"VOLTEI AQUI PORQUE PRECISO FALAR ISSO": A GRANDE VIRADA da INTELIGÊNCIA ARTIFICIAL [com FABIO AKITA

Context and meta

• The presenter describes a turning point in practical usefulness of LLMs (late 2025 → early 2026). After more than 500 hours of hands‑on experiments he open‑sourced and documented the work (≈16 GitHub repositories, ~300k LOC) and published many blog posts and step‑by‑step guides.
• Emphasis: moving from theory and futurology to reproducible, open tutorials and runnable projects so others can inspect, reproduce, and contribute.

What changed (technical pivot)

• Tool calling / harnesses Models are now trained to call external tools/functions instead of returning only text. LLMs can emit structured function calls (for example create_file) that a harness executes, and results are fed back into the model loop. This is the practical enabler for automation beyond natural language output.

• Architecture separation

  LLM = “deep thinking” / planning; harness = dumb executor that runs functions and returns results. This separation enables persistent multi‑step workflows, parallel tasks, and real execution on local machines or CI systems.

• System prompts / skills Persistent system prompts (the harness context) and user‑created skills (persona/config files) allow encoding recurring behaviour and improve consistency across sessions.

Practical product comparisons and behavior

• Claude (Anthropic) / “Opus” / cloud‑style tools

  • More disciplined and predictable.
  • Better at decomposing complex plans into parallel tasks and continuing them if interrupted.
  • Reliable for trusting long‑running tasks and context continuity.

• Codex / OpenAI

  • Historically focused on parameter scaling.
  • Less reliable for multi‑task or parallel interactions — may forget subtasks and require reminders.
  • Newer versions improve, but behavior varies by model and harness fit.

• Google (Gemini / related / “antigravity”)

  • Supports tool calling and integrations and uses its own ecosystem/harness language.

• Cross‑vendor notes

  • Vendors use different tool‑call formats and function names (for example create_file vs file_new vs NewFile). Models are trained against a specific harness language; a mismatch between LLM and harness degrades performance.
  • Open‑source harnesses attempt to map between formats, but fidelity can be lost if the LLM wasn’t trained for the target harness.

Developer experience, guides, and tool notes

• The presenter published detailed, step‑by‑step coverage (including a teardown of a leaked Cloud Code), showing real code and how to use tool calling, skills, and session handoff.
• Examples of useful automations demonstrated:
  • Organizing a downloads folder.
  • A ROM organizer.
  • Generating UI/UX files via a saved skill that can be reloaded for later tasks.
• Session continuity:
  • Harnesses/apps expose session IDs so you can resume context across sessions or switch harnesses by exporting/importing the context file.
• Vendor lock‑in and economics:
  • Subscription models and token limits affect choice and feasibility. Some providers restrict usage patterns or block non‑approved integrations, complicating a unified tool approach; attempting to route a model through an unsupported framework may result in bans.
• Code quality note:
  • The Cloud Code leak revealed messy, spaghetti code — a sign teams are moving fast to release features.

Key technical takeaways

• Tool calling and trained function outputs are the practical enabler for real automation (not just text).
• Harnesses execute functions; LLMs must be trained to speak the harness’ “language” for reliable results.
• Different vendors’ function schemas cause integration and performance issues; mixing LLM + harness often worsens output unless mappings are well aligned.
• Skills (saved persona/config) plus system prompts let you persist specialized behaviour across sessions and projects.
• Session IDs and resume support maintain project context without rewriting prompts.
• Open‑sourcing experiments and publishing step‑by‑step guides is valuable for reproducibility and community contribution.

Reviews, tutorials, and artifacts referenced

• Extensive benchmarks and step‑by‑step writeups on the presenter’s blog.
• 16 open‑source GitHub repositories (public, ~300k LOC) with runnable code and contribution opportunities.
• Coverage and analysis of the Cloud Code leak that explains internals and code quality.
• Comparative, hands‑on testing across:
  • Anthropic Claude / Opus / Cloud
  • OpenAI GPT / Codex
  • Google Gemini / antigravity (and related)
  • Various CLI and cloud integrations

Main speakers and sources mentioned

• Main speaker: Fabio Akita (presenter of the video).
• Platforms and vendors referenced: Anthropic (Claude / Opus), OpenAI (GPT, Codex), Google (Gemini / antigravity / Cria), plus mentions of Alibaba, Tencent, and various cloud/harness ecosystems.