Fable or Astra: The Choice for AI Developers

Theo clarified his position on AI, stating he is "very pro in the software dev world" because "writing code as a human is hilariously inefficient." He noted that this inefficiency is why AI has been so successful in software development. However, he expressed caution about AI's application in other areas, particularly generative media, due to a less favorable balance between economic upsides and real-world downsides. He pointed out that many creators conceal their use of AI tools to avoid negative reactions from their audience.

This serves as a foundational argument for the necessity of AI assistance in managing intricate coding environments.

He explained his intention to experiment with this new version to assess its capabilities. The primary focus of his evaluation was how effectively it could manage sub-agents and handle cross-thread communication, crucial aspects for complex AI-driven workflows.

Theo firmly stated that the core problem in coding is "staring at the code," not a lack of autocomplete, defining the solution as moving beyond direct code interaction. He highlighted that T3 code leverages remote systems and VS Code over SSH to boost efficiency. His focus remains on the orchestration layer, minimizing reliance on editor-based autocompletion for development tasks.

Theo outlined a strategy to mitigate the risk of platform bans by integrating T3 code widely within the developer ecosystem. He posited that the more popular T3 code becomes, especially among users of cloud code and CodeX, the less likely providers like Anthropic are to impose restrictions. He believes this increased adoption shifts the user base from 'power users' to 'average users,' making it less appealing for providers to ban essential third-party tools.

Theo issued a public challenge to detractors, inviting them to anonymously prove their current salary through his moderators in exchange for a substantial payout. He presented his consistent refusal of high-value sponsorship deals as evidence that he is not influenced by corporate interests. He expressed his frustration with users who accuse him of compromising his principles for financial gain, reinforcing his commitment to independent content creation.

Theo provided an update on his health, clarifying that his injury is a ligament tear from an old fall, not carpal tunnel syndrome. He described the surgery he underwent in August of last year as a complete thumb reconstruction, where a synthetic ligament was created and other material grafted to facilitate ligament regrowth. He is currently in a cast and faces a 50/50 chance of needing additional surgery by October. Despite his limited hand mobility, he noted that his ability to manage AI engineering workflows is crucial for maintaining his productivity.

Theo revealed that his hand injury has compelled him to adopt 'vibe coding,' a method heavily reliant on AI. He explained that this shift is not solely due to typing difficulties but also because navigating his computer has become harder. He clarified that voice-to-text is insufficient for coding; instead, he leverages AI to act as an agent to operate his computer. Theo emphasized that these workflow adjustments, while necessitated by his injury, offer universal benefits beyond his specific circumstances.

Theo expressed his strong conviction against using terminals for agentic development, stating, "Talking to your terminal is hell." He asserted that his hand injury has amplified his belief that terminals are unsuitable for voice-driven navigation. He now completely avoids manual terminal use, relying instead on high-level orchestration agents to manage tasks. This represents a significant shift in his workflow, as he now views terminals as a major bottleneck to efficiency.

Theo described a significant evolution in his workflow, where agents are now required to verify and validate their own changes. He instructs them to use computer resources for verification and validation, as well as to utilize AI code review bots within a repository to provide feedback and ensure code quality. This approach delegates testing, code review, and sub-agent verification directly to the AI system, reducing human overhead by ensuring agents achieve high confidence before requesting manual review.

Theo explained his increasing confidence in AI models to autonomously merge code, moving past previous concerns about 'safety.' He noted that AI agents, Astra and Fable, have merged over 150 pull requests (PRs) with only two minor regressions, both involving animation removals. Theo compared the success rate of these 'YOLO merges' favorably to that of human developers, indicating a high level of trust in the AI's judgment for code integration.

Theo advised against the manual copy-pasting of context between agent threads, stating that current models are proficient enough at building context using their own tools. He emphasized that advanced models can find necessary information independently, rendering manual intervention largely unnecessary. He proposed a challenge to the audience: to stop manually guiding agents and instead observe if the AI's output aligns with expectations, demonstrating the models' self-sufficiency.

Theo noted that both Fable and Astra models are capable enough for a wide range of tasks, asserting, "You can use any of these models to do anything you want at this point. They're both more than good enough." He encouraged the audience to await an upcoming dedicated video that will provide a detailed comparison of the two models.

He also mentioned that OpenAI effectively surpassed Anthropic's results shortly after Astra's launch, further underscoring Astra's competitive edge.

He revealed that he has dedicated a secondary machine exclusively for 24/7 Astra computer use tasks, demonstrating his reliance on its efficiency. This significant difference in speed makes Astra a preferred choice for tasks requiring rapid and extensive computer interaction.

He stated that relying on AI to edit content, manage audio, or handle AV pipelines in these ways is premature. Theo emphasized that video editing automation is not yet close to being usable for professional purposes, highlighting the current limitations of AI tools in creative production workflows.

Theo recounted Astra's performance in a complex project migration from TypeScript to Rust, where it improved test accuracy from 30% to over 80% relatively quickly, within approximately three days. However, Astra then stalled significantly at 82.6%, despite extensive sub-agent deployment. In contrast, Theo noted that Fable 5.1 provides a more stable and consistent development experience, suggesting it is better suited for maintaining momentum in such intricate projects.

Theo presented data showing that "Fable 5.1 writes way more mergeable code" than Astra. He detailed that Fable 5.1 averages two additional follow-ups from the time a pull request (PR) is filed to when it is merged. In contrast, Astra averages six follow-ups per PR, indicating it requires more post-filing corrections and manual verification. This makes Fable a more consistent producer of shippable code.

Theo noted that while GBD6 Astra is significantly closer to Fable 5.1 in performance, a gap still exists. He explained that he defaults to Fable for critical UI changes or quick bug fixes due to its reliability. Astra, despite its advancements, often enters "death loops," necessitating Fable to finalize the code and ensure stability in development workflows.

Theo detailed Astra's unique ability to spin up numerous sub-agents, which can pass messages among themselves and, crucially, provide updates back to Astra. Unlike Fable's upfront planning, Astra can dynamically fan out tasks, allowing for greater flexibility in complex workflows. This 'swarms' capability enables Astra to successfully manage intricate projects with many parallel agents, receiving real-time feedback to keep the entire system progressing.

Theo observed a significant difference in how Fable and Astra apply skills. He stated that "If the skill has a reasonable description and it's useful for the work going on, Fable correctly pulls it in and applies it." In contrast, he expressed frustration with Astra, noting that it frequently ignores active skills, even when they are directly relevant and in context. This inconsistency makes Astra less reliable for tasks requiring precise skill utilization.

He pointed out that Codex offers 12,000 units of inference compared to Claude's 8,000 for the same $200 price. Additionally, Claude's Fable 5.1 model is capped at 50% usage, further reducing its effective utility. He jokingly offered to hire anyone who could convince a user that Claude was a better deal, emphasizing Codex's superior value proposition for power users.

Theo articulated his strategy for combining AI models, stating, "I will be defaulting to Fable for code and I'll be defaulting to Astra for every other thing that I do." He explained that Fable is his preferred choice for primary coding tasks due to its output of more mergeable code. Conversely, Astra is utilized for general computer usage and managing sub-agent workflows, leveraging its strengths in dynamic task orchestration. This approach highlights both models as complementary tools rather than mutually exclusive choices.