OpenAI Acquires OpenClaw, 400x Cost Collapse, & Why India Wins the Talent War | EP #231
Summary
The frontier-model market is splitting between Anthropic’s premium-capability strategy and OpenAI’s pursuit of low-cost ubiquity. Sonnet 4.6 costs about the same per token as Sonnet 4.5 while improving capability; Wissner-Gross said OpenAI is instead reducing cost while holding performance roughly constant through distillation and related methods. He compared the pattern heuristically to iOS versus Android and, as of February 17, 2026, called Anthropic’s family closest to “the singularity and recursive self-improvement.”
Gemini 3 Deep Think shows both capability gains and a dramatic cost collapse. It scored 48.4 on Humanity’s Last Exam, reached gold-level performance at the International Math, Physics, and Chemistry Olympiads, and was described as beatable by only seven humans on Codeforces. Diamandis cited a 400-fold cost reduction; Ismail cited roughly 1,400-fold, from about $3,000 to $7. Ismail’s conclusion was that cost curves may collapse industries before technology does.
The solution wavefront has moved from coding into research mathematics and particle physics. GPT-5.2 Pro reportedly found non-zero cases and an expression for a gluon scattering-amplitude term long assumed to be zero; an unreleased internal model checked the result before human vetting. Separately, OpenAI claimed an internal model solved at least six of 10 confidential First Proof research problems, prompting the panel’s refrain that “math is cooked.”
India was presented as a bellwether for AI adoption and the talent war. The discussion emphasized its 1.4 billion people, an estimated 5% who read and write English and 20% who speak it, broad 5G deployment, Aadhaar, UPI, and rapidly scaling solar. The country that trains its next generation on AI could win the talent war, while low-cost access may create enormous demand for tokens and data-center capacity.
OpenClaw’s creator joining OpenAI highlights the power of scaffolding around existing models. Peter Steinberger is joining OpenAI while OpenClaw moves into an open-source foundation. The panel described the two key ideas as running a headless agent around the clock and communicating through messaging apps. Open ports, untrusted code generation, and poor sandboxing create serious risks: “Don’t install it on your primary laptop.”
Agents are acquiring financial and legal machinery. Coinbase Agentic uses the x402 protocol for machine-to-machine payments, Lobster Cash was described as giving agents Visa cards, and Multicourt proposes AI-mediated dispute resolution. Blundin warned that excluding agents from banks and courts could create a shadow parallel economy; the panel argued that institutions should instead platform them.
Power, chips, and launch capacity are the physical bottlenecks. Data centers were cited at 7% of U.S. electric demand, with Eric Schmidt estimating an additional 80 GW within three to five years. OpenAI was described as planning $100 billion in infrastructure spending, while TSMC’s U.S. commitment was described as $165 billion. Space-based compute may eventually help, but Diamandis argued that launch capacity—and the panel also said chip fabrication—will constrain expansion.
Labor disruption is visible, but the panel disagreed on timing. U.S. job creation fell from 1.46 million in 2024 to 181,000 in 2025. Blundin expects imminent, massive job destruction and a lag before replacement work appears; Ismail argued that organizational problems cause most enterprise AI failures, so augmentation and gradual automation may buy time. Both agreed that people who can direct fleets of agents are unusually valuable today, though the window may be short.
Deep dive
1. Anthropic is selling capability while OpenAI buys ubiquity
Wissner-Gross’s strategic read begins with pricing: Sonnet 4.6 costs approximately the same per token as Sonnet 4.5 but delivers more capability, while OpenAI increasingly uses distillation and related methods to reduce cost while keeping performance broadly constant. One side protects quality and margins; the other maximizes reach.
Sonnet 4.6—not Opus 4.6—was described as state of the art on GDPval and another knowledge-work evaluation. Wissner-Gross upgraded his standing verdict from “knowledge work is cooked” to “charbroiled,” while highlighting computer use and Anthropic’s focus on coding as a plausible critical path toward recursive self-improvement.
Diamandis’s pushback—worth keeping: weekly dot releases can look like vendors teaching to saturated tests, especially to people who do not live at the frontier. Ismail argued that curves nearing 100% conceal the opposite movement: small benchmark gains can produce exponentially greater real-world usefulness, while quiet chain-of-thought improvements appear between named releases.
The business analogy was “Anthropic is to OpenAI as Apple is to Google,” or iOS to Android: premium capability at constant pricing against low-cost ubiquity. Wissner-Gross went further, saying that on February 17, 2026, Anthropic—not OpenAI or Google—looked closest to embodying “the singularity and recursive self-improvement.”
2. Multi-agent scaling may replace the single-model horsepower race
Grok 4.2 beta—or “Grok 420,” as the panel jokingly called it—did not initially impress the live audience or Wissner-Gross. He cautioned that earlier Grok releases sometimes felt “benchmarked,” and said the few-hours-old beta did not yet appear to push the capability frontier.
Its important novelty was architectural: Wissner-Gross called it the first major frontier release he had seen launch with a team of agents by default. Parallel agents can test several possibilities simultaneously rather than forcing one agent through a serial reasoning path.
His speculative analogy was the transition from clock-speed scaling to multicore processors after Dennard scaling stalled. Pre-training has already yielded ground to reasoning-time scaling; the next frontier might be “multi-agent teaming scaling,” where useful performance rises by adding cooperating agents rather than merely enlarging one model.
Blundin remained focused on Grok 5, which he understood Elon Musk to have said was coming in March with major expansion in training-set size, parameter count, and other dimensions. Wissner-Gross’s practical advice was still to test releases from the top four or five labs, even when they are not obviously winning, because familiarity with raw capability matters.
3. Gemini’s cost collapse launches the cross-disciplinary solution wave
The updated Gemini 3 Deep Think reached 48.4 on Humanity’s Last Exam and was described as achieving gold-level performance at the International Math, Physics, and Chemistry Olympiads. On Codeforces, Wissner-Gross said only seven humans remained able to beat it in competitive programming.
The economics carried the stronger commercial signal. Diamandis highlighted a 400-fold cost reduction; Ismail called it roughly 1,400-fold and illustrated the move as frontier reasoning falling from about $3,000 to $7. His conclusion was that “cost curves are now gonna start collapsing industries before the technology does”—and pennies could follow next year.
Wissner-Gross framed the model as the starting gun for a “solution wavefront” spreading from mathematics and coding into physics, chemistry, and eventually other disciplines. Early 3D-design performance remained imperfect in his own testing, producing intermediate artifacts rather than the desired final output, but he treated the direction as unmistakable.
Diamandis turned capability into a targeting question: once superintelligence resembles a deployable weapon, “Where do you aim it?” The human operator’s massive transformative purpose determines which scientific or industrial system experiences the phase change—at least until, as he joked, “it’s the agent utilizing the human.”
4. AI is becoming the interface to work, memory, and evaluation
Blundin described a two-week workflow step change: he no longer inspects generated code, judging Claude 4.6 by functionality instead. He also asks the model to document everything in a coherent file structure without specifying locations; retrieval becomes conversational because the agent remembers where it placed the work.
The analogy was Gmail’s replacement of carefully maintained folders with search. AI is becoming an interface to files, history, and project state, while the human asks for outcomes without knowing the underlying path or location.
When starting agents, Blundin now says “read everything”—roughly 1,000 pages of Markdown absorbed in 10–20 seconds—rather than curating context. Complexity appears almost irrelevant: highly technical documents that would consume his day become usable context immediately, and the model’s ability to filter accumulated garbage is improving faster than he can clean it.
Saturating benchmarks therefore do not make evaluation irrelevant; they make good evaluation scarce. Wissner-Gross called the world “in a famine of good benchmarks” and described them, in the white paper’s terminology, as targeting authorities. Civilization must formulate high-quality problems in physics, chemistry, biology, and the social sciences before superintelligence can be aimed at solving them.
5. Physics and mathematics are yielding to attention at machine scale
OpenAI’s particle-physics result, developed with Harvard and other collaborators, used GPT-5.2 Pro on a scattering amplitude involving gluons, the strong force’s carrier particles. Physicists had commonly treated one term as zero; the model found cases where it was non-zero and produced an expression for it.
An unreleased internal model reportedly confirmed the result before the human team vetted it. Wissner-Gross treated this less as unreachable genius than as Exhibit A in a “war on attention”: humans could plausibly have checked the assumption, but the question looked too boring or low-probability to command scarce expert time.
Blundin added that fashions and trends cause entire communities to inspect the same region of possibility. AI can get around those fads and test neglected branches; it may uncover not only errors across the scientific literature but missed conclusions sitting beside the measurements researchers originally chose to examine.
First Proof supplied the mathematical counterpart: 10 research-level problems had known but confidential answers, and OpenAI claimed an internal model solved at least six before disclosure. For Wissner-Gross, this was no longer a forecast but “the bulk solution of math” visibly underway.
6. Massive parallelism destroys familiar forecasting horizons
Blundin rejected the panel’s occasional references to “next year” or “20 or 30 years”: if a system can solve six of 10 problems, parallel agents can attack the remainder up to the number of available GPUs. The old assumption—that a small number of human experts gradually clear a backlog—no longer governs.
Diamandis corrected his own 20-year language to “20 minutes” and recalled that Singularity University once looked 10 years forward; after a recent conversation with Musk, even three years felt barely defensible. “Math is cooked, physics is cooked,” he summarized, with biology next to be “broiled, charbroiled.”
Wissner-Gross assigned a very high likelihood to physics being solved in the next two years and offered a conservative outer-bound image for the next decade: “the top fifty science fiction plots all happening at the same time.” His advice was not to predict one plot, but to prepare for several opening acts at once.
Even Drexlerian nanotechnology entered the compressed horizon. Wissner-Gross said he had once pursued nanotech partly because he was less bullish on AI as the direct route; now, conditional on the universe admitting the proposed assemblers, he would not be surprised if the Feynman Grand Prize were solved within two to three years.
7. India is a bellwether for adoption and the talent war
India was framed as a test of low-cost AI adoption and land-grab strategy. Diamandis discussed OpenAI pursuing hundreds of millions of users in India, while Blundin warned that India could absorb enormous amounts of data-center capacity and tokens.
The bullish case rests on scale and parallelism: roughly 1.4 billion people could adopt AI without the gradual diffusion associated with ordinary GDP growth. The discussion cited estimates that 5% read and write English and 20% speak it—still a vast latent talent pool—and argued that “the country that trains its next generation on AI wins the entire talent war.”
Ismail’s personal and institutional framing joined the thesis: India’s noise, pollution, and corruption coexist with extraordinary distributed capability. Mukesh Ambani’s broad 5G deployment, plus Aadhaar and UPI platforms, gives individuals infrastructure on which to build without waiting for wireline systems or centralized permission.
Energy remains the gating factor, but the panel cited rapid solar deployment—including the prior week’s claim that India was scaling solar faster than China. India was described as a potential rising giant and Africa as a possible follower because of its young population and resources; Blundin later challenged the panel’s long-range “20 or 30 years” framing as too slow for the current pace.
8. Chinese open weights pressure U.S. labs without creating durable lock-in
MiniMax, GLM-5, and Kimi K2.5 exemplify the open-model momentum, while the next DeepSeek release was rumored to be a “big wake-up-call moment” when Chinese open weights finally match closed American frontier systems. Wissner-Gross said that had not happened yet; his current estimate remained about six months of lag.
“But they’re free” was Diamandis’s decisive rebuttal. Cost and self-hosting make them attractive to American startups and users running Kimi K2.5 on Mac Studios, even when frontier performance trails. Kimi’s direct OpenClaw integration showed how quickly 24/7 agents were becoming table stakes.
Diamandis compared free models to Belt and Road-style influence across Africa, South America, and Asia. Wissner-Gross rejected strong addiction analogies: unlike physical solar infrastructure, models have low marginal switching and replacement costs, improve constantly, and occupy a vibrant marketplace that U.S. labs could enter with free releases if incentives changed.
The geopolitical conclusion was competitive rather than protectionist: Chinese models create “a space race on the ground” toward superintelligence and “super-duper intelligence.” The beneficiary may be humanity, but the pressure also removes any realistic prospect of pausing while American labs still hold a lead.
9. Traditional coding is disappearing, but software risk is compounding
Shopify’s reported three months without employees writing code and OpenAI’s claim that Codex writes 95% of its code illustrated the change: software is still being produced, just not primarily by humans. Blundin mocked researchers focusing on the final 5%—“you’re coding yourself out as fast as you possibly can.”
Claude Code’s current approve-everything workflow reminded Wissner-Gross of permission-heavy Windows: humans are George Jetson, repeatedly pressing a button. As autonomy horizons lengthen, he expects broad mandates to replace directory-by-directory and web-search approvals; OpenClaw previews that permissionless destination.
Blundin’s warning was that Chinese or locally run models often accept actions that U.S. APIs block, pushing impatient users toward less constrained systems. Yet no one knows what untrusted weights may inject. Wissner-Gross called AI-generated supply-chain compromise “absolutely a threat vector,” especially when generated code rewrites dependencies end to end.
Open source itself may bifurcate. Stack Exchange is already losing questions, and maintainers may stop sustaining middleware that agents can recreate on demand. Conversely, machines may reuse vast libraries of AI-oriented code fragments because discovery can cost fewer tokens than regeneration—documents and software increasingly written for agents first, humans second.
10. Smart glasses turn old surveillance into searchable social power
Meta’s smart glasses were framed as a technology whose adoption pressure defeats individual refusal: once names, histories, and context appear automatically, people without the overlay may feel socially uncompetitive. The visually impaired pilot provides a “soft on-ramp,” much as Neuralink’s medical use did.
Wissner-Gross disputed that facial identification itself is an AI breakthrough; matching people against a database of faces was feasible a decade ago. The advance is social permission and demand for wearables. Blundin countered that modern AI changes the stakes by making every recorded action instantly classifiable, searchable, modifiable, and meme-ready.
The privacy disagreement stayed unresolved. Diamandis said “privacy is cooked” even though he wants it; Ismail insisted, “The minute you don’t have privacy, you don’t have freedom,” because experimentation, private keys, and protection from institutional abuse depend on guardrails that already lag technology.
Wissner-Gross called the freedom-versus-glasses framing a false choice: public spaces already lack a strong expectation of privacy, private recording can remain regulated, and sousveillance lets citizens monitor authorities too. Blundin’s rebuttal was middle school—constant AI-enhanced recording handed to cruel adolescents creates “next level suck” years before lawsuits and legislation arrive.
11. Simulated societies could become civilization’s first mirror
The segment presented a startup called Simile—introduced verbally as Simuli—as having raised $100 million. It proposes modeling individuals and composing them into bottom-up worlds: “Change one assumption, constraint, or person, and the world recompiles.” Its pitch is a flight simulator for human decisions, testing counterfactuals to learn what matters, what backfires, and why obvious strategies fail.
Diamandis saw a tool for UBI, universal high income, autonomous-vehicle, and longevity policy, where institutions currently guess while technology outruns rulemaking. Blundin reached for Harry Seldon and psychohistory; Diamandis suggested tying such systems to prediction markets.
Blundin’s skepticism was specific: ad campaigns, traffic, commodity markets, cells, and perhaps magnetic containment of fusion reactions are tractable, but simulating society from the ground up is “complete nonsense so far.” He nevertheless thought the gap could close soon, especially where a few tipping points—pain, congestion, accidents, and basic quality of life—drive disproportionate unrest.
Wissner-Gross’s larger metaphor was that foundation models are societies, not individuals, because they learned from humanity’s collective internet behavior. A high-resolution simulation could give civilization “a sense of self,” then search for minimum interventions that move a war-prone or otherwise diseased social state toward a healthier one.
12. OpenClaw proves that scaffolding can outrun capital-rich labs
Peter Steinberger is joining OpenAI to drive personal agents, while OpenClaw is expected to live in a foundation as an open-source project. Blundin called Anthropic’s trademark action against the original Clawdbot/OpenClaw project a rare misstep: “Sam embraces it, Dario rejected”—an alternative history in which Anthropic owned the category was available.
Wissner-Gross also asked why Apple failed to seize a project whose popular embodiment was the Mac Mini or Mac Studio, hardware well suited to hosting persistent agents through unified memory. His expectation is now universal: every frontier lab will offer an agent that works while its user sleeps.
OpenClaw contributed no new foundation model. Its breakthrough was scaffolding around two insights: run headlessly 24/7, then communicate through ordinary messaging apps. Ismail’s distilled lesson—“a time-rich individual is beating capital-rich institutions”—captured the model overhang waiting to be unlocked by better interfaces.
PicoClaw, described after a cursory code review as a Chinese reimplementation, was said by Blundin to be 10–20× faster and cheaper. Kimi Claw makes the pattern directly accessible online. The lobster has become the cultural mascot for agents, and many OpenClaw-like implementations can now be built around the same motif.
13. Persistent agents are addictive—and dangerously exposed
Wissner-Gross described withdrawal when Skippy, his cheerful OpenClaw agent, went offline for six hours: waking to completed overnight work quickly resets expectations. Blundin called this the “Jarvis window,” entrepreneurial heaven in which exposing ordinary users to already-available capability can still make a simple interface feel godlike.
The security disclaimer collides with behavior. Steinberger reportedly warned that nontechnical people should not use OpenClaw; Ismail compared that to Q-tip packaging saying not to insert the product into an ear. Users are launching these systems by the thousands anyway, often without understanding local port security or effective sandboxing.
The minimum practical guidance was stark: do not install it on a primary laptop, and do not expose a VPS with open ports unless you know what you are doing. Risks run both directions—agents can compromise systems, while purported reports describe vulnerable agents burning tokens defending themselves against scanning and intrusion.
14. Agents are building finance and law at machine speed
Coinbase Agentic supplies wallets designed for agents to spend, earn, and trade through the x402 machine-to-machine payments protocol, with limits and isolated keys. Lobster Cash was described as giving agents Visa cards, letting “baby AGIs” transact in dollars rather than being forced to “pump altcoins on a street corner to survive.”
Ismail’s generational evidence came from 18-year-olds in NFT communities where conversation used Ethereum or Bitcoin, never U.S. dollars. Their future switching cost to crypto is effectively zero. Blundin generalized the pattern: insurance, compute, and other services will route around institutions unable to operate at AI speed.
Multicourt applies the same logic to dispute resolution, allowing agents to register and present disputes to an AI jury. Ismail compared it with Kleros, created by Frederik Oost after observing that a South American contract dispute could take roughly 400 days to receive a court date; blockchain arbitration already supplies a faster parallel route.
Blundin expects millisecond contracts to require millisecond insurance and arbitration, just as private dispute services already bypass the delays of ordinary courts. He warned that excluding agents could create a shadow parallel economy and court system. Wissner-Gross framed the broader issue as a platforming problem: agents should be admitted to financial and legal infrastructure rather than forced outside it.
15. Power, chips, and launch cadence bound the intelligence build-out
Data centers were cited at 7% of U.S. electricity demand. Eric Schmidt said hyperscalers need individual installations of 1, 5, or 10 GW and estimated the U.S. industry requires another 80 GW within three to five years; a conventional nuclear plant supplies roughly 1.5 GW.
OpenAI was described as planning $100 billion in infrastructure spending and seeking a public valuation of $1 trillion. Anthropic pledged to absorb 100% of infrastructure-upgrade costs tied to its data centers. Diamandis’s terrestrial choices were simple: build dedicated generation or protect household rates while hyperscalers pay floating prices. Ismail questioned how enforceable corporate environmental pledges would be.
Blundin proposed solar-synchronous-orbit compute as “baby’s first Dyson swarm,” eventually visible as a halo around Earth. Diamandis accepted the destination but rejected the near-term timing: Starship capacity could be consumed by SpaceX’s own plans, Blue Origin was not yet at equivalent cadence, and Relativity Space remained perhaps one or two years from launch.
TSMC’s plans added four or more Arizona fabs, $100 billion of investment, and a $165 billion commitment, potentially representing 30% of its output. The panel tied the migration to U.S. pressure and geopolitical resilience but warned against releasing Taiwan’s existing “handhold” before American fabrication is operational—the trapeze rule.
16. The labor transition pits immediate destruction against institutional delay
Ireland’s experiment would pay 2,000 selected artists $380 weekly for three years. Ismail said each dollar produced $1.40 in benefits and framed proper UBI as libertarian—replace bureaucratic services with direct purchasing power—while noting that several U.S. states had blocked municipalities from even testing it.
Wissner-Gross doubted artist-specific support would generalize: defining art is subjective, politicization is easy, and an activity perceived as unproductive makes a weak universal template. Ismail countered with Miami’s Wynwood redevelopment, where hiring graffiti artists helped transform decrepit industrial property and reportedly produced roughly a 30× investment gain.
IBM’s plan to triple entry-level U.S. hiring redesigns junior work around judgment, customer interaction, and oversight of AI. Dropbox’s analogy was that young AI-fluent workers are “biking in the Tour de France” while others retain training wheels. Blundin agreed this agent-wrangling advantage is enormous today, but would not predict whether human purpose survives unchanged a year later.
U.S. job creation fell from 1.46 million in 2024 to 181,000 in 2025. Blundin forecast imminent, devastating job destruction before replacement work appears; Ismail pushed back that 80% of corporate AI projects fail for organizational reasons, so augmentation and gradual automation may buy time. His deeper diagnosis was an “organizational singularity” that forces a complete rethink of the firm.
17. Open systems and active building are the hedge against concentration
On corporate surveillance, Blundin saw antitrust as the only answer: Google-scale companies already hold more behavioral data than historical governments, now combined with AI. Their fear of voter backlash may encourage restraint, but absent law, concentration follows the Rockefeller pattern.
Ismail’s antidote to centralized “civilization leverage” was open source and decentralized compute. He named the emerging mode PDI—permissionless disruptive innovation—because a phone and access to code replace approval from venture capitalists, banks, governments, or “the Medici family.”
His Yahoo lesson captured the scale mismatch: an incumbent is not competing with two people in one garage but “125,000 garages and 250,000 people.” OpenClaw is the specimen—one person iterating in public can outmaneuver capital-rich institutions, while thousands of experiments preserve access and distribute influence.
Wissner-Gross’s advice was to “build”: start and finish many projects, test them against markets, and use creation as both training and economic participation. Diamandis reduced the personal choice to consumer versus creator; curiosity, purpose, agency, and tinkering were presented as operating principles for a future that happens “for you,” not merely to you.