Andrea Volpini

🏆 We won the @huggingface Build Small Hackathon! Sangue e Grafi, our soap opera experiment combining SLM with KG, was named Bonus Quest Champion from 600+ submissions. Built with @WordLift it + @NVIDIA Nemotron. An encouraging signal for our work on neuro-symbolic AI. 🩸

@lilyraynyc @dawnieando Have a fantastic birthday Dawn!

Happy birthday @dawnieando!!

@RedCardinal Agreed.

@cyberandy It's noteworthy that alongside this they claim to have the highest search figures ever. Yet they choose to provide no relative distinction for "all" these clicks they are sending from AI. It's such a shame they can't just be honest with publishers. It's all politics now.

🤔

@IntuitMachine For those interested in the topic, here’s a deeper dive: wordlift.io/blog/en/ontolo…

Microsoft open sources the Palantir $PLTR killer. github.com/microsoft/Onto…

Reality is far more complex than a single eval to optimize for. Memory itself is part of the equation. We learn by memorizing what matters, often before it even happened. AI agents grounded in graphs consume up to ~85-87% less compute in real benchmarks.

Didn’t I say this before good ol’ Pete? 😉

A question many people are asking today: Is the AI infrastructure I have in place efficient enough in terms of compute and token costs? Ultimately, semantics is about efficiency at the information-theory level: representing more meaning with less computation cc @TheYotg

Friday test 🧪 A frontier model working alone vs. a smaller model grounded in an ontology. AOOE delivered: → 21 vs. 4 first-pass validations → 22% lower cost per valid draft → 79% fewer repair tokens → 86.5% less avoidable compute linkedin.com/posts/volpini_…

The Semantic Medallion: Building a Knowledge Graph-Powered Data Catalogue. Veronika Heimsbakk Leaders and Innovators Defining the Agenda on Knowledge Graphs, Graph Data Science and AI, Graph Databases, Semantic Technology and Ontologies Veronika is a Knowledge Graph Specialist at Data Treehouse, working on turning unstructured and chaotic data into a wonder of semantic knowledge graphs. Veronika is the author of SHACL for the Practitioner, an international speaker on Knowledge Graphs and Semantic Technologies, and a guest lecturer at the University of Oslo. Highlighting Veronika's work: The Semantic Medallion: Building a Knowledge Graph-Powered Data Catalogue Every data engineer knows the medallion architecture: Bronze (raw), Silver (cleaned), Gold (business-ready). But what if your Gold layer wasn’t just “clean data in nice tables”? What if it were a knowledge graph, where every record understands its relationship to every other record, across all your sources? Traditional data catalogues describe data structurally. A knowledge graph-powered catalogue describes data semantically - what things mean, how they relate, and what you can ask. The relationships are in the data itself, not buried in join logic or SQL scripts. Veronika is a member of the Programme Committee for CDL26. She is helping define the agenda and evaluate contributions for The leading global technology conference for those using Relationships, Meaning, and Context in Data to achieve great things. Join us as we celebrate a decade of innovation in #KnowledgeGraphs, #GraphAnalytics, #DataScience, #AI, #GraphDatabases, #SemanticTech and #Ontology. The Semantic Medallion: Building a Knowledge Graph-Powered Data Catalogue - moderndata101.substack.com/p/the-semantic… Connected Data London 2026 | 11–12 November | Leonardo Royal Hotel London Tower Bridge 🎤 Share your work with the world's most passionate data community. The Call for Submissions is open. connected-data.london/2026-call-for-… 🎟 Tickets on sale now. Advance rate discounts up to 15%. 2026.connected-data.london/?utm_source=tw… 📺 For sponsorship opportunities. Contact info@connected-data.london for details. #GraphRAG #ConnectedData #GraphDB #Analytics #EmergingTec

Every millisecond counts for voice agents. Introducing Gemma 4 31B on LiveKit Inference! Optimized for real-time voice agents: - 354ms time to first audio - 192ms time to first token - Beats GPT-4.1 in agentic tool use (31B achieves 76.9% on tau2bench) Fast, capable, and efficient!

Still not cited by AI search? @AiswaryaMenon built the strategy that took @Hurree from invisible to cited in Google AI, Perplexity & Copilot, in 3 months. She's breaking it down live on Aug 20th. 👉 eu1.hubs.ly/H0w-s350

🙏 Thank you to everyone building SkillOpt with us. We’re approaching 13K GitHub stars—but the number we’re proudest of is >1/3: more than one-third of the code commits we’ve merged now come from external contributors. SkillOpt is becoming a genuinely community-built open-source project. We’re grateful for every PR, issue, experiment, and idea—and excited to advance agent self-optimization together. Our latest technical blog analyzes 499 completed run summaries across 3 benchmarks and 4 target models, organized around three updates: 1️⃣ More studies, clearer knowledge • Gating is not a free performance win. The combined “Accept Every Edit” policy often matched or exceeded the fully gated baseline, especially on gpt-5.5. Because this comparison changes more than the gate alone, we interpret it as a policy-level result—not proof that removing gating is universally better. • Gating is primarily a stability and safety control. Some exploratory policies reached higher best-observed scores but later drifted. Gating may constrain exploration, while helping bound harmful updates. • Best-on-validation is not the same as final. Finding a good skill during optimization does not mean later updates will preserve it. Reliable deployment needs checkpointing and rollback. • Dense feedback helps when exact correctness is too sparse. On SpreadsheetBench, partial test-case pass rates provided useful learning signals that all-or-nothing task success could miss. Soft and hybrid policies helped in several settings, but neither was universally best. • The best update policy depends on the model and task. Stronger targets appeared to tolerate more exploration, while smaller targets were often more sensitive. No single policy dominated every model and benchmark. 2️⃣ Introducing Skill-Aware Reflection Inspired by EmbodiSkill, SAR does not reflect on a trajectory in isolation. It interprets the evidence against what the current skill already contains before deciding the update direction. More reflection memory is not always better. Broad, unconsolidated memory showed large regressions for some smaller targets. In three matched comparisons, consolidation recovered +9.65, +15.00, and +2.32 percentage points relative to the same unconsolidated SAR memory. The key idea is: reflect against the current skill, preserve selective lessons, and consolidate memory before it becomes noise. 3️⃣ SkillOpt-Sleep SkillOpt-Sleep turns agent improvement into an overnight workflow: replay → propose → validate → stage → explicitly adopt Replay, recall, and diverse rollouts showed promising signals where models had headroom. But more compute was not automatically better: in one sensitivity study, adding more rollouts or more nights increased cost without improving the final gain. A deliberate stress case also illustrates why validation matters: an ungated run propagated a plausible but incorrect rule and fell from 55.4% to 2.6%, while a gated run remained flat. This is a downside-control example—not a claim that every ungated run will collapse. The broader lesson: Agent self-improvement needs more than additional reflection or compute. It needs controlled update policies, skill-aware memory, validation, and a safe path to adoption. 📖 Technical blog: microsoft.github.io/SkillOpt/blog/… ⭐ GitHub: github.com/microsoft/Skil…

Advanced US open-weight models like Inkling are a serious valuation overhang for closed-model labs approaching the public markets. Isn't this a rapidly commoditising infrastructure?

Kimi K3 incoming. Most likely today. Btw. banger clip.

The Open Knowledge Format: what Google really shipped, and why the interesting question is not SEO - A must read guide by @gfiorelli1 going through all what you need to know about it as an SEO: * What OKF really is, with the specification open in front of us * The search question, settled by the only source that can settle it * A shared way to store context is not a shared way to understand it * Who fetches the bundle? * Three different things are being called “OKF” * Where this fits And more! Check it out: iloveseo.net/the-open-knowl…

Image tokenization in Claude needs an update 👇👇

Google has introduced a new development: Google-Agent. This is a user agent built specifically for AI agents that browse the web and perform tasks on behalf of users. This represents a shift toward agent-driven search and introduces a new layer of optimization: Agent Search Optimization. ASO builds on the same foundation SEO has always required but adds legibility for machines evaluating your brand on someone else's behalf. Understanding where the web is heading – including emerging standards like WebMCP – is crucial for staying ahead of that curve. social.semrush.com/4wwHZfw.

Poster spotlight: @wordliftit has a sharp AgentCommunity member poster: agentcommunity.org/m/wordlift/pos…