CHAPTERS
- 0:19 – 1:51
Why patent law is a uniquely high-value AI domain (between code and legal search)
Ollie Cobb frames patent law as the intersection of two AI strengths: deep technical reasoning (like software) and large-scale document retrieval/synthesis (like legal work). He argues that the high value and high stakes of patent outcomes make the “right” product and workflow choices especially consequential.
- •AI helps software via abstract/technical reasoning; helps legal via searching massive document sets
- •Patent law sits at the intersection of these strengths
- •Because patents are high value, product decisions to unlock AI value become clearer
- •Sets up the talk’s central question: what should the application layer do beyond a general LLM tool?
- 1:51 – 3:23
Patents as a social contract + the four requirements for grant
The talk establishes foundational patent concepts: patents trade public disclosure for a time-limited monopoly. Ollie walks through the legal criteria an invention must meet and why each matters for examination.
- •Patent = disclosure in exchange for ~20-year exclusive rights
- •Four requirements: novelty, non-obviousness, utility, sufficient disclosure
- •“Person having ordinary skill in the art” as a key legal fiction for obviousness/disclosure
- •These requirements shape what must be written and proven in an application
- 3:23 – 4:56
Anatomy of a patent application: claims, spec, drawings, and special formats
Ollie explains the main components of a patent filing and emphasizes that claims define the legal boundary. He highlights the tight coupling between claims and supporting disclosure, plus extra representation requirements in chemistry and biotech.
- •Application includes: claims, description/specification, drawings, abstract
- •Claims are the legal scope; written with strict syntax and fallback scopes
- •Every claim element must be supported in the spec/drawings
- •Chemistry needs structural representations; biotech needs sequence listings
- 4:56 – 5:59
Prosecution: the multi-year negotiation with the patent office
After filing, applicants enter “prosecution,” an iterative exchange with an examiner who reviews claims against prior art. Responses and amendments create a permanent record that later affects interpretation.
- •Examiner issues office actions for novelty/obviousness/disclosure problems
- •Attorney responds by argument and/or narrowing amendments
- •Rounds can continue for years
- •All arguments become part of the file history, affecting later claim interpretation
- 5:59 – 6:29
Enforcement, litigation risk, and why early drafting decisions echo for a decade
Granted patents can be enforced against infringers, but can also be attacked later as invalid. Ollie stresses that drafting and prosecution choices are long-horizon bets whose consequences may surface years later.
- •Infringement requires matching every element of at least one claim
- •Attorneys aim for broad claims with minimal elements—balanced against prior art/support
- •Patents can be challenged in litigation as wrongly granted
- •Drafting/prosecution decisions have delayed, high-stakes consequences
- 6:29 – 8:00
Patent drafting workflow: technical understanding → novelty finding → legal framing
Ollie summarizes the attorney’s end-to-end task: understand a complex invention, assess novelty versus the universe of prior art, and craft claims/spec anticipating future objections. This motivates why AI seems promising here.
- •Must understand the inventor’s technical disclosure
- •Must identify what is truly novel relative to prior art
- •Must draft using patent-specific syntax while planning for prosecution/litigation
- •Problem demands both deep reasoning and ‘needle-in-haystack’ retrieval
- 8:00 – 9:31
Build a patent-specific app or use a general agent? Using software dev as the reference
He introduces the key product question: does the application layer add enough value for patent lawyers, or could they use a domain-agnostic tool like a general coding/agent product? Software development becomes the comparison case for the “delegation” model.
- •Core question: dedicated patent tooling vs general-purpose agent
- •Software dev success story: describe task, delegate implementation to agent
- •Delegation model works well where validation is fast and iteration is cheap
- •Sets up why patents break these assumptions
- 9:31 – 11:03
Why delegation fails (1): you can’t validate patent “correctness” like code
In code, tests and quick QA let users delegate long runs and verify outcomes cheaply. In patents, correctness is unknowable at authoring time because it depends on future examiners, competitors, and litigators, turning choices into risk trades rather than testable facts.
- •Software outputs are testable; patents aren’t runnable
- •Patent quality depends on adversarial future events (2, 5, 10 years out)
- •Decisions trade different risk types based on stakeholders’ risk appetite
- •Delegating without strong validation makes errors harder and costlier
- 11:03 – 13:36
Why delegation fails (2): patent decisions are tightly entangled and reveal late
Software decisions can often be revisited without unraveling everything; in patent drafting, claim scope, terminology, spec support, and drawings constrain each other. Because dependencies emerge as the document forms, attorney judgment must be applied continuously rather than upfront or at final review.
- •Software decisions are often loosely coupled; many are safely revisable
- •Patent claims/spec/drawings have strong mutual dependencies
- •Reframing a top claim can cascade into many downstream rewrites
- •Judgment must be ‘sequenced’ through the drafting process; nothing downstream reliably catches a bad call
- 13:36 – 16:43
Additional mismatch factors: OOD inventions, costly hallucinations, and non-text artifacts
Ollie adds further reasons patents are hard for pure delegation: each invention is by definition novel (out of distribution), hallucinations are harder to detect and more expensive, and key patent artifacts include drawings and domain-specific representations. These factors create room for application-layer innovation in representation and workflow.
- •Software benefits from reusable patterns and RL-friendly optimization
- •Patent tasks involve out-of-distribution inventions by definition
- •Hallucinations are harder to spot and potentially more damaging
- •Patents rely on drawings, chemical structures, and sequences—representation matters
- 16:43 – 22:49
Collaboration over delegation: a new model + three product principles
He proposes a collaboration-first approach: AI should surface the right judgment calls at the right time and then execute after the user decides. He introduces three guiding principles for building such systems, spanning UI/UX and the underlying AI layer.
- •Goal: help attorneys move through dependent, high-stakes decisions quickly
- •AI’s roles: (1) surface tradeoffs/decisions, (2) execute once aligned
- •Collaboration model differs from delegation in product shape and UX
- •Three principles: first-class citations; unify dedicated workflows with a general agent; parallelize alignment then sequence execution
- 22:49 – 25:51
Demo (Principle 1): first-class citations for auditability in prior-art comparison
In the Solve drafting module, Ollie shows an agent comparing an invention disclosure to multiple prior-art documents. The system provides clickable, source-specific citations so users can verify exactly what text informed each conclusion.
- •Project setup: disclosures + inventor form + prior art + application template
- •User asks: “How does my invention compare to the prior art?”
- •Agent summarizes invention and produces comparison table across references
- •Citations link to precise passages in PDFs and uploaded docs for verification
- 25:51 – 27:22
Demo (Principle 2): guided claim-drafting UI backed by the same general-purpose agent
He demonstrates drafting claims via a dedicated interface capturing user preferences (claim count, indentation, labeling, style guidance). The request is still executed through the general agent, preserving consistent reasoning and citation behavior while producing draft claim text.
- •Claim drafting has repeatable structure but many user preferences
- •Dedicated UI captures constraints and style requirements
- •System translates UI inputs into instructions for the general agent
- •Agent drafts around perceived novelty while maintaining legibility and traceability
- 27:22 – 30:54
Demo (Principle 3): parallel review analysis, then user alignment, then consistent edits
Ollie shows application review where multiple criteria are checked in parallel to generate comments, avoiding conflicting edits. After the attorney aligns (dismisses/accepts comments), the agent performs a sequenced editing run to resolve agreed issues consistently, again using the general agent plus tools and citations.
- •Runs multiple sub-reviews in parallel across review criteria
- •Generates comments first to prevent conflicting direct edits
- •Creates a human alignment point before executing document-wide changes
- •Agent then applies sequenced edits and explains changes with citations
- 30:54 – 32:15
Closing synthesis: where AI adds application-layer value in non-delegable domains
He recaps the thesis: the best opportunities are domains where AI is useful and where the app layer adds control, legibility, and workflow structure beyond a general agent. Patents illustrate how collaboration-centric design—supported by citations, unified agents, and alignment/execution sequencing—can unlock value in high-stakes, non-testable work.
- •Pick domains where AI is helpful and the application layer can differentiate
- •Non-delegable domains create opportunity for collaboration-oriented products
- •Legibility/control must be designed at UI/UX and AI-system levels
- •Principles recap and invitation to discuss after the talk
