Brightwave: financial research built from bounded AI systems

Seeing the economy through individual actions

A person changing jobs leaves a small digital trace. Hundreds of millions of job changes can reveal the movement of labor through the economy. Mike Conover describes a Nature Communications paper on talent movement that hierarchically clustered 500 million job transitions into a network predictive of next-quarter S&P 500 market-cap changes. The result illustrates his broader interest: individual actions expose systems people cannot see from their local perspective. He gives no prediction accuracy or investment returns.

That interest connects Conover’s research and financial machine-learning work. His DARPA-funded PhD used early Twitter data to study propaganda and misinformation networks. At LinkedIn, he worked primarily on newsfeed relevance and served as a research liaison for the Economic Graph Challenge. At Workday, his financial machine-learning work brought the same perspective to organizations: digital records reveal collective decisions about resources. Markets, in his view, are another record of that process.

He extends this perspective to pretrained language models, which he sees as compressions of the world at training time—including its beliefs and polarization. Gender–profession associations in Word2Vec provide his example of social assumptions becoming part of model structure. Broad exposure to information does not make a model independent of the society that produced it.

Turning financial documents into an investment view

In this 2024 conversation, Conover introduces [Brightwave](https://www.brightwave.io/), a research partner for finance professionals, and announces a $6 million seed round led by Decibel, with Point72 and Moonfire Ventures participating. Its premise is that finding overlooked information about mispriced assets strains human attention. He distinguishes working memory from reasoning skill: human demonstrations may not yield superhuman reasoning, but he argues that models already exceed people’s capacity to consider information together.

The team combines several kinds of expertise: - Models and retrieval: co-founder and CTO Brandon Kotara led semantic search across hundreds of millions of résumés and job listings at Workday. - Reliable infrastructure: Kotara’s background also includes serving as CTO of a federally regulated derivatives exchange. Conover considers predictability, repeatability, and observability as challenging as model methodology. - Financial judgment: domain specialists assess the depth, quality, and departure from consensus of an analysis—judgments that require more expertise than evaluating a grammatical correction.

Conover’s product example begins with NVIDIA’s GPU position and possible rare-earth shortages. Brightwave follows the question into Chinese restrictions on germanium and gallium transfers after A100 export controls, and Chinese control of mining operations in Africa and South America. He explicitly distinguishes germanium and gallium from rare-earth metals: they are semiconductor inputs. The intended research motion is to follow a premise several steps into a supply chain, including adjacent facts that change how the original question should be understood.

Financial research faces two linked constraints: reading filings and transcripts, then solving the blank-page problem—turning facts into an articulated view. Brightwave reports customers ranging from owner-operated registered investment advisors managing roughly $500 million to hedge funds managing tens of billions, with distinct uses: - Hedge funds: investigate investment theses through company value chains and supply chains. - Wealth managers: prepare for successive client conversations. - Investor relations and corporate strategy teams: develop a view of unfamiliar markets and themes quickly.

Larger context does not guarantee deeper synthesis

Brightwave’s architecture grew from a disappointed expectation: in the team’s experience, larger context windows did not reliably produce deep synthesis across disparate material. Recovering a fact and integrating evidence are different capabilities. A needle-in-a-haystack task tests whether a model can find a particular item in a long input. Success does not establish that it can build a useful investment argument from many filings.

A book report illustrates the bottleneck. A short response about an entire book leaves less room for individual passages than separate chapter synopses. Input capacity does not automatically increase output depth. An informal estimate in the conversation puts characteristic responses around 1,200 tokens and describes difficulty eliciting 5,000-token answers. Without specified models or settings, this motivates decomposition rather than establishing a universal output limit.

Brightwave instead describes specialized subsystems that perform bounded tasks and pass their outputs onward. Retrieval-augmented generation (RAG) supplies retrieved evidence to a model when it generates a response. Dividing documents into smaller reasoning units remains important even when the model can accept the entire document. Whether synthesis-oriented training demonstrations could solve this problem more robustly remains open.

An example involving drugs and the diet industry shows the required reasoning. If an investor expects a drug category to disrupt that industry, an unfavorable clinical-trial result needs interpretation: how should it change the investment thesis, the investor’s worldview, or the fund’s strategy? Retrieving the result supplies an input; synthesis connects it to a decision.

The interviewer raises a related hypothesis: emphasizing exhaustive recall might discourage models from ignoring repetitive filing boilerplate. Brightwave’s response is to select passages by their purpose and relevance. Domain knowledge helps identify what a clause means and whether it matters to the question. A sliding window over the corpus does not supply that judgment.

Check source support; learn the user’s priorities

Factuality requires several kinds of work. The team describes these approaches without detailing its full implementation: - Repeated assessment: revisit conclusions and compare different models’ outputs, potentially treating agreement as a voting problem. - Source entailment: train models to assess whether primary sources support a generated statement. - User-directed verification: allow additional compute to check a surprising finding. Bard’s double-check feature is offered as a product-design example.

Entailment asks whether the evidence supports the statement, a narrower question than whether an investment thesis will prove profitable. The discussion leaves open whether models trained through next-token prediction can guarantee factual correctness. Secure code generation and execution might help with quantitative reasoning, but the product still needs to accommodate fallible outputs.

Users also possess information outside the system: expert conversations, unstated investment theses, and beliefs that cannot yet be verified or falsified. Asking them to document their entire worldview is unrealistic. Brightwave suggests follow-up questions and observes which directions they pursue. These revealed preferences—interests expressed through investigative choices—could guide personalization at the individual, fund, and firm levels.

Evaluate automated judges against domain experts

Brightwave uses language-model judgments to evaluate multiple subsystems, including nongenerative components and work during fine-tuning. It also pays human annotators to assess outputs. A small expert-annotated corpus is the reference standard for deciding whether an automated judge or heuristic is useful.

One deliberately crude heuristic would prohibit generated numbers absent from the source material. Brightwave explicitly says it does not use this rule; it illustrates a check whose quality could be compared with human judgments. Combining several imperfect supervision sources may provide more value than any one alone—the discussion cites Snorkel as taking a similar perspective—but the combination still needs evaluation against people. Nuanced judgments cannot always be reduced to multiple-choice questions.

Making those human judgments reproducible is substantial work. Even skilled annotators can struggle to apply a detailed rubric consistently. Defining high-quality insight and organizing the people who assess it remain difficult to automate.

Databricks’ employee-created dolly-15k dataset introduces two organizational distinctions: - Sharing depends on incentives. The explanation offered for Databricks releasing the data is that wider access benefited the company; many firms would retain such a dataset as an exclusive advantage. - A dataset push is different from a production process. Repeatedly producing useful annotations requires sustained coordination and people management, which helps explain reliance on external data providers.

Preserve time, provenance, and source meaning

Financial retrieval makes small textual differences consequential. The questioner contrasts 50% and 30% gross margins: similar passages can describe materially different economics. Quarterly updates can also supersede earlier information. Brightwave assigns specific goals to subsystems, including temporal interpretation and connecting changing statistics to accurate narrative descriptions.

Temporal relevance depends on the question, not simply the newest document: - Breaking news requires evidence appropriate to the current event. - An 18-month history of Federal Reserve interest-rate policy requires evidence spanning that period. Retrieving the closest semantic matches without temporal or qualitative metadata can produce related facts that fail to answer either question. Brightwave describes multiple checks of generated statements against retrieved content because it considers convincingly wrong answers the worst failure mode. No measured error rates accompany that account.

Source handling also covers sensitive customer material. Brightwave says it anticipated controls over access, management, and treatment throughout the data lifecycle in a system serving multiple customers. Combining nonpublic material with public or semi-public evidence is part of its strategy for producing insight that differs from consensus; it is not a reported investment result.

Composable prompts adapt instructions to the retrieved material. Instructions might identify an uninteresting filing section or indicate that analysis comes from a favored source. Across chains of seven to ten inference calls, metadata should travel with the information so downstream components retain its origin and the confidence placed in its source. Seeking Alpha illustrates why this matters: contributors vary in quality, and users have preferences about whose analysis they value.

Extract reusable relationships and specialize proven behaviors

Brightwave reports substantial investment in knowledge graph extraction: converting documents into structured relationships that can be explored and reasoned over. A single extraction pass can produce reusable information, reducing repeated document analysis for each query. These artifacts supplement other retrieval methods and enter the RAG context alongside text.

The proposed graph goes beyond recording that a person works for a company. It could record that someone submitted congressional testimony critical of an organization, with the testimony’s language attached. Users could explore those relationships, and models could reason over them. Exceeding the granularity of Bloomberg or LinkedIn is presented as an opportunity rather than an achieved comparison.

The team distinguishes two model responsibilities: - RAG supplies explicit evidence: a known collection of facts is available when the model generates its answer. - Fine-tuning specializes behavior: once a task proves useful and repeatable, human-created demonstrations train a subsystem to perform that class of work. This is a preference for behavioral specialization, not a claim that fine-tuning cannot add facts. The speaker takes no strong position on how much new information it can add beyond pretraining.

The preferred agent architecture resembles a finite-state machine: the system operates within bounded modes of behavior and chooses which state to enter next to satisfy its goal. The team reports obtaining useful reasoning by defining tasks clearly, without elaborate human personas. It also uses classical machine learning and traditional deep learning where appropriate. The task determines the component.

Why spreadsheet generation is outside the product

Brightwave includes quantitative reasoning but explicitly excludes generating Excel spreadsheets from its goals. Its reasoning concerns both errors and the purpose of financial modeling: - A hidden formula bug can invalidate the work. A user may spend hours with a spreadsheet before discovering an error that undermines its assumptions. In a research dialogue, the user can challenge or discard an individual finding and continue. - Building the model develops the investor’s view. Deciding how a company works and which performance drivers matter is part of the value. Receiving a finished model may bypass that thinking, much as receiving an essay differs from writing one.

Tables remain essential inputs. Brightwave emphasizes retrieval over tabular data, though it does not disclose implementation details or quality measurements. The spreadsheet decision is explicitly a judgment about the technology in 2024. Future workflows could include models commenting on spreadsheet analysis or using financial-model outputs for qualitative reasoning. The expectation is that better tools expand analytical sophistication; who ultimately constructs the models remains open.

Generate investment theses; retain the decision

Systematic and semi-systematic hedge-fund trading groups are described as resembling large machine-learning teams. That is distinct from an autonomous AI operating a trading desk, which the speaker has no direct visibility into. The interviewer asks how predictive approaches might extend into thesis-driven investing, where new companies can lack useful historical precedent. The response expresses confidence in automating idea generation while leaving autonomous thematic investing unresolved.

A reported Brightwave example follows AI → energy → long-duration energy storage. Starting from interest in AI investing, it identifies energy as an adjacent opportunity that might be priced differently from GPUs, then storage as a bridge between renewables and data-center compute requirements. The speaker reports seeing reasonable second- and third-order theses, not investment returns.

The proposed human role is to test these ideas against fund strategy and knowledge from management conversations, then make the decision. Fully autonomous investing remains possible in this account, but the speaker declines to forecast it confidently because the possible outcomes diverge too widely.

Build private evaluations and learn to shape training data

The closing advice begins with private evaluation corpora for testing models on fresh examples. One concrete starting point is to assign each team member 250 annotations. Doing the work internally also forces the team to define its loss function: what behavior it wants, which mistakes matter, and why it prefers one output over another.

The economic argument reflects the conversation’s 2024 context. The speaker predicts fewer companies will pretrain independent foundation models as their period of competitive advantage shrinks. A hypothetical $5 million–$40 million effort illustrates the investment decision, explicitly excluding the larger scale of a 400-billion-parameter model. Similar behavior across models over the preceding few months raises the possibility of building on common foundations such as Llama; overlapping pretraining data is suggested as a possible cause, not an established explanation.

Mike sees a more accessible near-term opportunity in training data that induces specific behaviors. His experimental practice is concrete: 1. Compose a training corpus and fine-tune a model. 2. Inspect many generated outputs to understand their qualitative behavior. 3. Add or remove a set of demonstrations, repeat the process, and examine what changes. Repeated experiments build intuition about how corpus composition shapes behavior. He describes this experience as more accessible than a year earlier, when scarce high-performance GPUs made it expensive, and expects such skills to produce differentiated models and products.

Alessio argues that datasets retain value longer than individual models, citing the Pile’s continued role in training as older models lose their relative standing. Mike corrects his recollection that Dolly had been the best open-source model: it was never state of the art. Its contribution, he says, was demonstrating something many people had not considered obvious—a useful demonstration is a different achievement from leading a benchmark.

At the time of the conversation, Mike says Brightwave was hiring across AI engineering, classical machine learning, systems and distributed-systems engineering, front-end engineering, and design. His hiring philosophy was to fit the job to an exceptional person.