AIE Miami Keynote & Talks ft. OpenCode. Google Deepmind, OpenAI, and more!
TL;DR
AI Engineer Miami framed AI as leverage, not replacement — MCs Ethel and Iman opened by noting attendees came from 23 countries and joked that maybe 20 people in the audience had “realistic expectations” about being replaced, before arguing AI will multiply what people can do, not just eliminate jobs.
G2I founder Gabe Greenberg turned a deeply personal health story into the conference thesis — after years of mold toxicity and mercury poisoning, a community fundraiser boosted by Dan Abramov raised $22,000 for his treatment, and Greenberg said React Miami and now AIE Miami were built as a non-corporate response to that generosity.
G2I used the keynote to launch Orchestrator AI, a multi-agent engineering platform with aggressive benchmark claims — Greenberg said the system can spin up 16 agents per task and reported 100% path coverage and 92% semantic score on a large spec where a single-agent harness hit 22%, plus an 8.4% overall lift on SWE-Bench Pro over GPT-5.4 high alone.
Dax Raad’s core warning was that AI removed the friction that used to kill bad product ideas — his line was blunt: “you don’t have any good ideas,” and his point was that one-hour AI-built MVPs look done enough to bypass product, design, and engineering scrutiny, creating bloated products and faster software rot.
Dex Horthy argued the first wave of coding-agent advice was wrong, especially “don’t read the code” — after watching teams struggle with research-plan-implement workflows, he said long plans were a trap, instruction budgets are real, and the new rule is simple: read the code, because lights-off software factories create slop you’ll pay for later.
The OpenAI/Cloudflare/Modem panel landed on a surprisingly analog conclusion about taste — Max Brunsfeld, Sunil Pai, and Ben Vinegar said AI can accelerate exploration, but taste still comes from being human: watching good movies, reading good books, going to museums, and knowing when an agent solved a problem at the wrong layer.
The Breakdown
Miami opens with community, not hype
Ethel and Iman set the tone with a playful but grounded welcome: people had come from 23 countries, two companies had sent 12 engineers each, and the room was overwhelmingly full of AI engineers. The joke about who’d be “replaced by AI” got a laugh, but they quickly reframed the day around practical applications, personal motivation, and stories of people wanting to use AI for healthcare and education.
Gabe Greenberg’s origin story makes the conference feel personal
G2I founder Gabe Greenberg gave the kind of opening you remember because it swerved from Brad Pitt on Ryan Florence’s laptop at React Conf 2016 to years of debilitating illness. He described sleeping under his desk with migraines, finally getting diagnosed with mold toxicity and mercury poisoning, and then seeing Dan Abramov amplify a fundraiser that raised $22,000 and helped him recover. React Miami and AIE Miami, he said, were built in response to that support — not as “a corporate event for the profit,” but for the people in the room.
Orchestrator AI gets announced with big multi-agent benchmark numbers
Greenberg then pivoted into product launch mode, announcing Orchestrator AI at orc.ai as a multi-agent orchestration platform for complex engineering. He described roles like coordinator, implementer, auditor, reviewer, validator, and researcher, plus adversarial governance, self-pruning memory, and up to 16 agents per task. The punchline was benchmark-heavy: 100% path coverage and semantic score on smaller APIs, then on a much larger spec an improvement from a single-agent harness at 22% to Orchestrator at 92% semantic score, plus an 8.4% overall lift on 731 SWE-Bench Pro tasks over GPT-5.4 high.
Dax Raad’s “you don’t have any good ideas” talk was really about restraint
Dax came out with no slides and maximum menace, telling the room they’d go home rich, successful, and finally make their moms proud — except for one problem: “you don’t actually have any good ideas.” His point was that old engineering bottlenecks used to act like a filter; product and design had to refine ideas because implementation was expensive, and lots of bad ideas died before shipping. Now anyone can prompt out an MVP in an hour, and because it looks 90% done, the team stops questioning whether it should exist at all.
Why AI speed is making products rot faster
Dax argued this isn’t just causing product bloat; it’s scrambling team dynamics. Design is now behind engineering, polishing dozens of shipped features instead of shaping a coherent product, while engineers are more willing to jam hacks into systems because the agent absorbs the pain. His best line on this mindset was that “the agent will fix it later” is basically a faith-based approach — and the result is products that feel “post-private-equity-acquisition” old within months.
Dex Horthy on what RPI got wrong
Dexter Horthy followed by revisiting his own popular research-plan-implement guidance and saying, plainly, that parts of it were wrong. Teams got bad research when they handed the model a ticket instead of turning it into questions first, and they got bad plans because giant prompts with 85 instructions quietly skipped crucial interaction steps unless users knew the “magic words.” His diagnosis: models have a real instruction budget, and once you pile on system prompts, tools, MCP, and giant workflows, adherence falls apart.
The new coding-agent advice: split workflows, read code, don’t outsource thinking
Horthy’s updated playbook was more modular and more skeptical: break work into smaller phases like questions, research, design, structure, and plan; use control flow for control flow instead of giant prompts; and stop pretending long plans replace human judgment. He said reviewing thousand-line plans is basically reviewing the code twice, and his strongest reversal was the clearest one: “please read the code.” Otherwise, he warned, you end up with a lights-off slop factory, a bug the agent can’t solve, and a team that has to re-onboard itself to a codebase no one has actually read in months.
The panel on taste turned out to be anti-slop and pro-humanity
Eric Torelli moderated Max Brunsfeld, Sunil Pai, and Ben Vinegar on “taste,” and the group mostly agreed that AI can explore faster but still tends to solve problems at the wrong layer. Max called taste “opinions that make sense,” Ben said it’s a shortcut to assessing quality before metrics validate it, and Sunil pushed hardest on the human source of it: good movies, books, music, museums, brunch, friends. The practical takeaway wasn’t mystical at all — if you want better products in the agent era, don’t just prompt harder; become more human.
Ben Vinegar’s SSH setup was the most practical systems talk in the chunk
After the break, Ben came back with a grounded talk on running coding agents over SSH instead of on a local laptop. At Modem, he said, six engineers have built a roughly 270,000-line codebase that’s 99% codegen, but local workflows broke down because agents stalled on permissions, maxed out CPU during heavy test runs, and died on bad Wi-Fi. His answer was charmingly old-school: Linux boxes, Tailscale, TMux, and terminal-native tools, all framed not as hacker cosplay but as a way for a normal “Mac enjoyer” to get more reliable compute, persistent sessions, and better ergonomics for long-running agent work.