From: allin
The AI industrial complex has recently dominated discussions, particularly with OpenAI’s launch of ChatGPT 4o (Omni) on a Monday. This release occurred just three days after Sam Altman, CEO of OpenAI, appeared on the All-In Podcast, leading to a discussion about a potential strategic error in not postponing his appearance until after the announcement [07:40:00]. Altman had originally intended to discuss upcoming announcements but they were delayed [08:07:00].
GPT-4o: Omni-Modal Capabilities
OpenAI launched GPT-4o, where the “O” stands for Omni, indicating its omni-channel or omnivore capabilities [13:17:00] [24:09:00]. This new model is faster, cheaper, and offers significant improvements in user interface and experience (UI/UX) [13:21:00]. It can process and respond to audio, text, images, and even desktop and video inputs from a camera, functioning as a 360-degree AI producer [13:34:00] [24:17:00].
Key innovations highlighted include:
- Real-time Conversational Interaction The model addresses the “cold boot” problem seen with tools like Siri, where there’s a delay or interruption when users speak over each other [13:53:00]. GPT-4o demonstrates much smoother interaction, as shown in examples of counting exercises (speeding up/slowing down) and real-time translation between English and Italian [14:07:00].
- Adaptive Learning and Personal Coaching A desktop and iOS application of GPT-4o can observe a user’s screen or activity. An example showed Sal Khan from Khan Academy using the app to tutor his son in math, providing nudges and questions without directly giving answers [15:10:00]. The app acts as a personal coach, observing the user’s attempts to solve problems [15:34:00]. It can also participate in Zoom calls and explain charts [16:56:00].
Evolution in Model Architecture
Freeberg noted that GPT-4o represents an evolution in model architecture, moving away from large, bulky models released quarterly that are expensive to rebuild [17:16:00]. Instead, it leverages a “system of models” – a multimodal system where several smaller models work together, or are linked, to respond to inputs and generate outputs [17:42:00]. These individual models can be continuously tuned and updated, making updates more continuous rather than large, infrequent releases [17:57:00].
Performance and Benchmarking
Initial criticisms of performance degradation with GPT-4o were countered by independent assessments. Stanford University runs a “massive multitask language understanding assessment” that frequently publishes model performance scorecards [19:05:00]. This scorecard shows that GPT-4o actually outperforms GPT-4 [19:17:00]. While it still underperforms Claude 3 Opus on these specific charts, it dispels the narrative that performance was sacrificed for speed [19:31:00]. GPT-4o is significantly faster than its predecessor, GPT-4 Turbo, with reported speeds twice as fast and feeling 10 times faster in practical use [24:48:00].
Consumer Growth and Business Model Shift
Despite OpenAI’s innovations, web visits for ChatGPT have plateaued [21:51:00]. This suggests that initial “lookie-loos” have moved on, indicating a need for more concrete use cases [22:00:00]. Sam Altman noted that with new model releases, older versions tend to become free [22:21:00]. New models are also becoming significantly more efficient and cheaper, with costs decreasing by about 90% annually [22:29:00].
It is speculated that OpenAI may eventually offer basic services for free or close to free, potentially charging only for multiplayer or advanced versions [22:50:00]. The long-term business model might shift away from B2C subscriptions (which tend to have high churn and limited expansion) towards a B2B direction, monetizing applications built on top of their models via API usage [23:21:00].
Impact of AI on Industries and Startups
The rapid pace of AI innovation, particularly with models like GPT-4o, has significant implications for existing companies and startups:
- Obsolescence of Product Roadmaps: Startups focused on areas like virtual customer support agents, which spent years making agents more conversational, may find their product roadmaps suddenly obsolete as core capabilities are integrated directly into foundational AI models [26:20:00].
- Model Innovation vs. Product Innovation: A critical challenge for developers is understanding where core model innovation ends and their own product innovation begins [27:23:00]. Building features that foundational models will soon obsolete is a risk.
- Shift Towards Open Source: There’s a strong incentive for companies to push operationally necessary, but not core, solutions into open source. This allows the community to maintain and improve them, freeing companies from ongoing investment and enabling easier switching between different AI models (e.g., GPT-4o to Claude to Llama) [27:52:00]. This mirrors initiatives like Facebook’s Open Compute Project, which open-sourced data center hardware designs, dramatically reducing costs and fostering an open platform [28:37:00].
- Deflationary Tailwinds: AI is expected to be a massive deflationary tailwind. The declining cost of compute, coupled with the ability to automate and delegate tasks (the “80/90 rule” - automate, deprecate, delegate), will lead to lower operating expenses for companies and potentially smaller company sizes [20:17:00] [13:00:00]. This enables more people to create products, solve problems, and launch niche businesses with significantly less capital, fostering a new wave of entrepreneurship [21:16:00].
Google’s Response: AI Overviews and Search Evolution
Google, responding to the rapid advancements in AI, announced “AI Overviews” at its I/O conference, a major change to its search product [22:15:00].
- AI Overviews: This feature provides step-by-step guides with citations and links for queries, effectively integrating AI-generated summaries directly into search results, similar to Perplexity [22:48:00].
- Monetization and Content Creation: While this could cannibalize traffic to content creators, it’s also suggested that it will lead to more searches and engagement, potentially creating new monetization opportunities through targeted ads within the overviews [23:14:00]. The issue of content licensing and fair use is expected to lead to major lawsuits, potentially necessitating a new market for licensing AI rights [23:31:00].
- Google’s Strategy: Historically, Google’s “I’m Feeling Lucky” button aimed to provide direct answers, a concept now realized through AI-powered “one boxes” [24:40:00]. This shift, though Google was initially “dragged kicking and screaming” into AI by innovators like Perplexity, positions them strongly for the future, leveraging their existing monopolies and diverse portfolio of innovations (e.g., Isomorphic Labs) [25:24:00].