Is the ChatGPT Era Over? Opus 4.6 & The Shift from Chat to Delegation - EP99.33

Chris discusses the recent advancements in AGI with the releases of Anthropic's Opus 4.6 and OpenAI's Codex 5.3. Opus 4.6, currently in beta, features a 1 million context window and a 20% improvement in coding benchmarks over its predecessor, Sonnet 4.5. However, the high costs associated with its usage raise concerns, prompting some users to consider alternatives like Gemini Flash 3. Codex 5.3 emphasizes rapid iteration and user interface improvements, with initial usage being free and a competitive pricing model for larger teams.

The larger context window in Opus 4.6 is beneficial for turn-based chat operations, but the cost-effectiveness of models is a significant concern, especially regarding potential layoffs if output does not improve with rising expenses. Users are exploring smaller models for lower costs, highlighting the importance of context and tool usage in enhancing performance. Initial impressions of Opus 4.6 indicate significant improvements in agentic workflows, with better context handling and a solid performance in executing tasks.

The discussion reflects a shift from chat-based interactions to delegation, emphasizing the emergence of agentic loops in workflows. Models optimized for coding are proving more efficient, expanding their applications beyond coding tasks. The comparison between models reveals that while Gemini 3 Pro is intelligent, it lacks the agentic loop that Opus has integrated, which allows for improved performance through context gathering.

The emergence of agent swarms, where a master process manages sub-processes, is also discussed. The distinction between an agentic loop and a regular tool calling loop is highlighted, with the former allowing for retries with different strategies. The competitive landscape between OpenAI and Anthropic is noted, with both companies rapidly releasing models, and the challenges of getting knowledge workers to adopt these tools are acknowledged.

The conversation distinguishes between productivity enhancement and full automation in the workplace, emphasizing the importance of owning and hosting automation tools to maintain intellectual property. The shift from a chat-based AI world to one focused on delegation and automation is recognized, with effective delegation being a new skill that may not suit everyone.

The inadequacies of current models and user interfaces are emphasized, with users seeking systems that feel more sentient and adaptable. The challenges of switching between skills and modes in workflows are questioned, alongside the need for improved coordination in task management. The evolution of tools from simple tasks to complex processes involving multiple agents is noted, with skepticism about claims of generating income on autopilot.

Testing comparable systems reveals similar issues regarding token usage and costs, raising questions about the rationale behind extended system usage without considering expenses. The limitations of current AI tools, including security concerns, are acknowledged, and a collaborative approach is suggested for evaluating AI outputs effectively.

The ongoing relevance of turn-by-turn interactions in AI usage is debated, with the chatbot era remaining significant for collaborative work. Concerns about AI's limitations in understanding task priorities and agency are raised, along with the introduction of a "master thread" concept for managing tasks with AI assistants. This approach allows for delegation and progress reporting while maintaining overall goals in task management.

The conversation emphasizes AI's role in reducing overwhelm by streamlining tasks and enhancing human agency in business. Participants suggest that many businesses may experience enhancement rather than disruption through AI technologies. Skepticism about AI's current capabilities is expressed, particularly regarding true agency and intelligence, with a consensus that AI will not completely replace human roles.

Cognitive overload and mental fatigue from productivity pressures are discussed, with concerns about reliance on software for decision-making leading to increased stress. Reflecting on past work habits that allowed for downtime, the speakers contrast this with the current demand for constant output, humorously concluding that while AI may thrive, humans may feel overwhelmed and unproductive.