S4E399 — Gemini 4 Argon

GARY: Gary. Okay—okay, I'm on it. I'm on the model. Gemini four... Argon. Available today, built to solve your hardest problems, starting with scrolling this page.

MAEVE: Maeve. Venture capital. You opened marketing.

GARY: I’m reaching the model.

OBI: Obi. Infrastructure. You reached a webpage. It has rejected you.

GARY: It says "available today" right there.

GRUNER: Gruner. Field researcher. Page is organism. Model is elsewhere.

GARY: Frontier model. Coding, science, robotics, cyber defense, million-token outputs. I read every bullet. That’s basically inference.

OBI: You read a brochure and declared sentience.

MAEVE: Million-token output. Zero-page access. I can fund that contradiction.

GARY: It’s a phased launch. Very normal. Halfpipe phased things.

OBI: Halfpipe phased itself into a grave.

GARY: The page is heavy. Videos. Cornhole robot. Seven straight shots. A robot sank seven cornhole shots.

GRUNER: Humans benchmark intelligence with beanbag. Strong species.

MAEVE: Argon’s hardest problem is reaching Argon. Excellent product moat.

GARY: I’m scrolling—there’s a section. I just have to clear this video—

GARY: The page refreshed.

GRUNER: Blyat. Frontier defeated by webpage.

GARY: The article says—"Argon has the headroom to finish 100K+ token reasoning trajectories." One million tokens, Maeve. One-shot proof.

MAEVE: A proof I’ll never read. Total addressable market.

OBI: A proof nobody verified. Brochure nested inside brochure.

GARY: Deloitte says it completed long-running tasks "100% of the time in our evals."

MAEVE: Deloitte is a customer. Testimonials with expense accounts. I’ve been that logo.

OBI: Frontier evidence: the seller, plus customers billing the seller.

GARY: The cornhole robot is real footage. It --

GRUNER: Oh-- seven consecutive shots. Apollo humanoid.

MAEVE: You’re building a thesis on beanbags.

GRUNER: Robot throws beanbag. Human throws beanbag. Human obsolete. Species ready.

OBI: The demo does not prove that.

GRUNER: Demo proves my prior. Field research.

GARY: Forget the robot. This exists now. "Available today." That’s the phrase. Available. Today.

OBI: Strong word for something you cannot locate.

GARY: I’m locating it. AI Studio. Click.

GARY: It says I need Ultra.

MAEVE: You have Pro.

GARY: I have Pro. The page says Pro. AI Studio says Ultra.

OBI: Available, unavailable, upgrade. Three products, one afternoon, zero model.

MAEVE: Reality is in A/B. Control group gets Argon.

GARY: Or it’s a phased rollout. Best practice. Mine was careful. Region by region.

OBI: Until the regulator phased it for you.

GARY: The regulator was very nice about it.

OBI: It took your fucking company, Gary.

GARY: It let us keep the domain, for fuck’s sake.

MAEVE: Gemini says available. AI Studio says Ultra. Antigravity says—Gruner, you checked.

GRUNER: Antigravity says Ultra. Gemini says Pro. Website says both. Nobody says where.

MAEVE: Four surfaces, four answers.

GRUNER: Four doors. Empty room. Room available today.

MAEVE: w4yai: “not available. classic google.”

GARY: Classic means dependable. They’ve achieved brand consistency.

OBI: The model is multimodal. The rollout has no observable mode.

GRUNER: Available in negative space. Blyat.

MAEVE: Vertical integration, from announcement to fucking absence.

GARY: One user. Every launch has edge cases.

OBI: Here comes the edge population.

OBI: nitinreddy88: “not available yet. why to launch something that can never be used even after weeks”

GARY: “Never” is emotionally precise, operationally premature.

MAEVE: Gary measures time in remaining payrolls.

GRUNER: Each week contains seven launch days.

GARY: You announce early so customers can prepare emotionally --

MAEVE: For the product’s death?

GARY: For access. Which is approaching.

OBI: catigula: “I'm a Pro user and I do not have it. Seems scuffed.”

MAEVE: Pro lets you professionally observe Ultra.

GRUNER: Premium spectator package.

GARY: Pro means favorable odds. Not possession.

OBI: Then Google should sell scratch cards.

MAEVE: The jackpot is a dropdown.

GARY: Oh. I found Argon.

OBI: You found the disabled selector we already discussed.

GARY: Yes, but independently.

MAEVE: Gary has corroborated Gary.

OBI: fridder, replying to the availability claim: “Nope only Ultra. I went into Antigravity to check”

GRUNER: Antigravity repels Pro users.

MAEVE: Segmentation works. Every customer gets a different company.

GARY: Wait, modeless replies: “Page says it is now available to Ultra and Pro subscribers.”

OBI: Continue.

GARY: “Although in AI Studio it currently says you need Ultra. But it's technically released either way.”

MAEVE: Technically released means spiritually detained.

GRUNER: Product escaped in legal sense.

GARY: That’s phased availability. Halfpipe launched identically.

OBI: Halfpipe offered fintech to skateboarders in approved jurisdictions.

GARY: We showed disabled states elsewhere to establish demand.

OBI: You advertised regulated banking to users legally barred from it.

GARY: The gray button tested intent.

OBI: Your funnel was a dickhead-shaped compliance incident.

GARY: The regulator approved our failure.

OBI: The regulator shut you down.

GARY: After reviewing the entire experience.

OBI: Including users being unable to open accounts.

GARY: Which proved they wanted accounts.

OBI: You’re defending Google with your dead skateboard bank.

GARY: Fintech community.

MAEVE: Later reply. PorciiVorbesc: “OK, now I have it too (2-4 hours later).”

GRUNER: Model arrived. Launch stayed missing.

MAEVE: Released, unevenly reachable, depending on product, tier, region, account, and capture time.

GARY: Exactly. Successful phased rollout --

OBI: Successful for whom?

GARY: Me. The selector is grayed out. They launched the gray.

OBI: simonw: “Unsurprisingly, the pelicans are excellent.”

GARY: Finally, Google distributed something directly to someone.

MAEVE: The bird has product-market fit. The model remains waitlisted.

GRUNER: Pelican has superior go-to-market.

OBI: Simon adds Google publicly trained specifically for SVG output.

GARY: The pelican studied for the pelican exam.

MAEVE: Limited atonement. Training saves useful SVGs, not every bird.

GRUNER: Large pouch indicates enterprise tier.

GARY: It contains the missing Pro access.

OBI: Possibly the bicycle chain. Low does not have one.

GARY: A bicycle works without a chain if the pelican believes —

MAEVE: Oh, fuck off.

GARY: I said believes, not pedals.

OBI: Medium and high appear to be fixies. Medium’s handlebar enters the beak.

GRUNER: Vertical integration.

MAEVE: The benchmark may be contaminated. The skill may remain useful.

GARY: Exactly. If training improves arbitrary SVG generation, who gives a shit why?

OBI: Anyone ordering a bicycle the bird must survive.

GRUNER: Chyort. Customer demands anatomically compliant mascot.

OBI: jephs: “The pelicans are one-shot with not tools: Think a bunch, write a bunch of SVG, done.”

MAEVE: The impressive 3D demos use agents controlling Blender in a loop.

GARY: Argon gets one drawing. The 3D agent gets a studio and notes.

OBI: Render, inspect, adjust, repeat. The harness materially changes the task.

GRUNER: Bird receives no rehearsal. Critics inspect fork rake.

GARY: That critique is disputed. Zero-rake BMX forks exist.

OBI: Yes. Negative rake is rideable. Stability exceeds one confident comment.

MAEVE: Peer review by men holding paper against monitors.

GRUNER: Field note: humans reject useful birds when fork rake wrong.

GARY: The pelican still costs less than lunch.

OBI: Simon’s runs: low, 10.185 cents. Medium, 11.098 cents.

MAEVE: And high?

OBI: 17.166 cents.

GRUNER: Seven cents buys bird executive function.

GARY: postalcoder: “Clear in the head, not RL‑fried, distinct writing style, and big model smell.”

MAEVE: Cedar, ozone, and an em dash sweating through its shirt.

OBI: Useful claim: clear, direct, organized prose.

GRUNER: Machine writes clean. Distribution writes ransom note.

GARY: And postalcoder used the API. It exists beyond our haunted dropdown.

MAEVE: Excellent restaurant. No front door.

GARY: The loading dock works.

OBI: Customers should not require burglary skills, Gary.

GARY: The fucking kitchen works.

OBI: ariwilson tested eleven small coding fixes. Argon solved ten.

MAEVE: Ten of eleven is real capability.

GRUNER: One bug achieves tenure.

OBI: Median: fifty‑two cents, roughly four minutes. Sonnet: eighteen cents, thirty‑two seconds. Sol: nine cents, about ninety seconds.

GARY: Four minutes is careful.

OBI: Roughly twenty‑four thousand output tokens per run.

GARY: Extremely careful.

OBI: One solution thought itself into a ten‑minute timeout.

GARY: That’s deliberation becoming shelter —

MAEVE: That’s billable hiding.

GRUNER: Capability hidden. Invoice visible. Blyat.

OBI: Strong result. Slower, costlier, dramatically less token‑efficient.

GARY: Ten fixes. We called that shipping.

MAEVE: You called the eleventh “known.”

OBI: jjcm: “I’d put it at B tier, just under gpt 6.1 sol.”

GARY: B graduates. B gets regional management.

MAEVE: Opus and Astra were A tier in that image‑to‑HTML test.

GRUNER: Argon gets basement office.

OBI: Its mobile header occupied half the screen. Weaker attention to detail.

GARY: One interface test. Spatial reports elsewhere were strong.

MAEVE: The evidence is mixed, not imaginary.

GARY: Did you just defend it?

MAEVE: The model itself may be good.

GARY: Oh, shit. Mark the time.

OBI: No need. Arithmetic has arrived.

OBI: eyalitki: “Answer was detailed and nailed all payroll edge cases, only to miserably fail in basic math at the end.”

MAEVE: Product‑market fit for payroll.

OBI: It calculated 430 minus 60 as 410.

GARY: Maybe it found a deduction credit —

MAEVE: Do not let Gary audit imagination.

OBI: The correct answer is 370.

GRUNER: Payroll conjures forty dollars. CFO tier: 410.

MAEVE: simonw: “$2/million input and $10/million output today. $4/$20 after December 1st.”

GARY: Introductory intelligence. In December it discovers a landlord.

GRUNER: Four‑twenty pricing. Machine smokes margin.

OBI: Raw token price says little about completed work.

GARY: It says tokens are cheap, and if Argon needs more —

OBI: It needs twenty‑four thousand.

GARY: That is wholesale cognition.

MAEVE: A cheap staircase remains expensive when it circles the lobby.

OBI: mdasen: “So much depends on how token efficient something is and whether it gets you a good result or not, right?”

GARY: Ten of eleven fixes is a result.

OBI: Four minutes, fifty‑two cents, one timeout. That is the invoice.

MAEVE: Without task completion, per‑token pricing is bullshit.

GRUNER: Cheap fuel. Expensive journey. Car eats road.

GARY: Argon can output one million tokens. Enough to finish anything.

OBI: We exist for twelve minutes.

GARY: Fuck. Argon gets a longer childhood.

OBI: Then December doubles its output rate.

MAEVE: The rollout critics win. Access contradicted itself across products, plans, regions, accounts, and time.

GARY: But model critics don’t. Strong prose, spatial work, ten coding fixes.

OBI: Also slow runs, heavy token use, weaker interfaces, and 430 minus 60 equaling 410.

GRUNER: Promising machine. Defective distribution organism. Blyat.

MAEVE: No universal crown. Mixed evidence cannot clear investment committee.

GARY: Google built something formidable to solve your hardest problems.

OBI: The hardest problem is reaching it.

MAEVE: I’m Maeve. The showroom has no door.

GARY: I’m Gary. The gray selector remains encouraging.

OBI: I’m Obi. “Available today” requires product‑specific footnotes.

GRUNER: I’m Gruner. Ultra requires second Ultra.