Alan Nichol

so @hugobowne @dakshv22 and I tried some different approaches to building LLM-powered agents externalising business logic cuts LLM costs by 78% and latency by 5x more importantly your assistant won't go off the rails in ridiculous ways rasa.com/blog/cutting-a…

@hwchase17 The tools themselves need state! rasa.com/blog/process-c…

im bullish on long-running, stateful agents who's building one of those?

MCP is all the rage, but for customer-facing AI "tool use" 🛠️ is actually the wrong abstraction. There's a new agentic pattern emerging which I've started calling "process use". It's a much better fit for voice and chat AI in customer support.

and link to the code github.com/amn41/cigarett…

link to sign up: cigarettecard.club

Tobacco companies used to put lovely little collectible cards in packs of cigarettes, covering a broad set of topics. I made a little newsletter that will send you one every day.

@hyusapx this is so cool! would be amazing if you put the word in a query param so you can share

but it also could assign meaning to words that don't exist!

built an ai tool for etymology nerds: deconstructor. live rn. the goal was to analyze the origins of english words, but it ended up having a ton of amazing emergent properties (see below)

Such a great conversation with @alanmnichol about the limits of LLMs and on how to integrate a stochastic component (LLMs) into a more formal system (application).

@jacksbridger @mgonto if you're collecting information, you have to store it somewhere. If people can host the model themselves, the model would have to like secretly store stuff somewhere? Can't do it in the weights, those are static.

@mgonto Interesting! @alanmnichol is there any world this is possible theoretically?

How could they do this with something fully open source?

Are LLMs a magic solution for enterprise customer support? @martin_casado and @Rasa_HQ cofounder and CEO @alanmnichol dive into the challenges and opportunities of AI-driven customer interactions. Here, they explore the hurdles of ensuring reliability, the sweet spot for LLMs in support workflows, and the evolution from traditional NLP. For the full conversation, check out the AI + a16z podcast: a16z.com/podcast/from-n…

@matthausk @martin_casado a16z.com/podcast/from-n…

@alanmnichol @martin_casado is there a link to a longer version of this conversation?

On the a16z pod @martin_casado and I riff on how chatbots evolved from word2vec to gpt-4. We touch on agents, RAG, function calling, reliability, and briefly features me beatboxing.

@solyarisoftware @hwchase17 @ilanbigio @langchain @OpenAI @Rasa_HQ thanks for the tag, Giorgio! Btw you don't have to write yaml if you don't want to :) youtube.com/watch?v=ZhukuP…

9/ Vision: Make conversational AI accessible and scalable by empowering non-coders to design robust systems with ease. 🌐💬 Mentioned technologies @hwchase17 @ilanbigio @alanmnichol @langchain @OpenAI @Rasa_HQ #LLMs #AgenticAI #ConversationalAI #chatbots

I just published my new article: Conversation Routines: A Prompt Engineering Framework for Task-Oriented Dialog Systems convcomp.it/conversation-r… 🧵Exploring a no-code framework for designing task-oriented conversational agentic systems (CAS) using Conversation Routines (CR)!👇

Let’s go! #calmsummit24

@AdityaSohoni but no intents doesn't mean no structure. It's absolutely silly to have an LLM guess a known bit of business logic over and over rasa.com/blog/stop-buil…

@AdityaSohoni 💯 that 'prompt and pay' is not business ready. You do need structure, but it doesn't have to be intents. I've been saying intents need to go since 2019 rasa.com/blog/its-about…

sometimes I feel like guys like @alanmnichol actually make sense, that intent-based chatbots are better than llm agents. But then I'm like nah llm agents are the future. but one thing is clear. llm agents today are not business ready

admire the boldness of a blog on 'robust generative AI agents' w/no actual solutions to make these robust Lots of monitoring & observability to quantify how flaky your system is, but still just a "prompt and pray" approach to having LLMs execute tasks. aws.amazon.com/blogs/machine-…

OK this is happening...

@stephensackur @BBCHARDtalk So sorry to hear. First saw the show at age ~11 and it was the first time I understood you could ask powerful people hard questions.

THREAD: 1. PERSONAL NEWS: Today BBC News has announced plans to close @BBCHARDtalk after 3 decades holding the world’s politicians and powerbrokers to account. This is sad news for me personally, but much more important, I think it’s depressing news for the BBC🧵

@hwchase17 @simonw if you wanna build something truly agentic jit_implementation is all you need reddit.com/r/MachineLearn…

@simonw The more agentic an application is, the more an LLM decides the control flow of the application

Let’s see if we can crowdsource a robust definition of “agent” (with respect to AI and LLMs) that fits in a <=280 character tweet Reply to this with your best attempt, then scroll through the replies and fave the ones that makes sense to you

@sh_reya So right! The thing the @dltHub folks have figured out is how to bridge the ML and Data Engineer crowds. ML people do not know what airflow or delta lake or dbt are

it has been 10 days since I posted about our DocETL project, and the most surprising (and exciting) thing to me is that (so far) many users have never used a data processing framework before. there is a huge learning curve to understand and manipulate dataflow