Examples range from writing wedding speeches and speeding up coding to using AI for creative storytelling. One founder describes asking AI to add dark mode by describing the desired change. Another hopes creative tools will let anyone make an animated show from home.
40 AI Founders Discuss Current Artificial Intelligence Technology
#Generative Media#AI-Assisted Development#Search & Retrieval#Prompt Engineering#Fine-Tuning#AI Reliability#Human-in-the-Loop Systems
Unknown speakers in Y Combinator’s 40-founder discussion describe how they use AI, where it helps, and why human oversight still matters.
The founders describe conversational AI voices becoming much more natural and language models helping with semantic search over arbitrary text. They also explain that models need ongoing fine-tuning as data and language change, using fashion trends such as mermaidcore and balletcore as examples.
AI tools can provide an 85 to 90 percent solution, but teams still need to add structure, tune prompts, and build reliable workflows. Founders emphasize debugging and iterating as data and model quality change. A key engineering challenge is combining deterministic software with probabilistic outputs.
Variation can help entertainment models explore funny outputs, but it is riskier in settings such as driving or medical diagnosis. The founders describe hallucinations, trouble distinguishing fact from fiction, and inconsistent citations as limits on trust. They argue for human review and steering, especially when corrections affect real-world decisions.
The founders say AI systems need a human in the loop to check corrections and catch hallucinations. They frame the task as finding the right ways to steer these new tools while keeping technology in service of people and preserving human say-so.
Takeaways
- Treat AI output as a starting point that needs review and refinement.
- Break complex work into structured steps, then test and tune the prompts.
- Revisit workflows when your data or the underlying model changes.
- Match the level of human oversight to the consequences of an error.
- Use AI to support human judgment and connection, not to remove them.
Mentioned in this talk
Search inside this talk
What was claimed 8
- featurelarge language models: Can create creative storytelling work and help generate software from natural-language descriptions.
“they’re really good at what we thought they would be really bad at, which is creative storytelling work”
YouTube Unverified1 source· Talk
- feature
Voices: Conversational voices can sound indistinguishable from a human voice.“Now they’re just indistinguishable from a human voice.”
YouTube Unverified1 source· Talk
- featuresemantic search: Finds relevant information from arbitrary text and data.
“taking a random piece of text and finding relevant things”
YouTube Unverified1 source· Talk
- featureFine-tuning: Adapts models to new types of data to improve accuracy.
“fine-tuning all of our models on these new types of data to make them as accurate as possible”
YouTube Unverified1 source· Talk
- featurePrompting: Structuring a process as a prompt or sequence of steps can make it more reliable.
“engineer a prompt or a sequence of steps so that the entire process is even more reliable”
YouTube Unverified1 source· Talk
- featurelarge language models: Can produce variable outputs from the same inputs.
“you put in the same inputs and you might get some variation”
YouTube Unverified1 source· Talk
- featureHallucination: AI generates something that does not exist but looks as if it might.
“AI generating something that doesn’t exist but looks like it might or should exist”
YouTube Unverified1 source· Talk
- featureHuman in the loop: People assess corrections and supervise model outputs for hallucinations.
“There needs to be someone supervising and making sure there are no hallucinations.”
YouTube Unverified1 source· Talk
