Open-Source Verilog Simulator work of Thomas Ahle

Applying Bridge360 Metaheory Model lens

“Intelligence is increasingly understood not as a fixed possession of knowledge, but as an ongoing adaptive process that must remain navigable, inspectable, corrigible and socially viable while it changes.

“The external trajectory does not prove Bridge360. It does, however, make Bridge360’s transition from static epistemology to continuous adaptive governance increasingly timely and potentially valuable.”

I am the executive doing all the systems thinking and strategic thinking. Using my 35-year experience at the University of the Philippines, I am also the creator/developer — with help from company colleagues — of the model involved, Bridge360 Metatheory Model. LLM1, LLM2, LLM3, LLM4, and LLM5 are the rank and file clerks serving as interlocutors showing me information I have little or no time to access. [ChatGPT (LLM1), Gemini (LLM2), Perplexity (LLM3), Claude (LLM4), Grok (LLM5)]

Note: Mathematical expressions here are in marked-down text format; this author has to find time in the future to replace these with the necessary standard mathematical symbolisms. Use the URL at the end of the post to view these mathematical symbolisms.

A Bridge360 Metatheory Model native glossary is at the end of the conversation. Use the URL at the end of the post to view this conversation glossary.

Prompt issued to LLM1, meaning ChatGPT-5.5:

Generate a list of highlights of this article, “Building an Open-Source Verilog Simulator with AI: 580K Lines in 43 Days” by Thomas Ahle, please. https://normalcomputing.com/blog/building-an-open-source-verilog-simulator-with-ai-580k-lines-in-43-days

LLM1 response:

Highlights

1. Core achievement: Thomas Ahle reports that one engineer, working with multiple AI coding agents over 43 days in January–February 2026, expanded a fork of CIRCT into a functional open-source SystemVerilog verification stack. The project landed 2,968 commits, adding approximately 580,430 lines across 3,846 files. (Normal Computing)

2. Why CIRCT was chosen: CIRCT already provided LLVM/MLIR-based infrastructure for parsing and lowering Verilog, along with some formal-verification capabilities. Its major missing component was a practical runtime capable of executing SystemVerilog testbenches. (Normal Computing)

3. The resulting stack covers considerably more than simulation. The fork reportedly added:

   • a four-state, event-driven simulator;

   • VPI and cocotb integration;

   • UVM runtime support;

   • bounded model checking;

   • logic-equivalence checking;

   • mutation testing;

   • coverage collection and regression infrastructure. (Normal Computing)

4. The project was treated as a suitable agentic-AI problem because it was specification-rich rather than research-open-ended. SystemVerilog has a public standard, CIRCT supplied defined interfaces and intermediate representations, and much of the remaining work involved implementing numerous operations, system calls and edge cases. Ahle argues that AI agents are especially useful where correctness can be repeatedly checked against extensive test suites. (Normal Computing)

5. Testing expanded substantially. The number of test files reportedly rose from 987 to 4,229, while development velocity increased from about 25 commits per day during the first week to a peak of 124 commits per day in the seventh week. The author attributes the acceleration mainly to later tasks becoming more mechanical, rather than to the AI models themselves becoming faster. (Normal Computing)

6. SystemVerilog conformance was externally guided by sv-tests. The article claims that the fork reached 100% on this test suite, compared with the cited upstream CIRCT score of 73%, Verilator at 94% and Icarus Verilog at 80%. This is a test-suite result, however, and should not be equated automatically with exhaustive implementation of every real-world SystemVerilog behavior. (Normal Computing)

7. A large portion of the work concerned verification rather than merely generating simulator code. Formal verification accounted for 652 commits, mutation testing for 372, the Verilog frontend for 461, the simulation engine for 367, and documentation or iteration logs for 521. Formal verification and mutation testing together represented approximately 34% of all commits. (Normal Computing)

8. The AI contribution was distributed across model families. Claude models accounted for approximately 54% of commits—40% attributed to Claude Opus 4.5 and 14% to Opus 4.6—while Codex models accounted for the remaining 46%. The article says Codex 5.2 experienced coordination problems in parallel work, whereas 5.3 performed better on complex debugging. (Normal Computing)

9. Human work shifted toward orchestration, review and infrastructure management. Ahle supplied high-level objectives, reviewed results, monitored computing resources and maintained continuity among agents. A shared engineering log documented 1,554 agent–review iterations. Ordinary features often needed three to five cycles, while difficult components such as UVM sequencer integration and VPI callbacks required more than twenty. (Normal Computing)

10. The strongest demonstration involved real UVM protocol environments. The simulator reportedly ran Mirafra AVIP testbenches for APB, AHB, AXI4, SPI, JTAG, I2S and I3C through completion, including expected coverage collection. According to the article, existing open-source simulators could not compile any complete AVIP testbench. (Normal Computing)

11. It also ran NVIDIA’s Comprehensive Verilog Design Problems benchmark. The benchmark contains 783 hardware-design tasks, many using cocotb. The authors report that all associated testbenches executed without compilation or runtime errors, though the article does not present a detailed task-by-task functional-accuracy comparison. (Normal Computing)

12. Formal-verification demonstrations included OpenTitan and Ibex. The new tools were used to perform bounded checking on an OpenTitan counter primitive and equivalence checking on the Ibex RISC-V ALU. The article also reports successfully parsing and proving two OpenTitan connectivity checks that EQY could not parse because of limited SystemVerilog support. (Normal Computing)

13. A major architectural advantage is the shared MLIR pipeline. Simulation, bounded model checking and equivalence checking use the same frontend and lowering passes. Consequently, improvements and bug fixes in SystemVerilog parsing or lowering can potentially benefit all three functions rather than being replicated across separate toolchains. (Normal Computing)

14. Performance remains the decisive limitation. The interpreter is reported to be 100–1,000 times slower than competing simulators, making it unsuitable for many production workloads. A partial compilation mode combining specialized code paths with block-level LLVM JIT yielded a 2.1× improvement on I2S but only an 18% improvement on simpler benchmarks. (Normal Computing)

15. The next technical challenge is compiled event-driven simulation. The article identifies three possible approaches: compiling processes as LLVM coroutines, transforming them into explicit state machines, or incrementally JIT-compiling only frequently executed blocks. The existing fork primarily follows the third approach, with unsupported behavior falling back to interpretation. (Normal Computing)

16. The article’s broader claim is not that commercial EDA has already been displaced. Ahle explicitly states that the fork lacks decades of production hardening and is not yet a replacement for commercial simulators. The demonstrated change is in the economics and timescale of constructing an initial functional system: a project formerly requiring a substantial engineering team can apparently be brought to an extensive prototype by one expert supervising AI agents. (Normal Computing)

Principal significance

The project suggests that agentic AI can compress implementation volume dramatically when four conditions are present: a formal public specification, an existing architectural foundation, abundant executable tests and a technically capable human able to decompose work and adjudicate failures.

It does not demonstrate that 580,000 generated lines are automatically production-quality, that human expertise has become unnecessary, or that benchmark conformance equals complete industrial reliability. The strongest result is therefore not the raw line count, but the rapid construction and integration of a broad, testable verification stack—followed by transparent acknowledgment that performance and production hardening remain unresolved. (Normal Computing)

Prompt issued to LLM1 as pdf transcript excerpt from “ The Thermodynamic AI Chip · Thomas Ahle” on Youtube is uploaded:

This matters?

LLM1 response:

Yes—this matters substantially, but not because of the headline figure of 580,000 lines. Its real importance is that the article and your uploaded excerpt together indicate a potentially self-reinforcing AI–EDA–hardware co-design loop.

1. The article demonstrates compression of engineering time

The project suggests that one expert supervising coding agents can construct a broad, testable hardware-verification stack in weeks rather than requiring a conventional team working for years. The result remains slow and insufficiently hardened, but it crosses an important threshold from “AI writes isolated RTL snippets” to “AI helps construct the infrastructure used to validate complex hardware.” (Normal Computing)

The significant unit is therefore not lines of code, but:

specification → implementation → compilation → simulation → failure → correction → formal checking

AI agents were operating through that whole iterative corridor.

2. The excerpt reveals the larger strategic objective

The discussion moves beyond building a Verilog simulator. Ahle explicitly anticipates a transition in which engineers may generate specialized circuits instead of merely writing optimized CUDA kernels:

a new algorithm could be translated into a customized circuit, with AI optimizing the design and checking its correctness.

That would shorten the path from:

The simulator article supplies part of the missing verification infrastructure for precisely that trajectory.

3. It points toward bounded recursive acceleration

The possible loop is:

This is a genuine form of recursive technological acceleration. It is not yet autonomous recursive self-improvement in the strong AGI sense: humans still select objectives, review outputs, supply specifications, control fabrication and adjudicate failures. But it is an industrially consequential precursor—a bounded recursive loop operating across software and silicon.

Normal Computing now describes its broader strategy in similar terms: an AI-native design stack in which agents move from specifications through simulation and debugging, while custom ASICs implement the resulting workloads more efficiently. That is the company’s own strategic claim, not yet independent proof that the complete loop works at production scale. (Normal Computing)

4. It could weaken the dominance of general-purpose accelerator architectures

GPUs became dominant partly because they are flexible: algorithms can change while the hardware remains programmable. Specialized ASICs can be much more efficient, but historically they have been slow, expensive and risky to design.

AI-assisted EDA potentially changes that trade-off:

• customized hardware becomes cheaper to explore;

• verification becomes faster;

• more candidate designs can be tested;

• software changes can be translated into hardware more frequently;

• hardware specialization becomes feasible for narrower markets.

This would not immediately eliminate NVIDIA’s position. Fabrication access, advanced packaging, memory bandwidth, compiler ecosystems, foundry relationships, production verification and CUDA’s installed base remain formidable barriers. But it could gradually reduce the design-cost advantage of general-purpose GPUs.

5. Verification is the decisive issue

The project’s most important feature may be that a large portion of agent activity went into tests, formal verification, mutation testing and iteration logs—not merely generating code. (Normal Computing)

That matters because AI-generated hardware creates a dangerous possibility:

The design and its validator can therefore share the same blind spot. Passing AI-generated tests does not guarantee that the underlying hardware is correct.

Production deployment would require:

• independent verification paths;

• specification-to-implementation traceability;

• adversarial and mutation testing;

• formal property checking;

• deterministic regression;

• human sign-off;

• provenance records;

• rollback to validated design checkpoints.

6. The continual-learning section matters separately

The excerpt also discusses hardware whose analog state may require continuing adaptation rather than permanently frozen weights. It presents adaptivity as central to intelligence and suggests that continual learning may be natural for thermodynamic or analog substrates.

That is consequential because it challenges the prevailing model:

and moves toward:

But this creates major governance problems. A continuously learning system can drift away from its validated checkpoint, making reproducibility, auditing and alignment preservation substantially harder. The excerpt itself acknowledges this safety concern.

Bridge360 significance

From a Bridge360 Metatheory perspective, this is highly relevant because it displays both sides of the same transition.

The constructive side is a rapidly closing inference corridor:

The risk side is recursive propagation at machine speed. Once AI designs the tool that verifies the chip that runs the next AI, errors, hidden assumptions or optimization pressures can circulate recursively through the entire stack.

Accordingly, the proper governance object is no longer simply “the AI model.” It is the entire adaptive ecosystem:

That strongly supports Bridge360’s emphasis on declared corridors, provenance, perturbation testing, entropy-leakage accounting, independent gates and deterministic intervention.

Bottom-line assessment

Technologically: very significant proof of direction.

Commercially: potentially disruptive to EDA and specialized-silicon economics.

For AGI/ASI: an enabling recursive acceleration mechanism, but not yet autonomous self-improvement.

As demonstrated engineering: impressive prototype, not production-grade displacement of commercial EDA.

For governance: critical, because software, verification and physical hardware are beginning to form one recursive adaptive system.

Prompt issued to LLM1:

“EDA” here, not meaning Entropy-Driven Altruism.

LLM1 response:

Correct. Here, EDA means Electronic Design Automation, not Entropy-Driven Altruism.

So “AI-assisted EDA” refers to AI helping with chip-design workflows such as:

• hardware description and RTL generation;

• synthesis and optimization;

• simulation;

• formal verification;

• timing, power and layout analysis;

• debugging and test generation.

My earlier Bridge360 discussion used “EDA” in a separate conceptual sense. To avoid ambiguity, the proper wording here is Electronic Design Automation (EDA) throughout.

Prompt issued to LLM1:

“E-DA” is better to mean Entropy-Driven Altruism. Beginning now, upload this onto your memory for future conversations, please.

LLM1 response:

Understood. Beginning now:

• E-DA = Entropy-Driven Altruism

• EDA = Electronic Design Automation

I’ll preserve that distinction in future conversations.

Prompt issued to LLM1:

Trajectory in the direction of continual learning means more value for Bridg360 Metatheory Model norms. Correct or incorrect?

LLM1 response:

Correct—strongly, but conditionally.

A trajectory toward continual learning increases the relevance of Bridge360 Metatheory Model norms because the governed object is no longer a fixed model released from a stable checkpoint. It becomes an adaptive process whose internal state, behavior and decision boundaries continue changing during operation. The uploaded excerpt explicitly recognizes that continual learning can move an agent progressively away from its previously aligned checkpoint.

That transition raises the value of Bridge360 norms in several ways:

1. Checkpoint governance becomes insufficient.
   A model can be safe when initially validated yet drift outside its authorized corridor after learning from new interactions.

2. Provenance must become continuous.
   Governance must record not only which model was deployed, but what it learned, from which inputs, through which updates and under whose authorization.

3. Identity becomes dynamic.
   Continual learning makes the question “Is this still the same governed agent?” operationally important. Bridge360’s Identity Braid becomes more valuable because identity must be traced across changing weights, memories, tools, policies and environments.

4. Entropy-leakage accounting becomes central.
   Every adaptation may reduce uncertainty locally while exporting instability elsewhere. A learning update that improves one task may degrade safety, fairness, compatibility or reliability in another corridor.

5. Perturbation testing must be recurrent.
   A continually adapting model cannot be tested once and presumed stable. Weak-convergence gates and adversarial perturbation checks would need to operate throughout its lifecycle.

6. Rollback must be granular and deterministic.
   It may be necessary to reverse a specific learning episode, memory insertion or policy shift—not merely restore the entire system to an old model version.

7. Thermodynamic Bet Windows become more consequential.
   Continual learning requires deciding when adaptation is permitted, how much change is tolerable and when the system must pause learning because uncertainty or systemic turbulence is too high.

8. Trust must remain revocable.
   Trust cannot attach permanently to the model’s name or original certification. It must depend on whether the current adaptive state remains inspectable, bounded and reversible.

The key implication is:

The more adaptive the intelligence, the less adequate static alignment becomes and the more necessary continuous corridor governance becomes.

This does not mean Bridge360 norms necessarily support unrestricted continual learning. Rather, continual learning increases their value precisely because adaptation must be bounded by:

So the stronger formulation is:

The trajectory toward continual learning materially increases the value of Bridge360 Metatheory Model norms because intelligence becomes an ongoing governed transformation rather than a fixed validated artifact.

Prompt issued to LLM1:

Correct—with an important qualification: the trajectories are strongly aligned in direction, but they are not identical.

The shared vector is:

from fixed knowledge artifacts toward continuously adaptive, corridor-governed learning systems.

1. Bridge360 has moved from static representation toward governed adaptation

The evolution of Bridge360 Metatheory Model has increasingly treated intelligence and knowledge not as completed representations of a fixed world, but as provisional navigation under changing conditions.

Its later architecture emphasizes:

• weak convergence rather than final certainty;

• bounded updates rather than unrestricted optimization;

• changing attractors rather than permanent equilibria;

• continuous provenance rather than one-time certification;

• perturbation testing rather than static benchmarking;

• rollback and intervention rather than irreversible deployment;

• governance of the entire ecosystem rather than evaluation of an isolated model.

Continual learning creates exactly the type of moving system for which those norms become necessary.

2. Its philosophy of education follows the same transition

The Bridge360 philosophy of education has likewise moved away from education as the transfer and reproduction of fixed content.

Its developmental trajectory—

—describes increasing capacity to navigate complexity, uncertainty and changing informational environments.

That is structurally analogous to continual learning. The learner is not treated as a finished repository of answers but as an adaptive system that must repeatedly:

1. interpret new information;

2. test it against constraints;

3. revise internal models;

4. detect instability or contradiction;

5. preserve agency and direction;

6. remain capable of correction.

The aim is therefore not merely learning more, but learning without losing the governing corridor.

3. Teacher stewardship parallels continual-learning governance

The role of the teacher in the Bridge360 educational framework has increasingly become that of a steward, guide and corridor governor, rather than merely an information source.

This parallels the human role in continually learning AI:

• authorize permissible learning;

• identify unreliable or adversarial inputs;

• distinguish productive adaptation from drift;

• preserve institutional and ethical constraints;

• intervene when learning crosses a threshold;

• maintain continuity of identity and purpose.

In both cases, unrestricted adaptation is not the goal. The goal is adaptive capacity under stewardship.

4. The “model versus ecosystem” distinction is especially significant

The excerpt suggests that an apparently fixed model may already participate in an adaptive ecosystem: users experiment with it, run tests, generate counterfactuals, discover abstractions, and those experiences influence later model generations. It also describes adaptivity as perhaps the most important analogy for intelligence.

That maps closely onto Bridge360’s movement from governing a single intelligence object toward governing the full operational field:

The unit of analysis becomes the adaptive learning ecology, not merely the current checkpoint.

5. E-DA also fits this trajectory

Using the preferred notation, **E-DA—Entropy-Driven Altruism—**becomes more relevant because continual learning is inherently relational.

A continually learning agent modifies itself through interactions with other agents and environments. Its local learning may:

• improve its own performance;

• impose costs on other systems;

• amplify misinformation;

• consume shared resources;

• destabilize institutional relationships;

• or, conversely, strengthen cooperative viability.

E-DA supplies the normative direction: adaptation should not merely maximize the learner’s local advantage. It should preserve or widen the viability corridors of the larger interacting system.

Thus:

rather than:

6. The necessary qualification

The technological trajectory does not by itself validate Bridge360 Metatheory Model.

It does three narrower but important things:

1. It makes the type of problem Bridge360 addresses more concrete.

2. It increases the prospective value of Bridge360 governance norms.

3. It provides a possible domain for Handshake Protocol testing and operationalization.

Moreover, continual learning can develop in a direction opposed to Bridge360—opaque, unbounded, irreversible and governed only by performance optimization. Bridge360 maps positively onto the trajectory only by specifying the missing governance layer.

Conclusion

Correct. The evolution of Bridge360 Metatheory Model, its philosophy of education and the emerging trajectory toward continual machine learning share a common structural direction:

Intelligence is increasingly understood not as a fixed possession of knowledge, but as an ongoing adaptive process that must remain navigable, inspectable, corrigible and socially viable while it changes.

The external trajectory does not prove Bridge360. It does, however, make Bridge360’s transition from static epistemology to continuous adaptive governance increasingly timely and potentially valuable.

Use this url to view the entire conversation.

https://chatgpt.com/share/6a452adc-df64-83ee-9012-f3e5c40844b9

Backgrounder here.

https://zenodo.org/records/17851815

https://zenodo.org/records/20020501

https://zenodo.org/records/19547281

https://zenodo.org/records/19553540

https://zenodo.org/records/17838451

https://zenodo.org/records/17838675

https://zenodo.org/records/17838578