PunchTornado

Plenty of things. I am using it to create a dataset of fatigued drivers for the company I work. It keeps the starting frame / angle really well and is the best I tried so far. We are building a fatigue detection model for safety.

did you read the abstract?

goog released it just before the earnings call. anthropic and openai don't have these issues.

My view is that the llms aren't sticky. I switch from one to the other daily. frequently auto which selects the cheapest option. When codex came with sol the entire team switched instantly. Then opus 5.5, the entire team switched instantly. We have a channel on slack where somebody says which is the best cheapest model today and I usually use that.

They keep talking about gpu depreciation and I pay a fortune to rent A100. We have a policy at our company to never close them because demand is so high that we can't find resources to spin them back up when we need them. So msft doesn't even spend money cooling them.

How is this open weights? Open weights means anyone can use the weights. The license forbids certain categories.

If you expect some arguments in this text, you'll find none. It is just some absolutist language, x is y and y is z.

yes, but in this case, openai does a lot of PR how their models solve important math problems. if it goes unchallenged, parts of society would think that is true. what would happen if we scale it and 10k companies dump 10k papers every month claiming solved math problems. how is this scalable?

We need the companies to humanly review their papers. in the same way as at other companies we use humans to review the papers.

If you are a EU company worried about your data then mistral is your only option. Think about EU military companies. They can't use US and Chinese models.

how hard is it with AI to do an optogenesis surgery on some mice and do some experiments with neural networks? would the AI guide you step by step, tell you where to cut, what to put where? I have no medical background.

Doesn't everyone agree that recent rogue events are entirely the fault of management wanting to do some PR for their companies?

Is there anyone who genuinely believes that current models can't be contained if we want too?

What is LeCun saying here that is debatable?

I actually disagree. Why do you need a car for city trips? Most often you can take transport, a bike or walk. And in emergency situations take a taxi.

I have a car and use it only to go OUTSIDE the city, road trips and to the mountains / sea. That's why I think a hybrid is still the best option.

This site was painful to read on mobile. Each time I switched to another app and then back it automatically scrolled to the top and I had to painfully rediscover where I left.

Can't you argue that now, with llms, the written word speaks back?

Interesting that it is ok in the US. I also work in weapon detection for driver safety systems. Obviously it is impossible for a human or a machine to detect intent with accuracy. A security guard can enter the car and then shoot the driver. You never know. That is why our systems always send alerts no matter the situation.

But claude will not be able to debug the whole stack? I have no experience in the domain. But as an outsider, it seems to me, that if the rate of progress is maintained, i don't see why agents wouldn't be able to do it in a couple of years

But regarding this specific case, where is the misdetection? Why would you enter a taxi with a ghost gun?

Good point. One could counter with the fact that US transitioned to service based and software while EU kept the heavy industry (more) while doing the change

I remember i switched from evernote to obsidian in a few hours

Do you know what a ghost gun is? Where is the misdetection?

If i were a driver i would call the cops too. What do you mean? A ghost gun in a taxi is ok?

A lot of taxis have models detecting weapons and aggressive behaviours. I work on some models like this. I think they are necessary. For driver's safety.

Imagine writing an article in 5 days. Working hard. Posting it online and then everyone says it is ai and laughs and dismisses it because a tool says it is AI.

Why trust pangram? What are their accuracy scores in the wild? Ai detection is notoriously difficult.

loved it

Such a nice story to take us out of the news cycle of ai in math

You can do both. Release it, but ask for money.

Hopefully. Tired, as a shareholder, to see the giveaways that deepmind is doing.

What is astra?

I open the link and I see Flash 3.8 high at 0.58 and Sol at 0.95. I don't understand why you say that "Sol 56 high ranks smack between Gemini 3.8 flash medium and high" but that is clearly wrong.