IVAN ILIN

I've been researching Agents for the past 6 months and collected 40+ materials on the most capable architectures & implementations. The intent was to publish a comprehensive overview, like I did on RAG techniques, but been too busy with iki.ai, so sharing it here. There are some great intro lectures by Andrew Ng to start with. The following types of Agentic architectures are covered: 🤖 Chain of thought (Plan & Execute agent) 🤖 Tooling operators (An agent upon a set of tools, routing to them) - good for connecting external data storage & APIs, pretty fast and robust 🤖 ReAct (Thought - Action - Observation) - capable of iteratively executing complex tasks or answering complex queries 🤖 Self-Reflection - (Action - Observation / Evaluation - Reflection - Planning) - adds some quality and reasoning clarity compared to the ReAct scheme, might be slower 🤖 Agent upon agents (A multiagent scheme) - a quite complex setting, slow, but capable of executing very complex multistep tasks, not super robust as loops are a frequent issue. Most successful projects: @Auto_GPT, @AgentGPT, @MemGPT, GPT-Researcher, @crewAIInc, @MetaGPT_. There are also some arXiv papers & blog posts on the most important architectures. 🔗 All the materials are here: lnkd.in/ex2hE22k 🧠 The best part is there is a co-pilot to chat with all this knowledge! If you’d like to add some valuable publications on Agents to this collection - just share a link in the comments 👇

📟 Meet SWE-rebench-V2: the largest open, multilingual, executable dataset for training code agents! We at Nebius AI R&D are releasing the biggest open dataset of RL environments for training coding agents. We built an automated pipeline to extract real-world tasks at scale, and now we are sharing everything with the community. This release is designed for large-scale RL training. What’s inside: > 32,000+ executable tasks — every task is based on a real-world issue and comes with a pre-built Docker env. > 20 programming languages — moving beyond Python-only datasets (including less-represented ones like Lua, Clojure, etc.). > 120,000+ extra tasks derived from real pull requests. > High quality — tasks are filtered and labeled using an LLM ensemble. They are also enriched with metadata and tested interfaces to ensure solvability. We are also dropping a technical report with all the details on our extraction pipeline and model evaluations. 📄 Paper and dataset: huggingface.co/papers/2602.23… 👾 Discord (we are online there for any feedback/issues): discord.gg/wXYmWpMu We are open to research collaborations — feel free to reach out! 🔁 If you find this useful, please help us spread the word by sharing

@scade_xyz Looks great! Eager to try it

Meet Scade. The first full-cycle marketing system Forget the 30-person marketing team Build a full content plan for the month in 10 minutes All in one place: visuals, copy, trends, auto-posting, metrics Waitlist is LIVE: scade.xyz More below 👇

@karpathy Of course, RAG and Agentic patterns like ReAct are mainstream for 2 years, they are all about context, not just prompting. Sounds a bit trivial TBH

+1 for "context engineering" over "prompt engineering". People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step. Science because doing this right involves task descriptions and explanations, few shot examples, RAG, related (possibly multimodal) data, tools, state and history, compacting... Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and the LLM costs might go up and performance might come down. Doing this well is highly non-trivial. And art because of the guiding intuition around LLM psychology of people spirits. On top of context engineering itself, an LLM app has to: - break up problems just right into control flows - pack the context windows just right - dispatch calls to LLMs of the right kind and capability - handle generation-verification UIUX flows - a lot more - guardrails, security, evals, parallelism, prefetching, ... So context engineering is just one small piece of an emerging thick layer of non-trivial software that coordinates individual LLM calls (and a lot more) into full LLM apps. The term "ChatGPT wrapper" is tired and really, really wrong.

@Samarsky That’s a HUGE milestone! No bullshit, just traction. Congratulations, well deserved 🤝

We reached $100k in monthly revenue after 1.5 years

@omarsar0 It’s like a common knowledge for 2 years already

YC on the key prompting techniques used by the best AI startups:

@karpathy You can rephrase it another way - a system prompt is one thing that is not learning now, one parameter that is not updated after the agent gets more experience interacting with environment. Actually having a tool to update system prompt after some findings would work as v0

We're missing (at least one) major paradigm for LLM learning. Not sure what to call it, possibly it has a name - system prompt learning? Pretraining is for knowledge. Finetuning (SL/RL) is for habitual behavior. Both of these involve a change in parameters but a lot of human learning feels more like a change in system prompt. You encounter a problem, figure something out, then "remember" something in fairly explicit terms for the next time. E.g. "It seems when I encounter this and that kind of a problem, I should try this and that kind of an approach/solution". It feels more like taking notes for yourself, i.e. something like the "Memory" feature but not to store per-user random facts, but general/global problem solving knowledge and strategies. LLMs are quite literally like the guy in Memento, except we haven't given them their scratchpad yet. Note that this paradigm is also significantly more powerful and data efficient because a knowledge-guided "review" stage is a significantly higher dimensional feedback channel than a reward scaler. I was prompted to jot down this shower of thoughts after reading through Claude's system prompt, which currently seems to be around 17,000 words, specifying not just basic behavior style/preferences (e.g. refuse various requests related to song lyrics) but also a large amount of general problem solving strategies, e.g.: "If Claude is asked to count words, letters, and characters, it thinks step by step before answering the person. It explicitly counts the words, letters, or characters by assigning a number to each. It only answers the person once it has performed this explicit counting step." This is to help Claude solve 'r' in strawberry etc. Imo this is not the kind of problem solving knowledge that should be baked into weights via Reinforcement Learning, or least not immediately/exclusively. And it certainly shouldn't come from human engineers writing system prompts by hand. It should come from System Prompt learning, which resembles RL in the setup, with the exception of the learning algorithm (edits vs gradient descent). A large section of the LLM system prompt could be written via system prompt learning, it would look a bit like the LLM writing a book for itself on how to solve problems. If this works it would be a new/powerful learning paradigm. With a lot of details left to figure out (how do the edits work? can/should you learn the edit system? how do you gradually move knowledge from the explicit system text to habitual weights, as humans seem to do? etc.).

My guess is that MCP approach is just the beginning – LLM APIs would emerge to plug in services as tools instead of the traditional front-end of your webpage.

Consumer LLMs like ChatGPT and Claude are becoming the new gateways to the Internet, replacing traditional search engines. This will reshape the whole structure of the Internet, rewiring web traffic patterns – we'll get information, shop and even use other sevices through chats.

@ToporSB 😅

@jeff_weinstein Try IKI.AI – built for capturing knowledge from the web

I want “Save to memory” buttons all over the web. (And a share sheet item to send to memory.)

🚀 March has been quite a month for IKI.AI 🥳 📈 x2 MRR and growing with $0 marketing spend this month 🧠 Agent in collections now takes time to think and produces a highly precise and comprehensive result–perfect for automating analytical tasks. We’ve also introduced thinking step visualization, so you can see exactly how the AI reasons through your queries. Watching it reflect on your question is captivating! ✨ A ton of UI updates like a draggable co-pilot separator, code snippets highlights, markup support in notes, and many more small things, making the product enjoyable. 🥰 Most importantly, we're getting lots of positive feedback from our users, not on the idea, but on the product itself. That makes the whole startup journey and grind worth it.

@kevinrose @JustinMezzell @alexisohanian @digg Not an expert here, but for me it looks like we’ve got all the tech to do human / non-human verification right now with 99% accuracy. The other thing the post touches is the credibility of human claims - is there anything better than relying on reputation and expertise?

Working alongside @JustinMezzell and @alexisohanian on the new @digg, I find myself particularly focused on what I see as the defining challenge: the battleground for authenticity. Put simply: How do we know who's real in a world of perfect AI mimicry? Right now, we don't. And without intervention, we're heading for a complete collapse of online credibility. I believe the next evolution of trust goes beyond binary human verification (checkbox or not). What excites me is building systems where trust compounds through verified actions while preserving user privacy. With Zero-Knowledge Proofs, particularly zkTLS, we can authenticate claims without compromising personal data (e.g., someone cryptographically proving they own the Oura ring they're adamantly recommending in an upcoming digg comments thread). @sama was very prescient here, laying the groundwork with World ID (@worldcoin). I'm excited to see this extended from "this is a human" to "this human's claims are verifiable." This erosion of digital trust impacts everything we touch online, this is a big problem. If this resonates with you, I'd love to connect - we're seeking thoughtful collaborators who grasp the magnitude of what's at stake. cc: @OpacityNetwork @earnos_io @worldcoin

⚡️IKI.AI product video is organically going viral on Youtube ⚡️ - 12.5k views and growing - +2k users in the app - coming to 500 likes - almost 100 comments - +400 subscribers on the channel Curious how? Ask @TheMaxOr 🧙‍♂️ check it out: youtu.be/Zd4BYtb1Py0

@pmitu With IKI.AI – never more, just for rephrasing things

Do you use ChatGPT for your content?

@literallydenis Congrats, Denis! 🥳

Personal update: I've joined JetBrains! 💻 I've joined as a Group Product Manager – at JetBrains, I'll be helping develop AI Assistant tools for IDEs, turning my 8 years of AI expertise into valuable (and sometimes quirky) tools for developers. I want to dive deeper into language models, AI agents, and the future of developer tools P.S. neural love continues to thrive, and I'm staying onboard as co-founder P.P.S. Please feel free to share feedback about our AI whenever you'd like!

@iamfakhrealam There is a free access for the first month, would be happy to know your feedback!

@ivanilin9 If I had access, I may have left my feedback as well with upvote

@SergeiLavrukhin Hi Sergei! You can add a list of links at once, just need to export them from the other places. We’ll add YT playlist processing, that’s a great thing

@ivanilin9 I’ve installed IKI multiple times after each product update, but I still can’t find a way to add all my bookmarks from a YouTube playlist and Chrome browser. Do they still have to be added one by one, or is there a solution?

🚀 A big day for us – IKI 2.0 goes live on Product Hunt now producthunt.com/posts/iki-ai-2… 🚀  🎁 There is a nice discount - check it out 👀 Since our first launch 9 months ago, we’ve reimagined IKI as an LLM-native space for professional knowledge, gained thousands of paying customers, polished and scaled the platform, and now are happy to invite you to IKI 2.0. Our mission has not changed though - we’re here to equip knowledge workers with the best LLM-powered digital library—a true second brain, period. And we’ve done a lot to materialize this vision - in IKI 2.0 you will enjoy: ⭐ URLs, PDFs, YouTube, txt, docx, md, images & notes support ⭐ Spaces to collaborate with your teammates ⭐ OS-like navigation with a hierarchy, drag & drop and multi-select ⭐ SOTA reasoning LLMs, including Claude 3.7 Sonnet, o3-mini, and Grok-2 ⭐ Cutting-edge web search powered by Perplexity Sonar ⭐ Smart editor to write with IKI assistant ⭐ Deep comparative analysis & reasoning over collections – shipping this month! ⭐ Cloud integrations with Google Drive, Notion, Dropbox, and Obsidian are coming soon Happy to have you among our early adopters ✨

@TheMaxOr made it!

@pmitu Happy to have you as our trusted hunter, dear @pmitu! 🤝 Would not work without your support!

Let's support new launch 🎉 IKI 2.0 goes live on Product Hunt now: producthunt.com/posts/iki-ai-2… Product mission has not changed though - IKI is here to equip knowledge workers with the best LLM-powered digital library—a true second brain. And there is a nice discount inside, check it 👀✨