Mike Conover

Mike Conover is an AI researcher and entrepreneur who helped develop Databricks’ Dolly instruction-tuned language models and later co-founded Brightwave with Brandon Kotara. At Databricks, his team demonstrated modest-cost instruction tuning and released model weights, code and commercially usable training data. At Brightwave, he described dividing financial research into specialized tasks, grounding outputs in retrieved evidence and using fine-tuning to shape repeatable behavior.[3][4][2]

Conover led Databricks’ open-source large language models team working on Dolly. The March 24, 2023 announcement, coauthored with eight colleagues, described fine-tuning EleutherAI’s six-billion-parameter GPT-J with Stanford Alpaca instruction examples. Databricks reported a 30-minute run on one machine, later describing its cost as about $30. Those figures covered fine-tuning an existing model, not its pretraining. The demonstration showed a qualitative improvement in following instructions rather than parity with ChatGPT.[2][3][4]

In his April 2023 interview, Conover recalled requesting LLaMA weights without receiving them. GPT-J was accessible, familiar from his previous work and a practical basis for an experiment others could reproduce. He also credited accumulated engineering work that the team repurposed for Dolly. The choice of base model thus reflected both availability and reproducibility.[1]

Open weights alone did not resolve training-data restrictions. After Dolly’s release, prospective users repeatedly asked whether they could use it commercially. Databricks explained that Alpaca’s OpenAI-generated examples raised restrictions on competing models, prompting the team to collect a new dataset. Dolly 2.0, released April 12, 2023, used EleutherAI’s Pythia-12B and released weights, training code and roughly 15,000 instruction-response records across seven task families. The dataset carried a CC BY-SA 3.0 license.[4]

The team organized collection with task guidelines, nightly leaderboards and awards for the top 20 labelers. According to Databricks’ release account, the contest exceeded 15,000 results within a week and was closed over concerns about employee productivity. Answers were employee-written, while closed question answering, extraction and summarization also used Wikipedia reference passages. Original responses did not mean every passage in each record was newly authored.[4][5]

Conover argued that the composition of instruction data shapes how a model answers, contrasting brief classification responses with demonstrations of longer assistant-style writing. He also reported that GPT-J and the first Dolly had nearly identical benchmark scores despite markedly different instruction-following behavior, motivating better measures of the behavior users actually wanted. The Dolly repository explicitly describes Dolly 2.0 as not state of the art, with weaknesses in mathematics, programming, complex prompts and factual accuracy. Its dataset could contain errors and reflect both Wikipedia biases and an employee population unrepresentative of the wider world.[1][5]

Conover’s case for open models extended beyond customization and keeping sensitive data in-house. He argued that public models and datasets allow a broader community to investigate bias, ethics, safety and security by examining how systems are built and behave. Dolly’s contribution was a reproducible model-and-data release with clearer commercial-use permissions and explicit limitations.[1][4][5]

By June 2024, Conover had left Databricks and co-founded Brightwave, an AI research assistant for investment professionals, with Brandon Kotara. Conover identified Kotara then as CTO and emphasized his combination of financial infrastructure and information-retrieval experience. He described Brightwave as a “partner in thought”: a system for exploring investment theses and their implications while people supply fund strategy, management conversations and other knowledge outside the available documents.[2]

Conover described revising his expectation that large context windows would solve synthesis. His analogy was a whole-book report versus individual chapter synopses: if an answer remains roughly the same length as the input grows, it has less room for specific findings. He distinguished finding a particular fact in a long document from drawing useful conclusions across many documents. Brightwave’s approach, he said, was to build specialized subsystems and decompose material into smaller reasoning units.[2]

In that architecture, retrieval supplies facts at inference time, while fine-tuning shapes repeatable subsystem behavior. Conover favored a system resembling a finite state machine, moving among defined behavioral regimes to satisfy a goal. Once a behavior proved useful and repeatable, human-generated training examples could specialize a subsystem for it. He treated retrieval as a way to ground reasoning in available facts rather than relying on fine-tuning to insert new knowledge.[2]

Conover also described retrieval that responds to a question’s temporal intent: breaking news calls for different material from an account of 18 months of interest-rate policy. Context-aware prompts could vary with document type and source, while source metadata needed to travel through successive inference calls. He reported multiple evaluation steps to assess whether generated statements were factual given the retrieved content. His central concern was systems being “wrong in a convincing way.”[2]ⁿ¹

For Conover, generating ideas and deciding on investments were distinct tasks. He recounted Brightwave tracing an AI-investment premise through energy demand to long-duration storage, producing possible second- and third-order investment theses. A human still had to judge their coherence with fund strategy and other knowledge. In 2024, he also made automated Excel spreadsheet creation an explicit non-goal, arguing that subtle formula errors and the personal reasoning involved in financial modeling made dialogue a better product experience.[2]

Before Dolly and Brightwave, Conover studied social networks and worked on machine learning at LinkedIn, SkipFlag and Workday. His earlier research provides context for his experience analyzing relationships among people and firms: it examined political communication and economic structure through observed interactions.[1][8][7]

As an Indiana University PhD student, Conover was first author of the 2011 paper Political Polarization on Twitter, with Jacob Ratkiewicz, Matthew Francisco, Bruno Gonçalves, Filippo Menczer and Alessandro Flammini. The team examined more than 250,000 political tweets from the six weeks before the 2010 US midterm elections. Retweets showed pronounced partisan separation, while mentions connected ideological opponents more often. The distinction showed why the type of interaction matters when measuring political communication; cross-party contact did not itself establish agreement. Conover accepted the paper’s ICWSM Test of Time Award in 2021.[8][9]

Conover subsequently worked on LinkedIn’s homepage news relevance and at SkipFlag, an enterprise knowledge-graph company acquired by Workday. At Workday he became director of machine learning engineering, a role Indiana University’s June 2021 award announcement also identified. By his April 2023 interview, he was a staff software engineer at Databricks.[1][9]

In the collaborative 2019 labor-flow study, Conover shared data processing, analysis, interpretation and manuscript contributions; Jaehyuk Park, Ian B. Wood and Yong-Yeol Ahn led the writing. The study represented firms as nodes and job changes as weighted, directed connections, then recursively applied community detection to identify nested economic clusters. Its initial network contained approximately 130 million transitions among more than four million firms; hierarchical clustering used a filtered core of about 42 million transitions among 487,782 firms.[7]

The researchers found a stronger association between educated-worker inflows and market-capitalization growth when S&P 500 firms were grouped into labor-flow clusters than when grouped by geography or conventional industry categories. This was an observational association, not evidence that worker inflows caused higher valuations. LinkedIn membership skewed toward US users and younger people with technical backgrounds, and uneven adoption could affect clustering. Treating firms as homogeneous units and clusters as disjoint also simplified diverse workplaces and potentially overlapping economic communities.[7]

The study involved LinkedIn as both data provider and research supporter. Its disclosure identified Conover as a former LinkedIn employee, Wood and Souvik Ghosh as then-current employees, and five authors as partially supported through LinkedIn’s Economic Graph Challenge. The licensed employment-history data were restricted; replication data required a request and LinkedIn permission.[7]