minzi

Is it really that much better than cursor’s agent? I’m hesitant to try because it would be out of pocket and I get cursor for free (work). It’s hard to understand how it could be that different if both are using sonnet under the hood.

Its agent is a lot worse than Cursor's in my experience so far. Even tab edits feel worse.

My understanding is that these are not custom models but a combination of prompting and steering. That makes Cursor's performance relative to others pretty surprising to me. Are they just making more requests? I wonder what the secret sauce is.

I use Gemini inside cursor, but the web app is basically unusable to me. Of the big three, only Claude seems to have a sensible web app with good markdown formatting, converting big pastes into attachments, and breaking out code into side panels. These seem like relatively obvious features so it’s confusing to me that Google is so behind on the UI here.

Can you expand a bit on when it would make sense to use one versus the other?

Still no debugger. I know there is a branch open, but it’s surprising to me that there isn’t a more concentrated effort on getting that over the line. Major props to the folks working on it. I just wish they had more resources and help getting it done.

Good to know. I guess I'll just hold out hope that things are ok and avoid heavy work loads until I can measure things properly.

Are you using the splitter that nvidia provided or a 600w cable? Also, what PSU?

I've been using mine remotely, so trying to figure out how much I should panic. I'm running off the SF1000 and the cable it came with. Will be a few weeks before I can measure temperatures.

Yeah this is more what I was getting at. I took measures early on to reduce my own phone time, but it seems like most people I know can't or won't.

Is there any hope of turning this trend around or at least keeping it where it is? I don’t think that the months spent on my phone have benefited me or anyone I know.

Out of curiosity, how did you decide on the size of the device? Was it driven by cost considerations, preferences of the founding team, something else?

I think it’s a little too big for me, but I’m tempted.

Incredible! Obviously tons of credit to you individually, but also huge kudos to your teacher for encouraging your curiosity.

Embarrassingly, I’ve been writing Go for a while but never really thought about it. Now that it’s been mentioned I’m curious why this isn’t baked in by default for errors. Does anyone know?

I am always so conflicted about adopting new desktop environments. Every time I feel like I dump 10-20 hours into it and still end up using my mac more often. This looks so tempting, but I feel fairly confident the outcome will be the same for me. Maybe the only way to achieve this is to get rid of my macbook.

That stuff is definitely interesting! I’m thinking more about getting involved in things like PyTorch DDP and the internal equivalents I’m sure many companies massive scale companies are developing. Perhaps ML infrastructure is not exactly the right term for that…

Thank you for your insight! I was planning on checking out triton too once I got a baseline understanding of gpu programming in general. The problem you described is right in the realm of problems that I am interested in working on. I’m really looking to work on finding efficiency in distributed training and inference.

How are you using syncthing on mobile? I didn't know that was even possible.

I would be surprised if the original board’s reasons for caving in were not influenced by personal factors. They must’ve been receiving all kinds of threats from those involved and from random twitter extremists.

It is troubling because it shows that this “external” governance meant to make decisions for the good of humanity is unable to enforce decisions. The internal employees were obviously swayed by financial gain as well. I don’t think that I would behave differently were I in their shoes honestly. However, this does definitively mean that they are a product and profit driven group.

I think that Sam Altman is dishonest and a depressing example of what modern Americans idealize. He has all these ideals he preaches but will happily turn on if it upsets his ego. On top of that he is held up as some star innovator when in reality he built nothing himself. He just identified one potential technological advancement and threw money at it with all his billionaire friends.

Gone are the days of building things in a garage with a mission. Founders are no longer visionary engineers and designers. The path now is clear. Convince some rich folks you’re worthy of being rich too. When they adopt you into wealth you can start throwing shit at the wall until something sticks. Eventually something will and you can claim visionary status. Now your presence in the billionaire club is beyond reproach because you’re a “founder”.

> The future, it seems, will be decided behind closed doors.

I keep seeing so many comments about how the move was irresponsible from the board because it was a hyper growth company and blah blah blah. This board was not set up to care at all about shareholder value.

I’m not making any judgement about whether they made the correct decision. I’m just stating that everyone keeps talking about this as if it were a normal company structure and it absolutely is not.

Yeah that is a much more succinct way to put it!

It seems like lots of people are pointing out the inadequacies of the current state of the art models. It doesn't seem that unlikely that many of these problems will be solved in the coming years. I won't make any guesses about timelines, but there is certainly a large number of very well funded smart people working on these problems. I'm not sure if this current path leads to AGI, which is what I would consider the minimum requirement for truly replacing human jobs. However, it does suggest to me that building AGI is achievable.

In my mind there are two major architectural problems that still need to be solved if I am to be convinced AGI is close:

1. Medium term memory

2. The need to relive an entire conversation just to produce the latest response.

Here is another question: Suppose we develop AGI. What will the energy requirements be for that system? The scale at which ChatGPT is used today would be peanuts compared to a system that is supposed to replace millions of knowledge workers.

I think you are correct about the history of Razer products, but this latest line up has been very solid. I had issues with the first laptop I got. It turned out to be a software issue and I got support within a day and they patched the problem within a week. I should also add that they have battery limiting features now and a two year battery warranty. Those two improvements protect you from 90% of the bad experiences people report with these laptops.

Asus I have had consistently bad experiences with and that is why I tried Razer. Of course this is all just my personal experience and I doubt either is perfect.

As far as needing cloud connected software to configure Razer hardware: that is simply false. The laptop is configured via an embedded usb interface that you can send reports to in Linux. There are open source projects that make this particularly easy to do.

In my opinion, if you have the money, a Razer or Asus laptop is the highest quality machine that you can get to run Linux. Of course there is also Framework, but those are not as well built in my opinion. Personally, I have the new Blade 14 and it is great. Large trackpad, 16:10, 6 hours of battery life (pretty good for linux) and an RTX 4070. That said, it is absurdly expensive.

Sorry if what I wrote was unclear, but I mean the back button in the spotify player. The buttons that are used to "rewind", "go to previous song", "skip song", etc.

Bit of a tangential question here, but any recommendations on how to get started fine tuning this model (or ones like it)? I feel like there are a million different tutorial and ways of doing it when I google.

I don’t know much about security, especially at the hardware level. However, I have a question for those of you that do.

Suppose you were given a healthy budget, a team, and a few years. Would you be able to build network hardware that did not contain back doors? How healthy would the budget need to be? How skilled would the team need to be? I assume you’d have to assume most external vendors are compromised and rebuild whatever you needed from them. What would that take?

Can someone explain what they use structured output for? I’m just curious what kinds of use cases people have found for it.