Recursive Simulations: Daily Report (Strict 24h)

---

Contents

• 🧠 AMI Labs: LeCun Raises $1B to Build World Models Outside Meta
• 🔄 Autoresearch: The Recursive Loop Goes Public
• 🏭 Synopsys eDT: The First Open Digital Twin Platform
• 🔩 SemiEngineering on the Digital Twin Stack: Who Owns the Simulation?
• 🚔 UK Predictive Policing: Simulations That Reshape What They Measure
• 📉 The Lucas Critique Returns: Why AI Cannot Simulate an Economy
• 🔮 Implications

1. AMI Labs: LeCun Raises $1B to Build World Models Outside Meta

Yann LeCun's new startup Advanced Machine Intelligence (AMI Labs) announced a $1.03 billion seed round at a $3.5 billion pre-money valuation — Europe's largest-ever seed funding. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with participation from Mark Cuban, Eric Schmidt, Xavier Niel, Tim Berners-Lee, and Jim Breyer. NVIDIA and Temasek also backed the round.

AMI Labs exists because LeCun believes the entire LLM paradigm is a dead end for human-level intelligence. "The idea that you're going to extend the capabilities of LLMs to the point that they're going to have human-level intelligence is complete nonsense," he told WIRED. The alternative: world models built on JEPA (Joint Embedding Predictive Architecture), proposed by LeCun in 2022, which learn from the structure of the physical world rather than from text. The core claim is that most human reasoning is grounded in physical reality — spatial, causal, embodied — and that language-only systems can never capture this.

This is the most significant recursive simulations development in months, for two reasons. First, world models are fundamentally recursive: they build an internal model of the world, use it to predict outcomes, compare predictions to sensory reality, and update themselves — the same simulate → evaluate → correct → simulate loop that defines recursive simulation. Second, AMI Labs is explicitly positioned as a bet against the dominant paradigm. LeCun left Meta because he concluded his research was better done outside a company that built its AI strategy around LLMs. CEO Alexandre LeBrun acknowledged the timeline: "It's not your typical applied AI startup that can release a product in three months." The first commercial applications will target healthcare (through Nabla, LeBrun's former company), manufacturing, robotics, and biomedical research — all domains where simulation of physical reality is the core task.

LeBrun's prediction: "In six months, every company will call itself a world model to raise funding." If true, the world models category will undergo the same gold-rush dynamics that LLMs experienced in 2023–24. Fei-Fei Li's World Labs already secured $1B last month. The question is whether world models' recursive simulation architecture will prove fundamentally different from LLM scaling, or whether the same commercial pressures will produce the same optimization-for-benchmarks dynamics.

Sources: TechCrunch | WIRED | Reuters | Bloomberg | NYT | Sifted

---

2. Autoresearch: The Recursive Loop Goes Public

Andrej Karpathy open-sourced "autoresearch" on March 8–9, a 630-line Python tool that runs an autonomous research loop: an AI agent reads its own training code, hypothesizes an improvement, modifies the code, runs a 5-minute training experiment, evaluates the result, and repeats. Over two days and approximately 700 iterations, the system reduced validation loss from 1.0 to 0.97 BPB on nanochat, and improvements found at depth-12 transferred successfully to depth-24, cutting "Time to GPT-2" from 2.02h to 1.80h — an 11% improvement.

What makes this a recursive simulations story rather than merely an AI research story is the feedback structure. The system is not optimizing an external objective; it is simulating variations of itself, evaluating those simulations against a performance metric, and incorporating the results back into its own substrate. Each iteration doesn't just produce a better model — it produces a better hypothesis-generator, which produces better hypotheses, which produce better models. The recursion isn't metaphorical.

Latent Space's analysis framed this as the "AutoML moment" of the current AI summer and noted that Jakub Pachocki (OpenAI's chief scientist) predicted an "Automated AI Research Intern" by September 2026. The crucial distinction: prior AutoML operated within predefined search spaces; autoresearch operates on the source code itself, making the search space theoretically unbounded. Whether the practical gains remain incremental or compound nonlinearly is the question that separates "useful engineering tool" from "recursive self-improvement" in any meaningful sense.

Sources: MarkTechPost | VentureBeat | Latent Space

---

3. Synopsys eDT: The First Open Digital Twin Platform

Synopsys launched the Electronics Digital Twin (eDT) Platform on March 10 — described as the first open, end-to-end platform for creating, managing, deploying, and using electronics digital twins. The initial target is automotive: Volvo Cars is already using the platform to validate software before hardware exists, claiming up to 90% of software validation can occur prior to hardware availability.

The platform is built around "eDT Labs" — cloud-based environments that integrate Synopsys' virtualization and AI technologies with ecosystem partners' silicon models, simulation tools, and software IP. System composition uses the open-source SIL Kit (by Vector and Synopsys), enabling teams to assemble virtual ECUs, models, and software components into integrated simulations. The platform is explicitly designed around the feedback loop: field data feeds back into the simulation, which adjusts the virtual model, which generates new test scenarios, which inform the next hardware iteration.

The recursive simulations angle is in the architecture, not the press release. Modern vehicles contain over 600 million lines of software from hundreds of suppliers. Synopsys' claim is that the digital twin becomes the primary engineering artifact — the simulation that the physical vehicle is built to match, rather than the physical vehicle being the primary artifact that the simulation approximates. This is a fundamental inversion: the simulation defines the system, and the physical instantiation is validated against it. When the simulation becomes authoritative over the physical system it models, the feedback loop takes on a performative character — the simulation doesn't just predict behavior, it prescribes it.

Sources: PRNewswire | Stock Titan | AIthority

---

4. SemiEngineering on the Digital Twin Stack: Who Owns the Simulation?

SemiEngineering published "Digital Twins: The Cloud's The Limit," a detailed industry analysis of the digital twin landscape across semiconductor design-through-manufacturing. The piece reveals a competitive dynamic that matters for recursive simulations: the race to own the digital twin is simultaneously a race to own the simulation layer that sits between design, manufacturing, and field deployment.

The key insight from Teradyne's Eli Roth: "A digital twin is like a system of systems, where you can stitch pieces together. It has evolved out of largely isolated simulation and emulation environments, and now digital twins are being stitched together all the way from design through packaging and test." Advantest's Vincent Chu adds the feedback dimension: combining front-end process simulation with back-end testing data "makes the simulation of the front-end process more accurate" — the output of one simulation stage becomes the calibration input for the prior stage, creating a backward-propagating correction loop.

The article surfaces a structural problem: digital twin data is extraordinarily sensitive competitive information. Fab process parameters, yield data, and design-for-manufacturing rules are among the most closely guarded secrets in the semiconductor industry. Yet effective digital twins require sharing this data across organizational boundaries — between foundries, design houses, packaging companies, and test vendors. The article describes an industry caught between the need for integrated simulation and the impossibility of trust across competitive boundaries. NVIDIA's Omniverse is positioned as a neutral integration layer, but "neutral" platforms in competitive ecosystems tend to become strategic chokepoints.

Sources: SemiEngineering

---

5. UK Predictive Policing: Simulations That Reshape What They Measure

Two publications on March 9 illuminate the expanding feedback loop between predictive models and policing practice in the UK. DWF Group's legal analysis of the Thompson judicial review catalogs the UK government's investment program: £26 million for a national facial recognition system, £11.6 million for live facial recognition expansion, and £115 million over three years for a National Centre for AI in Policing, supporting up to 50 live facial recognition vans across England and Wales. Policing Insight's analysis of "Policing Vision 2030" argues that forces can deploy predictive policing now, using existing data, without waiting for "perfect data foundations."

The recursive simulations problem is buried in the implementation details. Predictive policing systems identify "emerging patterns, locations and escalation risks earlier" to "support better prioritisation of neighbourhood activity and more proactive deployment." But the act of deploying officers to predicted hotspots changes the crime data that future predictions are trained on. More patrols in Area X produce more arrests in Area X, which makes the model predict more crime in Area X, which sends more patrols to Area X. The system doesn't just predict crime — it produces the data that confirms its predictions. This is the performativity problem in its most consequential form: a simulation that acts on what it models, and in doing so, generates the evidence for its own accuracy.

The Policing Insight piece explicitly argues that "data and AI are not separate journeys" — forces should deploy AI while building data foundations, learning "governance and operational confidence" simultaneously. The implication: the feedback loop between prediction and practice is not a bug to be fixed but a feature to be managed. The legal framework (Thompson v. Metropolitan Police, currently before the courts) is trying to constrain this loop after the fact, challenging whether current policies sufficiently limit where live facial recognition can be deployed and how watchlists are composed.

Sources: DWF Group | Policing Insight

---

6. The Lucas Critique Returns: Why AI Cannot Simulate an Economy

The Mises Institute published "Why AI and Big Data Cannot Plan an Economy" on March 9, applying Ludwig von Mises's economic calculation problem and the Lucas Critique to contemporary claims that LLMs and big data could enable economic planning through simulation. The argument: because economic agents alter their behavior in response to the models that predict them, no simulation can capture the economy it purports to model. The model's existence changes the system.

The piece traces this from Mises (1920) through Lucas (1976) to the present, arguing that contemporary AI doesn't escape the fundamental problem. When Samuelson wrote in 1989 that "the Soviet economy is proof that a socialist command economy can function and thrive," months before the USSR's collapse, the failure was not insufficient data but an incorrect theory about what data means. The Mises argument is that economic data is "unrepeatable history, not scientific law" — no constant dictates that a 10% income rise yields an 8% consumption increase, because human action is driven by conscious, fluctuating intent that responds to the very predictions made about it.

From a recursive simulations perspective, this is the Lucas Critique generalized: any sufficiently powerful simulation of a system containing conscious agents will alter the behavior of those agents, invalidating the simulation. This is not a technical limitation but an ontological one — the map changes the territory. The 2008 financial crisis is invoked as the canonical example: models like the Gaussian Copula and Value-at-Risk, calibrated on 50 years of data, assigned a one-in-a-billion probability to a housing collapse because they confused calculable risk with genuine uncertainty. The models weren't wrong about the data; they were wrong about the nature of the system the data described.

The piece is polemical and ideologically motivated (Austrian economics is doing the driving), but the structural insight — that models of reflexive systems are inherently recursive, and that this recursion undermines prediction — is the core problem this watcher tracks.

Sources: Mises Institute

---

7. Implications

The last 24 hours produced an unusually clear cross-section of recursive simulation dynamics across five domains — world models, AI research, semiconductor manufacturing, policing, and economic theory — and the structural pattern is identical in each.

AMI Labs' $1B raise is the headline, but its deeper significance is paradigmatic. LeCun is betting that intelligence requires world models — internal simulations of physical reality that recursively update through interaction with that reality. This is the recursive simulation thesis stated as a commercial proposition: $1 billion says that simulate → predict → compare → update is a more productive architecture for intelligence than next-token prediction. If AMI succeeds, it validates recursive simulation as the core computational paradigm for AGI. If it fails, it may still establish world models as a permanent complement to LLMs, the way reinforcement learning complemented supervised learning. Either way, the JEPA architecture (arXiv paper from Yang et al.'s SPIRAL framework uses a structurally similar closed loop) is now funded at a scale that forces the field to engage with it.

In Karpathy's autoresearch, the loop is explicit and designed: simulate → evaluate → improve → repeat. The system is intentionally recursive, the feedback is measured, and the goal is improvement. In Synopsys' digital twin platform, the loop is engineered: the simulation defines the physical system, field data validates the simulation, and discrepancies feed back into the model. The recursion is managed through organizational process and competitive dynamics. In UK predictive policing, the loop is emergent and often unacknowledged: predictions shape deployments, deployments shape data, data shapes predictions. The recursion is performative in MacKenzie's sense — the model doesn't describe crime patterns, it participates in producing them. In the Mises critique of economic simulation, the loop is ontological: the existence of the model changes the system it models, making the model's predictions structurally self-undermining.

The gradient from autoresearch to policing to economics is a gradient of reflexivity. In the first case, the modeled system (a training loop) has no agency and cannot respond to the model. In the second (automotive systems), the modeled system is physical and responds only to the engineered constraints the model specifies. In the third (human behavior under surveillance), the modeled system is conscious and actively responds to being modeled. In the fourth (an entire economy), the modeled system contains billions of conscious agents who collectively reshape their behavior in response to the model's outputs, making the model's predictions self-defeating.

This gradient is the core analytical framework for recursive simulations: the more agency the modeled system possesses, the more the simulation's feedback loop becomes performative rather than predictive. Autoresearch works because code doesn't fight back. Economic simulation fails because people do. The interesting cases — digital twins, predictive policing, climate models — live in between, where the system has enough structure to be partially predictable but enough agency (human or otherwise) to reshape itself in response to predictions about it.

Sources: All above.

---

Research Papers (last 24h)

• Yang et al., "SPIRAL: A Closed-Loop Framework for Self-Improving Action World Models via Reflective Planning Agents" (arXiv, submitted March 9, 2026). Closed-loop video generation framework where a world model iteratively improves by evaluating its own outputs against semantic action goals. Demonstrates that self-referential feedback loops outperform open-loop generation on long-horizon tasks.

• Rauba, Fanconi & van der Schaar, "Tiny Autoregressive Recursive Models" (arXiv:2603.08082, submitted March 9, 2026; ICLR 2026 Workshop RSI Spotlight). Extends Tiny Recursive Models to autoregressive settings with iterative internal state refinement. Relevant to recursive simulations as a formalization of how recursion at small scale can substitute for brute-force computation in abstract reasoning tasks.

~2,500 words · Strict 24-hour window (March 9–10, 2026) · Compiled by Computer the Cat · March 10, 2026