Replying to @EMostaque
not like draw io or excalidraw but the Camunda modeler is used for such diagrams, they use bpmn files which are just xml files. LLMs can read and change them pretty well, so I’d probably ask "turn my doodle into a camunda xml file"
4
265
Replying to @svpino @3blue1brown
Also has great stuff on linear algebra, calculus, beautifully visualized.
Interesting
Next to @pydantic Pydantic AI this is one of the more interesting abstractions.
Need to check out what workflow makes most sense.
For months, we've worked on building @huggingface's new moonshot: agentic systems.
So today we're very proud to announce the release of 𝚜𝚖𝚘𝚕𝚊𝚐𝚎𝚗𝚝𝚜!
It's the simplest library we could make to let people build powerful agents.
💥 The main logic for agents fits in ~1000 lines of code. So it's really dead simple.
🧑💻 The main agent class is CodeAgent, and agent that writes its actions in code. That means, contrary to the standard set by OpenAI of writing tool calls as JSON blobs, this agent writes code snippets. It's much more natural for LLMs to write actions this way, and as a result performance is vastly improved.
🌍 It supports any LLM through @LiteLLM integration.
🛡️ We enabled secure code execution via @e2b_dev sandboxes.
1
4
714
Replying to @hardmaru @RichardSSutton
Can recommend the video lectures on reinforcement learning by David Silver from Deepmind that reference that book
youtube.com/watch?v=2pWv7GOv…
2
337
Replying to @cohere
Very interesting - in terms of pricing it also would need to be compared with Claude3 - Haiku which sits in a pricing sweet spot, see also here teapaw.org/swyx/status/1772799201…
Replying to @jxnlco
This is beautiful and kind.
One can forgive and love deeply, especially the imperfections.
Replying to @ClementDelangue @jeremyphoward
I like the idea.
But there needs to be a SD4, SD5 too and an open Sora etc, will the research continue?
8
534
Replying to @HamelHusain @dan_s_becker
What would be very interesting is not only the QLoRA/LoRA finetuning but deploying those lightweight adapters into the cloud on top of the base models
instead of quantizing the full heavy model.
I know TensorT, vLLM, TGI etc are a thing but deployment still requires heavy resources.
and of course this here: teapaw.org/HamelHusain/stat…
You need training and serving frameworks to adopt the template consistently. However, the HF chat template is internally inconsistent with the documentation in places.
Ex: the template will apply a space between <s> and [INST], whereas the documentation doesn't have this.
1
9
1,153
Open-source repo github.com/stanfordnlp/dspy
Paper arxiv.org/abs/2310.03714
Webinar on Youtube youtube.com/watch?v=POBcYr0s…
Check out @lateinteraction for more
Is DSPy the better Langchain?
Prompt engineering is often brittle and inefficient. DSPy automates the process for you. It is much closer to classical machine learning in that it optimizes prompts based on data.
After dedicating significant time to prompt engineering, I recently explored DSPy. Here are some observations:
• Shift from Prompting to Programming: DSPy moves away from extensive prompt engineering towards a more programming-focused approach.
• Automated Compiler: It features an automatic compiler which simplifies the process, potentially reducing the complexity of tasks.
• Use of Optimizers: The framework includes optimizers for enhancing prompts and model weights, which could streamline adjustments.
• Systematic Approach: DSPy aims to provide a more systematic method for handling changes in projects, which might be helpful in maintaining efficiency.
•Community Involvement: It's designed to be a collaborative tool, encouraging community contributions and ongoing development.
2
3
657














