Natural Systems in General

• A natural system is defined by a set of axioms
• The nature of natural axioms
  • If we can't know them, we can build approximate models for them
  • If we assume them precisely, they will be incomplete
• A natural system produces "observations"
  • "Truth" and "facts" are inherent in a natural system

Natural Systems - Examples

• There are only two examples of primordial natural systems

1. Our physical world
2. Concepts in pure mathematics
   • Number theory (e.g. Peano arithmetic)
   • Computation (e.g. Lambda calculus)
   • Formal logic (e.g. First-order logic)
   • Proof theory (e.g. Martin-Löf type theory)
   • Set theory (e.g. ZFC)

The pure study of natural systems is theoretical math / logic / physics / metaphysics / philosophy

Understanding Natural Systems

• Can we ever fully comprehend a natural system?
  • No: Gödel's incompleteness theorems
• Can we perfectly simulate a natural system?
  • No: Uncertainty principle
• Do we fully understand the physical natural system?
  • No: Cosmological constant problem

Engineered Systems

• The design of physical primitives is the core of applied science
• The physical manifestation of an engineered system is the point where abstractions are broken
• Engineered systems are perfectly simulatable with respect to their abstractions
  • Unlike natural systems, everything about an engineered system is known and complete
• Engineering research fits into these boxes
  • Study the implementation tradeoffs of a specification / task
  • Create new abstractions / models
  • Improve the design methodology and algorithms

The Missing Middle

???

Consider: biological systems, linguistics, artificial neural networks, symbolic AI

Evolved Systems

e.g. molecules, DNA, cells, linguistics

e.g. ANNs, animal husbandry

When is ML Powerful?

1. An evolved engineered system mimicking a natural system (i.e. generative)
2. An evolved engineered system acting within an engineered system (i.e. in/around-the-loop)

When is ML Powerful?

ML is powerful when it can wholesale break engineered abstractions to mimic naturally evolved systems

ML is weak when it is forced into a box defined by prior engineering efforts (e.g. surrogate functions, fixed abstractions)

ML is powerful when the problem is fuzzy and the solution can be too

ML is weak when the problem is well defined and it is trying to replicate an engineered algorithm

Ways to Use ML

• In-the-loop with an engineered algorithm
  • ML for deciding moves wtihin simulated annealing
  • ML for choosing branching variables within CDCL (SAT solvers)
  • In general, ML for heuristics and learned behaviors
• Around-the-loop with an engineered algorithm
  • ML for tuning hyperparameters of simulated annealing
  • Synopsys DSO.ai - hyperparameter tuning around CAD algorithms
  • In general, ML for embedding and regression
• "Generative" / "End to End"
  • ML goes from input to output, bypassing human engineered abstractions

The "Bitter Lesson"

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.

We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.

- Rich Sutton's "The Bitter Lesson"

Caveats of the "Bitter Lesson"

• Hypothesis: More data + compute + search + general techniques = better end to end performance
• The "Bitter Lesson" applies to mimicking natural systems
  • Not to mimicking engineered systems
• Deep ML works best for fuzzy solution spaces where errors are tolerable
• Out-of-distribution generalization is still hard

What do we do with the problems that are too hard to solve with the limited compute that we have? Lie down for 80 years and wait for compute to catch up? Or solve a smaller problem using specialized tricks?

The bitter lesson is nothing of the sort, there is plenty of space for thinking hard, and there always will be.

- Anon

Example - MNIST / Semantic Segmentation

Do you see the parallels with how ML is often "used" in EDA CAD?

Example - 3D Rendering

• What's the point of replicating an engineered algorithm?
  • The abstractions are unchanged! ML is weak!
• Training on data produced by an algorithm that encapsulates the entire problem isn't useful
• Instead, step back and train end-to-end models that learn novel representations of the input

More Examples

• Speech synthesis
• Language models (+ linguistics)
• Evolved antennas / RF circuits / silicon photonics / MEMS
• The evolution of Tesla FSD
  • From hand written state machines and control rules to a fully learned control blob

ML Surrogate Models for Engineered Algorithms

• Examples
  • What is the PPA of some block of RTL?
  • What is the max IR drop of this PDN?
• This is the most common type of ML for EDA CAD in academia
  • The data comes from an existing CAD algorithm
  • No new abstractions are learned
  • The out-of-distribution generalization is poor
  • These techniques can't ultimately improve QoR
• Why do people do this?
  • The CAD algorithm is too slow!
  • But how does this solve anything (algorithm-time moves to training-time)?
  • Is there a better way? (e.g. ML in the loop)

ML for RTL Simulation

• Single-shot, data flywheel, constant time! So cool!
• In principle, there is no reason you can't "one-shot" simulation
  • Many ML architectures are universal function approximators
• This is 'generative' AI, applied to an engineered system
  • An engineered system which requires exact solutions

This is just an illustrative joke

Ahh AlphaChip

AlphaChip was one of the first RL methods deployed to solve a real-world engineering problem, and its publication triggered an explosion of work on AI for chip design.

Nevertheless, as described in Sutton’s The Bitter Lesson, there is often reluctance to accept the application of machine learning to new areas, and ultimately this has led to some confusion around our work

- From the addendum

Why Didn't It Pan Out?

• Competing against classical algorithms based on force-directed placement and simulated annealing
• In principle, why couldn't it outperform human engineered algorithms? Nothing intrinsic about RL would prevent it from doing so.
• Inherent limitations in their approach
  • No simultaneous placement of soft and hard macros
  • No mixed-placement of standard cells directly with hard macros
  • Coarse placement grid for hard macros
  • Formulate macro placement as a game vs an optimization problem
  • Rewards only available at the end of a trajectory

The Big Mistake

The biggest mistake of them all is that they didn't build AlphaChip!

They built Alpha-HardMacroPlacement

• No new learned abstractions
• No learned novel representations of the input data
• Not even an attempt to put ML in/around the loop of force-directed mixed-placement

The solution is to go bigger. They weren't ambitious enough.

Altman's Plan? E2E Circuits?

• No pcells, no stdcells, no synthesis, no routing rules, no DRC, no LVS
• Learn a new fuzzy abstraction stack from scratch
  • specification → photomask → silicon → training data

LLMs for RTL

• It will continue to get better
• As of a few days ago: arXiv
• 

• Lots of startups in this area
  • ChipAgents, ChipStack, MooresLabAI
• Don't do this in academia, please collect your VC money
• Focus ML efforts on in-the-loop algorithmic enhancements and breaking engineered abstractions (the RFIC area is ripe)

The AI Scientist

• Has anyone tried this use case?
• My experience has been quite poor. LLMs will just regurgitate buzzwords
• High potential for drug discovery
  • Mechanical search, mechanical investigative process, mechanical hypothesis building
• Low potential for true discovery (not so mechanical)
  • Drug discovery, materials engineering, and mathematics seem like the first targets