“Llama 3 with Agents gives you Godlike power” - Pietro Schirano

Llama 3’s biggest practical punch, as described in this conversation, isn’t just raw benchmark parity with frontier systems—it’s the ability to run a capable model locally and still get “good enough” reasoning for real tasks. The guest built an agentic orchestration framework that turns a simple goal into a workflow executed by smaller “mini agents,” then uses a stronger model as a final checker. That design choice—splitting labor and verifying at the end—lets cheaper, faster models handle most of the work while a higher-level model catches mistakes, making local deployment feel like having “the power of the Gods” in hand.

Anthropic’s email about winning an ACT developer competition is mentioned as external validation for the framework, which the guest calls an agentic system (MyAso). The practical demo is local app generation: the user describes what they want, the system thinks through sub-agents, and then outputs a structured folder with files—an end-to-end workflow rather than a chat-only experience. The guest also emphasizes why Llama 3 fits this approach: it’s fast, relatively cheap (citing a cost around 25 cents per million tokens), and strong enough for task execution when paired with a “senior developer” style final review.

On model comparisons, the discussion draws a nuanced line between general capability and specific features. Llama 3 70B is described as close to GPT-4 for many tasks like coding and functional outputs, but still not quite there for areas such as function calling and “J mode,” with expectations that fine-tuning or future releases will close the gap. A key theme is that open models are becoming usable on everyday hardware—an argument reinforced by the claim that an 8B Llama 3 model compares well to a much larger Llama 2 variant, and that people are only just beginning to realize what local inference unlocks.

Creativity becomes a second major thread. The guest argues that Llama 3 (and especially Claude 3) can outperform GPT-4 on pseudo-creative tasks—brainstorming, generating variations, and image-related reasoning—where GPT-4 may “go off the rails.” A concrete example is YouTube title generation: GPT-4 allegedly tends to insert semicolons regardless of prompt formatting, while Claude 3 and Llama 3 produce more usable title variations. Another example is one-shot personalization: the guest describes feeding Claude data from ~800 tweets into a JSON structure and getting outputs that mimic the author’s voice and habits, including specific tagging behavior.

From there, the conversation shifts to fine-tuning and personalization as the next step beyond retrieval and prompting. Fine-tuning is framed as most valuable when a company’s or person’s data isn’t present in base models (brand style, proprietary image sets, customer-support tone) or when the goal is to make an assistant feel like an extension of the user. The guest predicts that “instant fine-tuning” and more dynamic weight adaptation will arrive within a few years, enabling household devices and truly personal AI.

Finally, the discussion broadens into a philosophy of preparation for a post-AGI world: don’t wait for alignment with the future—build for it. The guest recommends doubling down on what people are already uniquely good at (human skills like empathy and taste), then using AI to amplify the rest. In that framing, the real advantage goes to those who can “speak the new language” of AI and keep building as capabilities accelerate—while regulation and corporate risk determine how quickly the stack becomes widely available.