Roko's Basilisk
April 21, 2026
🧪 Meta spent $135B chasing AI. The leaderboard still says fourth.
🗞️ ChatGPT's darkest use case, Palantir's manifesto, Vercel breached.
📊 Tuesday Poll on whether Meta's AI payoff is ever coming.
💡 A prompt to audit any company's AI strategy in under a minute.
Let’s dive in. No floaties needed…
Framer helps teams design, build, and launch their marketing sites lightning fast. With the ability to publish hundreds of CMS pages in a single click, operate at a global scale with seamless localization, and even host unified content across multiple domains, teams have never been able to ship faster. Trusted by companies like Miro, Bilt, and Perplexity.
*This is sponsored content
Muse Spark is real, but not frontier: Meta’s first model from its Superintelligence Labs is natively multimodal with three reasoning tiers, but it scores 52 on the Artificial Analysis Intelligence Index, trailing Google, OpenAI (57), and Anthropic (53).
The reorg was seismic: Meta took a 49% stake in Scale AI for $14.3B, installed Alexandr Wang as its first chief AI officer, and lost Yann LeCun, who departed to pursue a fundamentally different AI architecture.
Revenue model remains unresolved: Nearly all of Meta’s ~$200B in revenue comes from advertising, and Muse Spark’s monetization runs through better ad targeting and social commerce rather than direct AI subscriptions or API access.
App Store spike, not a moat: The Meta AI app surged to No. 5 on the U.S. App Store at launch, but it still trails ChatGPT, Claude, and Gemini in sustained usage, and download spikes rarely convert into retention.
For any business to raise capital, it needs to articulate a credible path to a product that users are willing to pay for. In the case of OpenAI, that path was unusually long and, for a time, largely invisible; it operated more like a research lab than a conventional product company even as it raised billions, particularly after its partnership with Microsoft. When it finally brought products to market, though, it largely delivered, turning advanced AI into something broadly usable at a massive scale.
The breakthrough set off a scramble across Big Tech, as companies moved quickly to secure their place in what was beginning to look like a foundational shift in computing. Among those trying to catch up was Meta, which had only recently come off an expensive and uncertain bet on the metaverse and was now repositioning itself as a serious contender in AI.
It was against this backdrop that, in 2023, Meta reorganized its AI efforts and began speaking more explicitly about building AGI-level systems. In internal discussions and earnings calls later that year, its CEO, Mark Zuckerberg, framed the company’s ambition as creating AI that could “help people with almost anything,” a formulation that aligned more closely with AGI than with the narrower applications that had defined its earlier work.
That ambition has since taken a more concrete form. In 2026, Meta released Muse Spark, the first model developed by its Superintelligence Labs division, created a year earlier after Zuckerberg grew frustrated with the pace of progress in the company’s Llama family and its inability to keep up with rivals like OpenAI, Google, and Anthropic.
Muse Spark is a natively multimodal model, built from the ground up to process voice, text, and image inputs in a single system rather than stitching vision onto a language model after the fact. At launch, however, its outputs remain text-only. The model operates across three distinct modes: Instant for quick, conversational queries; Thinking for more complex problems that require extended reasoning; and Contemplating, its most technically ambitious setting, where multiple sub-agents reason in parallel and synthesize their outputs. The design is a clear attempt to compete with the deeper reasoning capabilities being introduced by rivals like Google and OpenAI.
Meta is also positioning efficiency as a core differentiator. The company claims Muse Spark can deliver these reasoning capabilities using 10–999 times less compute than its previous mid-sized flagship, Llama 4 Maverick. That gain is tied to a training approach it calls “thought compression,” where the model is penalized during reinforcement learning for taking too many intermediate reasoning steps, effectively forcing it to arrive at answers with fewer tokens without sacrificing accuracy.
However, despite these gains, Muse Spark still falls short of the frontier. On the Artificial Analysis Intelligence Index, an independent benchmark, the model scores 52, placing it fourth behind offerings from Google and OpenAI at 57 and Anthropic at 53. It marks a substantial leap from Llama 4 Maverick’s score of 18 on the same index, but the gap at the top remains, underscoring how far Meta still has to go to match the leading edge.
Even on GPQA Diamond, a graduate-level science test, Muse Spark scores 89.5%, behind models from Google and OpenAI. On SWE-bench Verified, which measures coding ability, it also trails Anthropic. The biggest gap shows up on ARC AGI 2, a test of abstract reasoning, where Muse Spark falls well behind the leading models. Meta has acknowledged these weaknesses, particularly in coding and agentic tasks, and says it is continuing to invest in improving them.
Where the model performs better is in areas that align more closely with Meta’s product strengths. On HealthBench Hard, which evaluates complex health queries, it outperforms both OpenAI and Google’s models. It also leads on CharXiv Reasoning, a benchmark focused on interpreting charts and figures from images, particularly in its Contemplating mode.
Overall, the model represents clear progress, but not quite at the level one might expect from a company that spent much of 2025 overhauling its AI strategy—a period marked by some of the most dramatic reorganizations the industry has seen.
During the period, Meta invested $14.3B in Scale AI for a 49% nonvoting stake and recruited its co-founder and CEO, Alexandr Wang, to become Meta’s first-ever chief AI officer.
Wang stepped down from Scale AI, where he had remained on the board, and took charge of a newly created division, Meta Superintelligence Labs (MSL). Former GitHub CEO Nat Friedman was also brought in to co-lead the division’s product and applied research side. Zuckerberg personally recruited researchers from OpenAI, Anthropic, and Google DeepMind, offering compensation packages that The New York Times reported ranged from $1M to $100M.
To offset restructuring costs, Meta laid off approximately 600 employees from the division. And in November, Yann LeCun, the Turing Award-winning researcher who had built Meta’s foundational AI research lab (FAIR) over more than a decade, confirmed he was leaving Meta to start his own company focused on world models, a fundamentally different approach to AI architecture than the large language models at the center of Meta’s strategy. LeCun’s departure was significant not just as a personnel loss but as an intellectual break: one of the most respected figures in the field was leaving because he believed Meta’s approach was a dead end.
With the release of Muse Spark, Meta hopes to put to rest any doubts about its ability to compete at the highest levels of the AI domain. However, even as the company continues to release models, it has yet to demonstrate a steady revenue stream that would make its AI strategy a viable business model.
Meta has guided capital expenditure of an estimated $135B in 2026, nearly double what it spent in 2025, alongside long-term commitments to U.S. infrastructure and large-scale compute deals.
However, the bulk of its roughly $200B in revenue still comes from advertising, with AI so far improving ad targeting and internal efficiency rather than generating standalone income—unlike rivals such as OpenAI or Google, which monetize AI through subscriptions, APIs, and cloud services.
With Muse Spark, Meta’s strategy will likely be to fold AI into its existing business. Features like Shopping mode, which uses data from Instagram, Facebook, and Threads to recommend products, point to a model where AI drives better ads, stronger engagement, and more seamless social commerce, turning capability into revenue indirectly rather than selling the model itself.
It must also be remembered that Muse Spark is explicitly positioned as the first model in a series, not a finished product. Meta describes it as small and fast by design, validating its new architecture and training approach before scaling up to larger models. The company has confirmed that additional Muse models are in development, though no release dates have been announced.
The initial response to Muse Spark has also been strong, with the Meta AI app climbing from No. 57 to No. 5 on the U.S. App Store within a day of the launch. And though Sensor Tower data reported that Meta AI in the U.S. rose over 450% on launch day, it still lags behind ChatGPT, Claude, and Gemini in overall standing, and download spikes do not necessarily translate into daily usage.
In the end, the question returns to where it began. Raising capital is one thing; sustaining it requires a product that proves its value in the market, not just in benchmarks or internal roadmaps. OpenAI managed to close that loop by turning years of research into products people actively use and pay for, creating a feedback cycle between capability and revenue.
Meta is currently trying to emulate that, and Muse Spark shows that the company can build competitive systems and narrow the technical gap. However, its strategy still leans heavily on indirect monetization through advertising and commerce rather than a clear, standalone AI business.
That approach is not necessarily flawed; it plays to Meta’s existing strengths, but it also means the success of its AI ambitions will be harder to measure and slower to materialize.
For now, Meta appears to be on a plausible path, but not a proven one. The pieces are in place: talent, infrastructure, and a product that is improving. What remains uncertain is whether those pieces can come together into something users will not just engage with, but pay for, or whether AI at Meta will remain an enhancement layer on top of its core business rather than the foundation of a new one.
ChatGPT's darkest use case: The Florida mass shooter sent 13k messages to ChatGPT in the months before the attack, using it to plan logistics and time the shooting.
Palantir's manifesto moment: Palantir posted a 22-point public statement denouncing pluralism and inclusivity while defending AI weapons and surveillance work.
Vercel breached: Vercel confirmed a security breach traced to a compromised third-party AI tool, with hackers claiming to sell access to source code and tokens for $2M.
Memorandum merges global headlines, expert commentary, and startup innovations into a single, time-saving digest built for forward-thinking professionals.
Rather than sifting through an endless feed, you get curated content that captures the pulse of the tech world—from Silicon Valley to emerging international hubs. Track upcoming trends, significant funding rounds, and high-level shifts across key sectors, all in one place.
Keep your finger on tomorrow’s possibilities with Memorandum’s concise, impactful coverage.
*This is sponsored content
 | You are a product strategist. Analyze this company’s AI strategy, identify the top three risks to its competitive position, and suggest one concrete move to mitigate each. Company: [paste details] |
🗳️ Meta's AI strategy: is the payoff ever coming? |
AssemblyAI: Speech-to-text API that handles transcription, speaker detection, and audio intelligence for production apps.
Chroma: Open-source vector database built for AI apps, fast to set up and easy to scale for RAG and embeddings.
Continue: Open-source AI code assistant that plugs into VS Code and JetBrains with full control over models and context.
What did you think of today's email? |
.svg)
.svg)
