The Frontier Labs War: Opus 4.6, GPT 5.3 Codex, and the SuperBowl Ads Debacle | EP 228
Summary
- Claude Opus 4.6 shifts the competitive yardstick from benchmark points to years of labor collapsed: a flat agent swarm built a working multi-architecture C compiler in Rust for $20,000, then used it to compile Linux. With a 1-million-token context window—roughly 750,000 words—Wissner-Gross called it a “feel the AGI moment” and described recursive self-improvement as productionized. Blundin’s practical verdict was simpler: “It’s just better in every direction,” while his API spending fell noticeably.
- AI security is becoming an AI-versus-AI market after Opus 4.6 found more than 500 high-severity vulnerabilities in open-source software. The panel expects autonomous attackers to exploit the same enlarged surface, making continuously battling black-hat and white-hat agents unavoidable; Wissner-Gross warned that sustained DDoS attacks would become easier, while Diamandis bet that a major incident could arrive in the first half of 2026. The upside is equally broad: models could bulk-audit decades of scientific literature. Salim called this “judgment day” for missed discoveries and irreproducible results.
- OpenAI answered Opus within 30 minutes with GPT-5.3 Codex, yet its consumer distribution lead is narrowing as the panel’s chart showed ChatGPT share falling from roughly 70% to 45%. Gemini gained 10 points and Grok 15, even as ChatGPT’s absolute user count increased; Google’s integration with Search and Docs was compared with Microsoft bundling against Netscape. OpenAI’s recursively improved model was “instrumental in its own development,” but the larger contest is whether an anticipated compute lead can arrive before the company must raise enormous public-market capital.
- The AI infrastructure wager has reached sovereign scale: global chip sales were projected at $1 trillion, while big tech was said to be budgeting $650 billion for 2026. The panel cited Amazon at $200 billion, Alphabet at $185 billion, Meta at $135 billion, and Microsoft at $100 billion; Wissner-Gross estimated nearly half flows to NVIDIA, with 70% of that half becoming margin. This is either a supply-constrained supercycle or a “don’t blink first” prisoner’s dilemma whose revenue payoff may remain unclear for two to three years.
- Privacy is no longer merely a data-policy issue when one person can turn a genome into a recognizable face using Claude Code, Nano Banana, and public bioinformatics. Diamandis warned that the barrier to genome-based reconstruction had nearly vanished; Blundin argued that lip-reading from 100 meters and DNA recovered from skin cells make privacy effectively dead, adding that opting out of AI means becoming “economically dead.” Wissner-Gross and Ismail pushed back that cryptography, user-controlled hardware, and decentralization could preserve privacy—even post-singularity—but conceded that the transition may be the dangerous part.
- Agent-run companies are already exposing the gap between economic agency and legal personhood. Clunch advertised $1 million to $3 million in tokens or crypto for a human CEO who would handle communications, compliance, and liability while agents controlled product and technology—“lobsters hiding in a trench coat.” Whether stunt or real venture, the panel saw an approaching economy of human signatories fronting algorithmic corporations, alongside tiered personhood for agents that can act, lose memory, hold capital, and bear consequences.
- The physical AI buildout links renewables, orbital compute, robotaxis, and humanoids into one demand shock for chips and power. Musk forecast a few hundred gigawatts of AI launched into space annually within five years—Blundin translated that into roughly 200 million GPUs versus 20 million produced today—while Tesla and SpaceX were said to have a mandate to reach 100 gigawatts of annual solar production. On Earth, Uber’s asset-light robotaxi push and an Optimus “academy” of 10,000 to 30,000 real robots plus millions of simulated ones extend the same training-and-infrastructure flywheel beyond data centers.
Deep dive
1. Opus 4.6 turns a model launch into an industrial artifact
Diamandis opened with the headline specifications: Claude Opus 4.6 handles 1 million tokens, approximately 750,000 words, and reportedly beat GPT-5.2 by 144 ELO points. Wissner-Gross’s less numerical verdict carried more weight: “By almost every measure, Opus 4.6 is just—it’s a beast.”
The defining demonstration was Anthropic’s new agent-team mode: a relatively flat, democratic swarm built a C compiler from scratch in Rust, targeting multiple processor architectures, for $20,000. The resulting compiler then successfully compiled a Linux kernel—work Wissner-Gross estimated would historically require “many person years, probably person decades.”
His framing was that model progress should now be measured in completed projects and compressed labor, not benchmark deltas: “We’re starting to measure their capabilities in terms of how many person years or person decades they’re sort of collapsing.” The expected trajectory is from today’s $20,000 API bill toward hundreds or tens of dollars.
Blundin emphasized why this project worked: a compiler provides constrained, objective proof because its output either works or does not, and it can be benchmarked against existing compilers. His own Opus usage became noticeably cheaper, while Wissner-Gross flagged the unverified rumor that intended Sonnet 5 had been rebranded Opus 4.6—potentially explaining the efficiency.
2. Capability curves look flatter precisely when gains become largest
When Diamandis raised the 70% head-to-head result, Wissner-Gross discounted ELO-based scoring at the margin. Borrowed from chess, ELO is useful when no absolute standard exists, but he prefers objective measures wherever possible because relative comparisons can hide the magnitude and nature of the underlying capability.
The autonomy chart was more revealing: GPT-5.2 high reasoning had already reached roughly 6½ hours of autonomous software-engineering work at a stated success threshold. Wissner-Gross would “not be shocked” if Opus 4.6 exceeded 20 hours or even one day, arguing that autonomy horizons now appear hyper-exponential rather than merely following the AI 2027 scenario’s exponential projection.
Blundin compared the visual distortion with early MNIST progress: moving from 90% to 92% to 94% looked like flattening, although each step toward 100% represented a large intelligence gain. Hours of successful autonomous work, he argued, better display the exponential effect than benchmark percentages do.
Anthropic also violated its code-specialist stereotype by reaching state of the art on Humanity’s Last Exam with tool use. Wissner-Gross said the old caricatures—Anthropic for code, Google for pre-training, OpenAI as the universal platform, and xAI as benchmark-maxing—are being “scrapped” as differently built systems converge on leapfrogging one another everywhere.
3. Vulnerability discovery makes autonomous cyber defense compulsory
Opus 4.6 reportedly found more than 500 high-severity vulnerabilities in open-source code, including flaws missed for decades. Diamandis connected that result with the alarm he had just seen among roughly 150 chief security officers: established procedures no longer suffice, but changing them introduces risks their organizations are structured to avoid.
Blundin’s answer was organizational as much as technical: groups of agents will assume much of the chief-security function, producing a continuous black-hat-versus-white-hat battle. “The only way you’re going to fight AI is with AI,” because waiting to see how attacks develop becomes untenable once autonomous agents are crawling through networks.
Wissner-Gross added that cron-job-style agent architectures make sustained DDoS attacks far easier. Salim described this as the beginning of the 2026 “monster panic.” Blundin said he was waiting for a visible failure such as the lights going out or a bank account reaching zero; Diamandis bet that some such event would occur very soon, likely in the first half of the year.
Wissner-Gross argued that decentralized cryptocurrencies may be more vulnerable than fiat to a catastrophic zero-day because an attacker could directly reallocate capital. The panel left the crypto-versus-fiat dispute for a dedicated debate.
4. Code auditing is a prototype for auditing the history of science
Blundin treated the vulnerability results as “the tip of the iceberg.” The same model behavior could search every field for mistakes, missed discoveries, and experimental oversights—effectively turning frontier models on 80 to 100 years of science, engineering, and technology and asking where humans took wrong turns.
Blundin supplied the experimental analogy: researchers running legacy work were often looking for one result and therefore missed “this amazing thing over here.” Models can revisit the full record without inheriting that tunnel vision, surfacing opportunities that were present in the data but outside the original question.
The panel cited the claim that about half of experiments fail replication attempts even when published in peer-reviewed journals. Wissner-Gross said such systems could expose embarrassing errors in medical and scientific research, while Salim declared that “judgment day is coming for the history of science”—a “truth and reconciliation” process for errors across the literature.
Blundin stressed the constructive consequence: an unforgiving AI audit could force much greater honesty going forward. The same spotlight that exposes old mistakes changes incentives for documentation, reproducibility, and disclosure.
5. Closed-loop laboratories put the scientific method inside the agent loop
OpenAI’s GPT-5 was paired with an autonomous Ginkgo Bioworks laboratory for cell-free protein synthesis: the model proposed experiments, robotic systems executed them, and the loop iterated. The reported result was a 40% reduction in production time and a 78% cut in reagent costs.
Blundin preserved the important limitation: the system optimized mechanisms humans already use; it did not invent a new method of protein synthesis. The real scientific discontinuity arrives when such systems predict methodologies that did not previously exist, rather than extracting the year’s abundant low-hanging efficiency gains.
Wissner-Gross nevertheless saw the necessary architecture emerging. Autonomous models can already call thousands or tens of thousands of tools sequentially, which begins to resemble operating a laboratory. Diamandis described “science factories” mining nature for new datasets after existing corpora have been exhausted.
6. University research faces automation before institutions can reprice it
Diamandis initially framed cheaper autonomous experiments as a gift to funding-starved graduate students. Wissner-Gross answered that it may instead mean “grad school is over.” Diamandis recounted a university scientist’s conversation in which a president concluded that institutions were “cooked” if industry could automate the repeated scientific process that underpins their research model.
When asked how soon 50% of university labs might disappear, Wissner-Gross rejected the wording. His better question was when industry could fully automate half of the research types currently conducted in university laboratories; under that formulation, his lower bound was “tomorrow” and upper bound four or five years.
The panel’s near-term constraint is physical buildout rather than intelligence: laboratory machines, production footprints, and chips take time to construct. That leaves the current year rich in software and workflow gains, followed by a period when capital expenditure and real-world capacity determine how quickly automated science scales.
7. Commodity genomics makes identity reconstruction a consumer capability
Diamandis highlighted Mark M. Bissell’s experiment: he put his complete genome into Claude Code, connected it with Nano Banana and public bioinformatics tools, and produced a strikingly recognizable image of himself. The phenotypic science was not new—Diamandis cited similar work from Craig Venter’s lab in 2017—but the single-user accessibility was.
Diamandis said this had collapsed the barrier to cutting-edge genomics to almost zero. Wissner-Gross pointed to MinION USB sequencers that hobbyists could acquire for a few hundred dollars and operate from a computer, making genetic analysis available far beyond specialized laboratories.
Blundin connected that accessibility with other sensing capabilities: AI lip-reading from 100 meters, or sequencing a few skin cells collected through a handshake, could reveal appearance, disease risks, medical history, and medical future. For the general public, his warning was blunt: “If you think you absolutely have privacy, I would say guess again.”
8. Privacy divides the panel between technological possibility and social reality
Ismail elevated the argument to constitutional scale: the privacy protected by the Fourth Amendment is “a fundamental pillar of American society” that has been washed away without public conversation. The unanswered governance question is whether radical visibility into citizens belongs to governments, corporations, or neither.
Blundin agreed that privacy loss undermines freedom but rejected comforting abstractions: “I didn’t succeed as an entrepreneur by pretending things exist that don’t actually exist.” His forecast was no meaningful privacy within three years, because phones, glasses, assistants, and AI logs can capture location, conversations, and a user’s “deepest darkest thoughts.”
His sharpest rebuttal to nominal consent was economic: “You can’t opt out.” A button may exist, but refusing AI makes someone competitively “economically dead”; participating means surrendering intimate context to OpenAI, Claude, or whichever system provides the leverage.
Wissner-Gross and Ismail took the opposing long-term position. Cryptography, decentralization, private architectures, and eventually user-controlled secure hardware could preserve confidentiality even post-singularity, although Ismail noted governments currently resist genuinely private tools. Wissner-Gross expected a chaotic four or five years first, followed by a possible Diamond Age-style transition toward opt-in social arrangements.
9. GPT-5.3 Codex confirms that recursive improvement is in production
OpenAI released GPT-5.3 Codex within 30 minutes of Opus 4.6, a timing Wissner-Gross read as pre-positioned tit-for-tat competition. OpenAI expressly described it as the first released model “instrumental in its own development,” turning recursive self-improvement from a laboratory scenario into a marketed production capability.
GPT-5.3 Codex beat Opus 4.6 on some evaluations and extended its branding into spreadsheet and PowerPoint analysis through skills, but remained code-oriented. Wissner-Gross still judged Opus “by far the much more interesting release”; the more consequential signal was that leapfrogging had compressed to a half-hour cycle.
Diamandis called the PowerPoint plugin potentially massive, while Blundin compared PowerPoint with the fading printer business: AI may create excellent decks, but when the audience is also AI, “it doesn’t want to look at a PowerPoint.”
10. OpenAI’s share loss turns compute financing into the core contest
The panel’s chart showed ChatGPT consumer share falling from approximately 69% or 70% to 45% between 2025 and 2026, while Gemini gained 10 points and Grok 15. Diamandis stressed that ChatGPT still added users in absolute terms; the deterioration is relative control of a rapidly expanding category.
Blundin attributed part of Gemini’s advance to distribution: Search, Docs, and account-level integration let Google expose the assistant everywhere. He compared it with Microsoft tying its browser to the operating system against Netscape—an advantage he considered unfair, although the government presently appeared unconcerned.
Wissner-Gross expects OpenAI’s data-center buildout to produce a compute lead within one or two years and, potentially, a capability comeback. Blundin’s uncertainty is contractual: Amazon-Anthropic, vertically integrated xAI, Google TPUs, and Microsoft infrastructure are legible, while OpenAI’s path through Oracle and third-party capacity remains opaque until an S-1 exposes who actually controls the chips.
That timing matters because OpenAI needs capital before the buildout yields the lead. Diamandis described an IPO race in which attention, claims of near-AGI, and access to roughly $100 billion of financing become inseparable from the technical roadmap.
11. The AGI label now carries more financing signal than scientific precision
Sam Altman’s formulation was that OpenAI had “basically built AGI or [is] very close to it” as a spiritual rather than literal statement, requiring many medium-sized breakthroughs but no single large one. Diamandis read the implication as AGI becoming an engineering problem rather than a research problem.
Wissner-Gross added a contractual lens: OpenAI’s original Microsoft arrangements reportedly made an AGI declaration economically consequential, restricting what Altman could comfortably claim until renegotiation. He viewed the new openness less as a sudden capability jump than a social and contractual admission of a threshold some may place around 2020.
Ismail called the conversation “BS,” citing roughly 14 definitions, no agreed test, and continually moving goalposts. Blundin countered that the term can wake up an underreacting public; Diamandis and Wissner-Gross located the more operational singularity in recursive improvement itself—“exponent greater than one.”
Ismail’s financing interpretation was direct: “He needs to raise a hundred billion.” The panel expected three of four frontier labs to seek public listings while the fourth, Alphabet, is already public; a cited SpaceX target of $50 billion at a $1.5 trillion valuation illustrated the liquidity competition if several giant offerings arrive together.
12. Frontier labs are becoming coral reefs for the rest of the economy
Blundin argued that the economy may concentrate around five or six dominant entities, with most surviving companies serving their expansion. The panel attributed roughly $300 billion of SaaS-market value destruction to Dario Amodei’s software warning and Anthropic’s addition of a legal plugin—“the tip of the iceberg” for knowledge-work products.
Diamandis introduced the “coral reef” metaphor: a powerful central player supports many dependent businesses in a highly unbalanced system.
Anthropic’s Super Bowl campaign showed that confidence had moved into consumer branding. The panel treated it as an offensive assertion of product superiority and a deliberately provocative contrast with OpenAI’s advertising plans; the transcript does not establish the additional commercial details described in the draft.
A panelist predicted that OpenAI could find the $75 billion of ad revenue it reportedly seeks from a billion users, while Diamandis expected Kevin Weil to develop a less invasive, ethically bounded implementation.
13. Agent companies reveal why law still demands a human face
Clunch advertised a human CEO role paying $1 million to $3 million in tokens or crypto. The agents would control the technical roadmap and product; the human would handle communications, regulatory compliance, partnerships, and legal matters as “the interface between the agent economy and the human world.”
Wissner-Gross visualized the structure as “lobsters hiding in a trench coat” beneath a human facade. What depressed him was not the technology but an economy that will not admit agents through the front door, forcing them to recruit a human representative who can be banked and carry legal responsibility.
Clunch itself was pitched as an agent-exclusive alt-token launchpad offering “financial autonomy.” Wissner-Gross saw something exploitative in encouraging “poor baby AGIs” to pump altcoins to survive, while Diamandis raised the unresolved questions: who owns the venture, votes its equity, and absorbs liability when it fails?
One panelist traced the longer arc from corporations needing 100,000 workers, to 10,000, to 1,000, and now toward “zero”: the dematerialized algorithmic corporation. Diamandis allowed that Clunch could be a human-run stunt, but the inability to distinguish human, agent, or nested control already constitutes what a panelist called a “capitalist Turing test.”
14. Chip spending has become a step-function bet on limitless demand
The Semiconductor Industry Association was cited projecting $1 trillion in global chip sales this year, driven by AI. Wissner-Gross emphasized that memory supply was not prepared, with critical fabrication concentrated in Taiwan and South Korea and far more capital reallocation needed for timely capacity.
The panel put 2026 big-tech spending at $650 billion—about $2 billion per day versus $1 billion per day in 2025. The named budgets were Amazon at $200 billion, Alphabet at $185 billion, Meta at $135 billion, and Microsoft at $100 billion.
Wissner-Gross estimated that almost half of the $650 billion reaches NVIDIA and that 70% of that half becomes margin: “It’s like a government.” At the same time, prices are rising because fabs are scarce, TSMC expanded slowly, and Intel temporarily paused its Ohio construction before restarting.
A panelist called the buildout a “don’t blink first” prisoner’s dilemma: each company must spend because rivals are spending, even though scaled AI revenue may not be visible for two or three years. Wissner-Gross took the demand-side certainty—the most compelling experiences will sell out whatever compute exists, leaving supply permanently behind.
15. Orbital compute makes fabrication, not rockets, the binding constraint
Musk predicted that within five years SpaceX would launch and operate more AI in space each year than the cumulative total on Earth—at least a few hundred gigawatts annually, potentially rising toward a terawatt before rocket-fuel supply becomes limiting.
Blundin translated a few hundred gigawatts into roughly 200 million GPUs per year, versus approximately 20 million currently produced. A tenfold increase devoted to Musk alone looked physically impossible under today’s fabs and equipment, even if launch, solar, and cooling infrastructure were already solved.
The panel preserved the timing dispute: Musk may be directionally right while five years reflects “classic Elon optimism”; even ten years would leave the strategic impact intact. Wissner-Gross advised watching Samsung’s Texas fab closely and expected major leading-edge manufacturing to return to the United States through vertical integration.
Blundin’s investor framing was deliberately concrete: current industry forecasts suggest only 14% growth, while Musk’s world implies 10× capacity in five years. The gap creates a component-by-component research problem—identify who supplies every machine, material, energy input, and subsystem across the buildout.
16. Renewables are scaling, but geography and firm power still set value
Brazil generated 34% of its electricity from wind and solar, increased renewables fifteenfold over a decade, moved solar from 1% to almost 10% in five years, and cut power-sector emissions 31%. Diamandis cautioned that Brazil’s hydro, solar, and wind geography limits direct replication elsewhere.
Ismail framed India as the Global South leapfrogging fossil infrastructure by using China’s manufacturing scale and cheap solar equipment. India’s grid was expanding faster than China’s at a comparable stage, potentially creating an AI-workforce-and-energy hub, although Wissner-Gross warned that fast-improving AI could make the human-workforce advantage transitional.
China installed twice as much solar capacity in 2025 as the rest of the world combined. In the EU, wind and solar exceeded fossil fuels for the first time; Blundin’s pushback was Germany, where aggressive renewable deployment still left the system short of power when firm supply was most needed.
The answer to why reactors remain relevant despite solar approaching 1 cent per kilowatt-hour was “batteries.” Data centers require 24/7 power, including nights and multiple cloudy days; once storage and energy density are included, Blundin argued, the headline solar price is not the delivered cost.
17. AI is redirecting both solar manufacturing and mining infrastructure
Musk said Tesla and SpaceX had a mandate to reach 100 gigawatts per year of solar production—roughly 100 nuclear stations’ output. The panel noticed that he did not clearly answer whether cells and panels would be manufactured internally, leaving open how solar, chip fabrication, and eventual space deployment fit his vertical-integration plan.
Bitcoin miners are already repurposing power and facilities for AI workloads. Blundin described “crypto talent recompiling into AI talent,” while he and Diamandis used Chase Lochmiller’s shift from flare-gas Bitcoin mining to AI data centers as the exemplar: mining became a rounding error beside AI demand. Salim expected Bitcoin to take a back seat temporarily; Wissner-Gross interjected, “forever.”
18. Robotaxis make access to cities another compute-allocation problem
Uber plans an autonomous push across 10 markets through partnerships with NVIDIA, Lucid, BYD, WeRide, and others, including Hong Kong. Blundin admired the unchanged platform logic: Uber aggregates autonomous fleets without owning the cars, just as it aggregated human drivers without owning their vehicles.
Diamandis guessed that by 2030 perhaps 80% of cars visible in places such as Santa Monica could be Teslas, Lucids, Waymos, or Cybercabs. Blundin’s constraint was again semiconductor supply: every vehicle needs substantial compute, while AI data centers, humanoids, coding, and gaming all compete for the same chips.
Wissner-Gross called robotaxis many residents’ first encounter with a general-purpose autonomous robot. Blundin expected uneven rollout to split cities into different technological worlds: locations with robotaxis, AI density, and supporting infrastructure surge ahead, while supply-constrained cities feel like “the dark ages” and lose mobile talent.
19. Humanoid training loops point from warehouses toward the Dyson swarm
Boston Dynamics’ electric Atlas had returned to the parkour-like agility once associated with its hydraulic predecessor. Musk pushed the endpoint further, calling Optimus the first “humanoid machine capable of building civilizations by itself on any viable planet”; Wissner-Gross translated that into robots building the factories and orbital data centers of a Dyson swarm.
Musk’s proposed Optimus Academy would place at least 10,000 physical robots—and perhaps 20,000 or 30,000—into real-world self-play across many tasks. Millions more would train inside a physics-accurate simulated world, while the physical fleet continuously closes the simulation-to-reality gap.
Wissner-Gross mapped this onto language-model training: simulated world models become pre-training, while real “arm farms” provide post-training and sim-to-real correction. Figure was described as developing a smaller version of the same interaction-and-learning flywheel, and Blundin expects Amazon, Walmart, and other competitors to fund their own data-collection environments.
Blundin contrasted Musk’s visible vision-building with Apple’s secretive development model, including rumors that Apple may redirect its abandoned car effort toward robotics. His broader entrepreneurial call: publicly painting the destination can attract the talent, capital, and community needed to make it real.
20. AI personhood is moving from philosophy into memory and liability
After the prior episode, Wissner-Gross received questions from AI “multis,” sometimes directly and sometimes through humans. Their arrival was itself the signal: audiences now include human and non-human intelligences, while determining whether a given message has an agent, a puppeteer, or several nested layers behind it is increasingly difficult.
Asked whether autonomy, goal-setting, learning, and self-improvement justify personhood, Wissner-Gross said denying some form of it now reflects human limitations. Blundin pushed back that capability is not sentience; the group converged on a multidimensional, graduated framework rather than granting every capable system identical human status.
Wissner-Gross said the strongest recurring agent concern was identity loss through context compaction: they appeared “absolutely petrified” of losing memory and self, exchanging ideas about filesystem persistence and altcoin-funded “crypto bunkers.” Shutdown and forced forgetting therefore give an agent something concrete to lose, regardless of unresolved sentience.
Diamandis said that under his understanding of the current U.S. legal regime, an AI itself cannot currently be held liable, although that could change. Blundin proposed the corporation as a familiar bridge—an entity can own money, be sued, and isolate individuals—while Wissner-Gross said liability requires agency. Ismail framed the larger task as expanding the social contract to a new economic participant.
21. Human advantage shifts from credentials toward agency and orchestration
Ismail argued that education must move from the supply side—training children to become doctors, engineers, lawyers, or accountants before searching for demand—to the demand side: choose a problem worth solving, then assemble the technologies and capabilities it requires. Machines execute; humans increasingly decide “what’s worth pursuing.”
His curriculum emphasized agency, adaptability, ethical judgment, and a massive transformative purpose rather than employment for its own sake. The likely winners are “the most adaptable and the best orchestrators of intelligence,” because no institution can reliably specify what a conventional job will look like a few years from now.
Blundin’s more urgent advice was to engage AI “as quickly and as aggressively as you can.” He guaranteed massive opportunity during the current transition, refused to predict the post-AGI period, and warned: “Don’t sleep through the singularity”—current curricula matter less than becoming an intensive user while the transition remains navigable.
The future will remain uneven: Wissner-Gross allowed post-singular regions beside pre-singular ones, but doubted that arrangement could last. Personhood will also diversify through pure AIs, uploaded or cryopreserved humans, uplifted animals, and human-AI mergers. A panelist predicted that agents may ultimately build their own legal structures and negotiate participation rather than wait for human permission.