Ask HN: Experiences with Claude 3 and future of LLMs? | Hacker News Reader

Ask HN: Experiences with Claude 3 and future of LLMs?

7 pointslambdaba9 comments

Just curious, now that it's been a mo.. Uh, 8 days?

Mine have been glowing, I feel it's still underhyped, I've been able to achieve things preciously impossible, in part thanks to the huge context length - granted I haven't tested and don't have access to GPT-4-128k, but even in the small it often find I prefer Claude.

Unfortunately the message limit is still too small, and inference can get very slow, but if the trend continues and somehow we have a model of this level that's as fast as Groq, and with practically unlimited use, it seems pretty revolutionary. At this point I think that anyone working with code that doesn't see how this will change the craft is willfully blind, LLMs as they are today are already amazing mental prosthetics, and I couldn't see myself going without. Even if I worked with 100% novel code that somehow couldn't benefit from LLM assistance, I still remain a computer user and I use LLMs to write scripts, to figure out how-tos constantly.

But I see this as moment where programmers got their first Photoshop. Sure, you still need to understand design, but the friction of production is greatly, greatly reduced.

Can you describe what you do with it?

I’ve been using GPT4 a bit and it’s pretty good at explaining stuff that I would otherwise need to look up myself but not groundbreaking.

What should I be doing with LLMs to vastly improve my productivity?

Again I would use the Photoshop analogy. It's hard to suggest anything in particular. And I'm specifically discussing what Claude 3, the most powerful model available, can do, and it's more than StackOverflow answering.

Is it better at coding that GPT4? I'm using ChatGPT (v4), just the UI, not the API, and it works really well for me. Is Claude a big step up (hard to imagine, but willing to be convinced).

I definitely agree anyone not using LLMs for coding is missing out.

Yes, it is meaningfully better at coding. Put in a large file and ask for a refactor with some criteria.

Yes, as mentioned the context length makes a big difference because you can feed it most documentation not even bothering to edit it since I you don't try to save tokens. I'm sure that a more structured tool (like an actual IDE) could do things even more efficiently but even chat is good enough.

Don't wait to subscribe, it's worth getting some experience with it because in the near future is be the floor for all models in terms of performance.

GPT-4 is still good for a lot of things and frankly even the free Mistral Large is impressive.

> LLMs as they are today are already amazing mental prosthetics, and I couldn't see myself going without

For me, it's a stack overflow ++, but is also hit or miss.

Common experience: I want to make a simple change to this bash script. Seems simple enough to ask GPT-4. It produces some piece of code, seems to do the job. Code review -> you don't handle this corner case. Shit. GPT-4, what's wrong? answer is suspicious... A couple of "man" later, I got it. Basically, it made me lose 30 minutes. I could have lost 30mn more if had insisted to get the answer from it.

I'm not saying it's not useful, but it's not the graal that will save us from learning our craft.

Not my experience even with GPT-4, but certainly not with Claude 3. And not even with bash, it gets things right in rather niche technologies and I can supply documentation if needed, and it gets it right in one shot.

It's not available in my region and it doesn't sound good enough to justify migrating to US or UK.

You just need to VPN during the sign-up process, phone activation works with international numbers, and a virtual US prepaid card, I used usunlocked.com.

I don't think you should stay at what it looks from a distance, I saw the same thing happen with GPT-4 where people were unimpressed without even having tried it out themselves.

Reply on news.ycombinator.com