Fabian Schreiber

@omarsar0 @yrzhe_top @karpathy I include queries that I want answered in the prompt for my llm when I task it to write a summary. As a guidance. Makes sense?

That’s important. Like when developing a skill the best approach I have found to optimize the thing is to test it on actual tasks. And improve it progressively. Indexing everything all at once might be a trap like you said. Like with everything I feel like the test and verification are important for it to be of any use.

A few notes on how to get started with building LLM Knowledge Bases. @karpathy popularized it but most people don't know where to start. Everyone should be creating LLM Wikis. Live session tomorrow. Shared a repo example and a Skill coming soon. academy.dair.ai/blog/how-to-bu…

@dwddao haha, epic and so true.

I created a team of coding agents that were productively building and coordinating with each other. However, once I added a "manager" agent, it started pinging the engineers every 10 seconds for product feedback, interrupting them in their flow 😂 I don't know what is more real

@Rosamygale @IBE_Barcelona she is the bestest! :-)

Second day of our workshop @IBE_Barcelona learning about genome annotation with BRAKER, TSEBRA, GALBA, Tiberius & more with the one and only Katharina Hoff! 🧬🧩🪱

This came unexpected! @OpenAI released Swarm, a lightweight library for building multi-agent systems. Swarm provides a stateless abstraction to manage interactions and handoffs between multiple agents and does not use the Assistants API. 🤔 How it works: 1️⃣ Define Agents, each with its own instructions, role (e.g., "Sales Agent"), and available functions (will be converted to JSON structures). 2️⃣ Define logic for transferring control to another agent based on conversation flow or specific criteria within agent functions. This handoff is achieved by simply returning the next agent to call within the function. 3️⃣ Context Variables provide initial context and update them throughout the conversation to maintain state and share information between agents. 4️⃣ Client run() initiate and manage the multi-agent conversation. It needs an initial agent, user messages, and context and returns a response containing updated messages, context variables, and the last active agent. Insights: 🔄 Swarm manages a loop of agent interactions, function calls, and potential handoffs. 🧩 Agents encapsulate instructions, available functions (tools), and handoff logic. 🔌 The framework is stateless between calls, offering transparency and fine-grained control. 🛠️ Swarm supports direct Python function calling within agents. 📊 Context variables enable state management across agent interactions. 🔄 Agent handoffs allow for dynamic switching between specialized agents. 📡 Streaming responses are supported for real-time interaction. 🧪 The framework is experimental. Maybe to collect feedback? 🔧 Flexible and works with any OpenAI client, e.g., Hugging Face TGI or vLLM-hosted models.

@cartalop You are very right, Carlos!

@HamelHusain timely AF! I have a customer meeting next week and want to think about evals right from the start

I've condensed my advice about how to think about LLM evals into this 15 min talk. I follow the same process on every project to improve AI products. We show how we use evals to create a data flywheel to move from demo to production-ready products.

Open Source AI/ML is on fire today! 🔥 Multilingual (29) Qwen 2.5 just dropped w/ 128K context too! The 72B rivals Llama 3.1 405B and beats Mistral Large 2 (123B) ⚡ > Trained on an extensive dataset containing up to 18 trillion tokens > It surpasses its predecessor, Qwen2, with significantly higher scores on MMLU (85+), HumanEval (85+), and MATH (80+) benchmarks > Excels in instruction following, generating lengthy texts (over 8K tokens), and understanding structured data like tables. It also shows significant progress in generating structured outputs, particularly JSON. > Supports over 29 languages, including major global languages, and can handle up to 128K tokens, with a text generation capacity of 8K tokens. They release specialised models as well: 1. Qwen2.5: 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B 2. Qwen2.5-Coder: 1.5B, 7B, and 32B on the way 3. Qwen2.5-Math: 1.5B, 7B, and 72B. Kudos to @Alibaba_Qwen team for shipping high quality model checkpoints! 🐐

Inspired by the new o1 model, I hacked together g1, powered by Llama-3.1 on @GroqInc. It uses reasoning chains to solve problems. It solves the Strawberry problem ~70% of the time, with no fine tuning or few shot techniques. A thread 🧵 (with GitHub repo!)

@kakape @pandemiapodcast Cool, dass ihr da ne aktuelle Folge zu macht. Höre ich mir gern an!

Wir haben gestern Abend noch eine kurze Folge @pandemiapodcast aufgenommen, um über die aktuellen Entwicklungen bei #mpox zu sprechen. Die kommt dann Anfang der nächsten Woche. Aber zur Vorbereitung hab ich mir noch einmal unsere mpox-Folge von Ende Juni angehört und ganz ehrlich da ist eigentlich alles drin. Also hört euch das ruhig (nochmal) an: podcasts.apple.com/de/podcast/pan…

@ECMOKaragianni1 @LeonDraisaitl what a comeback so far!

Cologne is keeping its fingers crossed for Leon Draisaitl @LeonDraisaitl and his #letsgooilers

@northumbriana The perfect place for "location, location, location"

Pleased to report that much of Northern England looks absolutely idyllic this evening.

@katharina_hoff Sorry to hear that! I fully support you, especially the concept of family friendly meetings. I mean, even those 1% have families, don't they? Keep fighting, it will be worth it!

@EckerleIsabella @hug_ge @UNIGEnews @MIMOL_UNIGE @gcevd Viel Erfolg!

@koehrsen_will +1 for VsCode's debugging tools. They make debugging so easy (as compared to jupyter notebooks).

You can do much better than print with no extra effort. `import IPython; IPython.embed()` drops you into an IPython session so you can inspect local values and run code one line at a time. Or use any of VSCode's built-in debugging tools.

Brain cells vs viruses vs bacteria. Microorganisms size comparison: #neuroscience

@eriksonnhammer @Cancerfonden Grattis Erik!

Vårt projekt "Spatially and temporally resolved gene regulation networks to understand liver cancer formation" var ett av de 224 som fick anslag från @Cancerfonden! Tack alla som skänker gåvor till forskningen och gör detta möjligt! #tillsammansmotcancer #cancerfonden

@BernhardQ @DanielWefers Wäre super, wenn ich auch einen haben könnte!

Wer hat noch nicht?

@miguelrausa ja, ich auch!

möchte hier noch jemand zu bluesky rübergerettet werden?

@roter_pander ich suche auch noch

Solltet ihr einen Code für die andere App haben, oder noch einen suchen, schreibt es mal in die Kommentare. Hat letztes Mal gut geklappt und vielleicht kriegen wir so noch ein paar Leute rüber 😊 Löscht bitte euren Kommentar, wenn ihr euren Code habt/losgeworden seid

Here, in full directly on Twitter, is "A Hackers' Guide to Language Models". This 90 minute tutorial is designed to be the one place I point coders at when they ask "hey, tell me everything I need to know about LLMs!" It covers both @OpenAI models and open source ones in depth.