Collection of notebooks showcasing some fun and effective ways of using Claude | Hacker News Reader

Collection of notebooks showcasing some fun and effective ways of using Claude

github.com

128 pointsadif_sgaid44 comments

Two Claude accounts of mine got blocked (one of them was a paid account for which I got a refund following the block), without explanation, very normal use.

I loved using Claude, I think it did better job than other LLM. Attempts to appeal failed with no response.

You think you can help me figure it out?

I've seen this kind of comment a few times, I was considering building some tools on top of Claude but this strongly puts me off. I don't want to invest the engineering time only to randomly have the account blocked with no warning and no explanation.

This puts me off using any remote LLM. I would pay per month to have the model run locally.

Like you shouldn't lock yourself to a specific cloud provider, you shouldn't lock yourself to a LLM as a service provider.

You could build your tools with a generic access to the best LLMs of the market. Today Claude is great, but perhaps Gemini, Mistral, ChatGPT, Command-R, or some wizard-dolphin-mixtral-carrots-merge-v3 can be descent replacements.

Totally, spending time developing the muscles to build more advanced RAG workflows is first and foremost a data modeling and data engineering challenge -- developing the hierarchical data modeling expertise, taking advantage of the structure and flexibility of the document model, chunking, and real time memory requirements combined with the power of vector embeddings and advanced query and filter capabilities is a skill that will last, irrespective of the popular model, service provider, cloud platform or framework du jour.

> is a skill that will last

As a user, why would I bother to learn these things, aside from a hobby? If it works well, I expect a product to come along soon.

You wouldn't bother as consumer. The advice above was for developers I think

Where do you start learning these workflows, as in do you have any you would recommend?

Why not use something like http://openrouter.ai? Pay as you go and you can select any model you want. Heaven!

I’ve been using Claude opus pretty heavily since it came out. In the last few days I’ve built a Mac app that can remove/replace image backgrounds in bulk with CoreML, despite not knowing much about Swift development. It works startlingly well - better than the model shipped with the system.

The notebooks are OK, but I would be much more interested in seeing an actual iteration engine that demonstrated tool use at scale, and how to set up the processes working it.

Right now it's all "here's our array of context data", "here's our tool function" and a linear flow through the process. Very little error handling. No consideration for all the failure modes involved in reaching out to actual data sources. No consideration for network errors, timeouts, retries (oh yeah, because there will be a lot of retries), very little (to nothing) regarding validating data schemas except one Pydantic example that doesn't even account for retries very well.

Every notebook is a happy path. A short one, too.

I made a website where you can save your prompts with Claude. I found myself reusing prompts a lot: templates.chat

EDIT: And you can use it from the EU

I just host LibreChat on US server.

This looks pretty good.

Anyone knows when the Claude paid version will be available in France ? using a vpn get's you around it but it's very bug prone.

I'm absolutely no expert, but I wonder why Claude is not available for ollama? Is it a license issue, a different architecture etc.?

What are people's experiences with function calling on Claude? Does it work well for anything reasonably big?

I’ve tried Claude, Bard and ChatGPT 4 and 3.5 for programming copilot.

3.5 is generally the best.

4 is ok.

Nothing else even comes close.

I'm interested in your experience finding 3.5 better than 4. Could you elaborate? Thanks

Depends on the use case. A lot of times I am just looking for a quick snippet of code, 3.5 is generally more than capable and blazing fast.

I agree openai models are by far the best, regardless of papers claiming otherwise according to some strange tests and metrics.

We are also using pagemaker hosted open source models, fine tuned for specific use cases. But even then they can't really compete with standard openai models.

Is there any open source multimodal LLM one can play with for free?

It may be a good resource, but as someone living in the EU, I am "protected" from using all the shiny new AI tools to increase my productivity. And then they wonder why the EU is so behind in innovation and startups.

You can use gpt4 which is at ~the same level freely in the EU, and you can use claude via other services.

Get a virtual card from usunlocked.com, and choose "international" at the phone verification prompt.

It makes no sense to just wait, as this situation is likely to go on for a while until the EU realizes the harm that's being done.

> I am "protected" from using all the shiny new AI tools to increase my productivity.

Can you elaborate? What’s happening in Europe that is “protecting” you from AI? America is clamoring for protections, maybe they should avoid whatever mistakes have been made in Europe.

They are open for business in Namibia though, but not in EU and Canada. It'a like when Meta stopped carrying news from Spain.

Claude's lawyers should find a way because they are lagging all other providers in this department. It's such a pity to close the model to a large number of countries. They probably expect to get sued for copyright infringement or something.

They went overboard a bit. I noticed yesterday they turned up the knob a bit further even. I gave it some lyrics (made by ChatGPT!) and asked it to rephrase certain things. It didn't want to comply because of possibility of copyright infringement. It didn't budge even with a prompt stating it's not copyrighted. They set themselves on a march to death that way and I can't blame them.

I got a "Content Filtered" error parsing a receipt on which was a product titled "Damen Sportbekleidung" (Women's sportswear), seems chatGPT is not much smarter.

    "sexual":{"filtered":true,"severity":"medium"}

sure agree - but it is a "march to compliance" .. to creative and free-thinking people, this is like Death.. but not the same. The leadership at Anthropic have made it very clear that they intend to partner with Amazon and other ultra-legal corporate aggressors.. to take and run the corporate AI services space.. and comply with every legal ruling.. (no mention in the public about Defense work?.. Mr Schmidt with Schmidt Futures has shown the way there)

Anthropic has the playing field to build an empire in the classical sense

'protected' how?

They don't increase your productivity any more than non-AI tools do. At most they make you feel as if. But you are probably just looking for a scapegoat anyways.

Just sign up and use the developer console. It works just fine in the EU.

You can't add payment methods if you're in the EU.

Yes you can, I had no problems adding balance with a credit card issued in Finland. No need for a VPN or anything. Been using the API for three weeks now.

I did.

What makes you think so? Isn't it more "AI" providers being afraid of legal consequences (e.g. due to copyright or false output), so they protect themselves by not offering certain things in the EU?

Are you able to access Claude 3 via AWS Bedrock or GCP Vertex AI? I haven't used Vertex AI, but I know that several US regions have Claude 3 access through Bedrock.

Try this, works from Spain: https://duckduckgo.com/?t=ffab&q=eu&ia=chat#

anyhing similar for chatgpt or other llms? claude is not availible in my country yet

OpenAI cookbook: https://github.com/openai/openai-cookbook

Btw, the Claude API might be available even if the chatbot is not.

Thanks for sharing!

Reply on news.ycombinator.com