hmottestad

At the pace that DeepSeek is developing we should expect them to surpass OpenAI in not that long.

The big question really is, are we doing it wrong, could we have created o1 for a fraction of the price. Will o4 cost less to train than o1 did?

The second question is naturally. If we create a smarter LLM, can we use it to create another LLM that is even smarter?

It would have been fantastic if DeepSeek could have come out with an o3 competitor before o3 even became publicly available. That way we would have known for sure that we’re doing it wrong. Cause then either we could have used o1 to train a better AI or we could have just trained in a smarter and cheaper way.

I don't use it much for programming, since I mostly use IntelliJ for that.

Where Sublime shines for me is opening large files, and opening pretty much any file. So many times MacOS wants me to open a json file in Xcode, or a txt file in TextEdit, when all i want to do is open it in Sublime! And I know that i can open files that are multiple GB without issues, just takes a few seconds :P

And the Ctrl-D shortcut and multi-cursor in general is so neat. I know other editors have it too now, but when I showed my wife how useful it was for simple tasks like formatting a list of emails...she ended up making her company buy her a license and teaching others how to use it.

Mac Studio has up to 196 GB of memory.

I wonder if Lockdown Mode would help?

Not sure what you mean, but are you saying you don't ask it anything related to work at all?

Memory requirement for the 7B model with full context is 120GB, so you would need 5 3090 GPUs, not 2-4. Do you know if you can get a motherboard with space for 5 GPUs and a power supply to match?

I bet that 5 3090s will smoke a Mac Studio. Can't find anyone in Norway with any in stock though. Or any 4090s with 24GB of memory.

You can get a nVidia RTX 5000 with 32GB of memory, there are two webshops that have those in stock. You'll need to wait though, because it looks like there might be one or maybe two in stock in total. And they are 63 000 NOK, and you need 4 of them. At that price you can buy two Mac Studios though.

I see people selling 3090s with 24GB secondhand for around 10 000 NOK each, but those have been running day in and day our for 3 years and don't come with a warranty.

Not sure what you mean. Cutting edge computing has never been cheap. And a Mac Studio is definitely within the budget of a software developer in Norway. Not going to feel like a cheap investment, but definitely something that would be doable. Unlike a cluster of H100 GPUs, which would cost as much as a small apartment in Oslo.

And you can easily get a dev job in Norway without having to run an LLM locally on your computer.

More than other models? I thought that context used a lot of memory on all models.

And I’d hardly call it obscene. You can buy a Mac Studio with 192GB of memory, that should allow you to max out the context window of the 7B model. Probably not going to be very fast though.

O1 doesn’t seem to need any particularly specific prompts. It seems to work just fine on just about anything I give it. It’s still not fantastic, but often times it comes up with things I either would have had to spend a lot of time to get right or just plainly things I didn’t know about myself.

If you run one of the distill versions in something like LM Studio it’s very easy to edit. But the replies from those models isn’t half as good as the full R1, but still remarkably better then anything I’ve run locally before.

I ran the llama distill on my laptop and I edited both the thoughts and the reply. I used the fairly common approach of giving it a task, repeating the task 3 times with different input and adjusting the thoughts and reply for each repetition. So then I had a starting point with dialog going back and forth where the LLM had completed the task correctly 3 times. When I gave it a fourth task it did much better than if I had not primed it with three examples first.

I tried two questions that I had recently asked o1 pro mode.

The first was about setting up a GitHub action to build a Hugo website. I provided it with the config code, and asked it about setting the directory to build from. It messed this up big time and decided that I should actually be checking out the git repo to that directory instead. I can see in the thinking section that it’s actually thought of the right solution, but just couldn’t execute on those thoughts. O1 pro mode got this on the first try.

Also tried a Java question about using SIMD to compare two CharSequence objects. This was a bit hit or miss. O1 didn’t do great either. R1 actually saw that it’s possible to convert a char array to a short vector, which was better than o1, but they both failed to understand that I don’t have a char array.

Also tried a maven build problem I had the other day. O1 managed to figure that one out, and R1 also managed on the first go but was better at explaining what I should do to fix the issue.

The comment says: Every time you click on a link your external IP addresses is exposed, is this a vulnerability? Being online without a VPN / proxy is inherent consent to have your external IP & other required items to be shared with services / middlemen.

The fact that a user's IP is exposed when they click on a link is only relevant to the original post if a user would do this automatically and without realizing. The original post alleges that they can send someone a message on Signal and have the user automatically and somewhat unknowingly load a resource from a server. Sure, the author doesn't claim they have much control over the resource or the server, but they do show how you can check which server the user accessed and how that leaks information about the location of the user to a certain extent.

If I know someone on Signal I can now check if they’ve left the country.

Or send this to a bunch of signal users whom you suspect one of them being a particular person, and if you know that the person you are looking for is going to travel you can send it once before and once after. Then see which of these users were in the home city and subsequently in the destination city.

If I can send you a link and be guaranteed that you click on it. Then that’s definitely a security issue.

That gut feeling approach is very human like. You have a bias and even when the facts say that you are wrong you think that there must be a mistake, because your original bias is so strong.

Maybe we need a dozen LLMs with different biases. Let them try to convince the main reasoning LLM that it’s wrong in various ways.

Or just have an LLM that is trained on some kind of critical thinking dataset where instead of focusing on facts it focuses on identifying assumptions.

Well, it’s not like anyone sued a Dutch university for finding security vulnerabilities in their NFC transport card solution…NXP would never rush their implementation to market and then try to sue their way out of any vulnerabilities.

https://freedom-to-tinker.com/2008/07/15/transit-card-maker-...

They specifically write that this only supports UID based authentication. So, the card answers with the same unique ID every time.

UniFi has support for this, but seemingly not by default.

This solution also doesn’t allow you to clone an existing card. You actually need the admins to add the UID of your Chinese transport card to their system.

We’ve had at least one cyclist killed since that article came out.

Regular voting is usually what affects things such as the transportation infrastructure in your country or city. It’s a slow proceed though.

Here in Oslo there has been a lot of investment in bike lanes, but just because one part of the local government builds more bike lanes doesn’t mean that other parts of the government will follow suit. Police still doesn’t care about cars illegally blocking the bike lanes. The people ploughing snow see bike lanes as the last thing that should need ploughing, preferably no earlier than 2 weeks after it snowed. A dedicated bike path I use to work is supposed to be ploughed within 2 hours of snow, but it took a week before anything was done and now three weeks later it’s still not to the standard that the government has set.

It's not properly grounded. I think I need to change out the power cable for one with ground, since the power cable that comes with the stereo only has two wires. There is a grounding point on the back of the chassis, but I don't have anything to ground it to except for the wall socket and at that point I should just use a proper power cable instead.

I have a 40-year old amp and speakers at home. My dad’s old stereo. Works fine, but it has so much static buzz.

Sounds great though. I’m just a bit hesitant to leave it powered on all the time which means I don’t use it much compared to my Bluetooth speakers. Old power supplies are not very efficient. They can get hot. There is also a decent amount of dust. And then there is the risk of some old cap blowing.

I assumed it was part of the migration plan to Apple silicon. Rosetta 2 makes x86 apps work on Apple silicon, but I had assumed that Apple could only really get 64-bit x86 apps working smoothly and that’s why they removed 32-bit support a few years earlier.

I tried to search for info about an update that broke iPhone 7 phones. Couldn’t find anything. Can you point me to some more info?

Norway works on a precedent system.

When electric scooters were reclassified as «small electric vehicles» they suddenly came under the same drunk drive laws as motorbikes and cars. So the lower courts ruled a bunch of drunk driving of electric scooters as severely as they would drunk driving of a 2000 kg car that can go 200 km/h. Essentially they just followed the precedence of previous rulings on drunk driving without taking into consideration the intent of the law. People got huge fines and lost their car licenses for several months on the assumption that if they were careless enough to drink and drive an electric scooter they would be just as likely to drink and drive their car.

Eventually a case went all the way to the Supreme Court where they actually thought it through and and decided that there wasn’t any reason to assume that a person would drink and drive a car just because they did so with an electric scooter.

The post mentions that the bots were crawling all the wiki diffs. I think that might be useful to see how text evolves and changes over time. Possibly how it improves over time, and what those improvements are.

I guess they are hoping that there will be small changes to your website that it can learn from.

That's really cool! Thanks for sharing :)

I assume that the model couldn’t run locally for some reason. Probably either uses too much power or needs too much memory.

I agree that this isn’t necessary and that Apple should have asked for consent from the user.

I do want to point out that Apple doesn’t get to see your photo. Homomorphic encryption is really cool in that way. Apple isn’t able to decrypt the photo and the results they produce are also encrypted. That’s the beauty of homomorphic encryption. I can send you an encrypted spreadsheet and you can compute the sum of a column for me. I’m the only one who can see the actual sum since you can only see the encrypted sum.

Your ISP, a bunch of routers and switches and the servers run by Signal can also see your encrypted photo. You don’t really get to decide who sees the encrypted photo. You do get to decide which photo you encrypt and send though.