Bucket baseball players into three training categories
Players are bucketed into bullpen guys, position players, and starters, each with distinct training systems; bullpen guys are 'up' (back-to-back), 'down', or in between.
33,353 results · Most relevant
Players are bucketed into bullpen guys, position players, and starters, each with distinct training systems; bullpen guys are 'up' (back-to-back), 'down', or in between.
Set up your test cases, evals, and examples, then do the search process from there. Find the cheapest model that passes most of your evals, then see how you iterate and fine-tune from there. It's also not that hard to just download 10 models from Hugging Face and try them as an alternative to using a provider.
You generally only want to pay for tokens-per-second speed when the model is in front of someone (developers or customers). Background agents running at midnight don't deserve premium speed because no one is waiting for the output.
If you ask an agent to write a poem or code and then ask it if the work is good, it will say yes because it did the work. Splitting generation and evaluation across separate agents (advisor pattern, generator-discriminator gap) gets better results; for massive problems, a generator-judge roster won't scale without separate and distinct checkpoints, and adversarial review should have its own isolated context uninfluenced by previous checks.
The speaker argues that removing one low performer (an 80) raises the team's average more than adding another high performer (a 140), so your bar will rise much faster by eliminating the duds than by adding new hires. He frames getting to a super high-performing team as coming from not letting people stay, and building a better team through elimination — get out everybody who sucks, and what you're left with is awesome. To increase the IQ per square foot, remove everything that is not high IQ. He notes this lesson requires boldness to act on.
When making hard people decisions, the speaker advises asking whether you want local empathy (for the individual) or global empathy (for the business and team). Shifting from local to global means you deal with the local pain while everyone else gets the global benefit; the reverse is that you get the local benefit and everyone else suffers the global pain. He encourages being willing to take the emotional pain of doing what's required to protect the high-performance culture you want to build.
The speaker argues for a slightly easier way in, because judging people from one or two conversations and a document is limited, whereas judging their output is much better. Assuming someone has gone through training, you can just look at their output to evaluate them. He advises shying away from allowing someone to explain away a lack of output, noting that mediocre people are often good at selling their excuses.
Because culture is built from the bottom up, the worst person on the team becomes the bar you have signed off on. One of the easiest ways to improve a team's culture is therefore to remove the people who perform poorly — the speaker frames this as the core mechanism of raising the bar.
The speaker defines the actual action of raising the bar as lowering your tolerance for people doing mediocre work. Practically, he advises not giving nine chances — give a few chances, and after that, that's it.
Harish Gaggar explains that sensitive data (documents, API keys) can be accidentally typed or uploaded into AI systems, so it should be blocked and redacted the moment it is ingested — above roughly 90 percent severity — so secrets never reach LLM models, the memory DB, tracking/eval tables, or tracing systems.
Harish Gaggar recommends a centralized MCP gateway where every tool call request — from a local client or an agent — is authenticated before reaching MCP servers, which then control which tool executes the query.
Extracting plain text from PDF tables formats them badly and can make the agent fetch wrong values (e.g., fees); transforming tables into CSV with spanned fields duplicated made a big difference in answer correctness.
The obvious retrieval move is dense embeddings — turning the question and every chunk into vectors and comparing distance — which is great for meaning. But rulebook codes, message names, and exact identifiers look like noise to a dense embedding, while to the member asking they are often the essence of the question. With BM25 it is the mirror image: exact tokens work but paraphrases would not. The team combined both in one query call and fused them with reciprocal rank fusion, whose nice property is that no score calibration between the two legs is needed. The pipeline takes dense top 50 and BM25 top 50, fuses to top 30, re-ranks for section coherence, and feeds the top 10 to the agent.
Eval questions come from real sources — rounds of user testing with actual members plus behavioral guidelines for what AI systems must and must not do (no advice, stay neutral about competitors, say clearly when something isn't present in the rule books). The team runs about 86 end-to-end tests with promptful: a full agent, an LLM judge against a rubric, a check that the right source document was cited, plus the original retrieval-only evaluation. LLM judge evals have a noise floor — you need to rerun before you trust the delta. Sometimes the rubric itself is the bug, not the model: a hallucinated date the judge kept punishing turned out to be real text sitting in the rulebook.
The section path stored on each chunk is used both to render citations and to fetch an entire section in order (sorted by chunk index) when the agent needs more context.
Dense embeddings capture meaning but treat rulebook codes and exact identifiers as noise, while BM25 handles exact tokens but not paraphrases; the team combines both in one query call, fuses with reciprocal rank fusion (dense top 50, BM25 top 50, fused to top 30, re-ranked for section coherence, top 10 to the agent), and notes RRF needs no score calibration between legs.
A lot of the important knowledge is structured within the rule books' ~3,000 tables, which the team initially did not realize. The initial table processing just fetched plain text from the table in the PDF, which did not format tables well: if table cell spans are not clearly represented in extracted text, the agent can fetch wrong information — especially dangerous with important things like fees. In a dummy example, the domestic field spans multiple rows, cross-border over two more rows, and the refund spans over both. The fix: transform tables into CSV format, duplicating spanned fields so the agent fetches the correct row it needs, which made a big difference for the correctness of the answers.
The core serving fix: pre-compute embeddings only for users and merchants, never per transaction, cutting the cache footprint from 60 GB to 2–3 GB — about 20× smaller and cheap enough to run on the same CPU class as the XGBoost model.
The host says you need friction to let people into your dating life, your company, and your friend circle, and that this friction can simply be asking people what they're about without tipping your hand. He advises letting people show up, show you their track record, and tell you what they're about instead of pitching and overselling. He warns that overselling makes people get excited just because you seem like you've got a mission and they feel like they could use one, and that getting so good at talking people into things that they enthusiastically assent in the moment but regret later creates more problems than you've solved. He describes his own former 'inspiring halftime speech' pitch — 'Here's what we're doing. It's going to be amazing. You should be part of it.' — and says his fake,
Prioritize selling brokerages over individual realtors: go to the head of brokerages (which have 5, 10, 20 agents) and offer a brokerage-level deal — free VIP upgrades for all their agents if you can be the first people they call for a move. One brokerage deal is worth roughly 60 houses a year; 20 agents each selling about a house a quarter equals about five houses a month, which at a $3,000 average is $16,000 a month per brokerage. Bring on a new brokerage every week, and each adds roughly five deals a month as it stacks. Brokerages are the long-term play where reviews compound.
Choose sources → to get a library of your own and a daily brief.