AI FOMO, Part 1: What Actually Compounds

All English wording below is an editorial translation/paraphrase of Italian auto-generated captions; nothing is a verbatim quotation.

YouTube also provides an English (US) dubbed audio track. This summary was checked against the original Italian captions.

Salvatore Sanfilippo's answer to coding-agent FOMO is to separate fast-expiring tool knowledge from capabilities that compound: domain and software knowledge, practical intuition about model behavior, and clear communication. His preferred workflow follows from that distinction: keep orchestration simple, communicate intent quickly, and reuse working implementations as context.

His learning-time estimates and orchestration curve are illustrative. The reported H3.c outcome and model preferences are first-person reports, not independent assessments.

What Fades, What Compounds

• Tool fluency fades; domain judgment compounds: Sanfilippo contrasts a developer experienced with AI tools but lacking database foundations with a database expert new to agents. Because the interface is conversational, he expects the latter to acclimate in days or a week or two and perform better on the database task. That timeframe and comparison are illustrative judgments, not measured results [00:02:08]-[00:05:18]. He extends the durable foundation to product-domain knowledge, algorithms, data structures, and software-quality principles [00:07:20]-[00:08:14].
• Release trivia fades; model intuition compounds: He argues that productive agent use does not require tracking every model release or architecture detail. The more durable knowledge is qualitative: common failure modes, the effect of test-time reasoning settings, and whether a new capability materially changes the work at hand [00:08:14]-[00:12:03].
• Clear communication is worth practicing: Sanfilippo recommends learning to turn a thought into a usable written brief quickly rather than repeatedly polishing its form while the agent waits. This fluency depends on understanding the domain well enough to state the intent [00:16:17]-[00:17:20].

Workflow Consequences

• Keep orchestration proportional to the work: Sanfilippo's illustrative curve is that a simple workflow captures most of the value, a little orchestration may add only a marginal gain, and excessive orchestration can make the result worse through context switching, divided attention, and merge overhead [00:04:14]-[00:06:24].
• Use working implementations as executable references: For H3.c, he reports supplying a short brief plus named projects as references for the interface, internal design, kernels, and SSD streaming. He says the agent produced a working base that he later extended, and that those references embodied choices accumulated through months of steering. This is his project account, not an independent quality assessment [00:12:03]-[00:16:17].
• Use models to explore, not certify: He recommends capable models as interactive tutors for unfamiliar technical concepts [00:17:20]-[00:18:23]. This supports exploration, not acceptance evidence; consequential claims still require independent verification.

Watch with YouTube's English (US) dubbed audio(opens in a new tab), or select the Italian original in the player. Then continue with Sanfilippo's follow-up on session overload, wealth, and professional relevance(opens in a new tab).