In November 2018, OpenAI showed a slide admitting that nearly all experts thought its approach could not work. Five months later, it beat the world champions.

These slides come from OpenAI talks given between 2017 and 2020, years before ChatGPT made the argument for them.

They matter because every founder with a contrarian thesis has the same problem: the evidence arrives after the pitch.

Read closely, they show five repeatable moves for selling a bet most of the room thinks will fail.

Below is the full breakdown, with the slides, a template you can use today and three decks to download for free.

1 Start With the Limit Everyone Has Stopped Questioning

2 Pair the Mars Slide With a Cheap Experiment

3 Answer "Why Now?" With One Equation

4 Quote Your Skeptics, Then Show the Scoreboard

5 Bring Up the Risk Before the Room Does

6 The Proof Slide Template and Your 3 Free Decks

Most pitches open with a market problem. The strongest of these talks opens with a limitation the whole field has learned to live with.

In August 2017, Dario Amodei put it in one line: AI behavior "can be very complex, but goal has to be simple." Chess and Go fit that rule. Real tasks, such as arranging the furniture in a room to look nice or planning a mission to Mars, do not.

→ The slide does not say "AI is the future." It says the current approach has a specific flaw.

The most telling detail is the next bullet: simple proxies for complex goals can produce "very unsafe" behavior. The audience now has a reason to care before it has heard the solution.

For founders: write one sentence that begins "Everyone currently accepts that..." If an expert in your field would nod and then frown, you have your opening slide.

A big vision alone sounds like science fiction, and a small experiment alone sounds like a feature. This deck uses both.

Its "Long Term Vision" slide shows a person asking an agent to "plan and execute a mission to mars," with the agent asking back what matters most: "Extra safe or fast?"

Then comes the experiment that makes the vision credible: a person picks which of two clips looks better, and that is the whole training signal. The slide's punchline is its cost: "Requires only a few hundred binary bits of feedback!"

→ Mars tells the room why it matters. A few hundred clicks tells it why it might work.

For founders: put your ten-year picture next to the smallest result you already have, and name what that result cost you to get.

Every contrarian bet meets the same objection: people already tried this. Ilya Sutskever's GPT-2 deck answers it with two rows.

Not enough compute plus LSTM: "Does not work." Enough compute plus Transformer: "Works." The past failure stops being evidence against the idea and becomes part of the argument.

He then backs the equation with numbers. GPT in 2018 used "8 Pascal GPUs for a month." GPT-2 in 2019 used "100 Voltas for a week."

Two slides later comes the conclusion, in plain words: "Bigger models works better." The chart runs from 117M to 1,542M parameters, roughly 13x, and the line keeps rising.

For founders: fill in this line, then defend each term: "Past: [old constraint] + [old method] = did not work. Now: [what changed] + [your method] = works."

Most founders hide the objection. OpenAI turned it into the slide title.

Look at how the Dota case is built. First the stakes: a game "more similar to the real world than other games," a $35M+ annual prize pool and, before OpenAI Five, no AI that did well at real-time strategy.

Then the result. On April 13, 2019, OpenAI Five beat OG, the reigning champions, in a best-of-three live match. The slide needs no adjectives, only four teams and a row of check marks.

→ Pick a test that outsiders already respect. Nobody needs a technical background to understand a world champion losing.

For founders: find your version of OG. It could be the incumbent you replaced in a bake-off, or the benchmark your category already quotes. Put your result beside it, on a date.

A founder with a risky thesis usually saves risk for the Q&A. These decks put it on a slide of its own.

The GPT-2 deck explains the decision to hold the model back with one line: "can always release later; can never un-release." OpenAI announced GPT-2 on February 14, 2019, released a 774M version that August and the full 1.5B model on November 5.

You do not have to agree with that call to learn from the framing. A year later, Sutskever came back with a retrospective and reported what they actually saw: "lots of beneficial uses, and evidence of malicious ones."

For founders: name the biggest risk first, then state which decision can be reversed and which cannot. A plan with a stated next checkpoint is easier to defend than a claim that nothing can go wrong.

The most reusable format in the whole collection is a three-part slide used for one breakthrough after another: Before, The result, After. It works because it shows the belief that changed, not just the number that moved.

The honest limit matters. The same deck describes OpenAI Five as "competitive with (not yet exceeding) top humans." That qualifier makes the claim easier to believe.

1. Human Preference Learning as a Direction in Safe AI, Dario Amodei, August 22, 2017 (20 slides). LINK

2. The power of large-scale generative models and large scale RL, Ilya Sutskever, February 20, 2020 (43 slides). LINK

3. GPT-2, Ilya Sutskever (33 slides). The cover reads March 28, 2018, but the slides cover GPT-2 (2019) and announce the April 13 finals against OG, so the talk most likely dates from March 2019. LINK

Three more decks, including the one that takes on the question directly: "Can the current AI boom scale to AGI?" It uses history to answer the skeptics, starting with an astronomer who called powered flight impossible.

One note before you decide: the third paid deck, "Introducing Chat GPT-4", is included as a teardown exercise. It closely mirrors a public Slidebean template, and we could not verify it as a deck written by OpenAI.