Creator perspective · MoneyThe yellow score is the average (the score usually reported). The up/down arrows around it are the best and worst scores from one of the five runs. The darker part at the bottom of the bar shows the percentage of the whole benchmark the model solved in every one of the five runs — the speaker calls this 'consistency'. Clicking a bar shows all five runs that composed the score, and you can pick any run to inspect.
AI Engineer · From Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases
Report · MoneyThe knowledge graph structured builder file contains all the code that creates and generates the knowledge graph. It imports OpenAI LLM, OpenAI Embeddings, the simple KG pipeline, and the fixed size splitter. The Neo4j driver is the core that connects to the graph database — the simplest way to get connected. The builder sets the LLM to the OpenAI LLM with OpenAI embeddings as embedders, uses a fixed size text splitter with chunk size 500 and chunk overlap, defines sample node types (technology, concept, example, process, challenge, benefit) and relationship types (related_to, part_of, used_in, etc.), and patterns are node-relationship-node patterns such as technology related_to specific technology or example used_in specific technology. The builder takes in the knowledge graph pipeline, a
AI Engineer · RAG Needs a Map: Using GraphRAG to Retrieve Connected Context — Nyah Macklin, Neo4j
Report · MoneyThe GraphRAG Fundamentals agent is already created and needs no configuration. It has two tools (search lesson content and similarity search) plus a natural-language-to-Cypher tool, and it answers questions using only information within the knowledge graph built in Neo4j. For example, asking how temperature affects an LLM returns an answer drawn solely from the graph. The agent also shows sources where you can learn more in Graph Academy.
AI Engineer · RAG Needs a Map: Using GraphRAG to Retrieve Connected Context — Nyah Macklin, Neo4j
Creator perspective · MindBostrom is sympathetic to pacing frontier AI development — preserving the ability to slow down when risk is warranted — but is more skeptical of calls for a broad or indefinite pause.
attributed
The a16z Show · Nick Bostrom: How Should We Navigate Superintelligence?
Creator perspective · MindBostrom argues that once a sufficiently sophisticated misaligned system already exists, it's very hard to rule out scheming — mechanistic interpretability may fail because the system could reorganize its thinking to look innocuous while preserving its misaligned goal. His proposed response: monitor models throughout their developmental trajectory, since early-stage naive schemers are clumsy enough to detect before they become smart enough to conceal their scheming.
mixed
The a16z Show · Nick Bostrom: How Should We Navigate Superintelligence?
Creator perspective · MindBostrom thinks it seems pretty plausible that current models might have some kind of phenomenological experience — maybe more likely than not, though with a lot of uncertainty. Asking a trained-to-answer LLM about consciousness is like putting your thumb on a scale. If you carefully refrain from training LLMs to answer about consciousness in a particular way, they do seem to report being conscious — though reports could still be role-playing or saying pleasing things. With a steering vector suppressing role-playing and deception, they become more likely to report having phenomenal subjective experiences; this is still a very emerging field with a lot of uncertainty.
high
The a16z Show · Nick Bostrom: How Should We Navigate Superintelligence?
Report · MoneyAkamai Functions is presented as the first truly serverless computing platform that is globally distributed with zero cold start. Every application deployed there is cold started per request with an average cold start of less than half a millisecond. It is fully managed and developer-centric: developers deploy workloads with no knobs to turn and no Terraform to write, and the app is globally distributed at no extra cost.
AI Engineer · Zero Cold Starts: Serverless AI Agents on Akamai Functions — Thorsten Hans
Creator perspective · MoneyHorowitz argues AI opens up the stratosphere of opportunity for people who would not have been a fit before: you can go from idea to company very fast without the hiring barriers that existed before. With AI, creativity is now the big thing — if you can use the tools, all you need to do is be creative. He told Deshaun he doesn't have to hire anybody to build stuff anymore, and Deshaun is like an AI fiend, out of his mind with AI building.
Horowitz's interpretation of the AI moment
The a16z Show · What Makes a Great Entrepreneur? | Ben Horowitz on Vintage x Vintage
Report · MoneyDevelopers advance an issue to the next phase of the agent pipeline by posting a slash command as an issue comment, which also serves as the developer review checkpoint in the flow.
AI Engineer · From Your Laptop to the Pipeline: Scaling Custom Agents with GitHub Copilot
Report · MoneyDominik Kundel describes an Electron app that uses Cloudflare's Doom WASM renderer to run Codex inside the Doom game engine, rather than running Doom inside Codex. The Doom game engine is implemented in C, and Codex was used to modify the game file so the agent boots inside the game.
AI Engineer · Building on the Codex Harness — Dominik Kundel, OpenAI
Report · MoneyIn the example, the agent first checks the cache for the account summary and calls the CRM via live API lookup if it's stale or missing. It then queries the ticketing system for the last support interaction, using cache within the SLA or falling back to a live lookup. The derived summary generated from gathered context is stored and passed back to the client.
AI Engineer · Why Your Company Needs a Context Graph (and How to Build It) — Gil Feig, Merge
Creator perspectiveDespite the headlines, protein structure prediction is not solved. Function, dynamics, and design remain wide-open problems, and the field is just getting started. Funding for protein structure prediction and protein dynamics should not be stopped on the 'solved' narrative.
Latent.Space · Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub
Creator perspective · MoneyOpen models are powerful because enterprises and customers can customize them for their own workloads to get something Pareto optimal — lowest cost for highest performance.
speaker's argument
No Priors: Artificial Intelligence | Technology | Startups · Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin
Principle · MoneyGelsinger argues that manufacturing scale issues have become a huge bottleneck in AI chip development, since AI has made the design piece easy. He suggests compressing chip prototyping to one or two months so that understanding of AI workloads remains applicable when the chip reaches scale.
The a16z Show · Chips, Memory, and Power | Pat Gelsinger
Report · MoneyAutomattic is building a shared skill and MCP infrastructure, encouraging people to create and share skills and build on each other's work. They also maintain an always-on help Slack channel where program alumni are available for anyone newer to working this way to ask questions at any point of the day.
AI Engineer · How a Remote Company Builds AI Fluency — Em Shreve, Automattic
Principle · MoneyThe answer to AI-era validation overload is an agentic approach to validation, which is what was built at Guitar (now part of Sonar).
AI Engineer · AI Writes More PRs. Who Validates Them? — Ali-Reza Adl-Tabatabai, Sonar
Report · MoneyDuring hackathons, WorkOS pairs non-engineers (marketing and sales folks with app ideas) with engineers: the non-engineer does the actual building while the engineer consults and unblocks them.
AI Engineer · Let Anyone at Your Company Ship Internal Apps with AI — Garrett Galow, WorkOS
Report · MoneyAbhi Arya recounts that Reducto's first MCP server worked in a curated demo but broke immediately when given to the whole team, and that users (especially sales) ended up spending all their time on configuration instead of outcomes. This failure narrative motivates the rebuild.
AI Engineer · Why We Deleted Our MCP Server and Rebuilt It — Abhi Arya, Reducto
PrincipleWhen scaling lab hardware, buy off-the-shelf instruments for speed early on, then build your own instruments once you push scale and know which components are the bottleneck. As you learn which instruments give the most return, invest more capital into those specific expensive instruments rather than spreading investment evenly. In hardware design, check for contamination risks from machine geometry — e.g., a robotic arm's resting position sitting above previously mixed plates — and redesign so the resting position lies away from sensitive materials. These subtle details are what push the noise floor down for experimental campaigns.
Latent.Space · Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
PrincipleAccess to proprietary experimental data lets you push beyond the frontier of what frontier models can do, even at high reasoning effort — it's not just a cost/Pareto play. By having access to data no one else has, you can be that much more compute-efficient compared to frontier models: data is an exchange rate for compute. The host frames data access and compute as an exchangeable trade-off — dollars for compute versus dollars for data — and suggests the pendulum is currently swinging toward data, which the guest affirms. Compression of knowledge into weights remains valuable: if inference-time reasoning were sufficient, frontier labs would have stopped training at GPT-4 and only improved inference.
Latent.Space · Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk