35,866 results · Most relevant
Workflow patterns like fan-out/fan-in (map-reduce with sub-agents), adversarial review, and loop-until-condition — as catalogued in Claude's Dynamic Workflows blog post — can all be expressed in a Prose program and run on any agent.
Dan Barrett describes that all runs are compared against a contract: think of it as a fresh session every time getting benchmarked against the contract you've written declaratively.
Dan Barrett endorses his own strong belief that language is a really powerful force of creation: you can write a sentence and the model makes that sentence true. The question to ask is what sentence you write now that you have a powerful machine to make it true. Once you can name the thing you want, any harness can self-shape, fan out, and loop to get you there.
Dan Barrett describes that shifting the spectrum from totally just a what to how to do it makes the workflow more deterministic, and the cost comes down as you do that.
Gary Vaynerchuk's principle: the best way to be a creator that pops is to be 100% focused on the audience.
Gary Vaynerchuk's principle: you are the niche. Talk about what you love or what you know. Content has to be you — either what you're really into or what journey you're going through.
Evan Conrad argues that compute hardware is not as fungible as people think, which is why SF Compute spent three years learning that physical settlement is required; building too much financialization before physical settlement creates a market that is easy to manipulate and full of basis risk, where the hedged price may have nothing to do with the actual price of capacity.
Because Monzo operates in a regulated environment, it separates agent non-determinism from the rest of its production infrastructure. Non-determinism is confined to a specific box — the agent running inside the sandbox — where it can perform everything it wants, while the surrounding architecture stays controlled. This creates an environment where agents can run loose and develop while remaining controlled.
With the MCP server giving agents safe data access and the orchestration system handling everything around it, any use cases that come after are much easier to build on top. Monzo connected Slack, GitHub, and Linear to the agent platform, plus more than 3,000 services as building blocks; once Agent Chip became a building block itself, multiple parts of the bank could leverage it, driving not only internal product features but concrete business outcomes.
This is the speakers' endorsed interpretation of why their approach compounded.
Code difficulty no longer functions as a moat in the age of AI. 'I wrote a bunch of code and it took a while' is not a defensible position, and neither is 'our software took 10 years to build' — it won't take anyone else 10 years to rebuild it with AI. If a product has 100 features but customers only use 5, a vibe-coder only needs to replicate those 5 features to replace it.
endorsed
More customers means more learning from customers, which improves the product, and a competitor coding from scratch doesn't get that learning.
Compliance requirements exist on a spectrum: some certifications seem hard to get but are actually easy, while others are genuinely painful to deal with. Organizations dealing with personal health records need HIPAA compliance.
The learning loop is the only mode likely to stay ahead of the expanding LLM capability frontier over time. It is not just a moat: as you build it, it should draw in more customers and build momentum.
Anar Volset describes his method for assessing SaaS defensibility: map out why customers keep paying for software, then evaluate each reason against AI capabilities to identify which moats are eroding and which remain durable. He synthesized commentary from the buy side (sources like Gartner and Morgan Stanley) to understand what investors and acquirers consider still durable.
described
The guests argue that real progress in recursive self-improvement cannot be judged by capability alone: measuring genuine improvement requires accounting for both the numerator (capability gains) and the denominator (the resources — compute, energy, money, and human effort — required to achieve them).
presented as the guests' argument, not an established fact
The speaker states that the defining factor for RSI improvement is what the slowest component in the loop is.
described with 0.85 confidence
Asked how to know what is legit in the sea of AI RSI efforts, the guests' technical answer is that you have to be really, really good at measuring things. They also note it's very, very hard to know what's legit, and that for a long time they had the question of how do we even know if something has taken off if we don't have the right metric — it's a question of what metric you use to view the progress.
presented as the guests' recommendation
Corey Gallon shares a method of agent-based review that catches agent errors before they reach work outputs, allowing the operator to trust agent reports without personally verifying them.
Karan Singhal notes that if you ask five physicians any question, you'll get about eight different answers. When physician consensus points one direction but the user wants comforting words even when the situation is not comforting, the resolution is a prioritization of safety. OpenAI reports dramatic model improvements across the board in these safety-first areas over time alongside improved physician-involved evaluation.
Consequential creator advice requires independent verification before Today can promote it.
Even with identical model weights, latency, throughput, and reliability can vary significantly depending on the inference provider you run them on — differences that are measurable in provider comparison snapshots like OpenRouter's.
Choose sources → to get a library of your own and a daily brief.