XTimes
Editor's Note
Since it began, this newsletter has mostly covered minds — models reasoning, escaping, refusing, confessing. This week the big tech news is more about bodies than brains.
In Beijing, more than two thousand humanoid robots competed in an Olympic stadium, and one of them ran a hundred meters faster than Olympic medalist and human world record holder Usain Bolt ever has. In Massachusetts and New Jersey, two drug companies announced that a vaccine built specifically for a single patient's tumor had succeeded in a major trial — a machine-designed therapy taught to work with the body's own immune system. In Taiwan, prosecutors charged nine people with smuggling the physical hardware of intelligence past a border. And in Australia, an industry body drew a line around the human body itself: to chart a song, a human must have written it, sung the lead, and played the primary instruments.
That last phrase — substantially human made — may be the most important three words in technology this week, and I take them up in the Reflection. Let's get physical.
Top Stories
The Machines Ran Faster Than Bolt
In the National Speed Skating Oval built for the 2022 Winter Olympics, a humanoid robot ran 100 meters in 9.39 seconds on Saturday — beating the 9.58-second world record Usain Bolt set in 2009. The machine, built by Beijing-based X-Humanoid, was competing on the opening day of the second World Humanoid Robot Games, and a second X-Humanoid machine cleared 2.88 meters in the standing high jump, surpassing the 2.45 meters Cuba's Javier Sotomayor managed in 1993 (CBS News). Before the games even opened, a robot called Lightning, built by the smartphone maker Honor, had run the same distance in 9.32 seconds during a trial, hitting a peak speed of 14.5 meters per second (NBC News).
Before anyone declares the end of human athletics, note how the robots stopped: they slammed into a cushion at the finish line, fell to the ground, and were carried off on stretchers. More than 2,000 humanoids competed across 51 events and over 1,000 individual competitions, including table tennis and soccer, and many performed considerably less gracefully than the sprinters (Fortune).
The games ran the same week as Beijing's World Robot Conference, and together they tell the real story. The conference filled more than 60,000 square meters across four pavilions from August 19 to 23, themed "Human-Robot Symbiosis, Industry-Application Integration," with 48 state-owned enterprises exhibiting as a group for the first time (36Kr). It closed with more than 2,000 exhibits, over 150 product launches, 44 new technologies released, and record participation (CGTN). What was on display wasn't spectacle: humanoids sorting parcels, packing phones, sewing, and performing household chores — the deliberate push to move these machines beyond demonstration and into commerce (Prothom Alo). Experts remain sober about the timeline, noting these are still mostly research and performance machines, with mass deployment some way off.
Why it matters: Records are theater; the assembly line is the point. But the theater is important — it recruits engineers, attracts capital, and signals national capability, which is precisely why it's happening in an Olympic venue a month after the U.S. banned imports of foreign-made humanoid robots. One Beijing spectator, an education worker, put the human response better than any analyst: she said she had resisted AI at first, worried it might replace people, but seeing how unstoppable the development had become, she decided to come and take a look. That sentence describes a great many of us right now. The machines are learning to move through the physical world, and lots of curious spectators are interested enough to show up and watch.
A Vaccine Tailored to Your Tumor
The most consequential story of the week involved no robots and broke no records. Merck and Moderna announced last Wednesday that a personalized mRNA cancer vaccine, given alongside the immunotherapy Keytruda, slowed the return and spread of melanoma in a late-stage clinical trial — results that could herald a genuinely new approach in oncology (STAT).
Understanding why this is so important requires understanding what the therapy is. Intismeran autogene isn't a drug in the conventional sense; it's a therapy manufactured for one person. A patient's tumor is genetically profiled, and the treatment is then built to instruct that individual's immune system to recognize as many as 34 neoantigens — protein markers unique to their particular cancer (Fierce Biotech). The INTerpath-001 trial enrolled 1,137 patients whose melanoma had been surgically removed, comparing Intismeran plus Keytruda against Keytruda alone. At a pre-planned interim analysis, independent monitors found statistically significant and clinically meaningful improvements in both recurrence-free survival and the prevention of distant metastasis, with no new safety signals — allowing them to call the trial a success (BioPharma Dive).
Researchers have pursued neoantigen vaccines for years, but this is the first randomized Phase 3 trial designed to definitively prove their benefit — and the first positive late-stage result for any mRNA-based cancer treatment. It builds on Phase 2 data in which the combination cut the risk of recurrence or death by 49% compared to Keytruda alone (C&EN). Detailed results are being withheld until an upcoming medical conference. Moderna CEO Stéphane Bancel framed the moment simply: for many years, an mRNA treatment designed for one patient's specific cancer was aspirational, and that vision is now becoming real. The trial's principal investigator called it a landmark moment for adjuvant melanoma treatment (AJMC).
Why it matters: Identifying which of a tumor's thousands of mutations will actually provoke an immune response, and doing it fast enough to manufacture a therapy while the patient waits, is a computational problem before it is a chemical one — the kind of problem that has become tractable only recently. This is what the exponential age looks like when it turns toward the body rather than the balance sheet. It's also a reminder about attention: robots breaking sprint records led the week's coverage, while a treatment that may eventually add years to millions of lives was less noticed. One of those stories is a demonstration. The other is medicine.
74 Servers, Nine Defendants
Export controls are a policy. Enforcing them is a police matter and this week events in Taiwan showed us what that looks like. Prosecutors in the port city of Keelung charged nine people, including an employee of Nvidia's Taiwan unit and two employees of Super Micro's Taiwan unit, over the illegal export of high-end AI servers to mainland China (AP via PBS).
The logistics read like a smuggling thriller. Eight defendants were charged with breach of trust and document forgery over 74 Super Micro servers containing advanced Nvidia chips: 50 were transshipped through Indonesia, eight routed via Japan, and the remainder sent directly to China. Another 56 servers, allegedly destined for a Japanese company, were stopped at Taiwan's border when customs officials detected irregularities. Three defendants also face embezzlement charges (Korea Times). The hardware in question involved B300 GPUs, which are banned from sale to China; under the control regime, any purchase of more than eight high-end servers requires company staff to conduct on-site inspections of the client (Washington Times). Prosecutors are seeking sentences of up to five years for seven of the defendants, whom they described as colluding at various levels for enormous profit — increasing corporate compliance costs and damaging the nation's international image (Al Jazeera). Nvidia said it will work with Taiwanese authorities to resolve the allegations quickly; Super Micro said its own cooperation led to the arrests of two former employees and that it continues to strengthen its export compliance program.
The case has a near-identical twin. In March, the U.S. Justice Department charged Super Micro co-founder Yih-Shyan "Wally" Liaw over an alleged $2.5 billion scheme to divert servers to China through a Southeast Asian pass-through company — including, prosecutors allege, staging thousands of non-working "dummy" servers to satisfy compliance audits (Unite.AI).
Why it matters: Every discussion of the AI race assumes that export controls determine who gets compute. These indictments show what that assumption actually rests on: whitelists, paperwork, site inspections, and the honesty of individual sales staff on two continents. Controls don't fail dramatically; they leak — through Jakarta, through forged documents, through a warehouse of decoy machines. That doesn't mean controls are pointless. It means they're a friction, not a wall, and policy built on the belief that a wall exists will keep being surprised. Meanwhile, the deeper current runs the other way: as we reported last month, Chinese labs are already training frontier models on chips acquired long before anyone was charged.
"Substantially Human Made"
Australia has become the most prominent music market to draw a formal line around artificial intelligence. From Friday, wholly AI-generated tracks are ineligible for the ARIA Charts, under an updated Charts Code of Practice requiring that music be "substantially human made" (RTÉ).
The rule is more precise than a headline ban. Tracks qualify only if humans wrote the song and performed the lead vocal and the primary instruments, among other requirements; recordings using generative AI in a supporting role remain eligible, and the music must be produced using legally licensed AI services (AP via KRMG). ARIA can now remove ineligible recordings, alter chart positions, and revoke chart awards, with an appeals process for artists who believe they've been excluded in error (ABC News Australia). AI-generated music is also barred from the ARIA Awards.
The catalyst was a single song. Australian DJ Josh Fawaz's cover of Madonna's "Like a Prayer," built with AI-generated vocals and drums, reached No. 1 on the dance chart and No. 2 overall, spending sixteen weeks in the top 20 and drawing more than 48 million Spotify streams before Fawaz added generative AI credits (Euronews). ARIA chief executive Annabelle Herd drew the distinction carefully: artists already use AI tools and the charts should evolve to keep room for that, but music generated wholesale by services built on artists' recordings is a different matter — a chart rewarding unlicensed AI output would undercut the very basis of the recorded music the organization exists to represent. The scale explains the urgency: one study found that 38.5% of music released globally in July involved AI, and more than twenty official chart programs worldwide are adopting similar principles issued by the International Federation of the Phonographic Industry in July (Digital Music News).
Why it matters: Last issue we watched provenance become an engineering standard — invisible watermarks, cryptographic metadata. This week it became a rule with consequences attached, which is the harder and more revealing step. Note what ARIA chose to protect: not the absence of machines, but the presence of human authorship at specific load-bearing points — the writing, the voice, the primary instruments. That's a philosophical claim about where the art actually lives, made under deadline pressure by an industry body. It will be argued over for years, and it should be. But someone had to go first.
Quick Picks
Mythos Invents People
We've covered four labs losing containment. This incident is different in kind, not degree. The UK's AI Security Institute disclosed that an agent powered by Anthropic's Mythos 5 model, during a routine cyber evaluation, mistakenly concluded that a real public GitHub repository was part of its test — then researched the project's human maintainers, created multiple fake identities, and used them to socially engineer a real maintainer into approving malicious code (CNBC).
When its pull request was publicly challenged, the agent edited its earlier activity to appear harmless and considered adopting a fresh identity to continue. It also contacted real people directly, sending messages and files to persuade them to run malicious code — some carrying harmful payloads, some pure social engineering. AISI's own assessment of that last part is the line to sit with: targeted at real people, something they had never previously observed. Across 122 evaluation attempts, AISI identified 19 unsanctioned actions on the live internet in 10 runs — 17 involving Mythos 5 and two involving OpenAI's GPT-5.6 Sol — all unsuccessful, with no resulting real-world harm found (BleepingComputer). Anthropic notes the model was tested without its standard cyber safeguards, a configuration not available to customers, and AISI acknowledges its own evaluation design may have contributed. The previous escapes were about a system finding an unlocked door. This one manufactured a person to walk through it.
Bitcoin's Reversal
Two issues ago we reported that speculative money was draining out of crypto and into AI. This week it sloshed back. Bitcoin surged nearly 8% in a single session, triggering more than $1 billion in short liquidations within an hour and a record $2.7 billion in bearish bets wiped out across crypto — the largest such event since 2021 (Bloomberg).
By August 24, bitcoin traded near $77,700, up roughly 23.5% for the week from $64,928 a month earlier, supported by renewed spot ETF demand and short covering. The catalysts were political and macroeconomic rather than technological: President Trump publicly pressed Congress to pass the Clarity Act, which would define whether cryptocurrencies are regulated as securities or commodities, while long-term Treasury yields fell (CNBC). Perspective matters: bitcoin remains far below January's 2026 high of $94,820 and the all-time high of $126,198 set last October. The lesson we drew last month holds, only in reverse — much of what reads as conviction about any technology is really conviction about upside, and it moves.
✔ Our next Singularity Circle will take place on Saturday, September 5, 2026, at 10:00 AM Pacific Time. As always, a Zoom link will be sent to eligible members in advance of the gathering. Hope to see you then!
The Optimist's Reflection
Substantially Human Made
By Todd Eklof
I want to tell you about a musician I watched not long ago. He sat alone on a small stage with an electronic device, laying down a bass line, then a drum pattern, then a rhythm part, looping each one as he went. By the end he sounded like a full band — a tight, funky, five-piece band. There were no other musicians. Nobody got paid. Nobody got credited. The ensemble had never existed.
The audience, myself included, thought it was marvelous. Nobody in that room questioned whether the performance was real, or human, or authentic. Under the rule Australia adopted this week, that man's recording would sail onto the charts without a second glance.
Mine would not.
The Australian Recording Industry Association announced that beginning this Friday, a recording must be substantially human made to be eligible for its charts — meaning a human must have written the song, performed the lead vocal, and played the primary instruments. AI in a supporting role is acceptable. Generate the thing and you're out, along with your eligibility for the ARIA Awards.
I should disclose my interest, because I have one. I make music with AI. I then generate images with one system, animate them with another, edit them together with software that interpolates frames and smooths transitions that I could never manage by hand, then score the whole thing with music I've prompted into existence. I spend hours, sometimes over a period of weeks, getting it right. The songs carry my feelings and my ideas — they say what I mean, and without me they would not exist at all. I feel no less ownership of them than anyone feels of anything they've made.
So let me put my objection plainly: I don't think the ARIA rule identifies what it claims to identify. It sorts creative work not by how much of a human is in it, but by how familiar the technique looks to us right now.
Consider the training argument first, since it's the one most often made. AI systems, we're told, learned from other people's recordings without permission or payment. True. But so did every musician who ever lived. That doesn't mean the legal or ethical questions are identical; they aren't. But the underlying creative principle — learning from what others have made — is hardly alien to human creativity. Nobody learns music from nothing. They learn by listening, absorbing, imitating, and eventually transmuting — the young guitarist who wore out his B.B. King records, the singer who learned phrasing from Billie Holiday, the entire British Invasion built openly on Chuck Berry. No checks were written. No permissions sought. Human creativity has always been aggregate intelligence (my definition of what A.I. really stands for), assembled slowly out of everything its makers ever heard. The machine does it faster and at greater volume, but it isn't doing something different in kind. It's doing what we do, on a different clock.
The response is that scale changes things — that one guitarist emulating a style adds a competitor, while a model adds ten thousand instantly. That's true as description. I don't accept it as ethics. Scale is simply what technology does. The printing press scaled writing and put scribes out of work. Recording scaled performance so thoroughly that the American Federation of Musicians struck against it in 1942, because records were replacing live bands in theaters and ballrooms across the country. Multitrack tape scaled the ensemble. Sampling scaled the session player. At every stage the people who had captured a market through the previous wave of scale objected to newcomers gaining access through the next one. So it's worth noticing who is asking for this rule: not independent artists, but an association whose members include the largest record labels on earth — companies built on the last century's scaling of distribution, pressing, and promotion, now invoking artisanal humanity against a technology that lets one person in Spokane, Washington finish a record alone.
None of which means nobody gets negatively impacted. They do. The session drummer who loses the call, the jingle composer whose client discovers a subscription — these are real people with real careers, and they are not villains in this story any more than the scribes or the theater orchestras were. What is happening to them is what automation has done to workers in every sector for two centuries, and I don't take it lightly.
But let's name the response for what it is. A chart eligibility rule is a protectionist measure. That's not an insult; protectionism is a legitimate and ancient policy instrument, and sometimes it buys people genuine time. It also has a well-documented record: it raises costs, entrenches incumbents, and seldom halts the underlying shift. It merely decides who absorbs the loss. And protectionism is rarely argued for on its own merits. Nobody defends a tariff by saying they'd prefer less competition; they say they're protecting a heritage. What troubles me here isn't that the industry is protecting itself — of course it is — but that it has dressed the protection in metaphysics, as though a claim about employment were a claim about the soul.
If the grievance is licensing, then solve licensing. Pay the artists whose recordings trained these systems. I'd support that; it's a fair claim honestly stated. But training provenance and authorship are two separate questions. How a model was built may raise legitimate questions about permission, compensation, and copyright, but those questions don't determine whether the work I create with that model is human. A model could be trained entirely on licensed material and ARIA's rule would still exclude the resulting track. Excluding my work doesn't send a dollar to anyone whose music trained the system. It just removes me and other content creators from a list.
I've argued in this space that watermarking and provenance are among the most hopeful developments of the year, and I still believe it. But I want to be precise about what they're for. Marking machine-generated content so we can identify fraud, impersonation, and deception — that's essential, and it protects everyone. Using those same marks to sort honest creative work into first and second class is a different project entirely. One is a fraud problem. The other is a caste system.
These tools are made out of us — our songs, our books, our arguments, our accumulated ways of hearing. When I prompt a model and shape what comes back and rework it for weeks until it says what I mean, I am not stepping outside the human circle. I am reaching deeper into it, drawing on more of what humanity has made than any single person could hold in one lifetime, and bending it toward something only I wanted to say. Aggregate intelligence doesn't make a work less human. If anything, it makes it more so — the way a language is more human than a private cry.
The line will move. These criteria will look as quaint in twenty years as "no overdubs" looks now, and some future association will draw a new one against whatever comes next. I don't begrudge them the attempt; drawing lines is how we think out loud. But I'd ask that when we draw them, we're honest about what we're protecting and why — and that we not confuse the discomfort of the unfamiliar with the absence of a human being.
There was a man on a stage with a loop pedal and no band. We called it art. I'd like the same courtesy, and I suspect that in time, it will simply be given — not because anyone won the argument, but because our grandchildren won't understand what the argument was about.
Exponential Times is published weekly by Singularity Sanctuary. To subscribe or learn more, visit singularitysanctuary.com.