Haiku is the biggest release since GPT4 from a cost/performance perspective. Glad you got to dig into it, always enjoy your videos.
@samwitteveenai4 ай бұрын
Thanks! I agree it really seems to open up a lot of opportunities.
@joflo59504 ай бұрын
Thanks for the video! I would really love to see the follow-up video on function calling you mentioned.
@pokerandphilosophy83284 ай бұрын
My main use of Claude 3 currently is to have extended philosophical discussions with it, discuss texts and have it help me rewrite my own papers and drafts. I often begin the conversation with Opus to maximise quality. But when the context gets longer, I sometimes switch to Sonnet or Haiku. Haiku very often surprises me with how smart it is. When its responses are informed by the longer context, including the prior responses from Opus, this serves as something similar to a many-shot prompting method with explicit examples and it boosts Haiku's intelligence. Furthermore, Haiku's slightly more unfocused or meandering intellect enables it to make relevant connections between various parts of the conversation that Opus often misses due to its more focussed attention to user instructions and strict adherence to prompt. As a result of that, Haiku's responses sometimes are more intelligent, insightful and (broad) context sensitive even if it is slightly more prone to error than its bigger siblings.
@nas83184 ай бұрын
It's meandering intellect may be due to a higher temperature setting. You may wanna look into that.
@pokerandphilosophy83284 ай бұрын
@@nas8318 I'm interfacing all three Claude 3 models through the Anthropic workbench with the temperature set to zero. So, it's really something else that is at play.
@paulmiller5914 ай бұрын
Thanks, Sam. I had played with Haiku previously but have not done this optimised prompting. Jumping onto doing this now. Cheers.
@davidtindell9503 ай бұрын
Great overview and multimodal examples from the Anthropic Claude Cookbook using Haiku! I ran some of the examples multiple times with variations and the cost so far is less than a U.S. $1.00. We should definitely consider Haiku for personal and business apps where the tradeoff between quality and cost must be balanced: e.g. summarizing a large volume of papers and documents and creating and maintaining a large database of vector embeddings to support documents Q&A.
@AdamTwardoch4 ай бұрын
It seems to be that Haiku is a distilled / sparsified / quantized Opus. When it works, it gives results that are quite similar to Opus - while Sonnet gives very different results, so it looks it was trained independently. This is great: I often prep few-shot with Opus and then hand or over to Haiku for scale.
@amandamate91174 ай бұрын
cant wait for the crewAI + Haiku video !!! it would be nice to have a superagent thats using OPUS and small agents only haiku.
@ehza4 ай бұрын
Thank you for this. Quite helpful to me!
@samwitteveenai4 ай бұрын
Glad it was helpful!
@walterpark88244 ай бұрын
What a great model for local use. Thanks for showing it so clearly.
@JD-hk7iw4 ай бұрын
I had written off Haiku after testing my use case with it using the same prompt I use for opus/gpt-4. Totally unusable. After watching this, I revised the wording & format of the system prompt and added three examples. Well I'll be damned. Touché Haiku, touché. Not as nuanced and focused as opus/gpt-4, but definitely serviceable. The combination of the 200K context window and the pricing really is what makes this model special. Thanks for the informative video showing the proper way to leverage Haiku.
@samwitteveenai4 ай бұрын
This is awesome to hear! I have found since the input tokens are so cheap I have been using 20 examples for some things and getting really good results in changing style and tone too.
@xemy10104 ай бұрын
Haiku might be the perfect model to label / caption an image dataset at scale using natural language. Dalle-3's paper makes it clear that generating detailed natural language captions for each image was a big part of the magic behind its ability to understand and follow prompts so well at inference. SD3 only used a 50:50 mix of CogVLM-generated captions and captions from the original images. I think a Haiku-captioned training dataset would be a big step up for training these models.
@kenchang34564 ай бұрын
Cheaper works for me as I'm in the learning/experimenting stage. Looking forward to you Claude 3 based function calling video. Thanks for sharing.
@samwitteveenai4 ай бұрын
Thanks Ken
@brandonwinston4 ай бұрын
One of the big challenges I’m having is plugging haiku into all the places OpenAi APIs are accepted
@samwitteveenai4 ай бұрын
checkout LiteLLM you can use that as proxy that takes OpenAI inputs and can reroute them
@alchemication4 ай бұрын
So far wasn’t able to get anywhere with haiku for any production quality use case, but the idea with using many examples sounds promising. Will test out. Thx for inspiration to try again 😊
@jayhu60754 ай бұрын
What a amazing explanation how to do with vision, xml and other stuffs with Haiku. Hopefully more in the future about Agents what you all mention with Crew AI. Many thnx.
@aa-xn5hc4 ай бұрын
Looking forward to your next video on CrewAI and Haiku
@EyadAiman4 ай бұрын
Impressive tutorial as always SAM I suggest you make a tutorial on how to build RAG system with agents using claude 3 haiku
@silvacarl4 ай бұрын
I look forward to every one of these videos. Can you do more langchain or RAG examples with open source LLMs?
@vivekpadman52484 ай бұрын
Thanks for this
@EmadElazhary-tt8tl4 ай бұрын
Thanks for the video! Please Please, Function Calling using Haiku in Langchain
@hendoitechnologiesАй бұрын
Can you post full course video about Claude 3.5 sonnet model
@UTubeGuyJK4 ай бұрын
I hadn’t heard of the xml tag prompting with Claude before.
@carterjames1994 ай бұрын
Please go over function calling asap really looking forward to it, from my test haiku is amazing with a few examples but still has some issues when I go upwards of 4 functions that can be called
@lancemarchetti86734 ай бұрын
DBRX just launched their new model on huggingface
@micbab-vg2mu4 ай бұрын
great video - thank you
@CookerSingh4 ай бұрын
I think there is no function calling feature and for now only be used in a wrapper based applications.
@carterjames1994 ай бұрын
There is function calling
@carterjames1994 ай бұрын
Its just not as mature as OpenAI function calling
@CookerSingh4 ай бұрын
@@carterjames199Is there any way I can add function calling or use available proxies out there for all LLMs.
@samwitteveenai4 ай бұрын
The function calling is in a different format than OpenAI they use xml which can be nested. for proxies etc checkout LiteLLM but I don't think that will convert function calls yet
@ShoomonPerry4 ай бұрын
I want to use claude haiku to process large documents...but end up running out of output tokens. Is there a simple hack (in langchain?...) to get a multipart response?
@samwitteveenai4 ай бұрын
The problem is Claude will only output 4k tokens (pretty sure from my memory). This is where getting a system to do multiple calls can be really useful. In langchain you can do it with MapReduce but it can be a bit hit or miss. Another way is to write your own splitting and prompting and run it in parallel. Let me try to think of a good use case I can show and I will try to make a video about it. It is certainly an issue.
@ShoomonPerry4 ай бұрын
Thanks Sam. What if you chained the queries, passing in the response of the first query into the context window of the second....and telling the LLM to pick up where it left off?... @@samwitteveenai
@jdallain4 ай бұрын
I’m not sure if it’s possible, but please use haiku with the more advanced crewai video that you mentioned making
@jdallain4 ай бұрын
Commented before watching until the end 😊
@sauravmohanty39464 ай бұрын
can you share the link to the notebook you explained in the video ?
@samwitteveenai4 ай бұрын
It’s the colab in the Video description
@RaitisPetrovs-nb9kz4 ай бұрын
Out of curiosity tested all 3 models got dogs correct with 1st shot without prompting
@samwitteveenai4 ай бұрын
😀 Thats interesting. Maybe their examples was planted to make it look wrong for the first one? For me it came back with 8 when I tried their fancy prompt but very quickly changed back to 9 with a few small changes.
@RaitisPetrovs-nb9kz4 ай бұрын
@@samwitteveenai
@clray1234 ай бұрын
The "fancy prompting" messing around with the count of dogs output is actually a glaring example of why all these models are crap. Lack of trustworthiness, making big mistakes on trivial tasks AND what's more, those mistakes depend on minute details of how you arrange the input! It reminds me of a fine tuned model claiming at the same time that it loves and hates tomatoes! Or claiming that its favorite animal is a tomato, unless, of course, you ask beforehand whether a tomato is an animal. This is simply ridiculous and inconsistencies of this sort highlight the simple fact that today even the most sophisticated of these models are still imitators of intelligence, rather than intelligence. Building castles on sand.
@andrada25m464 ай бұрын
Personally I haven’t had issues with haiku, it’s much better than GPT3.5 you just have to prompt it well.
@Quin.Bioinformatics4 ай бұрын
google bard is trash why was it rated so highly? it sucks at generative coding and variant annotation.