What's the technical difference between GPT-4o and Gemini 1.5 regarding their architecture?

GPT-4o primarily uses a text-based Transformer model with added modules for speech and video, akin to bolting on existing technologies like Whisper and DALL-E. Gemini 1.5, however, is a true 'mixture of experts' model trained from the ground up to activate different network paths for various data types, making it potentially more energy-efficient and integrated.

Will the massive context windows in models like Gemini 1.5 make RAG infrastructure obsolete?

It's unlikely RAG will become obsolete. While large context windows enable more data to be processed at once, RAG remains crucial for data privacy, permanent memory storage, and precise retrieval. Many users and enterprises still prefer storing sensitive data separately and controlling access.

How can startups avoid being 'roadkill' by large AI companies like OpenAI and Google?

Startups should focus on building valuable, perhaps 'unsexy,' features that larger companies are less likely to prioritize or demo on stage. Identifying niche markets, especially in B2B sectors or areas involving legal/PR risk, offers opportunities where incumbents may be hesitant to tread.

What are some promising B2B opportunities in the AI space?

B2B AI opportunities lie in automating complex, day-to-day tasks in regulated industries like fintech (e.g., KYC, compliance) and healthcare. Companies focusing on niche processes like expediting construction permits or streamlining compliance also show strong potential, often with significant ROI for clients.

What makes consumer AI opportunities unique in the current landscape?

Consumer AI opportunities can flourish in 'edgy' areas that larger companies, due to PR or legal risks, avoid. This includes applications involving deepfakes for satire, virtual companions with long-term memory, or tools that allow creative, potentially controversial, content generation.

Which recent AI updates most excited the speakers?

One speaker was excited by the emotional expressiveness of OpenAI's new voice generation, finding it far superior to robotic text-to-speech. Another was thrilled by the live translator demo, seeing its potential to break down language barriers personally and globally. A third noted the potential for robotics advancements driven by unified models and cost reductions.

Better AI Models, Better Startups

Y Combinator

Science & Technology4 min read42 min video

Jun 6, 2024|86,143 views|1,722|65

YC Y Combinator

Save to Pod

Key Moments

TL;DR

AI models advance, offering startups opportunities and challenges against tech giants.

Key Insights

Recent AI model updates (GPT-4o, Gemini 1.5) present both opportunities and existential threats to startups.

The increasing capabilities and multimodality of AI models benefit startups by enabling more sophisticated applications.

Competition among major AI providers (OpenAI, Google, Meta) is beneficial for startups, fostering a marketplace.

While large context windows are promising, Retrieval Augmented Generation (RAG) remains crucial for data privacy and precise retrieval.

Startups should focus on building valuable, unsexy B2B products or niche consumer applications that incumbents may avoid.

Edgy or legally/PR-risky areas, which large companies shy away from, present significant opportunities for startups.

The advancement of AI models is also paving the way for practical robotics and integrated personal assistants.

THE DUAL NATURE OF AI ADVANCEMENTS FOR STARTUPS

Recent significant updates to foundational AI models, such as OpenAI's GPT-4o and Google's Gemini 1.5, have created a climate of anticipation and concern within the startup ecosystem. Founders often anxiously await these releases, assessing whether the new capabilities might render their existing products obsolete or create new avenues for innovation. The core tension lies in the potential for large AI companies to rapidly integrate advanced features, while startups must strategically leverage these developments to avoid being outcompeted.

EMBRACING MODEL IMPROVEMENTS AND MULTIMODALITY

The continuous improvement in AI model performance and the expansion of their capabilities, including multimodality (handling text, audio, video, and code), offer significant advantages for startups. Enhanced reasoning, better structured output like JSON, and the ability to generate code that executes actions can lead to breakthroughs. These advancements allow startups to build more sophisticated and powerful applications on top of these foundational models, often just by upgrading their underlying AI integration without fundamental product changes.

THE STRATEGIC ADVANTAGE OF A COMPETITIVE AI MARKET

A landscape with multiple powerful and comparably capable AI models from different providers (OpenAI, Google, Meta, Anthropic) is a safer environment for startups than one dominated by a single entity. This competition fosters a marketplace, preventing monopolistic control and ensuring more reasonable pricing for APIs. It allows a wider array of startups to choose the best model for their needs, build unique products, and maintain healthier gross margins, contrasting with a scenario where only a few giants achieve trillion-dollar valuations.

THE ENDURING ROLE OF DATA PRIVACY AND RAG

Despite advancements like Gemini 1.5’s massive context window (up to 10 million tokens in research), Retrieval Augmented Generation (RAG) pipelines are expected to remain essential. For data privacy, user control, and the need for precise, fine-tuned information retrieval, RAG offers a more secure and reliable method than relying solely on a model's context window. This layered approach, akin to complex computer memory architectures, ensures that sensitive data is managed securely, especially crucial for enterprise applications.

IDENTIFYING OPPORTUNITIES BEYOND THE OBVIOUS

While major AI players like OpenAI focus attention on general-purpose AI assistants and multimodal consumer experiences, lucrative opportunities exist in less glamorous, unsexy B2B niches and specialized consumer applications. Areas like construction permit processing (PermitFlow), financial compliance (Greenlight, GreenBoard), and identity verification in fintech are inherently valuable but unlikely to be demo-worthy for large tech companies. Similarly, edgy consumer applications involving legal or PR risks, such as deepfake video generation (Infinity AI) or highly personalized AI companions (Replika, Character AI), are areas where incumbents tend to tread cautiously, leaving room for innovative startups.

THE CYCLICAL NATURE OF TECH DISRUPTION AND STARTUP SURVIVAL

The current AI landscape echoes historical tech cycles, particularly the era when Google and Facebook dominated, presenting similar anxieties for startups. The key to survival lies in avoiding direct competition with the core offerings of tech giants and instead focusing on verticalization, specialized services, or entirely new market categories. Startups must anticipate the trajectory of AI development and focus on building differentiated products with unique value propositions, rather than merely iterating on features that major players are likely to adopt.

THE PROMISE OF PRACTICAL ROBOTICS AND PERSONAL ASSISTANTS

The convergence of unified AI models, falling costs, and advancements in specialized hardware is accelerating the development of practical robotics. Innovations like the bipedal robots from companies such as Unitree, coupled with increased efficiency in AI processing, suggest that sophisticated robotics, potentially integrated into personal devices, might be closer than anticipated. This trend points towards a future where AI extends beyond software into the physical world, offering new frontiers for innovation and startup development, moving away from constant cloud dependency.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●Concepts

●People Referenced

Common Questions

New advanced models can be a double-edged sword for startups. While they offer enhanced capabilities and can implicitly make any startup using them 'smarter,' there's also a concern that dominant players like OpenAI or Google might release features that kill their startup's niche. Having multiple powerful models, however, creates a healthier market.

Topics

Gemini 1.5 B2B AI

Mentioned in this video

softwareOllama

An open-source platform mentioned as a way for users to run local AI models and experiment with personal data.

softwareKayak

A travel search engine mentioned as a successful example of a vertical engine.

companyGreenboard

A fintech company focused on compliance for banks, another example of successful B2B AI applications.

companyBronco

A company operating in AR (Augmented Reality) and payments, representing B2B AI opportunities.

softwareTPUs (Tensor Processing Units)

Google's custom hardware accelerators, crucial for training large models like Gemini 1.5 efficiently.

softwareReplika AI

An AI company focusing on conversational companions (AI boyfriend/girlfriend), considered an edgy consumer opportunity.

companyPermit Flow

An AI company that expedites the construction permit application process, highlighting B2B niche opportunities.

companyUnitree

A Chinese company that announced a $116,000 human biped robot, relevant to the advancement of practical robotics.

organizationYC (Y Combinator)

A startup accelerator mentioned as the affiliation of the speakers, who fund and support companies.

companyInfinity AI

A company that turns scripts into movies using AI, allowing for the creation of content with famous characters, seen as an edgy opportunity.

companySegment

Found this useful? Build your knowledge library

Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.

Try Summify free