The 2028 Countdown: Anthropic Says AI That Builds Better AI Is Now More Likely Than Not

TL;DR

• Anthropic co-founder Jack Clark has put a 60% probability on recursive self-improvement (RSI) — AI that can autonomously build a better version of itself — occurring before the end of 2028.
• The prediction, published May 7 as part of The Anthropic Institute's new research agenda, is the first time a major frontier-lab leader has attached a specific probability and deadline to RSI.
• Anthropic simultaneously shipped "dreaming" — a scheduled process that lets Claude agents review past sessions, extract patterns, and curate memories to improve over time without human intervention.
• The company also signed an exclusive compute deal with SpaceX for the Colossus 1 data centre in Memphis (the facility xAI built in 2024 to train Grok), giving Anthropic a massive hardware expansion.
• These three developments — the forecast, the capability, and the compute — form a single strategic signal: Anthropic is betting the company on self-improving AI, and it's telling the world on what timeline.

What Happened

On May 6–7, 2026, Anthropic made three announcements that, taken together, represent the most explicit statement any frontier AI lab has made about the near-term trajectory of AI capabilities.

First, at its Code with Claude developer conference in San Francisco, Anthropic unveiled "dreaming" — a system that lets Claude agents learn from their own past sessions. The mechanism works by scheduling a process that reviews agent sessions and memory stores, extracts patterns, and curates memories so agents improve over time. The original memory is never modified; instead, Claude creates a separate output memory store that developers can review, keep, or discard. ¹

Alongside dreaming, Anthropic moved two previously experimental features — outcomes and multi-agent orchestration — from research preview into public beta. ²

Second, Anthropic co-founder Jack Clark published a new research agenda for The Anthropic Institute, shared first with Axios. In it, Clark stated the company is seeing signs of "AI contributing to speeding up the research and development of AI itself." His central forecast: "By the end of 2028, it's more likely than not that we have an AI system where you would be able to say to it: 'Make a better version of yourself.' And it just goes off and does that completely autonomously." ³

Clark had previewed the 60% figure on X (formerly Twitter) on May 4, writing: "I think there's a 60% chance that recursive self-improvement (RSI) will occur before the end of 2028." He clarified he doesn't expect it in 2026, but predicts a proof-of-concept on non-frontier models within one to two years — a model that can train its own successor end-to-end. ⁴

Third, on May 6, Anthropic and SpaceX signed a partnership giving Anthropic full and exclusive access to Colossus 1, the Memphis, Tennessee data centre that xAI built in 2024 to train Grok. ⁵

What It Actually Means

These aren't three separate stories. They're one story told across three instruments.

The forecast (Clark's 60% by 2028) is the thesis statement. It tells investors, competitors, regulators, and enterprise customers: we believe the curve is about to bend, and we're naming the date.

The capability (dreaming + multi-agent orchestration) is the evidence. Anthropic isn't just predicting RSI — it's shipping the architectural primitives that make it possible. Dreaming is a mechanism for agents to improve without human retraining. Multi-agent orchestration lets those agents coordinate. Together, they form the scaffolding for the kind of autonomous improvement Clark is forecasting.

The compute (Colossus 1) is the fuel. You don't sign an exclusive deal for a data centre built by a competitor unless you expect to need enormous training runs. The Colossus deal signals that Anthropic is scaling its hardware footprint to match its ambition.

The through-line: Anthropic is no longer competing on the current generation of models. It's positioning for the generation after — the one where AI systems participate in building themselves.

The Clark Forecast in Context

Clark's 60% probability deserves careful reading. He's not saying there's a 60% chance of AGI by 2028. He's saying there's a 60% chance of a specific, narrower capability: an AI system that can autonomously produce a measurably better version of itself.

That distinction matters. RSI is a mechanism, not an outcome. But it's the mechanism that, if it works, makes all subsequent capability curves unpredictable. Once AI can improve AI, the bottleneck shifts from human researchers to compute and data — both of which scale differently than human ingenuity.

Clark's track record gives the forecast weight. He's been tracking AI progress publicly for years through his "Import AI" newsletter, and he's known inside the industry for calibrated, non-hype assessments. When he says he arrived at 60% after reading "hundreds of public data" points and was himself surprised by the conclusion, it's worth noting that this isn't a marketing number — it's a researcher updating his own priors.

The Hype Check

Let's be precise about what this is and isn't.

What it is: A specific, falsifiable prediction from a credible source with a track record, published alongside a research agenda that commits Anthropic to transparency about its own AI-accelerated R&D. Clark explicitly says he doesn't expect RSI in 2026 and frames the 2028 date as "more likely than not" — not certain, not imminent.

What it isn't: A product announcement. A guarantee. A claim that Claude itself is recursively self-improving today. The "dreaming" feature shipped this week is a scheduled memory-curation process — useful, novel, and directionally significant, but not the same thing as a model training its own successor.

The naming convention — "dreaming" — has drawn justified scepticism. WIRED published a piece titled "I Am Begging AI Companies to Stop Naming Features After Human Processes," and the criticism has merit. Anthropomorphising AI features creates confusion about what the system actually does. Dreaming is a batch processing job that reorganises memory stores. Calling it "dreaming" is evocative but obscures the mechanism. ⁶

That said, the naming debate is a sideshow. The architecture underneath — persistent memory, scheduled self-review, pattern extraction across sessions — is genuinely new in production AI systems. Whether you call it dreaming or "scheduled memory consolidation," the capability is real.

Stakeholder Landscape

Cross-Layer Implications

The SpaceX connection is stranger than it looks. Colossus 1 was built by xAI — Elon Musk's AI company — to train Grok. Anthropic now has exclusive access to it. This means the compute that was built to compete with Anthropic is now powering Anthropic. The deal suggests either that xAI has moved on to newer infrastructure, or that SpaceX (which owns the facility) found a more attractive tenant. Either way, the AI compute map just redrew itself.

Open-source agents are running the same play. Nous Research's Hermes Agent overtook OpenClaw for the #1 spot on OpenRouter's global rankings on May 10 — three days after Anthropic's announcements. Hermes Agent is also self-improving: it runs persistently with memory across sessions and builds reusable skills from experience. The self-improving agent paradigm isn't proprietary to Anthropic. It's emerging simultaneously in open-source. ⁷

The "dreaming" architecture has enterprise implications beyond AI. The pattern — scheduled batch review of session logs, pattern extraction, curated memory output that humans can audit — is a template for any enterprise deploying agents at scale. It solves the problem of agent drift (agents getting worse over time as context accumulates) by creating a structured review cycle. Expect this pattern to become standard in enterprise agent platforms within 12 months, regardless of vendor.

What This Means for You

If you're an AI/engineering leader

The 2028 clock is now the planning horizon. If Clark is right, the AI systems you deploy in 2027 will be qualitatively different from those you deploy in 2029. Build your architecture to be model-agnostic and agent-framework-agnostic. The vendor that leads today may not be the vendor that leads post-RSI.

Specific actions:

• Audit your AI procurement contracts for lock-in clauses that assume linear capability improvement.
• Evaluate whether your agent infrastructure supports memory consolidation patterns (the "dreaming" architecture) — this will become table stakes.
• If you're building on Claude's Managed Agents platform, the dreaming feature is in research preview. Test it on non-critical workloads now; the architecture pattern is what matters, not the specific implementation.

If you're in policy or governance

RSI needs its own risk category. It is not the same as "frontier model capabilities" or "agentic AI." It is a meta-capability — the ability to improve capabilities autonomously — and it requires distinct monitoring, reporting, and intervention frameworks. The 30-month window Clark describes is short enough that regulatory work needs to begin now, not after RSI is demonstrated.

If you're an investor

The Colossus deal signals that the AI infrastructure buildout is entering a second phase: not just building new data centres, but reallocating existing ones as competitive dynamics shift. The companies that control large-scale compute — even if they didn't build it — will have leverage. Watch for more compute-reallocation deals between AI labs and infrastructure owners.

If you're a general reader

Nothing about your daily life changes today. But Clark's forecast is worth filing away: it's the most specific, near-term warning from a credible source that AI progress may be about to enter a phase where the systems start contributing to their own improvement. When that happens, the pace of change stops being linear. The 2028 date gives you a benchmark against which to measure what actually unfolds.

Uncertainty Ledger

Bottom Line

Jack Clark just put a number and a date on the most consequential question in AI: when will the systems start improving themselves? His answer — 60% probability before the end of 2028 — is the most specific, credible forecast any frontier-lab leader has made. It's not a guarantee. It's not marketing. It's a researcher telling the world what he sees in the data, and what he sees is a curve that's about to bend. The dreaming feature and the Colossus deal are the infrastructure bets that back the forecast. Whether Clark is right or wrong, the 2028 clock is now ticking — and every AI strategy written before May 7, 2026 needs to account for it.