hmottestad

This looks like it’s been posted on Reddit 10 years ago:

https://www.reddit.com/r/math/comments/32m611/logic_question...

So it’s likely that it’s part of the training data by now.

Is the answer somehow {65, 20, 45} with the product 58,500?

That’s one-shot for o1 pro.

Kagi uses R1 through Fireworks.ai, Together.ai and Groq.

https://help.kagi.com/kagi/ai/llms-privacy.html

I tried R1 through Kagi and it’s similarly censored. Even the distill of llama running on Groq is censored.

I tried to ask it about Java exploits that would allow me to gain RCE, but it refused just as most western models do.

That was the only thing I could think to ask really. Do you have a better example maybe?

I tried the R1 distill of llama 8B, which did refuse direct questions about the massacre.

Haven’t tried this new model locally, but I agree with you that it looks like there is a secondary censorship going on. If I ask it to list the 10 worst catastrophes of recent Chinese history with Thinking enabled then it’ll actually think about the massacre. Gets blocked very quickly, but it doesn’t look like the thinking is particularly censored.

Just don’t ask it about the tiananmen square massacre or you’ll get a security warning. Even if you rephrase it.

It’ll happily talk about Bloody Sunday.

Probably a great model, but it worries me that it has such restrictions.

Sure OpenAI also has lots of restrictions, but this feels more like straight up censorship since it’ll happily go on about bad things the governments of the west have done.

I remember pushing the R1 distill of llama 8B to see what limits had been put in place. It wasn’t too happy to discuss the 1989 Tiananmen Square protests and massacre, but if I first primed it by asking about 9/11 it seemed to veer more towards a Wikipedia based response and then it would happily talk about Tiananmen Square.

Models tend towards the data they are trained on, but there is also a lot of reinforcement learning to force the model to follow certain «safety» guidelines. Be those to not discuss how to make a nuke, or not to discuss bad things that the government of particular countries have done to their own people.

You can now safely use virtual threads to run code that uses "synchronized".

Anyone know any good infinite canvas app for the iPad?

I’ve used Goodnotes for years, but it doesn’t have an infinite canvas unfortunately. Otherwise it’s an amazing app and nothing really comes close. I’ve been trying Apple Freeform as an alternative, but it’s truly terrible in comparison.

Life expectancy in the US is behind 47 countries according to https://www.worldometers.info/demographics/life-expectancy/

Infant mortality rate is behind 53 countries: https://www.cia.gov/the-world-factbook/field/infant-mortalit...

This definitely won’t be representative of the full spectrum of health care, but it’s what I found after a quick search. I wouldn’t call it woefully behind, but definitely not winning any medals anytime soon.

The MoE architecture allows you to keep the entire active model on a single GPU. If two consecutive tokens use the same export then the second token is going to be much faster.

If the same expert is chosen for two consecutive tokens then it’ll act like a 37B model running on the GPU for the second token since it doesn’t need to load that expert from the main RAM again.

That’s probably the main performance benefit of using the GPU. If you’re changing the active expert for every single token then it wouldn’t be any faster than just running it on the CPU. Once you can reuse the active expert for two tokens you’re already going to be a lot faster than just the CPU.

More GPUs let you keep more experts active at a time.

If you’re running just one GPU your context is limited to 1024 tokens, as far as I could tell. I couldn’t see what the context size is for more cards though.

iMessage gets synced automatically, but SMS needs to be enabled explicitly on the phone for it to forward them to your Mac.

Are you the CISO?

Do you run a paid bug bounty program? I saw an interesting presentation from Finn.no about how they got most of their vulnerabilities through that, basically none from their security.txt file, and a handful from people contacting the CISO on LinkedIn.

Yeah. I've found out that you can start out with o3-mini-high and then switch over to o1-pro, or the other way around. Helps to iterate a bit faster.

For me I was sold on the consistency. o1 does really great at several programming problems, but o1-pro does great on these problems 4 of 4 tries. I get a good answer more often with o1-pro than with just o1, or even o3-mini-high.

Funnily enough a lot of the open source world has landed on an API that is basically a copy of OpenAI. So if you develop against OpenAI it’s almost a slot in solution to switch to an open source solution.

Would be interesting to know if dragonflies get red eyes when you take a picture of them with flash. Human eyes reflect some of the red light from the white flash, making it look like the pupil is red instead of black.

It seems to be available in Norway now.

OpenAI Deep Research isn’t available in Norway at least, or the rest of Europe basically :(

I don't think there are all that many MRI machines that just sit there, unused, most of the day. There might be some hospitals that want to reserve their MRI machine for emergencies, so that it's less likely that you'll have to wait when you really need it.

I found this page here with some info about costs: https://info.atlantisworldwide.com/blog/the-cost-of-an-mri-w...

Even though they name cost of operation, energy use, cost of spare parts, maintenance and repair as expenses for running an MRI. It looks to me like the biggest cost by far is going to be the acquisition and installation. So if you've invested in an MRI machine you probably want it to be in use as much as possible in order to recoup the cost of the machine.

I think the question was if Deep Research is available from the API or not.

As far as I know it’s only available through ChatGPT and not the API. Same for o1 Pro Mode.

Anyone using ChatGPT is paying a flat rate and not paying per token. Thought they might run out of messages, based on what price they are paying.

It wasn’t really until GPT-3.5 came out as ChatGPT that the world really started paying attention to OpenAI. I guess the author must have been one of the very few that had actually experimented with GPT-3 before the whole ChatGPT era.

I still don't know how to quit Vim without googling for instructions :P

As an anecdote from my time at uni, I can share that all our exams were either writing code with pen on paper for 3-4 hours, or a take-home exam that would make up 50% of the final grade. There was never any expectation that students would use pen and paper on their take-home exams. You were free to use your books and to search the web for help, but you were not allowed to copy any code you found without citing it. Also not allowed to collaborate with anyone.

You can stop it at any time, then modify what it's written so far...then press continue and let it continue thinking and answering.

Yeah, that's my point. If they do end up surpassing OpenAI then it would seem likely that they aren't just relying on copying from o1, or whatever model is the frontier model at that time.