Unsupervised Learning

Unsupervised Learning is Redpoint’s AI content hub, combining an interview podcast with transcripts and original analysis of AI technologies and their business implications. As of September 18, 2026, its About page identifies Redpoint partners Jacob Effron and Patrick Chase as its leaders. The podcast examines model breakthroughs and practical application building for builders, researchers and founders.[1][2][3]

The publication pairs podcast transcripts with investment theses and observations about AI trends. Its podcast also distributes episodes through YouTube; guests listed in the March 2025 crossover include Noam Shazeer, Bob McGrew, Noam Brown, Dylan Patel, Percy Liang and David Luan. In April 2026, Effron described the show as a “nights and weekends project” alongside his work as a Redpoint investor, crediting listeners’ subscriptions and sharing with helping it attract guests.[1][2][3]

Before the 2025–2026 crossovers, Unsupervised Learning had published practical deployment analysis and a builder survey. That earlier work supplies context for the show’s later discussions of model choice, customer requirements and specialized training.[4][5][3]

On February 27, 2023, Effron and Chase published Using Large Language Models Effectively, synthesizing conversations with ten founders and executives who had deployed LLM features. Their eight lessons covered user co-creation, lower-stakes tasks, safety layers, embeddings, richer prompt context, tool augmentation, hybrid deployment and training smaller models with larger ones. The authors emphasized that user experience could matter as much as the underlying model: Descript’s Kundan Kumar described allowing users to correct text-to-speech mistakes by retyping, while Canva’s Ahmad Iqbal stressed users’ desire to retain creative control.[4]

The analysis explained how companies supplied information absent from a general model. Mem cofounder Dennis Xu described using embeddings—vector representations of data—to retrieve customer context, making a request such as “send an email to John” more specific by identifying the recipient. Ada used semantic search to find similar previous conversations and return stored responses, an approach intended to reduce hallucinations. Elementl CEO Pete Hunt used LangChain to combine GPT-3 with knowledge from Dagster’s documentation for a GitHub support bot.[4]

Other examples illustrated deployment tradeoffs. Sourcegraph’s Beyang Liu and Rok Novosel described considering third-party LLMs for some customers and an open-source model for self-hosted customers unwilling to export data. Neeva CEO and cofounder Sridhar Ramaswamy wanted answers to begin in under 1.5 seconds. The article reported that he used summaries generated by OpenAI’s DaVinci model as training data for a smaller summarization model that met Neeva’s latency requirements.[4]

Effron and Tony Shi supplemented the podcast’s builder interviews with the AI Builder Survey, published August 21, 2024. Conducted that June, it surveyed 40+ leaders at organizations building AI products, including traditional technology companies and startups with more than 50 employees, about application development and tooling.[5]

The authors reported that about one-third of surveyed builders had switched models in the preceding three months; two in five expected to switch or expand providers in the next few months. In the respondent sample, 61% used fine-tuning and 56% used retrieval-augmented generation. Usable data was the largest reported development pain point, at 49%, followed by model quality at 24% and evaluation at 12%. Internal applications were being developed by 93% of respondents and external applications by 59%; these categories overlapped.[5]

Unsupervised Learning collaborated with Latent Space on a crossover published March 29, 2025, discussing open models, product-market fit, application defensibility and infrastructure. A second crossover, published April 23, 2026, followed AIE Europe. Effron called the collaboration an “annual tradition”; Latent Space credited Jacob and the UL production team with hosting and editing that episode.[2][3]

The 2026 discussion documented a change in Effron’s judgment about quality-focused post-training. He had doubted its value when general models might deliver equivalent improvements within three to six months. He was becoming more open to it when it was the best way to improve a customer’s outcomes for that period, even if the resulting model was later replaced. He kept that test attached to alternatives such as adding data through connectors or improving backend engineering. Guest swyx added, “you, you throw out the results, but you don’t throw out the raw data.” Effron agreed that retained data could support another training run. This was an evolution in the show’s analysis of application development.[3]