Latest / Silicon & Shale Podcast / TotalEnergies x Mistral: Frontier AI for Reservoir Exploration

Transcript

TotalEnergies x Mistral: Frontier AI for Reservoir Exploration · Silicon & Shale Podcast

• 0:00
  SPEAKER_01: So I'm at this vendor breakfast Thursday, right?
• 0:02
  SPEAKER_01: And the slide behind the oatmeal says,
• 0:03
  SPEAKER_01: our agentic subsurface twin finds the sweet spot
• 0:05
  SPEAKER_01: before your geologist finishes coffee.
• 0:07
  SPEAKER_00: Before coffee.
• 0:08
  SPEAKER_01: Before coffee.
• 0:10
  SPEAKER_01: And the guy presenting it has never sat a well review in his life.
• 0:13
  SPEAKER_01: I asked what data the twin trains on.
• 0:15
  SPEAKER_01: He said all of it.
• 0:16
  SPEAKER_00: All of it.
• 0:17
  SPEAKER_00: The most expensive three words in E&P.
• 0:19
  SPEAKER_01: I wrote them down.
• 0:20
  SPEAKER_01: I almost believe the next paragraph.
• 0:21
  SPEAKER_01: That's on me.
• 0:22
  SPEAKER_00: Hold that thought.
• 0:23
  SPEAKER_00: Because this week,
• 0:24
  SPEAKER_00: an actual IOC put more than 100 million euros behind a version of that sentence.
• 0:28
  SPEAKER_00: Except they named the partner, the lab, and the job.
• 0:31
  SPEAKER_01: Total energies and mistrawl.
• 0:33
  SPEAKER_01: September 15th.
• 0:34
  SPEAKER_01: Frontier models for reservoirs, not a chat window over PDFs.
• 0:37
  SPEAKER_00: Exactly.
• 0:38
  SPEAKER_00: Welcome to Silicon & Shale, episode 16.
• 0:40
  SPEAKER_00: I'm Packer.
• 0:41
  SPEAKER_01: Derek, oil patch meets the algorithm.
• 0:43
  SPEAKER_01: And this week, the algorithm has a European accent and a joint lab badge.
• 0:46
  SPEAKER_00: Last week,
• 0:47
  SPEAKER_00: we were in Venezuela's data mess,
• 0:48
  SPEAKER_00: Bangert's chief of AI seat, and GPUs starting to trade like a commodity.
• 0:52
  SPEAKER_00: Different movie.
• 0:53
  SPEAKER_01: Different movie.
• 0:54
  SPEAKER_01: Titles and futures desks don't generate development scenarios.
• 0:57
  SPEAKER_01: This deal is trying to.
• 0:59
  SPEAKER_00: So, let's walk the press release like adults.
• 1:01
  SPEAKER_00: Paris, September 15, 2026.
• 1:03
  SPEAKER_00: Total Energies and Mistral announce a three-year joint program.
• 1:06
  SPEAKER_00: Investment, more than 100 million euros.
• 1:09
  SPEAKER_01: Not up to, more than.
• 1:11
  SPEAKER_01: That's the kind of floor that makes a CFO flinch in a good way.
• 1:14
  SPEAKER_00: The stated aim is a new generation of frontier AI models,
• 1:16
  SPEAKER_00: including Agentic AI,
• 1:18
  SPEAKER_00: to support Total's geoscience people on exploration,
• 1:20
  SPEAKER_00: characterization, and development of oil and gas reservoirs.
• 1:23
  SPEAKER_01: Support, not replace.
• 1:25
  SPEAKER_01: Sticky note that for every mid-cap that just got a deck titled Autonomous Reservoir.
• 1:28
  SPEAKER_00: They say they'll combine nearly 10 petaflops of data with nearly a century
• 1:32
  SPEAKER_00: of total energy's knowledge in geosciences and reservoir engineering.
• 1:35
  SPEAKER_01: Century of scars, 10 petaflops of whatever survived the migrations.
• 1:39
  SPEAKER_01: That's the honest inventory.
• 1:40
  SPEAKER_00: Objective language from the release.
• 1:42
  SPEAKER_00: Models that integrate,
• 1:43
  SPEAKER_00: process,
• 1:43
  SPEAKER_00: and interpret large volumes of data to generate multiple scenarios,
• 1:46
  SPEAKER_00: for exploration opportunities,
• 1:48
  SPEAKER_00: and for optimizing and extending the life of existing projects.
• 1:51
  SPEAKER_01: Multiple scenarios.
• 1:52
  SPEAKER_01: That's the tell.
• 1:53
  SPEAKER_01: They're not selling one magic answer.
• 1:55
  SPEAKER_01: They're selling a scenario factory with humans still in the loop.
• 1:58
  SPEAKER_00: They also stand up a joint scientific laboratory,
• 2:00
  SPEAKER_00: Total subsurface teams plus Mistral's science and tech people,
• 2:03
  SPEAKER_00: to build solutions for Total's businesses.
• 2:05
  SPEAKER_01: Joint lab means desks in the same building, not a slack channel and a hope.
• 2:09
  SPEAKER_00: And the framing is proprietary models inside a European digital ecosystem.
• 2:12
  SPEAKER_00: Total keeps strategic subsurface data close.
• 2:15
  SPEAKER_00: Mistral gets to grow as a European AI champion that respects IP.
• 2:18
  SPEAKER_01: Operator-side IP, say it plain.
• 2:20
  SPEAKER_01: The barrels and the segi don't leave to train somebody else's public demo.
• 2:24
  SPEAKER_00: Patrick Puyane's quote is blunt.
• 2:26
  SPEAKER_00: Exploration and reservoir engineering are where AI can create the greatest value,
• 2:29
  SPEAKER_00: and they want tools that help experts analyze complex data and decide.
• 2:33
  SPEAKER_01: Help experts decide, not decide for them.
• 2:35
  SPEAKER_01: Puyane still sounds like someone who's sat through a dry hole review.
• 2:38
  SPEAKER_00: Arthur Mensch from Mistral leans on customizable AI
• 2:41
  SPEAKER_00: that protects intellectual property and handles hard industrial science.
• 2:44
  SPEAKER_01: IP protection is the sales feature operators actually buy, the rest is scenery.
• 2:48
  SPEAKER_00: Distinction 1,
• 2:49
  SPEAKER_00: before we get cute,
• 2:50
  SPEAKER_00: episode 15 was Venezuela data that barely holds together,
• 2:54
  SPEAKER_00: Bangert as chief of AI, org scaffolding, and GPUs looking like a commodity.
• 2:59
  SPEAKER_01: This is not a commodity ticker and not a new C-suite nameplate.
• 3:02
  SPEAKER_01: This is an IOC writing a nine-figure check into frontier models aimed at reservoirs.
• 3:06
  SPEAKER_00: Distinction 2,
• 3:07
  SPEAKER_00: earlier this season we chewed on Petrona-style investment platform framing,
• 3:11
  SPEAKER_00: discovery to first oil clocks on a platform story.
• 3:14
  SPEAKER_01: Different animal.
• 3:15
  SPEAKER_01: Total and Mistral are talking frontier models on geoscience data and a joint lab,
• 3:19
  SPEAKER_01: not an investment marketplace logo.
• 3:21
  SPEAKER_00: Rails versus brains.
• 3:22
  SPEAKER_00: Both can matter.
• 3:23
  SPEAKER_00: They're not the same purchase.
• 3:25
  SPEAKER_01: If a salesperson conflates them, hang up politely.
• 3:28
  SPEAKER_00: Now the uncomfortable part?
• 3:29
  SPEAKER_00: I need to confess something about us.
• 3:31
  SPEAKER_01: Oh no, the quiet part.
• 3:33
  SPEAKER_00: Last spring,
• 3:33
  SPEAKER_00: I sat through a deck that said their model would just read the Seagy
• 3:36
  SPEAKER_00: and propose Wells.
• 3:37
  SPEAKER_00: I nodded like a professional.
• 3:39
  SPEAKER_00: I wanted it to be true because it would make A&D committees faster.
• 3:42
  SPEAKER_01: You believed the fairy tale?
• 3:43
  SPEAKER_00: I did.
• 3:44
  SPEAKER_00: You named it later in the parking lot.
• 3:46
  SPEAKER_00: We were shopping for a shortcut around interpretation judgment.
• 3:49
  SPEAKER_01: Once, naming it once is enough.
• 3:51
  SPEAKER_01: Then we do the work.
• 3:52
  SPEAKER_00: Then we do the work.
• 3:53
  SPEAKER_00: What does this total misdraw stack actually change for someone running A&D or Reservoir?
• 3:57
  SPEAKER_01: If it works as advertised,
• 3:58
  SPEAKER_01: you get more coherent scenario sets earlier,
• 4:01
  SPEAKER_01: not a single AI pick,
• 4:02
  SPEAKER_01: but branches you can stress,
• 4:04
  SPEAKER_01: volumes, timing, facility choices, how hard you lean on an existing field.
• 4:08
  SPEAKER_00: Scenario generation is only valuable if someone with a PE stamp still kills the dumb ones.
• 4:12
  SPEAKER_01: Hardgate, agentic loops can propose they don't sign AFEs.
• 4:15
  SPEAKER_00: Let's slow down on frontier model,
• 4:17
  SPEAKER_00: because that phrase is doing a lot of marketing work in this release.
• 4:19
  SPEAKER_01: In plain English,
• 4:20
  SPEAKER_01: they mean a big general model adapted hard to total subsurface world,
• 4:23
  SPEAKER_01: not a thin classifier that only flags one curve shape.
• 4:26
  SPEAKER_00: And agentic here means the software can chain steps,
• 4:29
  SPEAKER_00: pull a class of data,
• 4:30
  SPEAKER_00: propose a correlation,
• 4:31
  SPEAKER_00: draft a scenario set, maybe iterate when something doesn't fit.
• 4:34
  SPEAKER_01: Chain is not authority.
• 4:35
  SPEAKER_01: Chain is a junior analyst that never sleeps and never gets embarrassed.
• 4:38
  SPEAKER_01: You still review the work.
• 4:39
  SPEAKER_00: Midsize operator watching this from Midland or Oklahoma City.
• 4:43
  SPEAKER_00: What's the usable takeaway?
• 4:44
  SPEAKER_00: You're not writing a 100 million euro joint lab check next Tuesday.
• 4:47
  SPEAKER_01: Takeaway is the pattern.
• 4:48
  SPEAKER_01: Big operators are done pretending a chatbot on SharePoint is subsurface AI.
• 4:52
  SPEAKER_01: They're buying model stacks that sit on proprietary data with contractual IP walls.
• 4:56
  SPEAKER_00: So when a vendor shows up with,
• 4:58
  SPEAKER_00: we partnered with a frontier lab on slide two, you ask three questions.
• 5:02
  SPEAKER_00: Whose data?
• 5:03
  SPEAKER_00: Whose weights?
• 5:04
  SPEAKER_00: Who can walk out with what?
• 5:05
  SPEAKER_01: And fourth, show me where the engineer rejects a scenario.
• 5:08
  SPEAKER_01: If that button doesn't exist, you're not looking at a decision tool.
• 5:11
  SPEAKER_01: You're looking at a vibe.
• 5:12
  SPEAKER_00: Pouyane also ties this to European digital competitiveness.
• 5:15
  SPEAKER_00: We don't need a Brussels seminar.
• 5:17
  SPEAKER_00: Just notice the procurement logic.
• 5:19
  SPEAKER_00: They want capability that doesn't force strategic subsurface into a U.S.
• 5:22
  SPEAKER_00: cloud training free-for-all.
• 5:23
  SPEAKER_01: Sovereignty as a procurement checkbox.
• 5:25
  SPEAKER_01: Operators understand checkboxes.
• 5:27
  SPEAKER_00: Let's talk the brownfield clause, because exploration headlines always steal the ink.
• 5:31
  SPEAKER_01: Extending life of existing projects is in the release on purpose.
• 5:34
  SPEAKER_01: That's rent money,
• 5:35
  SPEAKER_01: base production,
• 5:36
  SPEAKER_01: workovers, how hard you push a water flood, whether a sidetrack earns its keep.
• 5:40
  SPEAKER_00: If a scenario engine is honest,
• 5:42
  SPEAKER_00: it should hurt your feelings about a pet development case you've loved since 2019.
• 5:45
  SPEAKER_01: Especially when the production history is uglier than the type curve slide.
• 5:49
  SPEAKER_01: Which is most of the time.
• 5:50
  SPEAKER_01: Which is most of the time.
• 5:52
  SPEAKER_00: Walk a concrete brownfield case.
• 5:54
  SPEAKER_00: You've got a 20-year water flood,
• 5:55
  SPEAKER_00: uneven pattern, a few wells you've always treated as character.
• 5:59
  SPEAKER_01: Character is what we call wells when we don't want to admit the completion history is fiction.
• 6:03
  SPEAKER_00: A useful scenario tool would force those wells into the open.
• 6:06
  SPEAKER_00: Show how sensitive NPV is if those headers are wrong.
• 6:09
  SPEAKER_01: And it would offer cases that cut rates,
• 6:10
  SPEAKER_01: re-drill selectively, or stop dumping money into the romance wells.
• 6:14
  SPEAKER_00: Romance wells are expensive hobbies.
• 6:16
  SPEAKER_01: We've all had one.
• 6:17
  SPEAKER_00: Failure modes, because press releases don't list them.
• 6:20
  SPEAKER_01: Failure mode 1.
• 6:21
  SPEAKER_01: The lab ship's beautiful notebooks and the asset teams keep using last decade's workflows.
• 6:25
  SPEAKER_01: Bangert problem without the title.
• 6:27
  SPEAKER_01: Capability that never touches the morning meeting.
• 6:29
  SPEAKER_00: Failure Mode 2.
• 6:31
  SPEAKER_00: Scenarios look diverse but share the same blind spot in the training data.
• 6:34
  SPEAKER_00: Old interpretations recycled as new confidence.
• 6:37
  SPEAKER_01: Failure Mode 3.
• 6:38
  SPEAKER_01: People stop calling counterparts because the dashboard felt 95% sure.
• 6:42
  SPEAKER_00: High confidence is when you look harder, not less.
• 6:45
  SPEAKER_00: We said that about classifiers before.
• 6:46
  SPEAKER_00: Same rule applies to reservoir scenarios.
• 6:48
  SPEAKER_01: Failure Mode 4.
• 6:50
  SPEAKER_01: It owns the lab calendar and subsurface owns the wells,
• 6:52
  SPEAKER_01: and they only meet at quarterly steering.
• 6:55
  SPEAKER_01: That's how joint labs become two slide clubs.
• 6:57
  SPEAKER_00: Put the subsurface people and the model people on shared milestones with
• 7:00
  SPEAKER_00: the same definition of done, or it dies of polite meetings.
• 7:02
  SPEAKER_01: Failure mode 5.
• 7:04
  SPEAKER_01: The pilot runs on the cleanest field in the portfolio,
• 7:06
  SPEAKER_01: then dies on contact with the messy assets that actually need help.
• 7:09
  SPEAKER_00: Demo field bias.
• 7:10
  SPEAKER_00: Classic.
• 7:12
  SPEAKER_01: If it can't survive the ugly asset, it can't survive Tuesday.
• 7:15
  SPEAKER_00: What would done look like in year one if we're being unromantic?
• 7:18
  SPEAKER_01: A tool and asset team actually opens on a Tuesday without a vendor in the room.
• 7:21
  SPEAKER_00: Scenario sets that show up in a decision memo.
• 7:23
  SPEAKER_00: Not a side PDF nobody attaches.
• 7:25
  SPEAKER_01: A documented reject path and a postmortem when a scenario set was wrong.
• 7:29
  SPEAKER_00: If there's no postmortem culture, the model becomes a blame sink or a trophy.
• 7:33
  SPEAKER_00: Neither helps.
• 7:34
  SPEAKER_01: Trophies look great in the lobby.
• 7:35
  SPEAKER_01: They don't move Eeyore per BOE.
• 7:37
  SPEAKER_00: Map a Monday morning.
• 7:38
  SPEAKER_00: You're a reservoir Aang with three development cases due Friday.
• 7:41
  SPEAKER_01: Old world.
• 7:42
  SPEAKER_01: You and a peer grind cases from the same tired assumptions because the clock is mean.
• 7:46
  SPEAKER_00: New world if this class of tool works.
• 7:48
  SPEAKER_00: You get a wider case set overnight with flags where inputs contradict each other.
• 7:51
  SPEAKER_00: Missing tests.
• 7:52
  SPEAKER_00: Inconsistent tops.
• 7:53
  SPEAKER_00: Production that doesn't match the completion story.
• 7:55
  SPEAKER_01: Then you spend your week killing bad cases and sharpening the two that survive.
• 7:58
  SPEAKER_01: That's leverage, not magic.
• 8:00
  SPEAKER_00: And your manager still asks where the analogs are.
• 8:03
  SPEAKER_00: Analogs don't retire because a French lab showed up.
• 8:05
  SPEAKER_01: Analogs are how we stay honest when the model is confident.
• 8:08
  SPEAKER_00: A&D angle.
• 8:09
  SPEAKER_00: You're screening a package.
• 8:11
  SPEAKER_00: Seller's data room is complete the way data rooms are always complete.
• 8:15
  SPEAKER_01: Which is to say, complete until you ask about the wells that made them uncomfortable.
• 8:18
  SPEAKER_00: A scenario engine trained on the buyer's own history might stress the seller's story
• 8:22
  SPEAKER_00: against patterns you've seen elsewhere in your portfolio.
• 8:24
  SPEAKER_01: Might.
• 8:25
  SPEAKER_01: It can also amplify whatever bias you already baked into your corporate lookback.
• 8:29
  SPEAKER_00: So you treat it like a hostile witness you hired.
• 8:31
  SPEAKER_00: Useful.
• 8:32
  SPEAKER_00: Not trusted on charisma.
• 8:33
  SPEAKER_01: Hostile witness on payroll.
• 8:34
  SPEAKER_01: I'll take that.
• 8:36
  SPEAKER_00: For land and commercial folks listening, this doesn't replace title or contract work.
• 8:39
  SPEAKER_00: It might change how fast Subsurface can tell you whether a package is even worth
• 8:42
  SPEAKER_00: the legal hours.
• 8:43
  SPEAKER_01: Saving legal hours by killing dead packages early is real money.
• 8:47
  SPEAKER_00: Exploration side?
• 8:48
  SPEAKER_00: Careful.
• 8:49
  SPEAKER_00: The release says generate multiple scenarios for exploration opportunities.
• 8:52
  SPEAKER_00: That is not the AI found a field.
• 8:54
  SPEAKER_01: It's here are coherent ways the data could be telling a story,
• 8:57
  SPEAKER_01: ranked by assumptions you can see.
• 8:59
  SPEAKER_00: If the assumptions aren't visible, it's astrology with GPUs.
• 9:02
  SPEAKER_01: Astrology with GPUs needs to be on a mug.
• 9:04
  SPEAKER_01: No, scratch that.
• 9:05
  SPEAKER_01: We are not making merch of our sins.
• 9:07
  SPEAKER_00: Agreed.
• 9:07
  SPEAKER_01: No merch.
• 9:08
  SPEAKER_01: Exploration committees already drown in cases.
• 9:11
  SPEAKER_01: More cases only help if they come with clearer assumption dials.
• 9:14
  SPEAKER_00: Dials you can turn in front of a room without a vendor translating.
• 9:16
  SPEAKER_00: How does this sit next to last week's GPU as commodity thread without rehashing it?
• 9:20
  SPEAKER_01: Clean handoff.
• 9:22
  SPEAKER_01: Last week was about renting flops and hedging them.
• 9:24
  SPEAKER_01: This week is about whether an operator spends serious money to own judgment fused to its data.
• 9:29
  SPEAKER_00: Renting Flops gives you a bill.
• 9:30
  SPEAKER_00: This deal tries to buy a capability that compounds.
• 9:33
  SPEAKER_01: And Bangert's seat was about org authority for AI.
• 9:36
  SPEAKER_01: Total's move is about a funded technical object, models, lab, IP, not a title alone.
• 9:41
  SPEAKER_00: Titles without budgets are press.
• 9:43
  SPEAKER_00: Budgets without workflows are souvenirs.
• 9:45
  SPEAKER_00: They're attempting both budget and workflow with the joint lab language.
• 9:48
  SPEAKER_01: Attempting, we'll know in three years.
• 9:50
  SPEAKER_00: Let's do a quick procurement workshop for the mid-caps because that's where
• 9:52
  SPEAKER_00: the inbox pain lives.
• 9:53
  SPEAKER_01: Vendor email subject line, Frontier AI for your reservoirs, partnered with a leading lab.
• 9:58
  SPEAKER_00: Step 1.
• 10:00
  SPEAKER_00: Ask for the data flow diagram, not the marketing one.
• 10:03
  SPEAKER_00: The one that shows where your CG and production histories go,
• 10:06
  SPEAKER_00: who can access them, and whether they're used to train shared weights.
• 10:09
  SPEAKER_01: If they stall on that diagram, you're done.
• 10:12
  SPEAKER_00: Step 2.
• 10:13
  SPEAKER_00: Ask who owns fine-tuned weights after the pilot.
• 10:15
  SPEAKER_00: You, them, or it's complicated.
• 10:18
  SPEAKER_01: Complicated usually means them.
• 10:20
  SPEAKER_00: Step 3.
• 10:21
  SPEAKER_00: Ask for a Reject Path demo.
• 10:22
  SPEAKER_00: Literally, watch an engineer kill a scenario and see what gets logged.
• 10:26
  SPEAKER_01: If killing a scenario requires a support ticket, that's not a tool.
• 10:29
  SPEAKER_01: That's a hostage situation.
• 10:30
  SPEAKER_00: Step 4.
• 10:31
  SPEAKER_00: Ask which ugly asset they'll run on first.
• 10:34
  SPEAKER_00: If they only want your cleanest field, you're paying for theater.
• 10:36
  SPEAKER_01: Theater has a high burn rate.
• 10:38
  SPEAKER_01: I'll admit another small miss.
• 10:40
  SPEAKER_01: I used to treat European AI stack as branding fog.
• 10:43
  SPEAKER_01: Watching operators put subsurface IP behind it made me update.
• 10:46
  SPEAKER_00: Updated beliefs count.
• 10:48
  SPEAKER_00: We don't need a parade.
• 10:49
  SPEAKER_01: No parade.
• 10:50
  SPEAKER_01: Just don't let me slide back into ignoring procurement constraints when a U.S.
• 10:53
  SPEAKER_01: vendor demo looks prettier.
• 10:55
  SPEAKER_00: Deal.
• 10:56
  SPEAKER_00: Pretty demos on clean synthetic fields are how we get into trouble.
• 10:58
  SPEAKER_01: Show me the messy carbonate with missing production tests.
• 11:01
  SPEAKER_01: Show me the brownfield where the well headers lie.
• 11:03
  SPEAKER_00: If it can't survive that, it can't survive Tuesday.
• 11:05
  SPEAKER_00: What about smaller AI for reservoirs pitches landing this week in inboxes?
• 11:09
  SPEAKER_01: Borrow total seriousness, not their budget.
• 11:12
  SPEAKER_01: Checklist,
• 11:13
  SPEAKER_01: named partner,
• 11:13
  SPEAKER_01: named dollar floor or clear compute commitment,
• 11:16
  SPEAKER_01: named technical object, named governance, named IP posture.
• 11:19
  SPEAKER_00: If a pitch is missing three of those five, you're being entertained.
• 11:22
  SPEAKER_01: Entertained is fine on a Thursday night.
• 11:23
  SPEAKER_01: It's expensive on an AFE.
• 11:25
  SPEAKER_00: Speaking of expensive, more than 100 million euros across three years is serious.
• 11:29
  SPEAKER_00: But it's not AI replaces exploration money.
• 11:32
  SPEAKER_00: It's build-a-capability money.
• 11:34
  SPEAKER_01: Capability spend looks boring in a keynote.
• 11:36
  SPEAKER_01: It's the only spend that changes who owns the advantage five years out.
• 11:40
  SPEAKER_00: Mid-caps won't match it dollar for dollar.
• 11:41
  SPEAKER_00: They can still refuse tools that hide the reject button.
• 11:44
  SPEAKER_01: Refuse tools that won't say who owns the weights after the pilot.
• 11:46
  SPEAKER_00: Refuse tools that need a data engineer living in your office just to keep
• 11:49
  SPEAKER_00: the dashboard alive, unless you're hiring that person on purpose.
• 11:52
  SPEAKER_01: On purpose is fine.
• 11:54
  SPEAKER_01: Accidental headcount is how AI projects become IT tax.
• 11:57
  SPEAKER_00: Let's talk people.
• 11:58
  SPEAKER_00: Puya Neh's line about helping experts.
• 12:01
  SPEAKER_00: If the experts aren't in the lab design,
• 12:02
  SPEAKER_00: you get a product that answers questions nobody asks in the morning meeting.
• 12:05
  SPEAKER_01: Morning meeting questions are ugly and specific.
• 12:07
  SPEAKER_01: That's the training distribution that matters.
• 12:10
  SPEAKER_00: Why is this well watering out faster than its neighbor beats summarize the field every time?
• 12:14
  SPEAKER_01: And the agentic loop should be allowed to say,
• 12:15
  SPEAKER_01: I don't know, you need another log, instead of inventing a tidy story.
• 12:19
  SPEAKER_00: Invented tidy stories are how we get confident and wrong.
• 12:22
  SPEAKER_01: Confident and wrong is the expensive quadrant.
• 12:25
  SPEAKER_00: Also, who gets blamed when a scenario set misses?
• 12:28
  SPEAKER_00: If the answer is the AI, you've already broken accountability.
• 12:31
  SPEAKER_01: Accountability stays human.
• 12:33
  SPEAKER_01: Tools don't get a signature block.
• 12:35
  SPEAKER_00: Alright, money where mouth is for majors.
• 12:37
  SPEAKER_00: What's a concrete 90-day move for a mid-sized shop that liked this episode and feels late?
• 12:41
  SPEAKER_01: 90 days,
• 12:42
  SPEAKER_01: pick one recurring decision,
• 12:43
  SPEAKER_01: work over candidates, sidetrack screens, or development case prep.
• 12:47
  SPEAKER_01: Write down where human reject happens today, then evaluate tools only against that path.
• 12:52
  SPEAKER_00: If a vendor can't plug into that path, they're a TED talk.
• 12:54
  SPEAKER_01: Second, inventory your data debts on that decision path.
• 12:57
  SPEAKER_01: Missing headers, sketchy tests, broken completion histories.
• 13:00
  SPEAKER_01: AI doesn't dissolve those, it spotlights them.
• 13:03
  SPEAKER_00: Usually by failing in a way that makes someone blame the model instead of the database.
• 13:06
  SPEAKER_01: Third, assign an owner who can say no to demos.
• 13:08
  SPEAKER_01: Without a no owner, every shiny pilot becomes a zombie.
• 13:11
  SPEAKER_00: We have seen the zombie pilot graveyard.
• 13:13
  SPEAKER_00: It's crowded.
• 13:14
  SPEAKER_01: Crowded and still somehow inviting new burials.
• 13:16
  SPEAKER_00: Fourth, schedule one ugly asset test before any steering committee victory lap.
• 13:20
  SPEAKER_01: Victory laps before ugly asset tests are how you fund souvenirs.
• 13:24
  SPEAKER_00: Okay, what this release does not say, so we don't fan fiction it.
• 13:27
  SPEAKER_01: It does not claim first oil timeline miracles.
• 13:30
  SPEAKER_01: It does not publish accuracy benchmarks.
• 13:32
  SPEAKER_01: It does not name fields.
• 13:33
  SPEAKER_01: It does not say humans step aside.
• 13:35
  SPEAKER_00: Good.
• 13:36
  SPEAKER_00: Anyone turning this PR into total solved exploration is writing fiction.
• 13:40
  SPEAKER_01: Our operating read.
• 13:41
  SPEAKER_01: Big operator pays up for proprietary scenario engines on real geoscience data with
• 13:45
  SPEAKER_01: a lab and an IP wall aimed at exploration and brownfield life extension.
• 13:49
  SPEAKER_00: And the human review boundary stays non-negotiable.
• 13:52
  SPEAKER_01: Non-negotiable.
• 13:53
  SPEAKER_01: If your process can't show where a person killed a bad scenario,
• 13:55
  SPEAKER_01: you don't get to call it engineering support.
• 13:57
  SPEAKER_00: Field takeaway then.
• 13:59
  SPEAKER_00: Mid-size shop, three people watching markets, and subsurface.
• 14:02
  SPEAKER_01: You probably can't replicate total mistrawl.
• 14:05
  SPEAKER_01: You can copy the shape,
• 14:06
  SPEAKER_01: proprietary data,
• 14:07
  SPEAKER_01: clear reject path,
• 14:08
  SPEAKER_01: shared ownership between subsurface and model people,
• 14:11
  SPEAKER_01: and no worship of a single AI answer.
• 14:13
  SPEAKER_00: And you can watch whether majors keep funding joint labs versus buying another dashboard SKU.
• 14:17
  SPEAKER_00: Capital allocation is the tell.
• 14:19
  SPEAKER_01: Capital allocation is the tell, always was.
• 14:22
  SPEAKER_00: One more distinction so we don't blur seasons.
• 14:24
  SPEAKER_00: Petronas-style platform stories were about investment rails and clocks.
• 14:27
  SPEAKER_00: This is about model weights and labs sitting on geoscience.
• 14:30
  SPEAKER_01: Rails vs.
• 14:31
  SPEAKER_01: Brains.
• 14:32
  SPEAKER_01: If a salesperson conflates them, hang up politely.
• 14:34
  SPEAKER_00: Politely.
• 14:35
  SPEAKER_00: Linger on IP for a minute because that's where mid-caps get soft.
• 14:38
  SPEAKER_01: Soft IP language is how you wake up
• 14:40
  SPEAKER_01: and find your field quirks inside someone else's demo six months later.
• 14:43
  SPEAKER_00: You want written answers on training use,
• 14:44
  SPEAKER_00: retention, deletion, subprocessors, and whether aggregate learning is allowed.
• 14:48
  SPEAKER_01: Aggregate learning is the phrase that hides a multitude of sins.
• 14:51
  SPEAKER_00: If counsel can't explain the clause in one sentence, don't sign it because the demo sparkled.
• 14:56
  SPEAKER_01: Sparkle is not a commercial term.
• 14:58
  SPEAKER_00: Joint lab mechanics.
• 14:59
  SPEAKER_00: What good looks like?
• 15:00
  SPEAKER_01: Shared backlog, shared definition of done.
• 15:02
  SPEAKER_01: Subsurface leads on problem selection.
• 15:04
  SPEAKER_01: Model folks lead on method.
• 15:06
  SPEAKER_01: Both show up when it breaks.
• 15:07
  SPEAKER_00: Cadence measured in ship decision support, not papers.
• 15:10
  SPEAKER_01: Papers are fine.
• 15:11
  SPEAKER_01: Papers aren't production.
• 15:12
  SPEAKER_00: And rotate asset teams through so the lab doesn't become a boutique for one VP's favorite field.
• 15:17
  SPEAKER_01: Favorite field boutiques are how the rest of the organization learns to ignore the lab.
• 15:20
  SPEAKER_00: Let's dig one layer deeper on scenario quality because multiple scenarios can still be theater.
• 15:25
  SPEAKER_01: Theater looks like five cases that are really the same case with a type curve slid around.
• 15:29
  SPEAKER_00: Real diversity means different geologic stories,
• 15:31
  SPEAKER_00: different facility constraints,
• 15:32
  SPEAKER_00: different commercial assumptions, and you can see which dial moved.
• 15:36
  SPEAKER_01: If the only dial is optimism, that's not a scenario set, that's a mood board.
• 15:39
  SPEAKER_00: Mood boards don't belong in AFE packages.
• 15:41
  SPEAKER_01: They belong on refrigerators.
• 15:43
  SPEAKER_00: Another mid-cap trap.
• 15:44
  SPEAKER_00: Buying a tool because a peer company put out a LinkedIn post about AI transformation.
• 15:49
  SPEAKER_01: LinkedIn is not due diligence.
• 15:51
  SPEAKER_00: Ask the peer what decision changed on a Tuesday.
• 15:53
  SPEAKER_00: If they can't name one, you learned something cheap.
• 15:55
  SPEAKER_01: Cheap lessons are the best kind.
• 15:57
  SPEAKER_00: On the total side of this deal, culture risk is real.
• 16:00
  SPEAKER_00: IOCs can fund labs and still route every output through 12 committees until
• 16:03
  SPEAKER_00: the insight is stale.
• 16:04
  SPEAKER_01: Stale insight is how you pay European AI prices for last quarter's PowerPoint.
• 16:08
  SPEAKER_00: So if you're benchmarking your own shop against this announcement,
• 16:11
  SPEAKER_00: benchmark decision latency too, not just model pedigree.
• 16:14
  SPEAKER_01: How long from new information to a killed or kept case?
• 16:17
  SPEAKER_01: That's the metric.
• 16:18
  SPEAKER_00: What about service companies watching this?
• 16:19
  SPEAKER_00: Baker, SLB, Halliburton folks in the headphones.
• 16:23
  SPEAKER_01: They're hearing a customer try to own more of the judgment layer.
• 16:26
  SPEAKER_01: That changes what partner means in the next MSA negotiation.
• 16:29
  SPEAKER_00: Operators will still need sensors, crews, and field execution.
• 16:32
  SPEAKER_00: They're less interested in renting the brain that sits on their Sigi.
• 16:35
  SPEAKER_01: So the service pitch shifts from our platform knows reservoirs to our platform plugs
• 16:38
  SPEAKER_01: into your proprietary stack without stealing the weights.
• 16:41
  SPEAKER_00: Plug in without theft.
• 16:42
  SPEAKER_00: Put that on the MSA cover page.
• 16:44
  SPEAKER_01: In bigger font than the liability cap.
• 16:46
  SPEAKER_00: Liability caps are their own comedy genre.
• 16:49
  SPEAKER_00: We won't go there today.
• 16:50
  SPEAKER_01: Thank you.
• 16:51
  SPEAKER_00: Data readiness, the boring villain.
• 16:53
  SPEAKER_01: Always, you can buy frontier models and still lose to missing deviation surveys.
• 16:57
  SPEAKER_00: Or production allocated with a shrug.
• 16:59
  SPEAKER_01: Or completion records that say frack and nothing else.
• 17:02
  SPEAKER_00: A joint lab will spend its first months arguing about those gaps, not generating glamorous maps.
• 17:06
  SPEAKER_01: And that argument is the work.
• 17:07
  SPEAKER_01: If a vendor promises to skip it,
• 17:09
  SPEAKER_01: they're lying or they're about to discover your database the hard way.
• 17:12
  SPEAKER_00: Hard way equals change orders.
• 17:14
  SPEAKER_01: Change orders equals CFOs learning AI vocabulary they didn't want.
• 17:18
  SPEAKER_00: Let's talk governance without sounding like a consultant brochure.
• 17:20
  SPEAKER_01: Please.
• 17:21
  SPEAKER_00: Who can promote a model version into the decision path?
• 17:24
  SPEAKER_00: Who can roll it back?
• 17:25
  SPEAKER_00: Who gets paged when a scenario set drifts after a data reload?
• 17:28
  SPEAKER_01: If the answer is the AI team, Subsurface will ignore it.
• 17:31
  SPEAKER_01: If the answer is Subsurface alone, the models will rot.
• 17:34
  SPEAKER_00: Shared on call for the decision path?
• 17:35
  SPEAKER_00: Ugly, effective.
• 17:37
  SPEAKER_01: Ugly and effective is our brand.
• 17:39
  SPEAKER_00: Investor relations angle for public operators listening.
• 17:42
  SPEAKER_00: This kind of announcement is also a signal to markets that AI spend is moving
• 17:45
  SPEAKER_00: from experiments to capitalized capability.
• 17:47
  SPEAKER_01: Markets will overread it.
• 17:48
  SPEAKER_01: They always do.
• 17:50
  SPEAKER_00: Your job internally is not to overread it with them.
• 17:52
  SPEAKER_00: Keep it tied to decisions in IP.
• 17:54
  SPEAKER_01: If IR wants a slide,
• 17:55
  SPEAKER_01: give them the joint lab and the euro floor, not a promise about discovery rates.
• 17:59
  SPEAKER_00: Discovery rate promises are how you get quoted in places you'll regret.
• 18:02
  SPEAKER_01: Regret has a long half-life in this industry.
• 18:04
  SPEAKER_00: Landman and A&D analysts, practical note.
• 18:07
  SPEAKER_01: If subsurface starts producing wider scenario sets faster,
• 18:11
  SPEAKER_01: commercial needs a way to consume them without waiting for a pretty PDF.
• 18:14
  SPEAKER_00: A table of cases with visible assumptions beats a narrated movie.
• 18:17
  SPEAKER_01: And please stop putting AI-optimized in the CIM if you can't show the reject log.
• 18:21
  SPEAKER_00: CM language is a separate episode of SINS.
• 18:23
  SPEAKER_01: We'll need more coffee for that one.
• 18:25
  SPEAKER_00: Security and export controls.
• 18:27
  SPEAKER_00: Light touch because we're not lawyers.
• 18:29
  SPEAKER_01: But if your data crosses borders into a model vendor's cloud,
• 18:32
  SPEAKER_01: someone in compliance should see the diagram before marketing does.
• 18:35
  SPEAKER_00: Marketing always wants to see it first.
• 18:37
  SPEAKER_00: That's how incidents start.
• 18:38
  SPEAKER_01: Incidents start with, it's fine, it's anonymized.
• 18:40
  SPEAKER_00: Anonymized subsurface is often neither.
• 18:43
  SPEAKER_01: Correct.
• 18:43
  SPEAKER_00: What would make us say this total mistrial bet failed two years from now?
• 18:47
  SPEAKER_01: Asset teams still don't open the tool.
• 18:48
  SPEAKER_01: Scenario sets never appear in AFEs.
• 18:51
  SPEAKER_01: IP clauses get loosened under delivery pressure.
• 18:53
  SPEAKER_01: Or the lab becomes a paper mill.
• 18:55
  SPEAKER_01: And success?
• 18:56
  SPEAKER_01: Ugly assets get better cases faster.
• 18:58
  SPEAKER_01: People argue about assumptions instead of vibes.
• 19:01
  SPEAKER_01: And a mid-cap can point to the pattern even if they can't copy the check size.
• 19:04
  SPEAKER_00: Pattern over parity.
• 19:05
  SPEAKER_00: That's the mid-cap strategy.
• 19:07
  SPEAKER_01: Pattern over parity.
• 19:08
  SPEAKER_01: Write that above the oatmeal slide.
• 19:10
  SPEAKER_00: One more operations example.
• 19:12
  SPEAKER_00: Gas plant bottleneck.
• 19:13
  SPEAKER_00: Midstream constraint.
• 19:14
  SPEAKER_00: Basin with nowhere to put molecules.
• 19:16
  SPEAKER_01: A reservoir scenario that ignores midstream is fanfiction.
• 19:19
  SPEAKER_00: So any serious stack has to let you pin commercial and midstream constraints into
• 19:22
  SPEAKER_00: the case set, or it's incomplete for real operators.
• 19:25
  SPEAKER_01: Incomplete tools create complete disasters in committee.
• 19:27
  SPEAKER_00: Harsh, but fair.
• 19:29
  SPEAKER_01: Fair is our other brand.
• 19:30
  SPEAKER_00: Before the pun, you still owe us.
• 19:32
  SPEAKER_00: One sentence on why this episode exists, instead of the heating oil cut we already had on disc.
• 19:36
  SPEAKER_01: Because the hard fork stack in the Total Mistral story beat another week
• 19:39
  SPEAKER_01: of inventory theater, even when inventory theater was already recorded.
• 19:43
  SPEAKER_00: Desk choice over sunk cost.
• 19:45
  SPEAKER_00: Rare and correct.
• 19:46
  SPEAKER_01: Sunk cost is how you ship the wrong episode with great audio quality.
• 19:49
  SPEAKER_00: We parked that audio.
• 19:50
  SPEAKER_00: It can stay parked.
• 19:51
  SPEAKER_00: Let's stay with capital allocation a little longer because that's where boards get sloppy.
• 19:55
  SPEAKER_01: Boards love a round number in euros.
• 19:56
  SPEAKER_01: They love it more than a messy workflow diagram.
• 19:59
  SPEAKER_00: So if you're presenting this kind of initiative upstairs,
• 20:01
  SPEAKER_00: lead with the decision path and the IP wall, then show the spend.
• 20:04
  SPEAKER_01: Spend without a path is a donation to vendor revenue.
• 20:07
  SPEAKER_00: Donation is the right word.
• 20:08
  SPEAKER_01: What about smaller public ENPs who feel pressure to have an AI story for earnings season?
• 20:13
  SPEAKER_00: Have a story about one decision you improved, not a story about platforms you licensed.
• 20:17
  SPEAKER_01: We kill bad work over candidates two days faster beats we embraced generative AI.
• 20:21
  SPEAKER_00: Every time.
• 20:22
  SPEAKER_01: Private equity-backed operators.
• 20:24
  SPEAKER_01: Different pressure.
• 20:25
  SPEAKER_01: Clock is shorter.
• 20:26
  SPEAKER_01: Tolerance for boutique labs is lower.
• 20:28
  SPEAKER_00: Then copy the checklist harder.
• 20:30
  SPEAKER_00: Reject button.
• 20:31
  SPEAKER_00: Ugly asset test.
• 20:32
  SPEAKER_00: Owner who can say no.
• 20:34
  SPEAKER_00: Skip the European joint lab cosplay if you can't staff it.
• 20:36
  SPEAKER_01: Cosplay is expensive in a hold period.
• 20:39
  SPEAKER_00: National oil companies listening outside Europe.
• 20:41
  SPEAKER_00: The sovereignty framing still maps.
• 20:42
  SPEAKER_00: Just swap the geography.
• 20:43
  SPEAKER_01: Same question.
• 20:44
  SPEAKER_01: Do strategic subsurface datasets leave your control to train someone else's foundation model?
• 20:49
  SPEAKER_00: If yes, you'd better be getting something that compounds on your side of the ledger.
• 20:52
  SPEAKER_01: Equity in the vendor is not the same as capability on the asset.
• 20:55
  SPEAKER_00: Equity is a hope.
• 20:56
  SPEAKER_00: Capability is a Tuesday morning.
• 20:58
  SPEAKER_01: Tuesday morning keeps winning these arguments.
• 21:00
  SPEAKER_00: Model evaluation.
• 21:02
  SPEAKER_00: How do you know a scenario engine is any good without waiting for five years of outcomes?
• 21:05
  SPEAKER_01: Blind retrospectives on all decisions.
• 21:08
  SPEAKER_01: Hide the actual choice.
• 21:09
  SPEAKER_01: See whether the tool surfaces the assumptions that mattered.
• 21:12
  SPEAKER_00: And score calibration.
• 21:13
  SPEAKER_00: When it says high confidence, is it actually right more often?
• 21:16
  SPEAKER_01: If high confidence is theater, turn that dial off until it's earned.
• 21:19
  SPEAKER_00: Human factors.
• 21:20
  SPEAKER_00: Engineers will game any score that affects their bonus.
• 21:23
  SPEAKER_01: So don't bolt scenario scores to bonuses in year one.
• 21:26
  SPEAKER_01: You'll train the wrong behavior.
• 21:27
  SPEAKER_00: Train people to interrogate assumptions, not to chase a green badge.
• 21:30
  SPEAKER_01: Green badges belong on hiking trails.
• 21:33
  SPEAKER_00: Change management without the soft language.
• 21:35
  SPEAKER_01: Please.
• 21:36
  SPEAKER_00: Pick two asset teams.
• 21:37
  SPEAKER_00: Give them the tool.
• 21:38
  SPEAKER_00: Sit in their meetings for a month.
• 21:40
  SPEAKER_00: Count how often output changes a decision.
• 21:42
  SPEAKER_00: Scaling zero gets you a bigger zero.
• 21:43
  SPEAKER_00: If that feels too formal, you will recreate the
• 21:46
  SPEAKER_01: same fight every sprint.
• 21:47
  SPEAKER_01: Formal is cheaper than
• 21:48
  SPEAKER_00: recurring fights.
• 22:07
  SPEAKER_00: Vendor Concentration Risk.
• 22:09
  SPEAKER_00: If your entire scenario stack sits with one frontier lab partner, what's the exit?
• 22:13
  SPEAKER_01: Exit means exportable weights,
• 22:14
  SPEAKER_01: or at least exportable features and prompt chains you can re-host.
• 22:17
  SPEAKER_00: If exit is, call a lawyer and cry, you built a hostage system.
• 22:21
  SPEAKER_01: Hostages don't belong in the subsurface toolkit.
• 22:24
  SPEAKER_00: Open source versus proprietary frontier stacks?
• 22:26
  SPEAKER_00: People will ask.
• 22:27
  SPEAKER_01: Open source can be great for tooling around the edges.
• 22:30
  SPEAKER_01: The differentiator in this total story is the proprietary data plus the funded lab,
• 22:34
  SPEAKER_01: not the brand name on the base model alone.
• 22:36
  SPEAKER_00: Base models get less special every year.
• 22:38
  SPEAKER_00: Your data and your decision process don't.
• 22:40
  SPEAKER_01: That's the uncomfortable truth for vendors selling only the base model glow.
• 22:43
  SPEAKER_00: Glow fades.
• 22:44
  SPEAKER_00: Decline curves don't.
• 22:46
  SPEAKER_01: Decline curves never got the memo about hype cycles.
• 22:48
  SPEAKER_00: Field development planning rooms.
• 22:50
  SPEAKER_00: The loud rooms with facilities and drilling and commercial arguing over a map.
• 22:54
  SPEAKER_01: If the AI case set can't survive that room, it isn't ready.
• 22:56
  SPEAKER_00: Survival means someone can attack an assumption and the tool helps you adjust
• 22:59
  SPEAKER_00: without a weekend of spreadsheet archaeology.
• 23:01
  SPEAKER_01: Spreadsheet archaeology is our national pastime.
• 23:04
  SPEAKER_00: And it should stay a backup sport, not the main event.
• 23:07
  SPEAKER_00: Discipline is allowed to be boring.
• 23:08
  SPEAKER_00: Training data provenance.
• 23:09
  SPEAKER_00: Another quiet killer.
• 23:30
  SPEAKER_01: If you can't say which surveys,
• 23:31
  SPEAKER_01: which vintages,
• 23:32
  SPEAKER_01: which interpreted surfaces went into a fine tune,
• 23:34
  SPEAKER_01: you can't debug a bad season of recommendations.
• 23:37
  SPEAKER_00: Providence logs are not optional in a serious lab.
• 23:40
  SPEAKER_01: Optional providence is how myths become models.
• 23:42
  SPEAKER_00: Myths become models.
• 23:43
  SPEAKER_00: Models become AFEs.
• 23:45
  SPEAKER_00: AFEs become postmortems with catered lunch.
• 23:48
  SPEAKER_01: I've attended those lunches.
• 23:49
  SPEAKER_01: The sandwiches were better than the root cause analysis.
• 23:51
  SPEAKER_00: Sandwiches usually are.
• 23:53
  SPEAKER_01: What do we tell a university geoscience department listening in?
• 23:56
  SPEAKER_00: Teach students to argue with model output.
• 23:58
  SPEAKER_00: If they only learn to accept plots, there'll be spectators in rooms that still need judgment.
• 24:02
  SPEAKER_01: Judgment is still the job.
• 24:04
  SPEAKER_00: Judgment plus better instruments.
• 24:05
  SPEAKER_00: That pairing is the whole show.
• 24:07
  SPEAKER_01: Always was.
• 24:08
  SPEAKER_00: One operational anti-pattern I want on the record.
• 24:10
  SPEAKER_00: Emailing scenario PDFs around until nobody knows which version died in committee.
• 24:14
  SPEAKER_01: Version chaos.
• 24:15
  SPEAKER_01: Instant classic.
• 24:17
  SPEAKER_00: Put cases in a system with IDs, owners, and a tombstone when rejected.
• 24:20
  SPEAKER_00: Reason required.
• 24:21
  SPEAKER_01: Tombstones teach the next team.
• 24:22
  SPEAKER_01: Inboxes don't.
• 24:24
  SPEAKER_00: Inboxes teach you fear of search.
• 24:26
  SPEAKER_01: Accurate.
• 24:27
  SPEAKER_00: Last deeper beat before your pun, compute location.
• 24:29
  SPEAKER_01: Keep sensitive fine-tuning close.
• 24:31
  SPEAKER_01: Inference can be flexible.
• 24:33
  SPEAKER_01: Training that sees the crown jewels needs a short leash.
• 24:35
  SPEAKER_00: Short leash, long memory of who held it.
• 24:37
  SPEAKER_01: Write the leash into the contract before the kickoff party.
• 24:40
  SPEAKER_00: Kickoff parties are where constraints go to die.
• 24:43
  SPEAKER_01: Then resurrect as incidents.
• 24:44
  SPEAKER_01: I still owe the room a pun and a wince.
• 24:46
  SPEAKER_01: You've been patient.
• 24:47
  SPEAKER_00: Make it quick.
• 24:48
  SPEAKER_01: If Total and Mistral actually teach a model to stress multiple development cases
• 24:51
  SPEAKER_01: before the committee meets, that's secondary recovery for bad first drafts.
• 24:56
  SPEAKER_00: Oh no.
• 24:56
  SPEAKER_01: I know.
• 24:57
  SPEAKER_01: That one can stay in the hole.
• 24:58
  SPEAKER_01: Moving on.
• 24:59
  SPEAKER_00: Buried.
• 24:59
  SPEAKER_00: Permanently.
• 25:00
  SPEAKER_00: Thank you for your service.
• 25:02
  SPEAKER_01: Never again.
• 25:03
  SPEAKER_00: Never again.
• 25:04
  SPEAKER_00: Back to the serious close before we get dumb on purpose.
• 25:07
  SPEAKER_01: Serious close.
• 25:08
  SPEAKER_01: This week wasn't a title story and it wasn't a GPU future story.
• 25:11
  SPEAKER_01: It was an IOC funding a frontier model joint lab aimed at reservoir decisions
• 25:15
  SPEAKER_01: with IP framed as a feature.
• 25:17
  SPEAKER_00: If you're evaluating AI for your own shop, steal the seriousness.
• 25:20
  SPEAKER_00: Leave the fairy tale at the vendor breakfast with the oatmeal.
• 25:22
  SPEAKER_01: Steal the checklist.
• 25:23
  SPEAKER_01: Steal the reject button.
• 25:24
  SPEAKER_01: Steal the ugly asset test.
• 25:26
  SPEAKER_01: Leave the oatmeal slide.
• 25:27
  SPEAKER_00: Especially the oatmeal slide.
• 25:28
  SPEAKER_00: Two quick listener scenarios, then closer.
• 25:31
  SPEAKER_01: Scenario A, you're a VP subsurface.
• 25:33
  SPEAKER_01: Board wants an AI strategy slide by month end.
• 25:36
  SPEAKER_00: Don't start with vendors.
• 25:37
  SPEAKER_00: Start with one decision path and the data debts on it.
• 25:39
  SPEAKER_00: Then the slide writes itself without fantasy.
• 25:41
  SPEAKER_01: Scenario B,
• 25:42
  SPEAKER_01: you're a CFO who just saw a totals number and wants
• 25:44
  SPEAKER_01: to know why we're not spending like that.
• 25:46
  SPEAKER_00: Because capability spend without a reject path and an owner is just a more expensive dashboard.
• 25:50
  SPEAKER_00: Match seriousness before you match euros.
• 25:53
  SPEAKER_01: I'd put that on a sticky note next to the oatmeal quote.
• 25:55
  SPEAKER_01: All right.
• 25:56
  SPEAKER_01: Newsblock stays serious.
• 25:57
  SPEAKER_01: Closer gets dumber.
• 25:58
  SPEAKER_01: House rules for week 39.
• 26:00
  SPEAKER_00: Messiest industry artifacts of the week.
• 26:02
  SPEAKER_00: Two or three.
• 26:03
  SPEAKER_00: No ceremonies, no mascots, no trains, no game shows.
• 26:06
  SPEAKER_01: Item one.
• 26:07
  SPEAKER_01: The vendor breakfast slide that promised a subsurface twin before coffee.
• 26:10
  SPEAKER_01: Confidence without a data diet.
• 26:12
  SPEAKER_01: Grade.
• 26:13
  SPEAKER_01: Dumped in the mud cleaner.
• 26:14
  SPEAKER_00: Item two.
• 26:15
  SPEAKER_00: Any AI platform pitch that says agentic five times and never once says who owns
• 26:19
  SPEAKER_00: the weights after the pilot.
• 26:21
  SPEAKER_00: Grade, recycle bin next to the expired jib.
• 26:23
  SPEAKER_01: Item three, midstream names that are three words of weather and zero words of molecules.
• 26:27
  SPEAKER_01: You know the ones.
• 26:28
  SPEAKER_01: Grade, politely lost in the corridor.
• 26:31
  SPEAKER_00: News was the total mistrial check in the lab.
• 26:33
  SPEAKER_00: Closer is where we put the glitter.
• 26:35
  SPEAKER_01: Glitter's contained.
• 26:36
  SPEAKER_01: Credits.
• 26:37
  SPEAKER_00: Silicon and shale.
• 26:38
  SPEAKER_00: Practical conversation where the oil patch meets the algorithm.
• 26:41
  SPEAKER_00: Siliconandshale at gemail.com.
• 26:44
  SPEAKER_01: If you sell to operators, landmen, engineers, and you want to be on the show, same address.
• 26:48
  SPEAKER_01: We use the product before we talk about it.
• 26:50
  SPEAKER_01: Which is its own kind of filter.
• 26:51
  SPEAKER_00: Go bring the reject button.