Harge says stronger models can improve a startup simply by switching to a newer model, while added capabilities such as coding and structured JSON output make it easier to build useful products. He argues that model progress can let applications handle tasks that were difficult to solve with prompting alone.
Better AI Models, Better Startups
#Company Formation#AI Benchmarks#Retrieval-Augmented Generation#AI Product Design#Multimodal Models#Product Strategy
Speakers
YC group partner Gary and fellow group partners Jared, Harge, and Diana discuss how better AI models affect startups on Y Combinator’s The Light Cone.
Gemini 1.5’s million-token context window raises the question of whether it could make retrieval-augmented generation obsolete. The panel says retrieval still matters for privacy, precise recall, permissions, and audit logs, especially in enterprise settings. Gary compares retrieval layers to caches and databases, which remain useful even as memory gets faster and larger.
The panel compares AI startups’ concerns about OpenAI to earlier worries that Google would copy startups. They suggest avoiding products that look like obvious extensions of a general-purpose assistant, while pursuing specialized workflows, vertical data, and less flashy tasks. Examples include construction permit workflows, fintech compliance, and B2B software that requires detailed customer feedback and sales.
The speakers describe AI applications in identity checks, bank compliance, and payments as ways to automate labor intensive work. Better models can support premium features and upsells as products do more for customers. One speaker says YC companies in the prior batch grew from $6 million in annual revenue to more than $30 million over three or four months.
For consumer AI, the panel points to products that incumbents may avoid because of legal or public relations risk, including AI companions and tools for making movies with familiar characters. They note that Character.AI has strong user retention and that some people spend hours a day with such products. The discussion closes with excitement about expressive speech, live translation, lower model costs, and possible uses in robotics.
Takeaways
- Track model releases and plan product improvements before competitors do.
- Keep retrieval for sensitive data, reliable recall, and access controls, even with larger context windows.
- Build for specialized workflows and customer needs that general-purpose assistants may not serve well.
- Use model improvements to add customer value and support higher priced features.
- Look for consumer uses that incumbents may avoid, while accounting for legal and public relations risks.
Mentioned in this talk
Search inside this talk
What was claimed 16
- launchedopen-source speech-to-text models
“after OpenAI released the Whisper speech-to-text open-source models”
1:04inBeyond Transcription: Building Voice AI That Understands ConversationsYouTube Verified2 sources· Talk
- feature
GPT-4o: Better structured output, including JSON“it’s better at getting JSON”
YouTube Unverified1 source· Talk
- featureOpenAI: speech and video processing added on top of textspeech and video processing added on top of text
“it added speech and video processing on top of text”
YouTube Unverified1 source· Talk
- feature
Gemini: Mixture-of-experts architecture with paths for different data types“it’s actually a true mixture of experts”
YouTube Unverified1 source· Talk
- metricGPT-4o: 128,000-token context window
“GPT-4o has 128,000”
YouTube Unverified1 source· Talk
- metricGemini: 10-million-token context window demonstrated in research
“in research, they proved it works with a 10-million-token window”
YouTube Unverified1 source· Talk
- featureGemini: Context window of 1 million tokens
“it has a context window of a million tokens”
YouTube Unverified1 source· Talk
- featureVector RAG: Supports retrieval with privacy, precision, permissions, and data control
“a RAG pipeline infrastructure is still very much needed”
YouTube Unverified1 source· Talk
- launched
Google launched 
Google Drive“Google was working on Google Drive but hadn't launched it yet”
YouTube Unverified1 source· Talk
- feature
Perplexity: Research workflow with specific links and source material“get specific links and YouTube videos”
YouTube Unverified1 source· Talk
- featurePermitFlow: AI-assisted construction permit application workflows
“They expedite the process for applying for construction permits”
YouTube Unverified1 source· Talk
- featureGreenlight: AI identity verification for consumers
“doing KYC using AI to replace the humans who do a lot of the validation of consumer identities”
YouTube Unverified1 source· Talk
- featureGreenboard: AI compliance work for banks
“Greenboard was also doing a lot of compliance work for banks”
YouTube Unverified1 source· Talk
- featureReplika AI: AI boyfriend and girlfriend companion experiences
“one of the top companies doing AI boyfriends or girlfriends”
YouTube Unverified1 source· Talk
- metricCharacter.AI: Users spend hours per day on the product
“people are spending hours per day using things like that”
YouTube Unverified1 source· Talk
- featureInfinity AI: Turns scripts into movies that can involve famous characters
“lets you turn any script into a movie”
YouTube Unverified1 source· Talk
