By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 2:38
Why AI matters to CEOs and why contact center attrition is so high
Steph Smith frames AI as a boardroom topic, with open questions on privacy, competition, cost and accuracy. She then asks Zayd Enam why contact centers see such high agent attrition, and he explains the stressful, low-paid nature of the job.
- •2022 was a breakout year for AI, and CEOs are weighing privacy, competition, cost and accuracy
- •Customers don't care about the underlying tech, so companies will differentiate in other ways
- •Contact centers see 30-45% first-year attrition, sometimes as high as 80%
- •Zayd Enam cites seasonal demand, low wages, inflation and high-stress calls as drivers
- 2:38 – 7:35
Lazy vs. creative AI: customer conversations as a data goldmine
Zayd Enam contrasts "lazy" AI, which automates an existing process end to end, with a creative approach that treats AI as a new building block. He says customer conversations hold insight about product, market and competitors, and tells the Andy Grove/Intel story to show the value of listening to customers.
- •Lazy AI automates an existing process end to end; creative AI reimagines the business around a new capability
- •Conversations are a goldmine of data about product, market and competitors
- •Telecom customers use conversation insight to inform pricing and packaging
- •LLMs can summarize and synthesize unstructured audio that once took hours to review
- •Andy Grove's pivot from memory to microprocessors began with salespeople noting customers were less respectful
- 7:35 – 8:41
Hex Magic: AI tools for data science and analytics
Steph Smith introduces Barry McCardel of Hex, a platform for collaborative data science and analytics. Barry describes Hex Magic, a set of AI tools inside the Hex UI that generate and edit code, such as SQL and Python, and document it.
- •Hex is a platform for collaborative data science and analytics
- •Hex Magic launched in closed beta earlier in the year
- •It generates SQL from plain-language questions
- •It can refactor Python code and document complex code written by others
- 8:41 – 11:26
Contextual understanding, brand risk and disclosing bots
Steph Smith raises the problem of AI lacking context, like chatbot loops, and unexpected responses that can hurt a brand. Zayd Enam answers that brands are built on trust, so he leans toward transparency about bots, and he raises accent masking as a related risk.
- •Chatbots that loop on the same question show the lack of contextual understanding
- •Poorly implemented models can create unexpected responses that affect the brand
- •Zayd Enam leans toward transparency because brands build long-term trust
- •Accent masking raises similar reputation risks if not disclosed
- 11:26 – 13:22
Sourcegraph's Cody and "context is king"
Beyang Liu of Sourcegraph introduces Cody, an editor extension that combats hallucination by supplying relevant code context. He explains that models can't know about private or post-training data and will invent answers unless given the right context.
- •Cody is a chat and search editor extension built on Sourcegraph's code search
- •Sourcegraph acts as a fact checker and context provider for language models
- •Models hallucinate when asked about things outside their training set
- •Most of Cody's quality issues come down to not fetching the right context into the context window
- 13:22 – 21:24
How Cody fetches context compared to Copilot, and the open source approach
Beyang compares Cody's broad context retrieval via the code graph with Copilot's more local, autocomplete-driven context. He also explains that Cody's context is introspectable and that Sourcegraph is building it as open source for a pluggable ecosystem.
- •Cody converts questions into search queries and surfaces relevant code and docs
- •Copilot, as Beyang describes it, relies mostly on recent open files and does not do Q&A yet
- •Cody shows which files it read, so users can spot and flag bad context
- •Cody is released under the Apache 2 license to enable a pluggable ecosystem of context sources
- 21:24 – 22:22
Proprietary datasets as a differentiator
Steph Smith introduces proprietary data as part of differentiation, and Zayd Enam explains Cresta's opportunity. Contact center work combines dialogue with actions in enterprise software, a dataset that doesn't exist on the web and could be used to train foundation models.
- •Web pages are a fairly static language modeling task
- •Contact center work combines dialogue and action in systems of record
- •Cresta collects a large dataset of this work across customers
- •This can enable training of foundation models that do things web data can't
- 22:22 – 26:33
Customization and prompt construction at Hex
Barry McCardel explains how Hex builds prompts using database schemas, past queries and project context, while keeping a human in the loop. He and Steph Smith discuss UI design, prompt-context tuning and why the real work happens before calling the model API.
- •Hex uses schemas, past code and chart formatting to build the right prompt
- •Hex's rule: it generates code and shows it, but doesn't run it for the user
- •UI choices such as a pop-out drawer and latency feedback matter
- •Too much context can confuse a model, so the right level has to be found
- •The hard work is everything before the API call, including LangChain and embeddings
- 26:33 – 29:47
Where is the moat? Data advantage and product experience
Steph Smith asks whether any advantage can't be copied, since competitors replicate UIs and features. Barry McCardel says model access is a commodity and points to Hex's data used for personalization (not model training), its UX, and being first to figure out how AI changes assumptions.
- •Competitors copy UI elements, so that has to be taken as a given
- •Model access is already a commodity
- •Hex uses data to personalize for a team but not to train models
- •Thoughtful UX and performance act as existing infrastructure
- •First movers who personalize may keep an advantage, as in a data moat
- 29:47 – 34:01
Privacy, security and a pluggable model strategy
Steph Smith raises privacy and security concerns, especially when competitors may share the same models. Beyang explains that Sourcegraph is self-hostable and treats the language model as one pluggable component, letting customers bring their own or choose among Claude, ChatGPT and others.
- •Enterprise customers are security and privacy sensitive, so Sourcegraph is self-hostable
- •The language model is one component alongside the code graph and developer tools
- •Customers can bring their own model, and Cody currently supports Claude and ChatGPT
- •Sourcegraph uses a fine-tuned open source embeddings model for semantic code search
- 34:01 – 37:27
Model pricing, Bing and search plus LLMs
Steph Smith brings up the cost dependency on model providers, citing Bing's price increase. Beyang argues the Bing integration shows language models need a retrieval engine, so search becomes more valuable, and he speculates this may relate to the price hike.
- •Different models carry different pricing, security and efficacy tradeoffs
- •Sourcegraph avoids being tied to one provider
- •Bing combined an OpenAI model with its search back end rather than shipping a plain chatbot
- •Beyang says language models make search engines more valuable, not unnecessary
- 37:27 – 39:30
The cost and risk of retaining data
Steph Smith frames data retention as having privacy and cost implications. Barry McCardel says deleting data eliminates knowledge, that retained data could have second-order value for models, and that Hex is already used to customers' concern for their data.
- •Companies may retain and collect more data as they see its value
- •Retaining data and running models are not free
- •Retention policies eliminate knowledge that could be useful later
- •Hex is used to customers being protective about their data and aims to keep earning their trust
- 39:30 – 44:50
The contact center in 2028
Returning to Cresta's domain, a speaker describes a vision of a touchless contact center where machines handle data entry and agents focus on relationships. Zayd Enam and Steph Smith discuss agents becoming strategic, account-manager-like roles, and customer interactions becoming frictionless or even fun.
- •The vision is an agent who isn't typing, summarizing or looking up records
- •Machines handle back-end work while humans focus on empathy and relationship building
- •Frictionless digital experiences for routine needs, humans for high-trust decisions
- •Agent roles become strategic and proactive, like account managers
- •Steph Smith imagines a shift from negative to positive NPS
