Krisztián Szűcs

Where are all the agents? We were promised autonomous intelligence and we only got chatbots and glorified workflows sprinkled with LLMs. Where is the autonomy if I still have to account for all possible edge cases? Why do I still have to deal with APIs, spaghetti logic, and esoteric protocols? Why can’t I just describe the outcome I want in plain english and the computer will figure out all the details? These are the questions we’ve been asking ourselves at Latitude for the past few months. Today we’re introducing a new way to build automation software. Building for autonomy requires upgrading your mental model of what software is. All you need now are prompts and tools. Code is out, the LLM is the backend now. Latitude is the first full-stack agent engineering platform built around this idea. It is a code editor but you won’t see any code, only plain-english instructions. We have all the advanced tools you need to design flexible behavior: agent orchestration, tool calling, evaluations, and monitoring. And the best part is you don’t even have to worry about any of it. We’ve built an agent—Latte!—that takes care of everything for you. You don’t have to be an expert to program computers anymore. We’re bringing joy to building software, and this is just the first step.

121

9 Comments

We just shipped a full Python data pipeline that runs entirely in your browser tab Meet the new dlt Playground → powered by Pyodide + WebAssembly. You can load data into DuckDB using dlt, without installing Python, setting up a notebook, or connecting to a backend. No infra. No accounts. No waiting. 🔍 So what makes this special? ✅ Runs Python in the browser (via WASM) ✅ No Python installed ✅ No cloud dependency ✅ No backend at all ✅ Even works offline after load It’s a Playground, limited and experimental, but it’s a glimpse at: Local-first, install-free data tooling Private-by-default analytics LLM-native notebooks that boot instantly 🔗 Try it now → https://lnkd.in/e7v-wtW9 #dlt #dataengineering #wasm #pyodide #duckdb #python #localfirst #llm #opendata #dltHub

328

20 Comments

Lots of questions in DMs about the new Typesafe Agents + parser model. They're a low-key release with incredible potential. If you're building anything beyond chatbots, your API will take JSON in and expect JSON out — Typesafe Agents enable this structured I/O by handling the heavy lifting for you. On top of that, it's a well-known secret that forcing structured output reduces reasoning capabilities (because the model is trying to adhere to the schema). Using a separate parser model lets you keep high reasoning capability, while still getting structured output from a smaller model. Basically, if you're building AI systems beyond chatbots, Agno's Typesafe Agents + parser model are best in class and the way to go! Here's an example: https://lnkd.in/ezUdCjiF

85

11 Comments

𝗙𝗦𝗤 𝗢𝗦 𝗣𝗹𝗮𝗰𝗲𝘀 𝗡𝗼𝘄 𝗣𝗼𝘄𝗲𝗿𝘀 𝗢𝘃𝗲𝗿𝘁𝘂𝗿𝗲 𝗠𝗮𝗽𝘀: 𝗔 𝗦𝘁𝗲𝗽 𝗙𝗼𝗿𝘄𝗮𝗿𝗱, 𝗕𝘂𝘁 𝗠𝗼𝗿𝗲 𝗪𝗼𝗿𝗸 𝗔𝗵𝗲𝗮𝗱 We're pleased to announce that Overture Maps Foundation has integrated a subset of FSQ OS Places into their dataset as part of their September release, with this portion being single-sourced from Foursquare and not conflated with other sources. This means all the updates, corrections, and enhancements you make through our Placemaker tools are preserved in their original form. While this development aligns with what we've long advocated, we still have a long way to go in realizing our vision for a comprehensive open Places dataset. 𝗪𝗵𝘆 𝗧𝗵𝗶𝘀 𝗠𝗮𝗸𝗲𝘀 𝗦𝗲𝗻𝘀𝗲 This integration shouldn't come as a surprise to anyone familiar with our respective datasets. FSQ OS Places delivers unmatched volume and update frequency compared to the existing Overture Places dataset. Our human-centric approach to Places data, built on real-world usage patterns and community contributions, combined with our continuous data refresh cycle and comprehensive coverage provide the foundation that mapping applications need to stay current and accurate. 𝗧𝗵𝗲 𝗖𝗼𝗻𝗳𝗹𝗮𝘁𝗶𝗼𝗻 𝗖𝗲𝗶𝗹𝗶𝗻𝗴 This move validates our key concern: conflation-based Places systems hit improvement ceilings when you run out of data sources. By consuming FSQ data directly, Overture implicitly acknowledges their approach's limitations and proves our hypothesis that a comprehensive Places dataset requires continuous community-driven curation. 𝗧𝗵𝗲 𝗕𝗶𝗴𝗴𝗲𝗿 𝗣𝗶𝗰𝘁𝘂𝗿𝗲 For over a year, we've engaged with Overture about adopting our Places Engine and Placemaker tools to generate open Places datasets collaboratively. While those conversations didn't initially succeed, largely due to steering committee concerns about perceived control loss, today's announcement represents progress. Thanks to Albi Wiedersberg for re-engaging with us. 𝗔 𝗖𝗮𝗹𝗹 𝘁𝗼 𝗔𝗰𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗖𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆 Today's announcement shows progress, but significant work remains. Overture included our dataset but disagreed to add links to our Placemaker tools, a foundational component to how we improve this dataset and enable community participation. We encourage the Overture members and the broader mapping community to advocate for a unified approach that creates a single, authoritative Places dataset maintained through continuous community curation, enabled by our Placemaker tools. Until then, we remain committed to improving our FSQ OS Places dataset. Become a Placemaker today (link in comments) and contribute to our open dataset.

99

8 Comments

AI programming’s worrying stat: 70% of developers trust the output. We need an intervention. New research from DORA confirms AI is ubiquitous. 95% of developers rely on it, and 80% report productivity gains. But there is a massive and dangerous disconnect. That trust is simply not supported by the facts. External research shows AI systems routinely make unsupported or fabricated claims, with failure rates ranging from 30% to 95%. When AIs are overeager and rush ahead with little insight, relying on them without scrutiny is a recipe for disaster. AI can and often will change the entire implementation after a small change, putting us in grave danger of losing the deterministic foundations of programming. We need to treat AI as an imperfect collaborator, not an infallible oracle. The core principles of software engineering are now MORE important, not less: 1. Work in small steps: actively constrain the AI. 2. Verify everything: after every small change, verify that the system is still working as expected. Successful AI adoption takes discipline. If we abdicate our responsibility to accurately specify and verify what we’re building, it is "unlikely to end well".

796

63 Comments

Most AI agent projects fail because teams delay or skip the most critical step: evaluation. Aurimas Griciūnas just shared a brilliant framework that shows why successful agentic systems require evaluation from day one, not as an afterthought. Most devs today are building agents, deploying them, and hoping for the best. But evaluation needs to be a P-0 - AI engineers need evaluation-driven development. Notice how Aurimas's workflow treats evaluation as the foundation, not the final step. You define performance metrics before you build, run evaluations on traces throughout development, and use human feedback to continuously improve the system (like with Galileo's auto-tuning via human feedback). Agentic systems are fundamentally different from traditional software. They make decisions, interact with tools, and produce non-deterministic outputs. Without proper agentic observability and evaluations from the start, you're essentially flying blind. The companies creating multi-agent systems are rapidly understanding this new paradigm. If you want to build AI systems that matter, they need to actually work reliably in production (and be able to prove it with data). Excited to talk evals + agentic observability with Aurimas next week 😎 #AI #AgenticSystems #AIAgents #AIEvaluation

49

4 Comments

From Dev Ittycheria’s callout yesterday from his MongoDB earnings call. If all you’re doing is vector search, you don’t have a knowledge graph. You need a graph DB, a vector DB, a data warehouse (a large SQL), and a super fast SQL engine on the edge (a small SQL). Now we can get conversational 💬 Dev: “DevRev, a well funded AI native platform with proven founders disrupting the help desk market built AgentOS, its complete agentic platform that autonomously handles billions of monthly requests on Atlas. DevRev accelerated development velocity, lowered costs and scaled globally with low latency by using Atlas. Agent OS also leverages Atlas Vector Search for semantic search enriching its knowledge graph and LMs with domain specific content.” Kash Rangan (Goldman Sachs): “it it's super interesting. You were talking about how some of the Silicon Valley AI startup founders don't have it have time to think about databases, but our good friend, Dheeraj at DevRev, seems to have made a wise choice here.” Dev: “Obviously, have so much respect for Dheeraj. He built Nutanix into a real great business, and he's gonna do the same at DevRev we feel good that someone like Dheeraj is betting early on MongoDB because that's a good signal for other founders who are thinking about doing the same…” ______ Thank you both, for the faith. We’re brewing something that will truly make us chatty 🔥

382

10 Comments

One of the most pragmatic CNN papers I have read. Most important conclusion points: 1. SOTA isn't all you need. Higher pre-training accuracy in ImageNet does not mean higher fine-tuning accuracy 2. Attention isn't all you need. Convolutional models strongly outperform transformers in resource-efficient low-data fine-tuning 3. Convolutions are powerful. ConvNeXt architecture consistently outperforms other models when fine-tuning on natural image datasets 4. Small models are good enough.　RegNet and EfficientNet models are excellent choices for fine-tuning across a wide range of image domains. 5. ResNets are getting old. Age of ResNet dominance is over.  6. Most of the top models in every domain retain their higher performance even with less training data

131

4 Comments

🚀 Meet SedonaDB: the first open-source analytical engine built with geospatial as a first class citizen. The Apache Sedona community just released SedonaDB - a single-node analytical database engine that makes spatial data the star, and it is fast. Why this matters: Built in Rust, SedonaDB is lightweight, fast, and spatial first Runs SQL and Python seamlessly with full support for spatial joins, CRS, and indexing Integrates directly with modern formats like GeoParquet, GeoArrow, and GeoPandas Purpose built for small to medium datasets where distributed systems can be overkill And it ships with SpatialBench which is a new benchmark to measure geospatial query performance across engines that goes beyond just a simple spatial join. This is a big leap forward in the geospatial data stack. Curious to hear what you think (or if you already tried it out). Comment SEDONADB and I'll send you the docs to get started - you are just a pip install away. 🌎 I'm Matt and I talk about modern GIS, earth observation, AI, and how geospatial is changing. 📬 Want more like this? Join 9k+ others learning from my newsletter → forrest.nyc

78

10 Comments

Geo-testing is great. But it's not enough. We're obviously big fans of structured incrementality testing – GeoLift, user-level tests, whatever you can run. You're probably not doing enough of these. But geo tests have limitations, and alone are not a substitute for a true incrementality system. First, the insights don't last forever: Here's a problem I’ve seen – a brand runs a lift test for Meta, sees a solid incremental ROI, and decides to scale spend. Two months later? Their creative is different. The algorithm’s changed. A competitor launched a new product. But the brand is continuing to scale, assuming that the incremental ROI is the same. Only once the business starts to miss its target does the brand go back and decide to re-test Meta, and suddenly sees that the incremental ROI isn’t as good anymore, and they’ve been dramatically overspending into an under-performing channel! Second issue: scale. If you're running 10, 15, 20+ paid media channels, you can't test them frequently enough to stay current across all of them. Even a team running one test per month, which is pretty aggressive, only revisits each channel every 12-18 months. That's not fast enough for quarterly planning, much less weekly budget shifts. Third, the results are often fuzzy. Maybe your TV experiment yielded a 2.8x-5.1x ROI. TikTok came back at a 3.5x-6x ROI. What do you actually do with that? Is one really better? Or are they both just pretty good? At Recast, we’ve built an incrementality platform that pulls in experimental results alongside ongoing media mix modeling so you can track performance trends in real time, understand uncertainty, and place smart bets even when the data is noisy. Experiments are essential, absolutely, but you need a system that ties it all together. For more information about Recast, check us out here: https://lnkd.in/e7BKrBf4

48

1 Comment

$400,000 in daily profit was being left on the table by an "optimized" crude oil blending process, and multi-agent orchestration uncovered it. The existing linear programming model had done its job: 20-plus constraints, predictable outputs, no alarms in the control room. Then we built a Composabl multi-agent system to optimize the situation. Same data. Same specs. No rules broken. It returned a new blending process that preserved every parameter and raised profit 15%. That’s not a control problem. It’s a decision problem. 🧠 Composabl orchestrates high-value decisions under constraint—the way seasoned operators adjust levers when the plant drifts. And when the decisions get smarter, the margins do, too. Industrial systems - especially those that require complex decision-making - now require that level of intelligence to make improvements to existing automation. Our live demos show you exactly how to design agents systems that produce these types of results. You can register for the next one here: https://hubs.li/Q03y0_Sn0

50

Nothing’s worst than seeing a code tutorial/demo and the presenter doesn’t share the code. 😡 Unless the code is proprietary, make it a point to share the sample with the audience. And if you’re still struggling to crack the code on how to create engaging technical content, then check out this advice! 📚 Lesson 2: Share the Code Unless you don’t want the audience to give your latest feature, API, MCP server, etc. a try, always strive to share the code that you show on screen. Otherwise, you’re leaving the audience in a state of trying their best to repeatedly pause the video, and copy + paste what they’re seeing on the screen. If you don’t already have yourself a GitHub account, get one today! If you create a lot of code tutorials, then create a template repository (💡 learn more here: https://lnkd.in/gudGG9hn) to both save yourself some time and establish consistency with your audience. Use short links (ex: bit.ly) whenever possible to link to your code samples. Long URLs can sometimes be clunky and are prone to typos. Share your link(s) on screen, in your content’s description (if it’s a video), or in a convenient place within your written content (if it’s a blog post or article). This week, I’m testing out some custom NFC tags that I made for our AI Tour. It’s so much easier to get someone to scan for repo access if they happen to stop me while walking around the halls. I’m all about convenience and most important…getting people access to information! ⭐️ If you found this helpful, then please tag/share this post with another budding technical content creator!

55

6 Comments