Instinct Raising $1B at $10B & Meta Launches Muse | Miro Sells for $1.36B After a $17.5B Valuation
Summary
- The 10% extinction claim turns AI safety from a lab debate into a governance and product-liability crisis. If Washington genuinely believed a San Francisco lab had a one-in-ten chance of ending humanity within a decade, the panel argued, it would shut the facility rather than merely suggest voluntary monitoring. The sharpest challenge to Dario: if loss of control is real, “your mission is to solve the problem”—and if it cannot be solved, stop building or replace the CEO.
- Markets priced the existential warning as a sector rotation, not an apocalypse. Semiconductor and AI-capex exposure dipped slightly, cyber names rose roughly 10%, and WCLD outperformed SOX; as one viral formulation put it, “even with news of a 10% chance of extinction, the market moved by only 0.1%.” Capitalism’s answer was effectively, “It will be fine,” even as $2 billion increasingly resembles a starting round.
- Meta’s Muse demonstrates that an incumbent with models, compute, storage, and distribution can finally ship a credible consumer agent. The product reportedly handled reservations, email, website rebuilding, and show preparation immediately, backed by roughly 500 engineers and infrastructure that peers may spend $3–4 per person to provide for such virtual machines. The unresolved question is whether convenience across many small tasks constitutes a killer app—or merely an attractive product used occasionally.
- Instinct’s $1 billion round at a $10 billion valuation is a bet that focus and cross-platform neutrality can overcome Meta’s balance sheet. The bull case rests on a highly regarded team, explosive consumer demand, Meta’s history of abandoning non-core products, and strategic acquisition value potentially reaching $50–60 billion; the bear case is that Meta now treats personalized AI as core after investing over $100 billion. Jack first recommended leading the round with carefully managed reserves, then said he would not participate; Harry likewise said he would not participate.
- The investment only works if $25 billion-plus outcomes become routine or Instinct creates defensible economics before its compute bill overwhelms it. Harry raised whether a portfolio could tolerate three highly volatile bets among 30; Jack said a $500 million investment could be rational even with a two-in-ten chance of a 20x outcome, but that it probably would not work at a $10 billion entry. Investing on the assumption of a 100% M&A exit is “gambling.” At roughly $50 million earlier in the year, the risk-adjusted return looked exceptional; at $10 billion only months later, much of that asymmetry had vanished.
- Miro’s fall from a $17.5 billion valuation to a roughly $1.35 billion Bending Spoons sale marks the liquidation of the 2021 SaaS vintage. Miro still reportedly has about $600 million of revenue, high-single-digit growth, and positive cash flow, yet recent investors received around 1x. Bending Spoons’ playbook is to cut hard, raise prices—potentially 40%—and retain the customers whose willingness to pay proves the product remains essential.
- Automattic’s board fight shows how excess venture capital and under-monetized open source can turn founder control into a “poisoned chalice.” Matt Mullenweg may have preserved authority over WordPress, but the panel sees a shrinking strategic island distracted by WP Engine litigation while AI-native website tools advance. Their counterfactual was a smaller, profitable company run on the founder’s own terms, rather than an overcapitalized organization whose investors and open-source community now want incompatible outcomes.
- Mistral and Adobe represent opposite responses to AI disruption: strategic sovereignty versus managed stagnation. Mistral’s €3 billion round and stated €1 billion year-end revenue ambition were framed not as proof it matches OpenAI or Anthropic, but as Europe’s Airbus-style decision that critical technology cannot remain wholly dependent on the US. Adobe, by contrast, may grow 6–9% and compound cash flow yet remain near today’s roughly $105 billion market value unless it discovers a genuinely transformative AI product: “If you don’t grow, you don’t win.”
Deep dive
1. A 10% extinction claim demands more than voluntary supervision
The opening dispute separated Dario’s stated concerns into cyberattacks, economic displacement, and loss of AI control. One panelist dismissed job-loss arguments as another iteration of technological change—agriculture once represented 73% of employment—and called cyber risk serious but already irreversible; only recursive self-improvement and loss of control produced a genuine “Hmm, I see.”
The fiercest test was institutional: if the federal government believed a downtown San Francisco facility carried even a 10% chance of destroying the world within ten years, “how many seconds would it take” before authorities shut it down? The absence of that response implies either an asleep government or officials treating the rhetoric as “the nonsense of excited teenagers.”
Harry’s qualification—worth preserving—was that Dario did not personally declare a 10% probability of doom; he responded to an employee’s claim. Yet words still create consequences: relatives outside technology were already texting, “Is the AI going to kill us?”, and the panel expects congressional hearings to consume the next 24 months.
A more tactical interpretation viewed the safety discussion as advance work for Anthropic’s eventual S-1 and a possible $2 trillion IPO: disclose the risk, show management addressing it, and neutralize it before roadshows. The complication is a second risk factor—employing highly motivated people who publicly believe their own company might destroy humanity.
2. Product liability makes the CEO responsible for stopping
The panel endorsed David Sacks’s formulation: “If the situation is that serious, Dario, it is your responsibility.” A company cannot externalize responsibility for a potentially catastrophic product; management must solve the danger, stop development, or install a CEO who can. Lina Khan’s agreement was striking because it placed two habitual Silicon Valley antagonists on the same side.
The proposed international solution looked politically implausible: begin with third-party monitors, make them mandatory across democratic governments, then secure agreement from China and Russia. Voluntary monitoring is a lab’s choice, but statutory supervision would be designed by government—“not your friends”—much like banking regulation, with no guarantee that its rules would match the industry’s preferences.
Jack’s synthesis, building on Harry’s framing, sharpened the dilemma: if the system can be controlled, continue; if it cannot, “it would be best to stop.” A frontier lab cannot “rub your hands together and lament, ‘I wish the world would stop me from destroying humanity’”; the very claim that its team is uniquely capable makes outsourcing the core safety decision incoherent.
3. Dual-use capability is undeniable, but the policy conclusion is contested
One side accused frontier labs of helping create the Chinese open-source ecosystem and then asking everyone else to bear the cost of slowing down. The rebuttal rejected granting any individual “god-like status”: China had capable researchers, the underlying technology was globally legible, and alternative paths to large language models probably existed regardless.
The atomic-bomb analogy produced a direct clash. One speaker called it crude, pride-inducing, and “like a Nazi analogy”; another insisted it fit because Soviet nuclear capability followed within three to four years, aided by stolen secrets, yet likely would have arrived eventually. Harry characterized Dario’s position as: AI has already become too powerful, so “it’s already happened, so I have no choice but to do it.”
Jay Kreps’s dual-use framing supplied the strongest concrete mechanism: superhuman coding implies superhuman hacking; structural engineering implies finding structural weaknesses; drug discovery implies designing undetectable poisons; curing viruses implies creating them. “All other things being equal,” an AI is equally capable in its bright and dark forms, while open-weight models lack safeguards.
The counterargument invoked Gutenberg and the internet: every information technology creates positive and negative uses, and authoritarian institutions predictably respond by seeking control. Society should manage the harms while developing the benefits—though the speaker immediately carved out a concrete restriction: “we will not sell chips to China for the time being.”
4. Capital markets heard “extinction” and bought cybersecurity
The immediate tape barely reflected existential panic. Semiconductor and AI-capex exposure softened without a sharp selloff, cyber-related stocks rose about 10%, and WCLD had a strong day relative to SOX. The memorable summary was that a 10% extinction warning moved markets only 0.1%: capitalism assessed the risk and said, “It will be fine.”
The panel compared the moment with the 1920s: limited regulation, enormous fortunes in reach, and no consensus about what an AI-era 1929 might look like. Capital abundance intensifies the danger—“$2 billion is just the starting round”—because a founder can finance both the constructive product and its dark counterpart with unprecedented ease.
They resisted treating Sam Altman, Dario, or Elon Musk as simple profit maximizers, arguing that each has legitimate motives and unusually strong operating ability. Harry’s claim was that 10–15% of founders have antisocial personality disorder, and the manipulation and obsessive drive useful for “creating something out of nothing” can become especially dangerous when paired with AI.
5. Muse proves Meta can ship; it does not yet prove daily necessity
Meta reportedly moved roughly 500 engineers onto Muse as a P1 project after the agent category accelerated, with teams answering public bug reports through the weekend. As software, the verdict was emphatic: “really, really, really great.” It handled reservations and email, prepared stories for the show, and logged into WordPress to rebuild a website with little setup.
The hidden advantage is infrastructure economics. Comparable builders may spend roughly $3–4 per person to provide a virtual machine, with a discussed configuration of two CPUs, two GPUs, 8GB RAM, and about 100GB storage; Meta already owns enormous compute and storage capacity and runs Muse on its own Muse LLM. That helps explain its speed relative to slower independent agents.
Harry’s challenge was not quality but intensity of use: booking a restaurant or a daughter’s ballet class is useful, but it is not obviously an eight-hour-a-day workflow like Claude Code or Codex. Muse’s Ideas tab and visual design were praised, yet he remained unsure whether it had a true killer app.
The opposing view was that no single killer app is required. Instinct already manages travel, reservations, shopping, and calendar invitations; many small superior experiences may aggregate into the product. Pre-AI conversational booking failed because browsing an airline website was easier, but backend intelligence can finally make the WeChat-style super-app thesis workable.
6. Instinct’s $10 billion round is focus versus incumbent power
Instinct reportedly progressed from a $50 million pre-money round less than five months earlier through $200 million and multibillion-dollar marks to a $1 billion raise at a $10 billion valuation. Consumer response appears genuine: after Harry posted an Instagram Reels video, he received more than 1,000 direct messages requesting invitation codes—an unprecedented response for his account.
The bull case is organizational. Meta cannot seamlessly span every platform and service, while Facebook itself reaches only a slice of users aged 18–55; WhatsApp and Instagram remain separate surfaces. Meta also abandoned Workplace despite strong architecture and reportedly the platform’s highest NPS, raising the question of whether it can sustain Muse-level intensity as long as a focused startup can.
The team drew exceptional reference checks, with Instinct founder Noah Shinn repeatedly described as elite. Their pitch behavior—playing World of Warcraft live and ranking around the top 10 and top 15 in League of Legends—was offered as an idiosyncratic signal of technical talent, competitive obsession, and unusual team quality.
The bear case is that personalized AI is not another Workplace. Meta has invested over $100 billion in AI, owns the model, GPUs, servers, distribution, and cost advantage, and is now building the application layer venture investors long feared. Instinct’s slow responses expose its compute burden. Jack first recommended leading the round with carefully managed reserves, then said he would not participate; Harry also said he would not participate.
7. Instinct’s return profile may require an extraordinary exit
A favorable acquisition case compared Instinct with Poolside: even if standalone cash flow disappoints, a much larger platform may want the models, users, and talent. OpenAI or another strategic buyer could find the asset compelling; given current prices for elite teams, Jack floated a $50–60 billion sale while explicitly admitting that estimate “might sound a bit exaggerated.”
The pushback came from venture discipline: never make an investment that requires M&A, because buyers, champions, and management teams change unpredictably. At $10 billion, acquisition optionality is no longer a comforting secondary path but close to the central underwriting assumption—“a really big gamble,” even if the expected value remains positive.
Harry raised whether perhaps three such high-volatility bets belong in a 30-company portfolio. Jack said a $500 million investment could be rational even with a two-in-ten chance of a 20x outcome, especially with a strong portfolio, but that it probably would not work at a $10 billion entry. The asymmetry was far better at the roughly $50 million valuation in April or May; after several rounds in months, much of the early risk-adjusted appeal had evaporated.
Menlo Ventures’ cited world model counts 81 technology companies worth at least $25 billion, versus 23 a decade earlier, and anticipates at least 100 more. If $25 billion becomes the new decacorn threshold, Instinct’s round can work—but a $25 billion exit from a $10 billion entry produces only a modest multiple, not classic venture power-law economics.
8. Miro’s sale liquidates a 2021 valuation without killing the company
Miro, the collaborative whiteboard company, was described as having reached a $17.5 billion valuation in 2021 before selling to Bending Spoons for roughly $1.3–1.35 billion. The company still reportedly generates about $600 million annually, grows at a high-single-digit rate, and produces positive cash flow; founders and employees made money, while recent investors realized only around 1x.
Jack called the outcome “inevitable,” not disastrous. The warning signs were visible in private-company data: Miro’s valuation had stalled at its prior high while its workforce continued growing rapidly. The discussion compared that pattern with Evernote, whose valuation had run ahead of its actual market size and growth, making an eventual clearing event unavoidable.
Late-stage investing’s peculiar protection is that a failed deal can return approximately 1x—still a bad venture result, but enough to “die rich.” Part of the consideration may have rolled into Bending Spoons stock, though the panel did not know whether that was required or elective; such a rollover offers another route to participate if the acquirer can improve the asset.
9. Bending Spoons buys willingness to pay, not growth theater
The Bending Spoons model was framed as single-owner discipline replacing a venture syndicate that cannot act ruthlessly enough. Its CEO’s controversial message—“We don’t care much about the title of founder”—means past creation earns no permanent exemption; present operating performance determines who remains.
Price increases and customer loss are features of the model, not accidental damage. Venture-backed sales and marketing may have manufactured marginal customers; Bending Spoons wants those who will dislike a doubled price yet stay because the software is essential. Harry estimated that tripling the price could produce 30–40% churn, while the retained base might produce better economics; customers should nevertheless “prepare for a 40% price hike.”
The panel’s musical-chairs analogy was bleak: Bending Spoons seriously reviews about 1,000 companies and buys only five to ten annually, leaving perhaps one or two seats for an entire vintage. If assets as solid as Miro or Airtable attract only roughly 2x offers, slower companies below them may find no buyer at all.
That is the “era of surrender”: businesses growing 0–5% must take the remaining seat or enter a zero-sum future where nobody wants them. The contrast with frontier AI is violent—stagnant SaaS clears below old valuations while new teams jump from tens of millions to billions within weeks.
10. The SBF detour separated proportional punishment from retroactive absolution
The case for reconsideration emphasized that creditors were repaid in full with interest under the bankruptcy court’s terms, SBF apparently did not extract material personal wealth, and a 30-year sentence plus roughly $11–12 billion could be disproportionate. Harry predicted the Supreme Court might hear an appeal concerning the facts and circumstances and eventually overturn the result by a narrow margin; Jack said that would at most create a chance to stand trial, not guarantee release.
The pushback was categorical on the underlying rule: returning money after a successful gamble does not legalize taking it. “I stole your money, but I bought call options and the stock price went up” cannot become a defense when the same trade had a 50% chance of loss. Thirty years might be excessive, but embezzlement still requires punishment to deter repetition and preserve equal treatment of white-collar crime.
11. Founder control cannot rescue Automattic from strategic drift
Automattic’s conflict with WP Engine was presented as both a governance fight and a failure to see the larger threat. Using influence over WordPress to extract more value from a hosting company was portrayed as conflicting with the spirit of the open-source ecosystem, while AI-native tools increasingly replicate website-building functions. Litigation and bylaws are secondary if the platform lacks a compelling five-year AI direction.
The panel’s board reconstruction was explicitly speculative: independent directors may have held enough votes to replace Matt Mullenweg, only to discover that the founder could replace them. The imagined response—installing “me, my pet dog, and my ventriloquist dummy” as the new board—captured the hollow victory: control of an ever-shrinking Fortnite island and a “poisoned chalice.”
Overcapitalization may be the original mismatch. A smaller Automattic could have employed roughly 80 people, generated $100 million annually, and let Mullenweg pursue the open-source mission without venture-style growth obligations. Instead, large financing created investors, directors, employees, and a community whose definitions of success diverged.
Jack’s harsher lesson was that venture-backed open source demands ruthlessness. Automattic left substantial hosting profits to providers and did not pursue WooCommerce against Shopify with comparable commercial intensity; generosity ceded parts of the ecosystem’s profit pools. Ownership permits that choice, but it cannot simultaneously promise venture returns and remain indifferent to commercial capture.
12. Mistral is a sovereignty project before it is a frontier-model rival
Jack dated the current cycle against his investing career beginning in 1993 or 1994 and the Y2K boom and still called it unprecedented. Y2K mixed wealth creation with fear that computers might stop; AI couples ten times the money with fear that systems might end the world. Its ability to make software feel human stimulates the imagination far beyond the internet cycle.
Mistral’s €3 billion financing—the largest European technology round cited—and stated goal of €1 billion revenue by year-end were treated as meaningful progress, but not evidence that it currently matches OpenAI or Anthropic at the frontier. “It is nonsense” to frame it that way; the real thesis is European control over a strategic technology.
Airbus supplied the analogy: Europe accepted years of apparent inefficiency because relying entirely on US aircraft manufacturing implied dependency, then built competitive capability over 10–15 years and eventually overtook Boeing over a much longer horizon. Mistral might remain smaller than US labs while still becoming strategically and commercially successful in Europe.
Jack recalled, with uncertainty, a US direction that Anthropic block every country’s access to a recent model called “Fable,” not merely adversaries. If Europe concluded that even allied access could disappear overnight, political action effectively created Mistral’s sovereign market. Samsung’s lead investment and ASML’s participation in the prior round reinforce that strategic logic, but do not constitute pure financial price discovery.
13. Adobe faces the brutal passage from growth multiple to cash-flow multiple
Adobe’s internal CEO succession, reached after roughly a year, was read as “a sign that nothing is going to change.” The critique was not imminent failure—Adobe remains enormous—but surrender: sacrifice the core product for margins, attach AI image features, emphasize free users and AI-linked metrics, and avoid confronting declining net-new ARR.
At roughly $105 billion of market capitalization, one forecast placed Adobe at only $120–130 billion three years later. The company could grow 6–9%, throw off abundant cash, and trade below ten times cash flow without disappearing; the investor problem is opportunity cost in a market expected to produce scores of new $25 billion-plus companies.
Canva provided the harder comparison. Harry said its growth had slowed from around 30% to 20%—“growth is slowing,” not stopping—and Jack said it had dropped roughly 30–40% when forecasts were lowered, while noting that it was not publicly traded. Adobe offers liquidity and durable cash flow; Canva offers faster growth but faces a hazardous repricing if investors cease treating it as a growth asset.
The cleanest valuation rule quoted from social media was: above 30% growth, use a revenue multiple; below 30%, use EBITDA. Box needed three to four years to cross that divide, because moving from six or seven times revenue to 20 times cash flow may require cash flow to rise by roughly 30% merely to preserve the stock price. “If people don’t pay, you don’t win.”