Token Drop

DaaX.ai
  • 5,0 (1)
  • Tecnologia
  • Semanal

A weekly, unscripted conversation from the DaaX team on the most interesting developments in AI. Sunil Baliga and Sajjad Khazipura (DaaX Co-Founders), along with Sam Pooni (DaaX Architect) and the occasional guest, explore, discuss, and debate new AI research, news, and real-world use cases. Built for developers and business leaders who want perspectives from experienced AI practitioners. All opinions expressed on this podcast are those of the panelists and do not necessarily reflect the views of their employers.

  1. há 1 dia

    Why AI Agents Forget: Graph Databases and the Agent Memory Problem

    AI agents forget, and they hallucinate — and after roughly a trillion dollars of investment, the stack still has no determinism. What's the fix? In this episode of Token Drop, the DaaX teami are joined by Arun Sharma of LadybugDB — a former Linux kernel committer at Facebook and ex-Google engineer — for a technical conversation about graph databases and the agent-memory problem. Arun introduces LadybugDB, an embedded columnar graph database that grew out of the University of Waterloo's KuzuDB project (later acquired by Apple), and explains why no single storage engine can solve agent memory on its own. The discussion covers the engine's DuckDB-influenced architecture and its unusual MMAP design — prompting war stories about dynamic linkers, the ELF format, and 64-bit file support — before turning to why columnar storage beats an LSM for read-dominant graph workloads, and how modular pieces (embedded database, network protocol, load balancer) combine into a distributed system. The group digs into the graph-vector hybrid at the heart of grounded retrieval (why similarity isn't relevance, and how edges disambiguate “dog bit man” from “man bit dog”), how time can be modeled efficiently in a columnar graph without a “time tax,” and the real problem these systems solve: externalizing the knowledge locked in an LLM's weights so a smaller model can query it through deep traversals. Real-world use cases include code knowledge graphs that cut token costs (with an Uber example and Git Nexus) and parsing SEC EDGAR filings, plus how Ladybug scales from a phone to a data lake via Grass Lake and the open IceBug format. Topics covered: graph vs. relational databases; RDF vs. label property graphs; agent memory; MMAP; columnar vs. LSM storage; graph-vector hybrid search; temporal knowledge graphs; and scaling from embedded devices to distributed data lakes. Arun Sharma Linkedin https://www.linkedin.com/in/arundsharma/ LadybugDB https://ladybugdb.com/ Chapters 00:00 — The Token Drop backstory & meet Arun Sharma 02:44 — What is LadybugDB? Graph databases, NoSQL & the relational debate 04:23 — Ladybug Memory: why one storage engine can't solve agent memory 06:42 — Inside the engine: DuckDB's influence and the MMAP surprise 08:31 — MMAP war stories: dynamic linkers, ELF & 64-bit vi 13:12 — Does “embedded” mean your knowledge has to be small? 14:41 — Modular pieces: partitioned tables & a Neo4j (Bolt) wrapper 16:40 — Why columnar instead of RocksDB / LSM 18:16 — What Ladybug Memory actually is (persistent memory for coding agents) 20:39 — A trillion dollars in, still no determinism in the AI stack 21:59 — Real use cases: code knowledge graphs, Uber's token costs & Git Nexus 24:44 — Scaling past the laptop: Grass Lake, the IceBug format & querying from Hugging Face 27:18 — Does adding time blow up a knowledge graph? 29:37 — When vectors and graphs disagree: “dog bit man” vs. “man bit dog” 33:38 — If the user is an LLM, not a human: what do you throw out? 34:27 — The real problem graph databases solve: deep traversals & externalizing LLM knowledge 35:50 — From phones to data lakes: how far LadybugDB scales 37:34 — Wrap-up

  2. 22 de ago.

    Domain Layers, Not Better Models

    Vinod Khosla says raw ChatGPT gets medical triage wrong 20–30% of the time — and that layering a domain system on top of the same model drives the error rate to zero. Half of that is investor shorthand. The other half is the most important architectural argument in enterprise AI right now. Sunil Baliga, Sajjad Khazipura, and Sam Pooni pull the two apart. Why the failure is structural rather than a training gap: the reward function rewards producing an answer, not a correct one, and fluent English isn't backed by provenance. Why RLHF and distillation can't close it — the cognitive surface is too large to cover every domain and every phrasing variant. And why the domain layer, not the model, is the asset that compounds: every builder can rent the same frontier model, so a product that's a model plus a prompt is a margin waiting to be compressed. Also covered: the progression from loop engineering to harness engineering to grounded knowledge, whether error can ever mathematically reach zero, the intent problem (is natural language even the right way to express what a user wants?), dark data in defense and finance, and why flash trading firms have quietly been running neuro-symbolic architectures for years. Full transcript: https://daax.ai/podcast/episode-21-domain-layers-not-better-models Vinod Khosla on YouTube & X as referenced in this episode https://www.youtube.com/shorts/k32VuZjbQls?app=desktop&ra=m https://x.com/vkhosla/status/2036453452641923496 CHAPTERS 0:00 Vinod Khosla's claim: does domain AI take error to zero? 1:41 Why build on top of a frontier model at all? 2:23 A patient, not a benchmark: the diabetic ketoacidosis case 4:05 Why LLMs behave this way 6:15 The reward function rewards answering, not being right 7:17 From loop engineering to harness engineering 8:07 Grounding answers: the outboard knowledge engine 8:51 Deterministic NLP generation as an alternative 9:38 Would better training (RLHF, distillation) fix it? 11:40 The model is a commodity; the domain layer is not 12:28 Why the domain layer is slow to build — and defensible 14:09 Can the error rate ever actually reach zero? 15:49 What LLMs don't capture: experience 17:00 Human-in-the-loop use cases vs. autonomous ones 18:32 The intent problem: is natural language even the right input? 19:37 Natural language vs. domain-specific languages 21:03 Dark data: why defense and finance are different 24:39 Bloomberg's abandoned LLM and neuro-symbolic trading 27:44 Wrap-up: converting silent errors into caught errors

  3. 15 de ago.

    How Knowledge Graphs Handle Time: Event Graphs, Scene Graphs, and Reification Explained

    Time isn't a timestamp you attach to a node — time is change, and most knowledge graphs were never designed to track it. Ontologist and knowledge graph architect Kurt Cagle joins Sunil Baliga, Sajjad Khazipura, and Sam Pooni to walk through the architecture that results when you take time seriously: a declarative knowledge graph for what's constant, an append-only event graph capturing what changed and why, and a scene graph holding the moving "now." Also covered: RDF 1.2 and reification as a way to attach provenance and confidence to any assertion, why information should essentially never be deleted from a graph, and why most enterprise ontology initiatives fail for a reason that has nothing to do with technology — you can't get people to agree on the definition of "customer." Guest: Kurt Cagle — Ontologist, Knowledge Graph Architect, Editor-in-Chief of The Cagle Report https://www.linkedin.com/in/kurtcagle/ Full transcript: https://daax.ai/podcast/episode-20-how-knowledge-graphs-handle-time Chapters (00:00) Introducing Kurt Cagle (01:33) Why time breaks knowledge graphs (03:17) Ontology vs. JSON vs. relational (05:53) Recording every transition (07:24) The event graph (09:40) The scene graph: the moving "now" (11:01) The open world assumption (13:01) Graph vs. graph database (14:30) The six-tuple knowledge unit (16:32) RDF 1.2 and reification (19:57) Is anything ever deleted? (22:29) Newton vs. Einstein: rescoping truth (25:29) The ontology fight in boardrooms (29:12) Why ontology projects fail (31:01) Reasoning is several processes (33:35) Becoming a defensive philosopher (34:41) AI as an epistemological engine

  4. 8 de ago.

    The LLM Is the Reader, Not the Library: Where Should Knowledge Actually Live?

    This episode examines a deceptively simple framing: the LLM is not the library, but the reader. Trained knowledge lives in the model's weights — distributed, entangled, and impossible to audit or edit — while new knowledge must live outside it. The panel traces what that split means for RAG, fine-tuning, knowledge editing, and why hallucination is a structural consequence of how neural networks store facts, not a bug to be patched.

  5. 1 de ago.

    Small Language Models: Great Translators, Terrible Knowledge Stores

    This episode examines the pitch behind domain-specific small language models — and why training one on proprietary data doesn't eliminate the two problems that plague every language model: knowledge changes (forcing costly retraining) and hallucination. The panel argues SLMs earn their place not as knowledge stores, but as narrow, testable translators sitting in front of a knowledge graph.

  6. 25 de jul.

    What Is Knowledge? Why Humans Can Guess but Agents Can’t

    Ranjit Singh Pabla of Stack Overflow joins Token Drop to unpack a deceptively simple question: what is knowledge, and why is serving it to an AI agent fundamentally harder than serving it to a human? The discussion moves from tacit knowledge and memory to trust scores, timestamping, and the frontier of "business as code.” Ranjit Linkedin https://www.linkedin.com/in/ranjitsinghpabla/

  7. 18 de jul.

    Context Engineering vs. Prompt Engineering: What Actually Drives a Good AI Answer

    This episode examines the shift from prompt engineering to context engineering — and why that shift is architectural, not semantic. Prompt engineering optimizes the wording of a single, stateless request; context engineering treats the entire input window — instructions, retrieved knowledge, memory, and tool outputs — as a systematic design problem with measurable tradeoffs in cost, precision, and failure mode.

  8. 11 de jul.

    The Invisible Layer: Why "Connective Tissue" — Not the Model — Is What Kills AI Projects at Scale

    Six months after the demo wowed executives, the AI pilot quietly stalls — and the model is rarely to blame. Theo Forbath, founder of Above the Model and former Accenture MD, joins Token Drop to unpack the invisible "connective tissue" layer most companies underbuild: shared context, working and long-term memory, tool contracts, evaluation frameworks, and observability. The panel walks through a real insurance-claim-adjudication case where hallucinations cascaded across independent agents, and breaks down five sharp questions for spotting a workflow dressed up as "orchestration." They close on a harder question: if AI increasingly handles the routine cases junior specialists used to learn on, who's left with real expertise when the toughest cases get escalated ten years from now — and what does the next generation of AI-fluent talent actually need to know? Theo's Linkedin https://www.linkedin.com/in/theoforbath/ Theo's Substack https://abovethemodel.substack.com

Ver tudo (22)

Sobre

A weekly, unscripted conversation from the DaaX team on the most interesting developments in AI. Sunil Baliga and Sajjad Khazipura (DaaX Co-Founders), along with Sam Pooni (DaaX Architect) and the occasional guest, explore, discuss, and debate new AI research, news, and real-world use cases. Built for developers and business leaders who want perspectives from experienced AI practitioners. All opinions expressed on this podcast are those of the panelists and do not necessarily reflect the views of their employers.

Informações

  • Criado por
    DaaX.ai
  • Anos de atividade
    2026
  • Episódios
    22
  • Classificação
    Livre
  • Copyright
    © ℗2026 DaaX Technologies Inc.
  • Site do podcast
    Token Drop