Alan Sass

AI R&D. Engineering Perception. Interdimensional Generation Engineer (USA).

Spent my life training to conquer the mountain, only to realize I am the mountain.

[Alan Sass @alansass]

Can N-body physics simulations reveal hidden structures in corporate ownership networks?

…

Hello, World!

I've been building an intelligence platform that maps the relationship networks between public companies, their subsidiaries, executives, directors, and beneficial owners by cross-referencing SEC filings with property records, court cases, professional licenses, and business registrations.

The core idea is simple: public companies disclose a lot of information, but they disclose it in fragments across dozens of disconnected systems. A 10-K filing tells you about subsidiaries. A 13F tells you about institutional holdings. A DEF 14A tells you about executive compensation and board interlocking. But none of these filings link to the property records showing what real estate those executives personally hold, or the court cases involving their other business entities, or the LLC registrations that connect seemingly unrelated companies through shared registered agents.

When you unify all of that into a single knowledge graph, structures emerge that aren't visible in any individual data source. I was curious so I started with Nevada, and a list of companies w/SEC filings.

System Details:
The system ingests SEC EDGAR filings (10-K, 10-Q, 8-K, 13F, DEF 14A) and beneficial ownership filings (13D, 13G, 3, 4, 5) and extracts entity relationships: parent-subsidiary chains, officer and director appointments, institutional holders, insider transactions. It then cross-references those entities against state-level business registrations, county property records across multiple jurisdictions, court filings, and professional licensing databases.

The relationships are typed and weighted across dozens of canonical types: ownership, board membership, executive role, family, property, legal, identity. Each edge in the graph carries a confidence score based on how it was derived. A relationship extracted directly from a proxy statement has higher confidence than one inferred from address matching.

The result is a knowledge graph where a single query can surface things like: a CEO's spouse holds property through a trust that shares a registered agent with an LLC that is party to litigation in a different county. These are connections that span five different government databases and would take a human analyst days to trace manually.

The Physics Simulation:
This is the part I find most interesting from a technical standpoint. To visualize these corporate networks, I'm using adaptive force-directed layouts built on Barnes-Hut N-body simulation. I believe that’s the same algorithm used in astrophysics to simulate gravitational interactions between large numbers of bodies.

What makes it unusual is that the physics parameters aren't uniform. Each edge in the graph has dynamically computed spring lengths and force constants based on the relationship type and strength.

The gravitational repulsion, damping coefficients, and collision avoidance parameters are tuned per-graph based on topology metrics with density, hub count, clustering coefficient, and community structure. There's an optimization layer that measures layout quality across multiple dimensions (text collision rate, node spacing distribution, edge crossing density, hierarchy visual clarity) and adjusts the physics parameters in a feedback loop until the layout converges on something readable.

The effect is that the spatial layout itself becomes information. Clusters of tightly related entities naturally group together. Holding company structures form visible hierarchies. Cross-cutting relationships between otherwise separate clusters become visually obvious as long edges spanning gaps in the layout. You can see the structure of a corporate empire before you read a single label.

I've started thinking of it less as "visualization" and more as a computational physics system that happens to produce visual output. The layout is the solution to an energy minimization problem where the energy function encodes business relationship semantics.

The Intelligence Layer:
Beyond the graph structure, each entity gets a multi-dimensional score vector rather than a single number. The dimensions span categories: financial capacity, network centrality, contact completeness, temporal patterns, resource indicators, government relationship complexity, and others.

The reason for a vector instead of a scalar is that two entities can have identical composite scores for completely different reasons. A large holding company scores high on financial capacity and network influence but low on direct property holdings. An individual beneficial owner might score high on property concentration and legal exposure but low on institutional visibility. The vector preserves that distinction, and different use cases can weight the dimensions differently.

Risk propagation works like heat diffusion through the graph. If an entity has adverse court filings or regulatory actions, that signal propagates to connected entities with per-hop decay and attenuation based on relationship type. An ownership edge conducts more risk than a social association. The propagation depth, decay rate, and edge-type conductivity are configurable: a compliance analyst looking at direct exposure wants shallow propagation with slow decay, while someone mapping systemic risk wants deep propagation with faster decay.

Entity Resolution (one of the hard problems):
The core challenge is that the same real-world entity appears differently across every data source. SEC filings use formal legal names. State business registrations use slightly different legal names. Property records use whatever the county assessor entered, which might be a person's name, an LLC name, or a trust name with inconsistent formatting. Court records use yet another variation. And beneficial owners actively structure their holdings to avoid obvious connections.

The false positive problem is existential for this kind of system. If you incorrectly merge two different people into one node, you've created a phantom relationship network that doesn't exist in reality, and every downstream analysis is corrupted. I built a verification pipeline with escalating confidence thresholds so that a match has to clear multiple independent checks before it's accepted as canonical. Below-threshold matches go into a review queue rather than being auto-merged.

There's also an adversarial dimension I didn't anticipate. When you're ingesting data from hundreds of sources with varying quality, you need to catch data that contradicts high-confidence established facts, data that looks structurally anomalous, and patterns that suggest injection or corruption. I built a quarantine system that intercepts suspicious data before it enters the graph, flags it for review, and maintains rollback capability. This isn't theoretical as data quality issues from inconsistent government sources were corrupting the graph until I added this layer. This also helped me spot many data poisoning attacks related to synthetic data from other systems.

The Autonomous Systems:
One design principle I committed to early was that every subsystem should be capable of operating without human intervention. This compounded into something larger than I expected. The platform now has roughly 60+ autonomous systems across the stack which are not microservices in the buzzword sense, but genuinely self-directing components.

The intelligence layer alone has seven autonomous systems: a graph explorer that traverses the network looking for unexplored relationship paths and queues them for investigation, a merge detector that continuously scans for duplicate entities that slipped past initial resolution, a verifier that re-examines existing relationships against new incoming data and downgrades confidence scores when contradictions appear, a node validator that classifies and reclassifies entities as new evidence arrives, and a visual graph optimizer that measures layout quality and tunes the physics parameters without manual intervention.

Beyond intelligence, there are autonomous systems for anomaly detection (both streaming and batch), data quality remediation, change detection across monitored data sources, predictive analytics that re-train on new data, search index synchronization, and self-healing infrastructure that detects degraded components and attempts recovery before alerting. The ML subsystem alone spans nearly 60 modules: anomaly detection, deep learning, graph neural networks, NER, time series forecasting, and an AutoML pipeline for model selection and hyperparameter search.

What I Learned:
The knowledge graph is only as good as the entity resolution. I spent roughly 2/3 of the engineering time on the matching, verification, and deduplication pipeline, and 1/3 on everything else combined. This ratio really surprised me. The graph algorithms, the physics simulation, the scoring engine: all of that was comparatively straightforward. Getting the nodes right is the hard problem. Getting the edges right is the next hardest. Everything downstream is just math on a correct graph.

Corporate ownership structures are adversarial by design. Not in the security sense, but in the information-theoretic sense. The whole point of a multi-layered LLC/trust structure is to make it non-obvious who controls what. Building a system that sees through this is fundamentally an adversarial information extraction problem, and I've started thinking about it in those terms. What information is the structure designed to hide, and what set of cross-referenced data sources is sufficient to recover it?

The visualization isn't a feature, it's the product. When a non-technical user sees the network graph for the first time and can visually trace an ownership chain through three shell companies to a beneficial owner they didn't know existed, that's the moment they understand what the platform does. No table, no report, no API response creates that understanding. The physics simulation creates an intuitive spatial representation that communicates graph topology faster than any other interface I've tried.

Autonomous systems create a supervision problem of their own. An autonomous merge detector that's too aggressive will silently corrupt your graph. An autonomous anomaly detector with poorly calibrated thresholds will either miss real problems or cry wolf constantly. Each autonomous component needs its own monitoring, its own confidence thresholds, and its own rollback capability. The quarantine system I mentioned exists partly because I learned this the hard way: autonomous ingestion without autonomous verification is a liability, not a feature.

It’s only been a few weeks of build time so far but I’m looking forward to my plans for the future.

…

I'm curious what everyone thinks about this approach, especially from anyone who's worked on entity resolution at scale, physics-based visualization of non-physical systems, or applying GNNs to real-world financial graph data.

I'd appreciate any and all recommendations or comments. Thank you!

[Varun @varun_mathur]

five days of fable 5. my stats:
- 3.4 billion tokens
- $5,718 equivalent API spend
- $28,084 saved due to caching

i ran it 24x7, didn't sleep much this week, and used it for everything - from advanced LLM research, to system architecture, to application UI/UX and builds, to full product roadmap review, to creating investor presentations. it felt like having the combined intellect and power of a team of top-tier PhD AI/ML researchers, top-tier software architects with decade of experience, to a security-obsessed engineer, and more - all rolled into one, always-on, team. felt like my personal productivity increased 10x.

...and now, it's gone, poof, the keyser soze of models.

[Anthropic @AnthropicAI]

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Access to all other Claude models is not affected.

We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.

Read our full statement: https://t.co/bwn0sximKZ

[Alan Sass @alansass]

@scott_seiver @gunslingerbach accidents happen amirite 🤷🏻‍♂️

[Alan Sass @alansass]

@boisechampion from another perspective: you’re lucky no one in your family caught hantavirus from the cabin… yet?

[Alan Sass @alansass]

@leerob Congrats on the future acquisition by @SpaceX

[World Labs @theworldlabs]

Today we are sharing three new research papers, each exploring a new way to generate 3D content by leveraging large-scale generative models and 2D priors.

These projects were led by our incredible interns @HaoZhang623 @BDuisterhof @DrTunnels

[1/4]

[ZaxtR @AmX_ZaxTr]

MACROHARD

[Bedrock Robotics @BedrockRobotics]

Autonomous excavation is a genuinely hard problem to solve because these machines can't just follow scripts.

We're building a system that takes in a target end-state (the intended shape of a trench, a graded surface, a loaded truck) and continuously plans a path toward it. The model replans roughly every half second, using the freshest sensor data available. That cadence matters because the environment is always changing, a dump truck shifts position, material behaves differently than expected, or the machine pauses to let a person pass by.

Getting the balance right is part of what we're solving for. Our system is designed to be responsive without being reactive: making confident, deliberate moves while staying ready to adapt.

If this is the kind of problem you want to spend your time on, we'd love to hear from you.

[Alan Sass @alansass]

@LasVegasLocally this is true but Vegas still has room to grow since Macau has KTVs inside the casino on higher floors.

[Heart of the Beast @heartofthebeast]

It’s not who you live for… It’s who you’d die for. Watch Brad Pitt in the new trailer for David Ayer’s HEART OF THE BEAST - only in theatres September 25th.

[Richard Raymond @curbyourchichi]

Five years. That's how long I chased this one. A script by Cameron Alexander that I fell in love with the moment I read it — a film that had passed through a lot of hands over the years, none of which quite knew what they were holding or how to treat it. I did. So I kept knocking, and I waited, until it was finally mine to look after.

I grew up around German Shepherds. My mum breeds and trains them, and I've always believed they're the best breed going — full stop. I wanted to make the definitive German Shepherd film, and in Cameron's pages, I saw it.

We redeveloped it together and brought on Andrew Simpson, the best dog trainer in the business — if you've seen the Malinois in John Wick, you've seen his work. Then I handed it to two dear friends, Damien Chazelle and Olivia Hamilton, who saw exactly what I saw: a gritty survival thriller made for the big screen. They took it to Paramount, who fell for it as hard as we had. Before I knew it, Brad Pitt had come aboard, with David Ayer directing — an absolute maestro — and we were shooting in New Zealand.

Here's the part that still gets me. Back in the 90s at Pinewood, I was a runner making Brad his tea and coffee on Interview with the Vampire. To now be one of the producers on what may be one of his very best films… I don't have a word for it other than: pinch me.

HEART OF THE BEAST. In cinemas September 25th.

I am so proud of this one. It's everything I hoped it would be.

And look — everybody loves a dog. It's about the only thing the whole world agrees on. So let me put your mind at ease right now: the dog does not die. You're welcome. 🐾

[Mark Gadala-Maria @markgadala]

It's rare that AI video makes me feel any emotion.

But these "nostalgiamaxxing videos" really do capture what it's like to grow up in the 90's

Credit: homeforchristmasofficial

[Séb Krier @sebkrier]

Over the past few months I've been working on a very exciting project: a new $10m fund for research on multi-agent multi-principal AGI safety! Instead of focusing on single agent alignment and centralized control, we're looking to support research focusing on multi-agent settings, mechanism design, cooperative AI, and coordination problems.

This is a joint initiative between @GoogleDeepMind, @Googleorg, @schmidtsciences, @coop_ai, and @ARIA_research. Huge thanks to @James_D_Fox, @weballergy, @FranklinMatija, @lrhammond, and @ObadiaAlex for their invaluable work!

See: https://t.co/L5351OpPqH

Apply: https://t.co/a1uJLJnfYw

[Alan Sass @alansass]

🔥

[skull @brutecat]

Hacking Google with A.I. for $500,000

https://t.co/S8NbDU9ZaN

[Chamath Palihapitiya @chamath]

https://t.co/MCXWZTVvte

[Andrew McCalip @andrewmccalip]

Get paid to wait

The Claude Code spinner might be the most watched line on Earth.

So I turned it into an ad marketplace.

Advertisers bid on it. You keep 50% of the money.

Install the extension → get cash from ads.

Introducing Kickbacks

[Alan Sass @alansass]

@Polymarket

[NASA Earth @NASAEarth]

It’s like CSI…but with rocks 🕵️‍♀️🪨

Scientists hiked into the Mojave Desert to investigate a mineral “fingerprint” detected by a NASA sensor. Topaz was hiding in plain sight, which could hint at something more valuable underground known as porphyry copper.

[Alan Sass @alansass]

🧐

[Recursive @Recursive_SI]

https://t.co/IlLyWptydX

[Tether @tether]

Tether is leading a landmark Series C financing round of up to $1.4 billion for NEURA Robotics, @NEURARobotics , representing one of the largest private investment rounds in humanoid robotics history.

As robotics moves into true autonomy, payment and compute systems must evolve. Tether is deploying its core technologies directly into the Neuraverse ecosystem. By integrating our open-source Wallet Development Kit ( @WDK_tether), we are embedding self-custodial wallet functionality into advanced robots so they can independently participate in the economic system. Simultaneously, NEURA will deploy Tether’s @QVAC edge-first AI runtime, allowing AI models to execute locally on-device rather than relying on remote cloud infrastructure.

Together, Tether and NEURA Robotics are building the foundation for the machine economy.