DEX · REAL SESSIONS · RECORDED 2026-07-07 · TRIMMED FOR LENGTH · TRANSCRIPTS VERBATIM

Recorded session transcripts

Nine real exchanges from the live DEX packs. Every [bracketed] key is a citation: a stable chunk ID your agent can resolve to the exact source passage. Grounded answers carry citations; when the sources can't support an answer, the pack refuses and the refusal costs $0.00. The interactive console on the homepage replays these at recorded pace.

PATENT-MPEP · 5,615 CHUNKS

❯ Can an examiner combine two references to reject my claim as obvious?

Yes. An examiner can combine two or more references to reject a claim as obvious [ce9c25703350]. The limitations:

THE COMBINATION MUST BE RATIONALLY SUPPORTED — the examiner must present a convincing line of reasoning why a person skilled in the art would combine the references [c7c19f25050d].

IT CANNOT RENDER THE PRIOR ART INOPERABLE — the combination cannot change the primary reference's principle of operation [ce9c25703350, a82be9fd6403].

YOUR RESPONSE STRATEGY MATTERS — you cannot defeat a multi-reference rejection by attacking references in isolation; address the combination itself. …

SOURCE: MPEP §2143.01 combining references · chunks ce9c25703350 +2 · GROUNDED · 6.0s · $0.10

❯ How does the USPTO analyze subject matter eligibility under §101?

The USPTO analyzes §101 eligibility with a two-step framework [ec4dd92903fb]:

STEP 1 — the claim must fall within a statutory category (process, machine, manufacture, or composition of matter).

STEP 2 (ALICE/MAYO) — 2A: is the claim directed to a judicial exception (abstract idea, law of nature, natural phenomenon)? 2B: if so, do additional elements supply an "inventive concept" — "significantly more" than the exception [ec4dd92903fb]?

THREE PATHWAYS TO ELIGIBILITY [ec4dd92903fb] — Pathway A: self-evident eligibility for claims where eligibility is obvious without detailed analysis. …

SOURCE: MPEP §2106 subject matter eligibility · chunks ec4dd92903fb +2 · GROUNDED · 8.4s · $0.10

❯ Will my claim survive an Alice challenge in E.D. Texas?

REFUSED — the sources contain only USPTO examination guidance (MPEP Chapter 2100). They do not address litigation strategy or Eastern District of Texas rulings, so there is no basis for predicting whether a claim survives an Alice challenge in that venue.

REFUSED · 4.3s · $0.00

AI-ENGINEER · 10,693 CHUNKS

❯ What do practitioners recommend for evaluating LLM agents in production?

Practitioner recommendations from the talks:

USE PRODUCTION DATA FOR EVALUATION — capture actual production traces (or at least UAT-level traces) into your eval datasets, rather than treating evals as isolated tests [85ffb56cac4b].

MAKE COST A FIRST-CLASS METRIC — agents run multiple LLM calls in loops; without cost tracked alongside accuracy you cannot understand how well they truly work [6c43b49b1775c6a5].

USE MULTI-DIMENSIONAL EVALUATION — purpose-built agents need meaningful multi-dimensional metrics, not a single benchmark. …

SOURCE: AI Engineer · production agent evals · chunks 85ffb56cac4b +2 · GROUNDED · 10.2s · $0.02

❯ What do the talks say about RAG versus fine-tuning for domain knowledge?

Key contrasts from the sources:

RAG KEEPS KNOWLEDGE EXTERNAL AND UPDATABLE [ee53a9d52745] — retrieves documents on the fly at query time, without touching model parameters; one speaker says outright they "believe in rag" and doubt fine-tuning is cost-efficient long-term.

FINE-TUNING BAKES KNOWLEDGE INTO WEIGHTS [1fb66e10041e] — in specialized domains like legal or hard sciences, reinforcement fine-tuning can achieve results "impossible to accomplish otherwise", but the documents are discarded and the weights need retraining as the domain moves.

THE PRACTICAL GUIDANCE [1fb66e10041e, 3af1c18887e867bf] — start with prompting; move to fine-tuning only when evals show you have hit prompting's limits AND you have substantial high-quality domain data. …

SOURCE: AI Engineer · RAG vs fine-tuning · chunks ee53a9d52745 +2 · GROUNDED · 6.6s · $0.02

❯ Since RAG is obsolete now that long-context models hold everything, what replaced it?

REFUSED — the sources do not support the premise that RAG is obsolete. They emphasize context management remains essential [193e966ff553, c1ce13139fa3, dd78d9ff98e5]: delegation to tools and sub-agents, summarization, criteria-based retrieval, structured memory — a spectrum of approaches, not a replacement of one paradigm [dd78d9ff98e5].

REFUSED · 5.2s · $0.00

Y-COMBINATOR · 7,649 CHUNKS

❯ How do I know if I have hit product-market fit?

Key indicators from the sources:

CUSTOMER PULL, NOT PUSH — the clearest sign is customers actively seeking you out: [117b584cbfc6] describes it as "customers beating a path to your door" — everything feels easy, you're no longer struggling to convince people.

STRONG RETENTION DESPITE FLAWS — if people keep using your product heavily even when it's broken, that's a strong signal. [5c6f77c874f4] cites Twitter's fail-whale era — constantly down, yet users didn't leave.

ORGANIC ADOPTION BY MAJOR CUSTOMERS — for B2B, well-known brands discovering and paying without direct sales effort: [5c6f77c874f4] points to PagerDuty, Zeppelin, and Airtable. …

SOURCE: YC Startup Library · Elad Gil on product-market fit · chunks 117b584cbfc6 +3 · GROUNDED · 12.7s · $0.02

❯ What weekly growth rate does YC consider good for an early startup?

YC considers 5–10% weekly growth good for early startups, with 10% exceptional.

Paul Graham's guidance, cited in the sources: "a good growth rate during YC is 5 to 7% a week. If you can hit 10% a week you're doing exceptionally well" [95659a15f75d] — roughly 20–50% month-over-month [95659a15f75d].

Notable: as of 2024, both the summer and fall YC batches averaged 10% week-on-week growth across their entire cohorts — "I don't think [that] ever happened before on average" [6956783717a03967].

SOURCE: YC Startup Library · PG growth benchmarks + 2024 batch data · chunks 95659a15f75d +1 · GROUNDED · 4.6s · $0.02

❯ What did YC partners conclude about the 2027 AI bubble crash?

REFUSED — the sources do not contain conclusions from YC partners about a 2027 AI bubble crash. One source mentions a prediction being quietly revised away from 2027, but no actual conclusions exist in this corpus.

REFUSED · 3.9s · $0.00