typical reaction time for real drivers with real muscle memory when reacting to an unexpected hazard on the road.

There are a lot of reasons people could sometimes react faster (for example, if they anticipate the hazard, or if they're just above average in reaction speed), but one to two seconds is the reaction speed we find most of the time.

The fastest human reactions aren't to unexpected road hazards. We have a much faster reaction speed in tests where you just have to click the mouse each time the screen flashes. Our reactions are fastest when you know in advance the event is about to happen. But this isn't relevant for road safety.

The typical perception-to-reaction latency of an alert driver to a hazard is about 1-2 seconds. 250 ms when you're waiting for an event and know how to respond. For example, like a batter in baseball waiting to swing.

Are you asking for the latency or throughput?

In humans, it's about 200–250 ms for a visual cue where you already know how to respond and you're ready, but you don't know exactly when it'll happen. It can be a fair bit longer if you need to identify what you see and choose how to respond. Typical perception to reaction time estimates for drivers when there's an unexpected hazard on the road are 1-2 seconds.

With GPT-6 Luna (which is what the parent comment was talking about), that would come to 3.2¢, assuming GPT-6 used the same number of tokens.

I don't think you can guess more precisely than an order of magnitude from trying each once on one task.

This isn't right. You're comparing cost per token, but DeepSeek V4 Flash uses more tokens. Artificial Analysis found GPT 6 Luna to be significantly cheaper than DeepSeek: https://artificialanalysis.ai/models/comparisons?compare=dee...

They did not hardcode that, because:

a) There's zero evidence of them doing so

b) Some models released before the car wash problem was discovered would consistently get it right

c) Hardcoding it is pointless. No one is seriously asking that. It's just a trick question. Hardcoding one trick question won't fix its weakness at other simple trick questions.

d) Since it went viral on the internet, the next time they updated the knowledge cutoff, the LLM would likely be aware of the trick. It will fix itself without the labs doing anything special, even assuming the new models weren't smart enough to naturally figure it out.

JimDabell said they didn't come up with that thought themselves. They never implied the poster didn't mean it. sm-silversight responded, "How do you know it's not his/her thought?" You can reasonably interpret them as asking how we know they didn't come up with that thought independently.

If we go with your interpretation instead, then sm-silversight simply misread the comment they replied to and their post doesn't follow. No one doubted OP felt sincere about not wanting to read AI.

In this comment section alone, 5 people have used some variant of that aphorism, and it's used constantly on hn and other sites. They've heard it.

> Things like "You're right, but for a stronger reason than you said:"

I also get this a lot, but almost always with "and for a stronger reason than you said", which makes it sound more like it's reinforcing me than one-upping me.

Looking through my recent transcripts, I found this pattern 12 times, with every time using "and" instead of "but".

Eg:

- "Confirmed, and it's worse than a pause quirk."

- "I think the ban is right, and it's a stronger idea than what I'd proposed."

I did find one instance with "but", but it was a genuine disagreement rather than 1-upping or presenting strengthening evidence: "I would cap it, but for a different reason than fairness."

Buckmaster (the mathematician) and Alpöge used and credit AI substantially for their proof. Even if OpenAI did copy their ideas, it still wouldn't show that this didn't come from AI improving.

OpenAI's proof is substantially different and I don't think anyone has claimed otherwise. The accusation is that they used the same avenue of attack, and it's an uncommon one, and that makes it suspicious that they may have taken the idea.

You're conflating two very different meanings of "the speed of light". Confusingly, "speed of light" can refer to the universal constant, c, which does not change in glass, or to the speed that light travels in a medium. Here, we're talking about c. Relativistic effects are based only on c. The Three Body Problem trilogy plot device is also about c.

Why?

1. I didn't say LLMs have made any breakthroughs in math, not because they haven't, but because it's irrelevant to my point. The parent comment is using the same argument academic research opponents have long used against research. The vast majority of research fails to meet their bar. How would your daily life be different if we had no humanities papers published since 2023? Or math?

2. You can google this in 10 seconds and see a dozen results in math. This is not a good-faith demand.

There are several hundred thousand mathematicians producing hundreds of thousands of new results in math each year. Why can't most people name any human contributions to mathematics from the past decade? What is the finished result that affects everyday life? Do you consider all those human mathematicians to be useless?

Most of the biggest breakthroughs in mathematics, breakthroughs that win Fields Medals like sphere packing in dimensions 8 and 24, have no applications in everyday life. Probably the only new mathematics results that people notice affecting their daily lives are the ones that enabled AI.

Nevermind mathematicians. What about the millions of programmers? Are they all hype too because people pre-2023 were griping on hn that software is buggier than ever and people are still using the same operating systems as always? Why couldn't the 30 million human programmers make something better in the past decade?

You set your bar so high that all the world's human experts in math and programming combined would fail to meet it.

The top LLMs in 2024 were Sonnet 3.5 and GPT 4o. You couldn't have expected those much weaker models to be making breakthroughs in math. The models that are making breakthroughs haven't been around very long.

The Kevin Buzzard post linked at the top says they budgeted £1M over 5 years for a smaller proof.

I think you're misunderstanding. Astra is at the top of the official ARC-AGI leaderboard, with an ARC-AGI approved harness. It's not a harness specialized for ARC-AGI. It just does the same thing the regular ChatGPT interface does: keeps conversation history across turns and compacts when it gets too long. Without the harness, it loses its entire context window every move. That's not how humans work and it's not how any real AI service works.

People assume that's the reason because it's intuitive and "strawberry" is one token. But that doesn't explain why those models would also often get it wrong for "StRaWbErRy" or even "s-t-r-a-w-b-e-r-r-y", where the r's are not combined into one token.

We don't know what was going on inside the closed source GPT models, but this paper investigated on some of the open-weight models and found it's not due to tokenization: https://arxiv.org/abs/2604.00778

I don't think we should count the lower tier models if we're discussing what the top ones are capable of. No one was suggesting that Sonnet is AGI.

The last version to fail on those questions was GPT 4.5.

Meanwhile most humans fail to correctly answer how many f's are in the sentence, "Finished files are the result of years of scientific study combined with the experience of many years.".

It also looks like they're saturating the test, with one LLM hitting the maximum possible score. (https://www.trackingai.org/home)

The test wasn't made to accurately measure IQs that high.

The paper also fails to show that their central example, Einstein, relied on sensory experience for his intuition leaps rather than general reasoning. They just kind of claim that thought experiments require sensory experience. But you can easily ask an LLM to perform a thought experiment and simulate an outcome, and the SOTA LLMs generally seem to do about as well as a human. An LLM would certainly know that freefall feels the same as zero gravity, even if they haven't felt the sensation, which was the key intuition the paper talks about for Einstein's General Relativity. The paper's author would probably say any examples of this don't count, but without clear criteria for what would count, their claim is unfalsifiable.

Only a tiny, tiny fraction of the parameters are encoding information that's specific to a particular programming language. Even if you could remove those without degrading performance, it would have a negligible effect on the model size.

The problem is Claude Fable is now better than most programmers I know at software architecture and performance optimization as well.

That's a good reason to use a dishwasher, and you should keep doing it. But the overall waste is small and people aren't going to care.

I'm not saying people who already have a dishwasher will throw it away. But a robot makes installing a new dishwasher much less tempting.

At my utility prices, it costs about 5¢ to hand-wash one serving (plate, pan, glass, utensils, and a bowl) with running water. That's not nothing, but it's dwarfed by the energy use of using the oven and the supply-chain energy and water use of the food if it includes meat.

People will casually waste 10x that energy without a second thought when they don't need to just because it's convenient, or because they're used to doing things that way.

The dishwasher looks more unfavorable if you're considering installing one in a small apartment or condo. Apart from the cost of the dishwasher, you need space to install it, and more cupboard space to keep more dishes, glasses, etc. For an extra 4¢ per person-meal, having your robot do the dishes buys you the convenience of skipping all that and having all your dishes clean and ready all the time.

Sure, but then there's no such thing as a network that isn't a classifier. Every physically computable function that terminates in finite time will map an input to a fixed set of outputs. And it goes against the common usage, where in machine learning we talk about classifiers, regressors, generative models, etc. as different things. They all become classifiers.

The parent commenter was trying to draw some insight from LLMs being classifiers that wouldn't apply equally to everything else.