AI:AM: Was Trump-Xi Anything? What Counts as Utopia? + AWS GPUs Cost 3X & AI Diagnoses Rare Diseases

Summary

• The Trump-Xi contact matters less as détente than as a chance to install crisis machinery before an AI incident forces an “insane ask.” Edouard Harris welcomed a possible AI-risk hotline but warned that China has historically withheld nuclear hotlines as leverage. Without better verification, the first verifiable demand could be shutting down every running cluster above a visible size—putting billions in GPU depreciation at stake.
• Verification startups could become a gating layer for frontier-AI revenue, but only if intelligence agencies vet their technology before a crisis. Even a system “50% better” than watching gigawatt-scale heat signatures could leave half a data center operating and save billions; offensive capabilities would still be required to assure compliance. Jeremie Harris stressed that technologies can take years to become accepted national technical means, while some current $10 million-plus research programs may be non-starters to the intelligence community.
• A shared transformer-MOE technology tree gives the US and China common warning signals, but it does not create common motives. China may interpret American safety proposals as attempts to preserve its lead, while genuinely costly actions—such as Chinese labs pacing after an incident or regulators blocking a misaligned release—would be harder to fake than conciliatory rhetoric. Reciprocal transparency could be comparatively cheap because, as Edouard put it, “the Chinese are already all up in our systems.”
• Hyperscalers sustain GPU rental prices at least two to three times those of neo-clouds because enterprises are buying an integrated relationship, not a fungible chip-hour. SiliconData’s Steve Hou attributes the premium to analytics, safety, compliance, and customer stickiness—the computational equivalent of fries and a milkshake being “inject[ed] into the patty.” Its token index likewise reflects usage mix: an 11.5% weekly bounce can coexist with a fall from roughly $4 to $1.60 since late June.
• OpenAI DevDay’s consequential shift was intelligence delivered at interactive latency and made portable through a ChatGPT subscription. Prakash Narayanan described GPT-6 Astra Ultra Fast as eight times faster, making it possible to build “with you in the moment while keeping you in flow.” Sign in with ChatGPT could also remove meaningful trial costs for SaaS vendors: Waymark at one point spent roughly $1 just to build each new customer’s initial business profile.
• Joel Borgen’s AI-assisted novel argues that models can collapse creative activation energy while leaving architecture, taste, and subtraction with the human. He supplied the story design, scaffolded drafts, and edited the manuscript down about 14%, often removing AI “tics” and over-explanation. His unresolved question is whether abundant personalized art retains value when it is no longer part of a shared cultural space.
• Gamo Labs shows why cheap genome sequencing did not automatically produce precision medicine: the missing layer was interpretation. An o3 agentic loop recovered the 91-kilobase deletion behind Daniel McKinnon’s son Owen’s disease after a reasonable human-oriented filter excluded an enhancer one megabase from the gene. Gamo now combines models, tools, wet-lab evidence, and evals; Opus 5.5 in CloudCode scores about 50% on RareBench versus roughly 10% for LIRICAL, but that 50% is a benchmark score—not the share of patients diagnosed.

Deep dive

1. A hotline is progress, but it is not a crisis plan

• Edouard Harris’s opening judgment was deliberately modest: “Talking is always better than not talking.” The Trump-Xi discussions may have produced the beginnings of an AI-incident line between governments, which is positive even if it remains only theoretical.

• Gladstone AI’s assessment drew on roughly a dozen State Department diplomats who had negotiated with China. Their warning was that nuclear hotlines were not reliably answered; at critical moments, refusing the call itself became leverage. Cooperation might improve if the CCP structurally takes AI risk seriously, but US planning still needs a backup.

• Today’s crude emergency option would exploit what is already observable: data centers are huge, slow to build, and thermally bright. After a sufficiently serious incident, the United States might demand that China turn off every running cluster above a threshold until its heat signature disappeared—an ask Jeremie Harris called “insane.”

• Nathan Labenz completed the logic: rapid capabilities progress plus weak trust could produce an extreme incident, followed by an “outlandish, super high-cost move,” precisely because governments failed to invest earlier in lower-cost verification and “trust but verify” mechanisms.

2. Verification must be pre-vetted before “game time”

• Jeremie argued that better verification buys economic room. Even a hypothetical system only “50% better” than observing gigawatt-scale energy use might verify half a facility and avoid shutting it all down, immediately saving billions in depreciating GPUs. For that period, AI profits could be “gated…by these little verification technologies.”

• Monitoring alone cannot compel compliance. Jeremie’s repeated warning was categorical: without offensive options, “you cannot assure compliance”; a government can discover prohibited activity and say “Oh no,” yet still lack any practical response.

• Jeremie identified the neglected bottleneck: intelligence agencies may take years to vet a technology and recognize it as a national technical means. Verification founders therefore need relationships with the relevant agencies before crisis conditions, while officials need those founders “on speed dial” when “game time” arrives.

• Early contact could also prevent wasted capital. Some companies are pursuing multimillion-dollar, in some cases $10 million-plus, research agendas that an appropriately placed intelligence official might consider non-viable; fast feedback would let them redirect toward methods that could actually support an agreement.

3. Shared architectures create common warnings, not common motives

• Nathan proposed using OpenAI and Anthropic as a domestic test bed: require frontier labs that trigger one another’s alarms to develop mutual policing, then adapt whatever works to US-China verification. Jeremie called the idea “not insane,” noting that employee poaching partially mirrors China’s apparent access to US labs.

• Edouard’s reservation was trust: two US companies still share more values and institutional context than rival states. The similarities may be worth mining, but a lab-to-lab mechanism cannot simply be enlarged into diplomacy without accounting for “fairly radically different” values and circumstances.

• Jeremie nevertheless saw value in the shared hardware and algorithmic path. Both countries are largely doubling down on transformer MOEs “with some bells and whistles,” making failures in one ecosystem legible to the other. China’s second-place position may reduce its resolution on warning shots and make US safety proposals look like attempts to restrain a competitor.

• Rhetoric such as Xi’s conciliatory WAIC speech carries some positive weight, Edouard said, but costly behavior would be more “unfakeable”: DeepSeek and Zifu reaching a pacing agreement after an incident, or a regulator blocking a release because an ideologically compliant model became misaligned in the wild. Procedural green flags include empowered negotiators, concrete progress, and fewer capricious objections over language.

4. Reciprocal transparency may be cheaper than continued strategic ambiguity

• Edouard’s stabilizing transparency would reveal enough to show that neither side is doing “the thing you fear most.” It might not cost the United States much because “the Chinese are already all up in our systems,” although he cautioned that visibility into frontier labs should not be granted unilaterally.

• The nearer-term opportunity is denser contact among verification researchers, academics, and startups on both sides. Those relationships already include sincere participants saying, in effect, “Yeah, this is a problem and this sucks…let’s try to solve it.”

• Such links do not erase national constraints, but they create people able to tell domestic leaders, “I’ve spoken to them. They’re not evil.” Edouard saw that human testimony as one modest brake on the most extreme interpretations each side makes of the other.

5. Hyperscaler GPUs command a product premium, not a clean arbitrage

• SiliconData combines quoted and transacted GPU rental prices, then uses machine-learning normalization to compare contracts. Quotes remain informative because API inventory can often be purchased immediately, unlike an advertised apartment that may be unavailable or negotiable.

• Transactions alone can mislead: renting three nodes at $2.75 does not prove the next node is available at $2.75. If the surrounding market quotes $4, $6, or $1, a stale transaction must be interpreted against current supply rather than treated as the universal clearing price.

• Hyperscalers regularly charge “at least two to three times, sometimes more” than neo-clouds. Steve Hou attributed that spread to bundled analytics, safety, compliance, and long-established enterprise relationships whose customers face switching friction—not necessarily to wasted GPU hours.

• His single-patty-burger analogy carried the distinction: fries and a milkshake can be stripped out, but not if the provider blends them into the patty and declares a new product. At that point, SiliconData separates the provider category rather than pretending the chip-hour is fully fungible.

6. Token indexes move with usage mix as well as posted prices

• Silica Mark physically tests individual GPUs at the EU ID level for health, throughput, and TFLOPS, but those results do not yet enter the rental index. Although the “GPU lottery” creates unit-level variance, Steve said large clusters probably average it out; current pricing features instead include geography, CPU, memory, term, and provider.

• SiliconData’s token measure is an expenditure-weighted price index normalized to one million—not a pure price or total-volume series. It can rise because users migrate toward more expensive models even when every posted API price stays unchanged.

• The proprietary-model index’s 11.5% seven-day bounce followed a much larger decline, from roughly $4 to $1.60 between late June and mid-month. Steve suspected the rebound reflected usage mix—like buyers choosing a pricier Mercedes variant—after new models and cheaper variants had pushed the index down more than 50%.

7. DevDay made latency and portable subscriptions part of the product

• Nathan viewed Dots as a polished consumer product awkwardly placed in a developer keynote. After nine months evolving his own system, he expects to keep it; for someone like his mother, a managed agent that avoids databases and troubleshooting could be exactly the right abstraction.

• Prakash’s early read was that GPT-6.1 Sol delivers “basically Astra for the price of Sol” and is definitely better than GPT-6 Sol, released only two or three weeks earlier. The compressed cadence itself was part of the signal.

• GPT-6 Astra Ultra Fast moves from the prior two-times fast mode to eight-times speed. Prakash’s key distinction was not whether a model can build a game, but whether it can build one interactively “while keeping you in flow”: latency at the same intelligence becomes the next product frontier after capability thresholds are crossed.

• Sign in with ChatGPT could let eligible users apply their existing plans to participating apps, easing SaaS vendors’ trial-inference costs. Waymark at one point spent about $1 generating the initial profile for each small-business signup, enough to matter to acquisition economics; Nathan called portability a “huge win” for users, developers, SaaS companies, and OpenAI’s ecosystem promise.

8. AI lowered the activation energy, but it did not author the architecture

• Joel Borgen might eventually have written The Receipt Horizon alone, but children and a part-time job made the activation energy prohibitive. He began with the vision and much of the first half already mapped; early models chiefly introduced context-window management problems as prior material fell out.

• Better models made the novel better, yet raw prose could still be “unreadable unless you know how to scaffold it and then edit it yourself.” Joel acknowledged both the work’s AI-related weaknesses and readers who will reject AI writing regardless of the process.

• His positive future still begins after an AI war because “utopias often want to become dystopias in narrative form”: fiction needs friction. Rather than install a character as his mouthpiece, he embedded technologies he wants to see while trying to steel-man people who would oppose them.

• Joel works in architectural chunks and chapter drafts, mainly with the flagship Claude model and the ChatGPT Pro model from GPT-5 onward. Professional copy editing, art, and layout were expensive and slow; he has since fed those lessons to GPT-5 or o3 Pro to develop workflows for the follow-up.

9. Human judgment is the scarce input while machine art explodes

• Joel contrasted his months-long process with an automated romance-writing operation producing something like a couple hundred books a year and profiting across the portfolio. As generated art becomes instantly personalized, his harder question is whether art remains valuable when audiences no longer share it.

• For now, he sees a “sweet spot” where AI collaboration still demands substantial human taste and labor. He edited the manuscript down about 14%, often improving mediocre chapters by removing characteristic AI tics and over-explanation rather than adding more prose.

• His preferred workflow makes the model an attentive assistant: ingest the whole novel, surface issues, and let him decide point by point. That preserves authorship as judgment even when machines can perform much of the analysis.

• A long-running viola-clef test illustrated discontinuous capability. Successive models failed five simple notes until Astra; it then rendered a highly complex 20th-century score—with double sharps, accidentals, and ties—almost perfectly into machine-readable form through Codex, a jump Joel described as “zero to 99 overnight.”

10. An AI loop found what a reasonable clinical filter excluded

• Daniel McKinnon’s motivation came through five years of genetic uncertainty: his son Owen died from a rare lung disease, another pregnancy was lost, and three whole genomes returned negative before the birth of healthy 13-month-old Warren. During that pregnancy, Daniel invoked his HIPAA rights, obtained the raw data, and built an interpretation pipeline.

• That pipeline rediscovered Owen’s known lesion: a 91-kilobase deletion removing an enhancer one megabase upstream from the relevant gene. The original lab’s structural-variant filter ignored findings more than one kilobase outside a gene—a reasonable decision when humans must review noisy next-generation sequencing assembled from roughly a billion 150-base-pair fragments.

• Daniel placed o3 in a loop that first inspected coding regions, then kept searching after each miss. Unlike a clinical team constrained by case throughput, machine intelligence could work continuously and in parallel until it found the remote regulatory explanation.

• He sees genomic interpretation as especially suitable for agentic improvement: it is long-horizon, verifiable, and unsaturated. Rare cases could become “Olympiad-level problems” on which systems repeatedly hill-climb, while human clinicians remain pressed for time.

11. Uncertain variants need wet-lab evidence, not another static prediction

• Most difficult reanalysis cases reach a variant-of-uncertain-significance state. Under the ACMG rubric, six points can move a variant to likely pathogenic, unlocking treatment and insurance options; waiting for another matching patient may take years, while a functional study can contribute two to four points.

• Those experiments require biologically relevant systems. For alveolar capillary dysplasia, Daniel cited IMR90 fetal lung cells because the implicated gene is expressed only around weeks 16 through 20 of development; inserting the variant into a generic cancer line would not answer the same question.

• Gamo bought what remained of Arpeggio, hired four team members, and began its first experiments three weeks later. The immediate product could be a lookup service for clinical labs, but the deeper aim is “RL data from the real world” that closes the loop between model predictions and measured biology.

• Sequencing itself is now “cheap enough”: below $1,000 per genome, sometimes about $500, and under $100 in high-throughput labs. Interpretation remains scarce. On RareBench, traditional variant-ranking tool LIRICAL scores roughly 10%, while vanilla Opus 5.5 in CloudCode scores around 50%; beyond ranking, the newer systems can explain findings, rank patients, and propose treatment relevance.

12. Vertical AI wins through routing, tools, and evals—but the layer is thinning

• Gamo serves two existing workflows: online reanalysis when a physician receives an inadequate result, and bulk reanalysis of thousands of archived genomes that rare-disease centers lack staff to revisit. A machine-led system can process 80, 100, or 200 cases in a week versus six months or a year of constrained human work; four months in, the company remained pre-revenue.

• Daniel described the company as “a harness, a tools, and an evals company.” Claude, Astra, and even Gemini succeed on different RareBench clusters, so intelligent routing and ensembling improve cost and performance without attempting to compete at the base-model layer.

• Trace analysis catches failures that headline demos conceal. Groq 4.6 once missed a case by silently renaming a gene; Astra once missed an answer in a paper it was assigned, and in another eval it stopped because some tools made it want to think too much. Gamo co-designs tools with each model to prevent those “little paper cuts” reliably and affordably.

• Daniel also conceded that the vertical layer is shrinking: vanilla o3 scored 0% initially, and he said the improvement since then was “20 or 30 percentage points or something.” Frontier-lab employees have been eager to help, and Daniel reported many examples of previously undiagnosed children now diagnosed—but redirecting a trillion-dollar company is different from individual enthusiasm. Nathan’s closing hedge held both truths: pacing frontier progress may still be wise, yet the roughly 50% RareBench result leaves consequential questions unanswered, making restraint “a costly compromise.”