Davos has a way of turning sci-fi titles into operational questions.

At the World Economic Forum, a session called “The Day After AGI” put two of the most influential AI builders on the same stage:

• Demis Hassabis — co-founder and CEO of Google DeepMind (and the “AI-for-science” builder behind AlphaGo → AlphaFold).
• Dario Amodei — co-founder and CEO of Anthropic (one of the key “safety-first” labs pushing frontier model capability while building formal guardrails).
• Moderated by Zanny Minton Beddoes (The Economist).

The interesting part wasn’t the buzzword “AGI.” It was the framing: what happens after AI stops being a tool that answers, and becomes a system that plans, executes, improves itself, and operates across the economy at scale?

That’s the “day after” problem.

---

1) First, who are these two people in AI history?

Demis Hassabis: from breakthrough demos to “AI for discovery”

Hassabis is one of the rare leaders who’s consistently pushed the idea that AI’s real endgame is not chatbots—it’s general problem-solving, applied to science and medicine.

A short timeline that matters:

• DeepMind (2010): founded around the ambition of building general intelligence systems.
• AlphaGo (2016): a proof point for deep reinforcement learning and complex planning under uncertainty.
• AlphaFold (2020+): AI-driven protein structure prediction that moved from “research milestone” to “global scientific infrastructure.”
• AI-for-science worldview: Hassabis tends to describe “AGI” as a system that can generate new hypotheses and accelerate discovery—not just automate office work.

If you want one sentence: Hassabis represents the “build general capability, then point it at the hardest problems” tradition.

Dario Amodei: frontier capability + safety as a core engineering constraint

Amodei represents a different arc: researchers who helped push large-scale model training forward and then centered the question, “how do we keep this safe as capability scales?”

A short timeline that matters:

• Former senior research leadership in the modern LLM era (pre-Anthropic).
• Anthropic (founded 2021): structured explicitly around building powerful models with safety methods treated as first-class engineering.
• Constitutional AI: a training approach where models learn to critique and revise outputs against an explicit set of principles—trying to make “alignment” more systematic than ad-hoc filtering.

If you want one sentence: Amodei represents “capability is accelerating; safety and governance must scale at the same speed.”

---

2) What WEF meant by “The Day After AGI”

WEF’s framing wasn’t “will AGI happen?” It was closer to:

• A credible pathway is increasingly visible because scaling + multimodal systems + agentic models are converging.
• But the “day after” depends on what we do about governance, safety, labor disruption, and security while we’re still early enough to steer outcomes.

That’s why this panel matters: it’s not a philosophical debate. It’s a strategic planning meeting for the global economy—just held in public.

---

3) The core tension: capability is rising fast, but the last mile is messy

Amodei’s stance (as it came across)

Amodei’s message was essentially:

• We’re approaching a zone of “incredible capabilities.”
• The next few years are the critical window where rules, evaluations, and deployment constraints either get built—or we lock in a chaotic race dynamic.

He tends to talk like someone staring at a compounding curve: once systems can produce useful work across domains, the slope changes socially (jobs, education, advantage) and geopolitically (security, misuse).

Hassabis’s stance (as it came across)

Hassabis sounded more like:

• The curve is real, but “AGI” isn’t a single trick.
• Certain tasks are moving fast (especially where answers are verifiable: code, math).
• But to get to robust, general intelligence you still need “missing ingredients” — the kind that show up when systems must operate in the open world, plan reliably, and generate genuinely novel insight.

In other words: capability acceleration is undeniable, but full generality is not just bigger models.

---

4) “Missing ingredients” — what that phrase really implies

One of the most useful ideas from this Davos discussion is the notion that we can be simultaneously:

• very close to massive economic impact
• and still missing pieces for the strongest definition of AGI

Here’s the practical interpretation of “missing ingredients,” translated into engineering terms:

A) Planning and long-horizon reliability

Getting “good answers” is not the same as:

• setting a multi-step goal,
• creating a plan,
• executing safely across tools and systems,
• recovering from failure,
• and doing it repeatedly with high confidence.

That’s what real-world deployment requires.

B) Strong world models (not just text prediction)

Many breakthroughs come when systems can represent:

• how the physical world behaves,
• how complex systems interact,
• what causes what,
• and how to test hypotheses.

That matters for “AI for science,” robotics, and anything involving uncertainty.

C) Continual learning without catastrophic failure

The day-after world is not a static benchmark. Models need to adapt safely without:

• forgetting,
• drifting,
• or getting exploited.

So the “missing ingredients” aren’t academic. They define whether AGI becomes:

• a reliable infrastructure layer, or
• a fragile power tool.

---

5) Agents, not chat: the real “day after” product shift

If you run a tech business, the most actionable part of the whole AGI debate is this:

The next platform shift is from “AI that answers” to “AI that does.”

Agentic systems mean:

• AI can operate software,
• coordinate workflows,
• run projects,
• write and ship code,
• handle customer operations,
• and supervise other AIs.

This is where both leaders implicitly converge: even if we disagree on the strict AGI definition, the agent era is already arriving—and it changes competition.

In the day-after world:

• Advantage comes from process + data + control, not from “prompt skill.”
• The best teams design systems where AI produces drafts, plans, experiments, and code—while humans make the final calls and own accountability.

---

6) Jobs: the near-term disruption is junior work

A striking theme from Davos coverage around this panel is that both leaders pointed at the same immediate pressure point:

entry-level knowledge work.

The logic is simple:

• junior work often involves “first drafts,” “first passes,” and “support tasks”
• those are exactly the tasks AI is becoming good at
• so organizations may hire fewer juniors—not because work disappears, but because the leverage per senior increases

Hassabis’s practical advice (as quoted in WEF coverage) was blunt: if he were addressing undergrads now, he’d tell them to become “unbelievably proficient” with these tools.

So the day-after response isn’t panic. It’s adaptation:

• education shifts from memorization → judgment,
• careers shift from “do the first pass” → “own the outcome.”

---

7) Governance: the window is now, not later

A key takeaway from Amodei’s side of the discussion (as reported through WEF’s Davos highlights) is that regulation and governance choices made in the next few years will shape the long-run outcome.

Translated into concrete needs, this implies:

• standardized model evaluations (capability + misuse risk),
• safety testing and red-teaming as deployment gates,
• incident reporting norms,
• secure deployment patterns for high-risk capabilities,
• and some form of cross-border coordination (because AI doesn’t respect borders).

The “day after” problem is not just building smarter systems. It’s building institutions and technical controls that can handle smarter systems.

---

8) What founders and tech teams should do in 2026 (non-hype checklist)

1) Treat AI literacy as a baseline skill

Not “prompting.” Real literacy:

• what models do well,
• where they fail,
• how to verify outputs,
• and how to build safe workflows.

2) Start building agent-ready processes

You don’t need “AGI” to benefit from agents. Pick 2–3 workflows where your company already has structure:

• support triage,
• sales ops,
• internal documentation,
• product QA,
• analytics summaries,
• code maintenance.

Then build:

• clear inputs,
• clear outputs,
• human sign-off,
• logging and audit.

3) Invest in your internal “system of record”

Agents amplify whatever you feed them. If your data is messy, your output is messier—just faster.

4) Update hiring: less junior volume, more leverage design

Plan for:

• fewer pure “first-pass” roles
• more roles that combine domain judgment + AI operation + verification

5) Build governance like you build security

Treat safety as engineering:

• access control,
• monitoring,
• evals,
• and explicit policy boundaries.

Because in the day-after world, “move fast” without controls becomes an existential risk—especially for companies integrating high-capability systems into real operations.

---

Closing thought

The real message of “The Day After AGI” wasn’t a timeline.

It was a warning about tempo:

• capability is accelerating,
• institutions adapt slowly,
• and the gap between the two is where disruption and risk live.

Hassabis represents the upside: AI that accelerates science and discovery. Amodei represents the constraint: if we don’t scale governance with capability, the transition gets ugly.

The day after AGI isn’t a moment. It’s a new operating environment—and the teams building for it now will define the winners in the next decade.