In April 2023 we released an episode named “Mapping the future of *truly* open source models” to talk about Dolly, the first open, commercial LLM. Mike was leading the OSS models team at Databricks at the time. Today, Mike is back on the podcast to give us the “one year later” update on the evolution of large language models and how he’s been using them to build Brightwave, an an AI research assistant for investment professionals. Today they are announcing a $6M seed round (led by Alessio and Decibel!), and sharing some of the learnings from serving customers with >$120B of assets under management in production in the last 4 months since launch. Losing faith in long context windows In our recent “Llama3 1M context window” episode we talked about the amazing progress we have done in context window size, but it’s good to remember that Dolly’s original context size was 1,024 tokens, and this was only 14 months ago. But while understanding length has increased, models are still not able to generate very long answers. His empirical intuition (which matches ours while building smol-podcaster) is that most commercial LLMs, as well as Llama, tend to generate responses most of the time. While Needle in a Haystack tests will pass with flying colors at most context sizes, the granularity of the summary decreases as the context increases as it tries to fit the answer in the same tokens range, rather than returning tokens close to the 4,096 max_output, for example. Recently Rob Mulla from Dreadnode highlighted how LMSys Arena results prefer longer responses by a large margin, so both LLMs and humans have a well documented length bias which doesn’t necessarily track the quality of answer: The way Mike and team solved this is by breaking down the task in multiple subtasks, and then merging them back together. For example, have a book summarized chapter by chapter to preserve more details, and then put those summaries together. In Brightwave’s case, it’s creating multiple subsystems that accomplish different tasks on a large corpus of text separately, and then bringing them all together in a report. For example understanding intent of the question, extracting relations between companies, figuring out if it’s a positive / negative, etc. Mike’s question is whether or not we’ll be able to imbue better synthesis capabilities in the models: can you have synthesis-oriented demonstrations at training time rather than single token prediction? “LLMs as Judges” Strategies In our David Luan episode he mentioned they don’t use any benchmarks for their models, because the benchmarks don’t reflect their customer needs. Brightwave shared some tips on leveraging LLMs as Judges: * Human vs LLM reviews: while they work with human annotators to create high quality datasets, that data isn’t just used to fine tune models but also as a reference basis for future LLM reviews. Having a set of trusted data to use as calibration helps you trust the LLM judgement even more. * Ensemble consistency checking: rather than using an LLM as judge for one output, you use different LLMs to generate a result for the same task, and then use another LLM to highlight where those generations differ. Do the two outputs differ meaningfully? Do they have different beliefs about the implications of something? If there are a lot of discrepancies between generations coming from different models, you then do additional passes to try and resolve them. * Entailment verification: for each unique insight that they generate, they take the output and separately ask LLMs to verify factuality of information based on the original sources. In the actual product, user can then highlight any piece of text and ask it to 1) “Tell Me More” 2) “Show Sources”. Since there’s no way to guarantee factuality of 100% of outputs, and humans have good intuition for things that look out of the ordinary, giving the user access to the review tool helps them build trust in it. It’s all about the data During his time at Databricks, they had created dolly-15k, a dataset of instruction-following records written by thousands of their employees. Since then, no other company has replicated that type of effort even though the data wars are in full effect. It’s been clear in the last year that the half-life of a model is much shorter than the half-life of a dataset. The Pile by Eleuther (see Datasets 101) came out in 2020 and is still widely used; if you had trained an LLM in 2020, you would have definitely replaced it by now as they have gotten better and cheaper. On the age old “RAG v Fine-Tuning” question, Mike shared a great example that we’ll just quote: I think of language models kind of like a stem cell, and then under fine tuning, they differentiate into different kinds of specific cells. I don't think that unbounded agentic behaviors are useful, and that instead, a useful LLM system is more like a finite state machine where the behavior of the system is occupying one of many different behavioral regimes and making decisions about what state should I occupy next in order to satisfy the goal. As you think about the graph of those states that your system is moving through, once you develop conviction that one behavior is useful and repeatable and worthwhile to differentiate down into a specific kind of subsystem, that's where like fine tuning and specifically generating the training data, like having human annotators produce a corpus that is useful enough to get a specific class of behaviors, that's kind of how we use fine tuning rather than trying to imbue net new information into these systems. There are a lot of other nuggets in the episode around knowledge graphs extraction, private vs public data, user intent extraction, etc, but we only have so much room in the writeup so go listen! And if you’re interested in working on these problems, Brightwave is hiring 👀