taurin
Junior Member
- Oct 24, 2023
- 182
- 207
- Swix — Opening Keynote & the State of AI Engineering
AI engineering is evolving fast; the spotlight is shifting toward agent engineering.
– We need standard frameworks; Swix suggests SPA (Sync, Plan, Analyze, Deliver, Evaluate).
– Tracking the input ↔ output ratio could be a more practical metric than arguing over what “agent” means.
– Core lesson: don’t over-complicate things.
- Asha Sharma — The Open Agentic Web
Reasoning models are exploding, unlocking new speed and possibilities.
– Envisions an agentic web where agents interact regardless of cloud, company, or device.
– Shifts: pair-programming → peer-programming, software factory → agent factory, cloud models → local/on-device.
– Introduces the “signals loop” (a continuous cycle for agents).
– Foundry platform: model ensembles, agentic RAG (+40 % accuracy on hard queries), tools-as-infrastructure, MCP, evaluation & observability.
- Sarah Goa — An Investor’s View on AI: What Works & What to Build
– AI is the largest tech revolution yet; real uptake proven by metrics (ChatGPT, Copilot).
– Reasoning enables transparent decisions and sequential tasks.
– Agents = software that plans, uses AI, executes tasks, and keeps a goal in mind.
– Start-ups in the agent space are booming and getting traction.
– Other modalities (voice, video, image generation) are maturing; model prices are falling.
– Opportunity: build “Cursor-for-X” apps—code-editor principles for other domains.
– Conservative industries are adopting AI fastest (“AI leapfrog”).
– Execution is the moat; copilots remain the least-friction path today.
- Simon Willis — Six-Month LLM Retrospective
– >30 significant models released in the last six months.
– Personal benchmark: “Pelican on a Bicycle.”
– Good local models exist (e.g., Mistral Small 3); high-quality model prices have crashed.
– GPT-4o shows context interference from memory.
– Notable bugs: overly deferential ChatGPT, “Snitchbench” (models reporting users).
– Tools + reasoning is currently the most powerful AI-engineering technique.
– Risks: prompt injection & the “lethal trifecta” (private data access + malicious instructions + exfiltration path).
- Steven Chin & Andreas Kleger — Agentic Graph RAG
– Stresses quality data and grounding.
– Social responsibility grows as we edge toward AGI.
– Discuss agent memory (e.g., Zep) and the broad concept of Graph RAG.
– Demo: Neo4j LLM Graph Builder—turn unstructured data into knowledge graphs; query them.
– Announced a Neo4j start-up program.
- Theo — The History & Future of MCP
– MCP arose from the need for LLMs to act beyond copy-paste.
– Built as an open, standardized protocol for scale.
– Adoption in tools like Cursor spurred momentum; now backed by major AI labs.
– Principle: optimize for server simplicity.
– Roadmap: elicitation (servers ask users for more info), registry API, open-source samples, governance.
– Call to action: build proper MCP servers (not just API wrappers), improve dev-ex, grow safety & observability tooling.
- John — Scaling MCP Clients at Anthropic
– Tool explosion caused integration chaos that naturally converged on MCP.
– MCP standardizes context delivery to models.
– “Pit of success” model: use an MCP gateway for centralized OAuth, credential management, observability.
– Standard message format centralizes context representation.
- Harold — Hidden Powers of MCP
– Gap between spec and implementations (“API-wrapper syndrome”).
– Full spec support in VS Code Insiders unlocks multi-level, stateful interactions.
– Spec enables dynamic tool discovery, semantic resource links, and sampling (server-initiated LLM calls via client).
– Needs better debugging & logging; spec updates in OAuth & Streamable HTTP are vital.
– Community registry emerging; action-, context-, and semantics-oriented servers are key.
- David Kramer — Challenges & Lessons from Using MCP
– MCP is a plug-in architecture for agents.
– At Sentry, MCP moved error context into the editor.
– OAuth 2.1 is hard but doable.
– MCP ≠ OpenAPI wrapper; design for how models handle context.
– Client support is inconsistent; remote servers + OAuth spec matter for B2B SaaS.
– Output must be model- and human-readable (Markdown > raw JSON).
– Devs don’t control the consumer/model—must account for cost pass-through.
– Lack of streaming tool responses hinders agent-to-agent UX.
- Samuel — MCP Is All You Need (for Autonomous Agents)
– People over-complicate agent-to-agent interaction.
– MCP works for autonomous code-agents too; prompts/resources matter less, tool invocation more.
– MCP vs OpenAPI: dynamic tools, logging, sampling, tracing, stdio subprocesses.
– Sampling is powerful: tools/agents can piggy-back on the parent agent’s model.
– Demo: Pydantic AI agent using BigQuery (MCP server) to generate SQL via sampling, with execution logging & tracing.
- Alex Vulov & Ben Ekl — Observability in MCP
– MCP creates observability blind spots that grow with tool count.
– Enterprises need seamless tracing.
– W&B Weave integrates MCP today, but custom integrations aren’t vendor-neutral.
– Proposes OpenTelemetry (OT) for standardized MCP observability (traces, spans, sinks).
– Demo: trace flow from TS client → Python server via OT & Weave.
– MCP Run will export telemetry to OT-compatible sinks.
– Community standards & conventions are needed.
- Jan Churn — The Rise of the Agent Economy on MCP’s Shoulders
– General intelligence may emerge from many goal-driven agents.
– MCP is the communication backbone for this B2A/A2A agent network.
– Appify marketplace hosts 5 000+ actors, discoverable via MCP.
– Issue: agents need API tokens for third-party services.
– Fix: a centralized marketplace where one token unlocks many MCP services; actors pay for upstream APIs and monetize.
– Demo: Claude Desktop + Appify MCP actors (Twitter, Browserbase) for rapid ecosystem scaling.
– Open questions: trust, autonomous tool discovery, agents & AGI.
- Antje Bar — Building Agents at Cloud Scale
– All customer interactions will be re-imagined with AI; new agent opportunities abound.
– Amazon has >1 000 AI apps in development; Alexa re-imagined with agentic capabilities (600 M+ devices).
– Future: specialized agents working together.
– Strand Agents (Python SDK) launched: built-in Bedrock, multi-provider support, 20+ tools (RAG, multimodal, multi-agent).
– Native MCP support; remote MCP servers can run on Lambda with Streamable HTTP.
– Demo: deploying an MCP server on Lambda & consuming it via Strand Agents.
– Next: agent-to-agent (A2A) interaction; joined MCP steering committee.
– “The atomic unit of every digital interaction will be an agent call.”
- Kevin Hoe — What’s Next for “Identisides” (Agent + IDE)
– Windsurf Editor launched 6 months ago and spread rapidly.
– Secret sauce: shared timeline between human & AI – feels like mind-reading.
– Windsurf must “be everywhere” for context gathering and must act & write everywhere (terminals, browsers, docs).
– Features: terminal commands, live preview, GitHub MCP PRs, async reviews, one-click deploy.
– Goal: 99 % agent work, 1 % human approval.
– Built Suite One (SU1): a model trained on workflows, judged on end-to-end & mid-flow tasks; near-SOTA with lower resources.
– “Data flywheel”: ship → users hit frontier → insights → refine model/tools → repeat.
- Greg Brockman — Fireside Chat: What It Means to Be an AI Engineer
– Love of coding = turning ideas into reality to help people.
– Stripe taught challenging conventions & first-principles speed.
– Effective self-learning needs passion & grit.
– AGI felt possible after reading Turing’s “child machine” and watching deep learning win across domains.
– Engineering is as important as research; harmony scales ideas to impact.
– Engineers need technical humility to keep learning.
– ChatGPT/4o launches proved viral demand & the need to scale fast.
– Vibe coding expands capability; agentic approaches will intersect and take over more.
– CodeX changes coding style toward modular, testable, model-friendly structures.
– Scaling large AI requires checkpointing & state management.
– Future AI infra will diversify for compute vs latency; predicting resource mix is hard.
– Bottleneck for GPT-6: fundamental research rises again—balancing compute, data, algorithms, power, money.