Three AI CEOs Shared a Stage and Said 2–5 Years. 'Learn Later' Just Stopped Working as a Plan.

The safest plan for the last two years has been "watch and wait." Let the technology mature. See what sticks. Keep your options open. On June 17, 2026, that plan became significantly harder to defend.

At a G7 summit in Evian-les-Bains, for the first time in public, Sam Altman, Dario Amodei, and Demis Hassabis sat at the same table. Altman described "systems of astonishing power within two years." Amodei said "1-2 years to models surpassing human capability across all fields." Hassabis put his own estimate at 3-5 years and called the current moment "the foothills of the singularity" at Stanford GSB the following day. The three also proposed a joint global standards forum — their first joint public initiative.

These are not public relations. They are internal operating assumptions from the people funding, hiring, and building based on them.

What this does not mean

It does not mean AI will replace your role. An analysis surfaced by Simon Willison on June 14 — drawing on work by Narayanan and Kapoor — found not a single company in New York's 2025 WARN Act filings attributed layoffs to AI. The bottleneck is not generating output. It is deciding, verifying, and understanding that output. That bottleneck does not disappear as models get more capable.

It also does not mean learning to code or becoming a power user. The professionals adapting fastest right now are not the best prompters — they are the ones who have a clear map of which decisions in their workflow require domain judgment and which do not. That map is built from experience, not from experimenting with new features.

What to do differently Monday morning

Two concrete things.

Map your decisions, not your tools. Which choices in your week require expertise that AI cannot reliably reproduce? Which are repeatable enough that AI can carry most of the load? That map is more valuable than knowing any specific AI feature.

Start building evaluation speed. Most AI training focuses on generating better output. The skill that compounds over the next two years is evaluating output — knowing in under 30 seconds whether an AI response is good enough to act on, or wrong in a way that requires your attention. That comes from working in your domain with AI on real tasks, not from demo exercises.

Lena: the account director who stopped waiting

Lena manages accounts at a 45-person creative agency. For eighteen months she tracked AI tool launches, tried them briefly, and set them aside. Her logic: invest seriously once things stabilize.

In April she looked at what she had actually built: she could use ChatGPT for email drafts and little else. Two junior account managers on her team had built specific, repeatable workflows for competitive briefing and client reporting that saved them three to four hours each per week.

The gap wasn't prompting skill. It was that they had spent months doing real work with AI, building a clear picture of where it was reliable for their specific clients and where it wasn't. That knowledge only comes from deliberate practice on actual work.

Lena's shift: she stopped broad experimentation and picked one workflow — competitive briefing — to make genuinely good with AI before moving to anything else.

James: the L&D manager who reframed his training budget

James runs learning and development at a 200-person professional services firm. After the G7 summit, his company's CTO asked for a skills investment recommendation for the second half of 2026.

His first instinct was to budget for tool-specific workshops. His revised recommendation: split the budget between tool fluency and a structured practice program that builds evaluation and verification skills — the ones people will need regardless of which AI tools they are using in eighteen months. Specific tools change every quarter. The ability to assess AI output in your domain does not.

His reasoning: judgment in your domain is different from judgment built on someone else's exercises. Tool demos teach tool use. Practicing on your actual work builds the evaluation instincts that survive the next tool refresh.

The question worth answering now

"I'll invest in AI skills when things settle down" assumed the settling would happen before the timeline compressed further. Three AI leaders publicly disagreed with that assumption last week, in the same room, at the same time.

The professionals best positioned in two years are not the ones who started prompting most fluently today. They are the ones who started building judgment infrastructure — the domain-grounded skill of knowing what good AI output looks like for their specific work. That infrastructure takes time to build. The G7 statements are a signal that the time available to build it is not indefinitely long.