https://discord.gg/botpress logo
Join Discord
Powered by
GPT-5 Level Research Agents
# 📖tutorials
q
Hi @gentle-engine-13076 👋 Hope you're enjoying the weekend! Maybe you have some ideas for this 💡 Let's hope this thread turns out to be an actual tutorial that has links to many chatbot projects built by numerous bot builders. Like I mentioned in a message in #business-section (don't read it, too long), if I feel something big work-wise is coming up, I start learning the needed skills way before. Now, I've got this feeling that in a year, most of our work / my work will involve using Agents. Hear me now, believe me later: So I decided to practice building Botpress Agents. It's going to involve chatbots talking to each other using something like socket.io (not sure if socket.io is the best choice or even needed here, but I need to test that anyway in other project). One chatbot acts as the boss (that's me, using it to send out orders), handing out tasks to managers, a chatbot that quickly thinks of a first plan using AI, then passes it to other chatbots (the team) who'll look for pros and cons of that plan, then report back. When the final plan is ready, they'll send it back to the boss chatbot or email it wherever it's needed. At first, these chatbots might just perform only simple or even silly tasks, possibly getting stuck in an infinite loop. But with time and practice, we could have a system where chatbots work together solving tasks for clients, making sure everything's right before sending it off. And when GPT-5 arrives, and if its focus is on using Agents as many believe, we will be ready! 🦾 🤖 https://cdn.discordapp.com/attachments/1226459863610294294/1226459864050569388/DALLE_2024-04-07_12.03.04_-_Imagine_a_futuristic_brightly_lit_office_space_with_sleek_modern_furniture_and_large_panoramic_windows_overlooking_a_bustling_cityscape._In_the_for.webp?ex=6624d8c2&is=661263c2&hm=5a5ca0e11e1d8779220668ace8758815661c256b9e2ba117b1164c5bc11a598a&
Like scientists, we start every problem by asking the first simple questions. Unlike scientists, who use research papers and still like to do it manually, we bot builders use YouTube videos and do it automatically using AI to create the best Q&As. Key topics covered include the advancements in GPT-5 and AI agents, the importance of early adoption, challenges related to hardware and access, the role of agents in various domains such as gaming, influencer marketing, and software development, the transformation of operating systems, the necessity of specialized agent teams, the potential for AI to automate tasks, and the future of AI agent development. Here are relevant Q&As for each topic: Advancements in GPT-5 and AI Agents Q: What is the main focus of GPT-5? A: GPT-5 is focused on calling AI agents to perform tasks autonomously. Importance of Early Adoption Q: Why is early adoption of GPT-5 and AI agents important? A: Early adopters can gain a significant competitive advantage as these technologies can enhance productivity and task completion quality. Challenges Related to Hardware and Access Q: What is the 'great hardware problem' in AI? A: The inability for general consumers to access the hardware necessary to run large AI models, making decentralized hardware essential. Role of Agents in Various Domains Q: How can AI agents impact gaming? A: Agents can play complex games like Doom using simple text representations, showing their potential in understanding and interacting with environments. Transformation of Operating Systems Q: How will AI agents change operating systems? A: Operating systems will evolve into agent interfaces, allowing AI to control computers, open apps, and perform tasks directly.
Necessity of Specialized Agent Teams Q: What are the two essential teams needed for building AI agents? A: A programming team to code functions and a prompt engineering team to write system prompts for future agents. Potential for AI to Automate Tasks Q: Can AI agents perform real job tasks today? A: Yes, AI agents like Devin are already completing real job offers on platforms such as Upwork and Reddit. Future of AI Agent Development Q: What will the future competition among humans likely be based on according to the text? A: The future will likely see competition based on who can build better AI agents, as they will perform most tasks. Influencer Marketing Q: How will AI agents transform influencer marketing? A: Agents will simulate followers to predict reactions to products, helping companies select influencers who can drive the most sales efficiently. Software Development Q: What is autode and how does it differ from previous AI agents? A: Autode is a Microsoft AI agent capable of performing complex software development tasks, operating with multiple agents for various tasks, unlike single-agent systems.
The key topics include the advancements and capabilities of AI agents, the significance of agent workflows, the improvement of agent capabilities over time, and the expected impact on various industries and societal aspects. Advancements and Capabilities of AI Agents Using GPT-5 Q: What new capabilities do AI agents have with the introduction of GPT-5? A: GPT-5 enables AI agents to perform tasks autonomously with improved understanding and interaction with text, speech, and video data. Significance of Agent Workflows Q: How do agent workflows enhance the performance of AI models like GPT-5? A: Agent workflows allow for iterative and collaborative processes, where agents can plan, review, and refine their outputs, leading to significantly better results compared to non-agentic workflows. Improvement of Agent Capabilities Over Time Q: How have AI agents improved over time, and what future advancements are expected? A: AI agents have become more capable of executing complex tasks autonomously. Future advancements include better reasoning, personalized and customizable interactions, and the ability to perform physical tasks in the real world. Expected Impact on Various Industries and Societal Aspects Q: What impact are AI agents expected to have on industries and society? A: AI agents are expected to revolutionize industries by automating routine tasks, enabling personalized customer experiences, and enhancing decision-making processes. In society, they could significantly affect employment, privacy, and security. Role of AI Agents in Enhancing Accessibility and User Experience Q: How do AI agents using GPT-5 enhance accessibility and user experience? A: By understanding and processing multimodal data (text, speech, video), AI agents can provide more intuitive and accessible interfaces for users, catering to diverse needs and preferences.
Key Topics: -Introduction to AI Agents -Early Adoption of AI Agents -Agent-Driven Workflow and Efficiency -Customization and Personalization through Agents -Development and Integration of AI Agents -Future of AI Agents and Learning Q&As: Introduction to AI Agents Q: What are AI agents? A: AI agents interact with the external world using tool usage, memory, planning, and taking actions. Early Adoption of AI Agents Q: Why is early adoption of AI agents important? A: It offers a unique opportunity to be ahead in technology, as building agents will become common knowledge in 5 years. Agent-Driven Workflow and Efficiency Q: How do agent-driven workflows enhance AI efficiency? A: They make AI models like GPT-3.5 perform tasks better than newer models by improving task execution flow. Customization and Personalization through Agents Q: What benefits do AI agents offer in terms of customization? A: Agents allow for personalized experiences, remembering user preferences and adapting to specific requirements. Development and Integration of AI Agents Q: What is essential for developing AI agents? A: Understanding the difference between human and model cognition, and focusing on flow engineering are key for developing effective AI agents. Future of AI Agents and Learning Q: How will the future of learning be impacted by AI agents? A: Learning how to create and interact with AI agents will become an integral part of education, shaping the future of technology engagement.
⬆️ Based on studies, if you're using GPT-3.5 with agent tasks/agentic workflows, it outperforms GPT-4. Does this mean that using GPT-4 with agent tasks/agentic workflows now outperforms the future GPT-5?! According to many AI experts, yes. Recommended playlist if you want to supercharge yourself with some other AWESOME knowledge related to AI Agents, and maybe cry some happy tears like I did

https://www.youtube.com/watch?v=pBBe1pk8hf4▾

https://www.youtube.com/watch?v=c3b-JASoPi0▾

https://www.youtube.com/watch?v=sal78ACtGTc&t=2s▾

https://www.youtube.com/watch?v=yinHx5UnYs0▾

https://www.youtube.com/watch?v=twiog1bzvYs▾

g
I feel like Neo, because the term “agentic AI” has been like a splinter in my mind. 😎 (Does that make you Morpheus? 🤔😆) Experimenting with Groq (and the Groq interview you shared) as well as Andrew Ng’s recent writing about agentic AI workflows hammering home the buzz that’s been in the scientific literature for a while now have all been percolating in my mind. The synergy between super fast inference and agentic workflows is very exciting. ⚡️ Specifically, what piques my interest is workflow efficiency and squeezing better results from smaller models (cost-efficiencies at scale). Sidebar: The buzz around GPT-5 is a little meh for me. (I mentally mentally earmarked GPT-V as the actual GPT-5 — since it’s the same in Roman numerals; and, to me, adding multi-modal support was the a game-changing level up from plain ole generative text.) Exploring agents as hyper-optimized prompt-engineered personas is something I was playing around with this time last year with GPT-3.5 and the fresh/untamed GPT-4. One of my AI obsessed engineer buddies has a robotics & orchestration background and was early to output optimization through iterative prompt-engineering and using multiple narrowly crafted (agent-like) personas. Ultimately, that influenced my experimentation led me down the rabbit hole to Botpress by way of the LangChain library based LangFlow and the visual-builder paradigm looking for fast & stable ways to build extendible agents with out getting bogged down in package management and half-baked open-source hacker projects. (Nothing against OSS hacker projects, but I’m not a god-level programmer to handle debugging or security issues that can accompany the bleeding edge in that category). I’ve got a few more to-do’s to check off my list before I can get into build-out mode and start iterating on a few flow designs I’ve been toying with in my mind. Definitely plan on sharing those here! What I’m currently thinking about mainly involves calling Groq API, but I’m also starting to imagine how to loop [Dust](https://dust.tt) into the workflow too (they’ve got a really interesting & powerful product.) If this all evolves the ways we’re thinking about it, I think the end-user-facing UI factor will play a bigger role than in previous chatbot era (defined by widgets). I’ve got some front-end experience, but as with bot/agent-building, I’m always looking for efficiency gains. My designer side has been crushing on Framer due to its yummy React based Motion library (and visual builder). I like FE coding directly w/o mock ups, but design workflows typically get segregated from dev work by way of Figma or XD etc. so I aim to find ways to tie bots/agents into Framer components and so forth (especially since Framer can also support Shopify API). Maybe Framer integrated bots/agents will be a parallel exploration to my agentic workflows (file under: medium-term goals for the year. Perhaps FlutterFlow integration as longer-term goal for the year TBD). Thanks for the ping. I’m glad you’re thinking about what comes next too! 🦾
PS Groq hosting an AMA on 4/16 re: new release of the tool use/function calling feature. 😗 The polling looks like it's gonna be 12pm PST (gonna miss office hours for this one 😆 sorry Robert.)
q
So, on Discord, if I prompt 'Hi [username] 👋 Hope you're enjoying the weekend! Maybe you have some ideas for this 💡', do I get these same GPT-10 level quality answers from other humans too? Is this the Matrix?
One thing I forgot to mention: Similarly, like how using GPT-3.5 with agent tasks can outperform GPT-4, combining different models to correct each other's errors in agent-like tasks — such as pairing GPT-4 with Claude — clearly surpasses relying on just one model for solving the same difficult task. So, for those use cases where GPT-4 isn't smart enough, leveraging multiple models might be a solid strategy.
g
Test the Matrix hypothesis!
Yes! That’s kinda what I’m thinking about with Dust because they’ve got multiple models and some really cool parallel processing jazz you can tap into. 💊
h
Following your recommendations has actually changed my YT content & proud to say it is because of you i have already came across and watched all the content you suggested in your playlist related to Agents
Everything you have said here is a game changer and i recommend the Andrew Ng on Sequioa Capital video. It has so much value and I learnt a lot from it. You can create agents via chain prompting Agents are the future and that video breaks it down so well. You can implement everything he is talking about right here on Botpress
Funny am just getting started "exploring agents as hyper-optimized prompt-engineered personas" just a few weeks ago and Carol was working on it a year ago🥶 This gives me hope that by next year I will be as good and as knowledgeable as Carol if i keep this up🔥
Thanks for the knowledge @quick-musician-29561 and for the detailed response @gentle-engine-13076. I am learning so much from you guys and just by going through this content I believe I will finish the year half as good as you guys!
q
"You can implement everything he is talking about right here on Botpress" That would be a JACKPOT 🛠️ 🚀 edit: I'll update the opening message always when we discover something worth mentioning there too.
g
If @hundreds-battery-97158 is talking about Agentic AI using Botpress, then 💯 that's exactly what I'm gonna start building this week. 🦾 My to-dos are 98% done and I'm pumped to shift gears! 👏🏼 Will try and keep the scope narrow for v 1 so I can share I think it's a relatively simple flow... or that's how it looks in my head
q
"The set of tasks that AI can do will expand dramatically because of agentic workflows." "We have to get used to delegating tasks to AI agents and patiently wait for a response." "Fast token generation is important. Generating more tokens even from a lower quality LLM can give good results." "If you're looking forward to running GPT-5/Claude 4/Gemini 2.0 (zero shot) on your application, you might already be able to get similar performance with agentic reasoning on an earlier model." -Andrew Ng-
Harrison Chase (LangChain), answering to these same questions and sharing his knowledge on the topic (via AI Summary): Key topics relevant to the integration and utilization of AI agents, especially in the context of enhancing their capabilities through agentic workflows, delegation, fast token generation, and leveraging existing AI models for improved performance. The topics revolve around agentic workflows, task delegation to AI, token generation efficiency, performance optimization using earlier models, and the evolving capabilities of AI agents in planning, user experience (UX), and memory. Agentic Workflows Q: What are agentic workflows? A: Agentic workflows refer to processes where tasks are delegated to AI agents, allowing them to perform a series of actions or steps autonomously based on planning and reasoning. Task Delegation to AI Q: Why is delegating tasks to AI important? A: Delegating tasks to AI agents is crucial for efficiency and productivity, as it enables automation of repetitive or complex tasks, allowing humans to focus on more strategic activities. Token Generation Efficiency Q: Why is fast token generation important for AI agents? A: Fast token generation is essential because it enables AI agents to provide quicker responses and solutions, enhancing their usability and application in real-time scenarios. Performance Optimization Using Earlier Models Q: Can earlier AI models achieve performance similar to newer versions with agentic reasoning? A: Yes, by employing agentic reasoning, earlier AI models can be optimized to achieve performance levels comparable to newer versions, making them more cost-effective and accessible.
Planning Q: How do AI agents use planning in agentic workflows? A: AI agents use planning to determine the sequence of actions or steps needed to achieve a goal, incorporating strategies for decision-making and problem-solving. User Experience (UX) Q: What role does UX play in the effectiveness of AI agents? A: UX is crucial for ensuring that interactions with AI agents are intuitive and efficient, improving the overall effectiveness and adoption of AI technologies. Memory in AI Agents Q: How does memory enhance the capabilities of AI agents? A: Memory allows AI agents to remember previous interactions, preferences, and outcomes, enabling them to provide more personalized and contextually relevant responses.
More Andrew Ng: The key topics revolve around the advantages and concepts of AI agents, agentic workflows, the significance of fast token generation, and the potential of combining various AI models to achieve high performance in task execution. These topics lay the foundation for understanding how AI agents can transform task management and execution, providing a more iterative, reflective, and comprehensive approach to problem-solving. Agentic Workflows Q: What makes agentic workflows superior to traditional non-agentic workflows in AI? A: Agentic workflows involve iterative processes where AI agents perform tasks, review, and revise their outputs, leading to remarkably better results compared to the one-shot answers of non-agentic workflows. Q: How do agentic workflows contribute to the expansion of tasks AI can do? A: By enabling AI to perform iterative, reflective, and complex task sequences, agentic workflows dramatically expand the range and complexity of tasks AI can effectively handle. Delegating Tasks to AI Agents Q: Why is patience important when working with AI agents? A: Effective delegation to AI agents often requires waiting for the AI to perform iterative refinements, which can take time but results in higher quality outcomes. Q: How does delegating tasks to AI agents change the interaction dynamics? A: Delegation requires adjusting expectations around response times, embracing a more managerial role where tasks are assigned to AI agents and outcomes are awaited.
Fast Token Generation Q: Why is fast token generation important in AI model performance? A: Fast token generation allows for quicker iterations over tasks, enabling AI agents to generate and refine outputs more rapidly, which is crucial for efficient agentic workflows. Q: Can lower quality LLMs be effective with fast token generation? A: Yes, generating tokens quickly with a lower quality LLM can yield good results, as the speed facilitates more iterations and refinements, compensating for initial quality. Combining Different AI Models Q: How can combining AI models improve application performance now? A: By employing agentic reasoning with existing models, applications can achieve similar performance to what is expected from future, more advanced models, demonstrating the efficiency of leveraging multiple AI capabilities in a cohesive workflow. Q: What is the advantage of using earlier models with agentic workflows? A: Earlier models, when used within agentic workflows, can outperform or match the effectiveness of more advanced models by leveraging iterative improvements and comprehensive task management strategies.
If you’re finding it hard to keep up with all the messages here (I also know that all too well), whether you’re part of the triple-A crowd (AI Automation Agencies) or still in a triple-B team like me (Beginner Bot Builders), here's one option. I converted the essence of these messages into an story-like audiobook. Like many, I enjoy a good AI-created audiobook, especially for science fiction or fantasy genres. But for non-fiction, the value is even greater when accuracy is guaranteed 100% (no errors, no AI hallucination), just like in this one. It’s perfect for catching up without the worry of missing out or misinformation. This wasn’t created by asking AI to create an audiobook about the topic; it was made by first providing AI with all the relevant information (for example, transcripts of expert talks) and then asking AI to create an audiobook based on that verified information only (presented in an adventure in a story-format) 🫡 https://cdn.discordapp.com/attachments/1226459863610294294/1227190897548988457/ElevenLabs_2024-04-09T09_22_17_Chapter_1.mp3?ex=66278196&is=66150c96&hm=12c120312574ab28174aabcfba608e4ecca5af4a4e72b0f74f4436c6ed304614& https://cdn.discordapp.com/attachments/1226459863610294294/1227190898077597726/Audiobook_AI_Agents.txt?ex=66278196&is=66150c96&hm=8e2fe6ded1709bc8bd4bdb66fb87fc66af478131cc170ef34425d68bf14f967b&
h
devmik already producing high quality and high value AI podcasts. YC has a company called infininity AI in its W24 batch and the company is generating AI video content from text
q
I forced one of my co-workers to listen to that 4-minute audiobook/podcast, chapter 1. He said he didn't understand anything 🤷🏾‍♂️ Then I forced him again to read all the summaries, then we took a 30-minute break, and after that, he listened again. Now he said it made perfect sense! 💡 Based on this lab test, with a sample size of 1 human guinea pig, the method for quickly understanding difficult or new topics, AI Agents in this case (listening, reading, taking a short break, and repeating) had a success rate of 💯
(Most likely everyone already knows this kind of stuff, but I'll add it here anyway) ⤵️ That's how I also learn to quickly understand everything I need to (in my work, for example). I make summaries of long, difficult, and important content and study that again and again until it clicks. If some part is really interesting, I dig deeper into it and read the original text. However, we can save lots and lots of time by focusing only on the summaries of quality content by experts in the field (quality content like their books, interviews or podcasts/videos) on the topic in the beginning, because most of the time, we need to understand it quickly and not waste days or weeks. And if I also need to learn how to do it, the actual skill and not just understand it (like how to use some new coding library), then the method of 'listening, reading, taking a short break, and repeating' also involves a massive amount of practicing and using that newly acquired knowledge to turn it into actual skills. Fast.
c
Multiple-AI Agent workflows is definitely interesting and is something I have been doodling around with for a while. It has a massive potential for SaaS. @User have you looked into Autogen(Studio) ?
Love hearing this since I have found this method to work for me instead. I personally use GPT to help me summarize texts/videos which I then log into my Obsidian database (which after a year of studying has become a library ha). For the last months I have been thinking about throwing my vault online on a forum, maybe it is an idea to combine our vaults and open it up for public contribution ( like an open Library )
q
I saw your message earlier, but I had to get this out of my head first https://discord.com/channels/1108396290624213082/1227955163373768714 Thanks for reminding about that. Autogen Studio has been something I've been meaning to check out, and now I'll dive deeper into it 🛠️ I've also been using Perplexity AI API a lot, which I really like and try to include that to my future Botpress projects.
c
using Perplexity AI API inside a bot can be interesting... especially with customized components so we can mimic how Perplexity presents it responses (very pretty )
q
I asked the following four questions -Can you explain what Botpress is and how it's used? -Could you provide information on neuroplasticity and why it's important? -What is Bitcoin, and how does it work? -Who is Elon Musk, and what are his notable achievements? and used four different chatbots: GPT-3.5 zero-shot, GPT-4 zero-shot, GPT-3.5 agentic workflow, and GPT-4 agentic workflow. Then, I sent those results to multiple LLMs (GPT-4, Claude-3, Llama 2, and Mistral 7b). I asked them to rate the responses in five categories from 0-100%: Freshness (up-to-date information) Helpfulness (guidance and clarity) Factuality (accuracy and verification) Holistic (user-friendly communication) Sounds Natural (likelihood of being a human answer) Here are the results: https://cdn.discordapp.com/attachments/1226459863610294294/1228416071912914964/Screenshot_from_2024-04-12_21-30-59.png?ex=662bf69e&is=6619819e&hm=4e0dc67607067fca106f18584fff34365725fb742feae0c2704845857b7eac04&
The next version uses seven research agents, including GPT-4, Claude-3, Mixtral-8x7b, and Llama2-70b. Here's a 1-minute, 16-second video https://discord.com/channels/1108396290624213082/1227955163373768714/1229064920155422750
Adding this to my 'summer hits' playlist:

https://www.youtube.com/watch?v=7LWTZqksmSg▾

@hundreds-battery-97158 This might be close to your earlier use case (checking government websites dealing with tenders). Months ago, when I first offered a simple customer chatbot to three companies owned by my friends, one in blockchain, one in nutrition, and one selling laptops, they weren't very interested, thinking it wasn't necessary or profitable. They thought a chatbot was just a nice new toy, but no one really needed it (at that point, I wasn't so good at building chatbots or AI tools) 10 days ago, I mentioned in this thread that when I feel a big work change coming, I start learning the needed skills way ahead. Since then, I've been all about building multi LLM workflows. Yesterday, I showed the first demos of those; a law firm and a blockchain company were really interested. The idea is simple (in my use cases anyway): if ChatGPT (GPT-3.5) or GPT-4 can’t solve a problem, I try building a tool using multiple LLMs to improve the solution, including error checks and, later, external APIs. And I'm trying the next step now for agent workflows. Instead of letting AI handle everything, users can now give feedback directly, no more 1-5 star ratings from AI (like in my first versions of these). Let real users decide what works by testing the code in actual situations, and give feedback if the solution was correct/perfect, or if it needs to be improved, while providing correction ideas or error messages to the chatbot and asking it to do it again. GPT-3.5 can sometimes do better than GPT-4, especially when you have it loop through tasks, refining answers multiple times. It’s not just cheaper but sometimes even faster.
I really like this comparison: Asking ChatGPT to solve a problem without refining it is like asking a writer to write a novel in one go, no edits, no pauses, just keep writing until it's finished. But using LLMs in a loop is like a writer drafting, getting feedback, and revising repeatedly until it’s just right. When I showed these my early AI agent workflows to business owners, they complained about the speed. So I told them, 'Imagine it's a critical task at your company. Would you prefer a quick, mediocre answer or the best possible solution, even if it takes a bit longer?' Just like employees or teams (from which it can take days or weeks), good AI needs a minute to deliver quality. That law firm I mentioned, they've tried many AI tools but struggle because the systems don’t handle their native language well. Even when the output is correct, it often requires too much manual tweaking to sound natural. I suggested a new test: let’s see if multiple LLMs can create a better starting plan based on earlier similar cases from the Supreme Court or preliminary rulings (if those are the right places; I’m no lawyer). They’re giving it a go, and we’ll know more after a few weeks of trials. The issue with the text not sounding natural, I'm betting that running the text through multiple LLMs for a language check will give much better results. I’d be surprised if it doesn’t.
I also think the tools they have used just try to find everything from the internet. That sometimes works (if you can provide the correct websites to look up information, preliminary rulings are public information), sometimes it doesn't. A much better solution might be to build really good Botpress Knowledge Bases, and even ask them to manually keep it always updated with 100% correct information (maybe with the help of Airtable or Google Sheets if needed), and explain that it requires more work, but the results (and profit) can be much bigger.
⬆️ It's good to mention again that this isn’t the optimal solution for a customer service chatbot, where fast answers from the company’s Knowledge Base are crucial. This setup is more suited for a company’s internal use (or as a paid service for their clients or on their website), because it costs more to deliver the best results (not when using GPT-3.5 at every step, which is also effective), and it takes more time (even up to 5 minutes, again not when using GPT-3.5). But if law firms or large companies can get really good results, they don’t mind if it costs $0.40 or $2, especially when it used to take a lot longer (days or weeks) and cost hundreds or thousands.
I don't actually recommend anyone use the new projects I've shared in #1132038253109837994 and #1120796649686573086 directly, they are the first simple versions and not that good yet. I wanted to share some ideas and provide starting points to encourage everyone to build their own versions, better suited to their use cases. Plus, if we get feedback from other bot builders who are testing and trying the same, it always improves our projects. Then sharing even the first versions becomes totally worth it! 🦾 🛠️ 🫡
c
that list must be massive innit ha, My own to-watch list just keeps growing and growing 😁
3 Views
PreviousNext
Powered by