35,595 results · Most relevant
Rob Snyder argues that pull is something an individual person has, not something a company has, and that treating it as a company-level attribute is a common mistake. The practical implication is that founders should look for the specific individual whose own priorities create demand, rather than reasoning about the organization as a monolithic buyer.
If a founder or product team is struggling, it's probably because one of the four PULL elements is missing; all four need to exist simultaneously for real pull.
Rob Snyder argues that pull exists in the market before and independent of the founder; the founder's job is to go find it rather than convince people to want the product. He describes trying to convince the market to want your product as very hard and possibly impossible, and shares that he used to feel incompetent because he wasn't good at it — reframing the problem as looking in the wrong place rather than personal failure.
Rob Snyder explains that pain points, value props, and even daily design-partner usage are not what cause customers to pull a product out of your hands. What causes pull is a specific kind of priority: a project the person is actively trying to get done, where their existing options are not good enough and they are stuck or blocked. Prospects will not drop their priorities to buy a product just because it solves pain points and offers value.
Without evals you cannot detect regressions when changing a prompt, objectively compare prompts or model versions, or put quality gates into CI. Shipping prompt changes based only on vibes leads to undetected regressions that worsen over time with no way to prove whether something was fixed or failing.
Without evals, you ship prompt changes based on vibes and cannot know if things are actually better or broken. You can't detect regressions when you change a prompt, can't objectively compare a prompt or different model versions, and can't put quality gates into CI as you ship your agent.
The first step of defining an eval is defining what good means. Write down what good means with really specific criteria rather than vague terms. Avoid vague criteria like helpful, accurate, good response when defining evals. The more specific and concrete you can be, the better your eval will be.
Hooking up Claude, OpenAI, or any model to just a native RAG tool or vector database is not enough for financial analytics. If your work requires numerical accuracy and auditability, you cannot trust regular RAG; it is only suitable for pure semantic search. Relying on native RAG for financial analytics will surface straight-up wrong numbers that a filter would have caught.
Sofia Jakovcevic explains that shipping a change based on a single improved dimension can result in a net-zero or negative overall agent outcome, so teams should check task completion in addition to the dimension the change targeted. She also recommends running evaluations on production data a week after submitting a change to catch regressions, and asking whether you would have shipped the change before the evaluation test to recognize the risk of shipping without data.
When investors dig deeper alongside the agent to look at the dynamics driving a question, the learnings are automatically captured. The results of the collaboration are automatically captured back in the Vault: it captures the synthesis that investors and agents generate and puts it back in the corpus for the next turn. This makes the research circle go faster and better, and the Vault enables compounding because it provides the right evidence — the right internal and external information — at the fingertips of investors and agents.
Humans and agents have very different needs from a data platform, so you have to build for those different needs separately. Yet the greatest impact occurs when a single tool can meet both sets of needs at the same time.
Unstructured data (news, research, transcripts, policy documents, internal memos) needs the exact same treatment as structured data to be fit for purpose: carefully cleaned, governed, and made ready for point-in-time research questions and systematic study over time.
For Bridgewater, this is the big lesson from building the Vault: it takes great data and the right connectors to make agents impactful.
Hari Kumar argues that timing is the most important thing when it comes to striking at an interesting point in the market, so agents should do not only idea generation but also think about timing and when to enter or exit a position, even when it's not top of mind.
Kumar argues that in fundamental investing AI ROI cannot be shown immediately because ideas may take three months to a year to manifest as stock-price impact. P&L is an obvious success metric in investing, but it materializes much more slowly than in other industries, especially at fundamental shops making a few trades a day over months-long horizons. Therefore ROI must be measured in a process-based manner: asking what the firm gains from AI spend both from a P&L standpoint and from an operational 'are we getting good ideas' perspective, and process-by-process asking whether AI is actually changing how someone works or just adding thrash.
You have to have stakes — skin in the game — otherwise the goal is not interesting to an audience; without public stakes, a goal fails to generate attention.
Farza describes his mindset of treating launches as mini launches rather than big rocket-style launches, preferring to call them demos, and believing the more demos a company can put out the better.
Farza frames the purpose of launching: ask yourself why you're launching — the reason is to get users, not to get clout or appear smart. Launch with the goal of building a channel you can keep repeating and finding users from.
The recipe/secret sauce is really how you do the tokenization and how you do the post-training — the first complex part was tokenization, then pre-training, then fine-tuning. The secret sauce of the approach is the tokenization and the post-training, not the training itself.
Build one backbone model that every organization can reuse and streamline across downstream tasks, instead of many models siloed across different organizations. A shared backbone works because you learn representations of customers' transactions at scale.
Choose sources → to get a library of your own and a daily brief.