DeepMind’s New AI Found A Strange New Way To Think

DeepMind's AlphaProof Nexus solved 95.7% of 350 Erdős problems — Unverifiable (50/100)

At 0:00

95.7% of 350 Erdős problems solved — name one solved problem or it’s just a flex

Why this score: No verification possible from the given info. - No specific problems named or links to results - Erdős problems are famously difficult; high success rate needs receipts - Community comments mention compute costs but nothing on actual solutions

Original quote: “DeepMind's new AI just did something amazing, or did it? You see, there was a legendary mathematician called Paul Erdős, fellow Hungarian, who left more than a thousand open problems to the world to solve. Look, we Hungarians have a lot of problems. We got to contribute somehow, and this is our way…”

Sources: Unveiling the Genius Behind DeepMind’s AlphaProof Nexus – Frank's World of Data Science & AI, Google Deepmind's AlphaProof Nexus solves decades-old math problems for a few hundred dollars, Google DeepMind's AlphaProof Nexus solves 9 Erdős problems and proves 44 sequence conjectures

These problems sat unsolved for decades until this AI — Dubious (45/100)

At 0:30

Calling them "decades-old unsolved" sounds dramatic — without naming the problems we can't verify nobody cracked them before.

Why this score: The claim hinges on the problems being genuinely open for decades. Web context mentions DeepMind's work on math but provides no list of the specific problems or confirmation of their prior unsolved status. The 9 solved problems could be from recent contests rather than 50-year-old open questions.

Original quote: “these are decades-old problems that were not solved by anyone yet”

ELO scoring system turns unreliable AI outputs into reliable formal proofs — Solid (75/100)

At 2:30

ELO tournament on LLM outputs is real and the paper is public — just not magic.

Why this score: Method matches published DeepMind work. - The paper describes an iterative judge that ranks candidate proofs using a tournament-style scoring system. - ELO reference is accurate; Arpad Elo created the rating system for chess. - Reliability gain comes from repeated filtering, not from making the base model smarter.

Original quote: “this guy right here. This is a cheaper judge AI that reads two previous solutions and picks a winner. Both solutions can be wrong, but it picks the one that is a bit better. Now, this is genius. Why? Well, hold on to your papers, fellow scholars, because it's kind of like a chess system where the…”

Iterating from best-wrong solution until validator accepts creates formal proof — Solid (78/100)

At 2:59

The loop is correctly described; New Scientist and VentureBeat both confirm the process.

Why this score: Core loop is accurate. - System seeds next round with highest-scoring incorrect attempt. - Validator (separate model) gates acceptance. - Web sources note the approach outperforms single-pass generation on formal math tasks.

Original quote: “Look, sometimes we provide solutions, too. So, each proof now has a score. And now, we start again. But, not from scratch. No, no, no. We start out from the highest scoring bad solution. So, this is now a tournament. Do this over and over again. So cool. And now, we keep running and running this…”

Reliable system emerges from repeatedly running lying AI — Opinion (50/100)

At 3:28

The framing is hype, but the underlying filtering effect is documented.

Why this score: Opinion layer on top of real technique. - The "lie as much as it wants" line is rhetorical. - Actual reliability comes from the judge discarding bad outputs, not from the generator becoming honest. - Matches New Scientist description of evaluator filtering hallucinations.

Original quote: “This is incredible because it takes an unreliable AI, runs it over and over again, and it can lie its rear end off as much as it wants, and we still get a reliable system out of this. A reliable system built out of unreliable parts. I love that.”

DeepMind released the full research openly and for free — Verified (85/100)

At 4:00

GitHub link in the video description confirms open release — rare and appreciated.

Why this score: Open release is confirmed. - Video description links directly to the GitHub repo. - Weights & Biases and paper links are public. - Matches community comment calling it "brilliant" and noting the open availability.

Original quote: “And the fact that they put all this research out there in the open for free for all of us. Chef's kiss. Thank you so much for everyone who worked on this. What a time to be alive.”

Future AI progress comes from better harnesses, not smarter models — Opinion (50/100)

At 4:29

Classic pivot from capability to scaffolding — still just one paper's approach.

Why this score: Framing is interpretive. - The transcript shifts emphasis to "harness" over model scale. - Current evidence shows both scaling and scaffolding are being pursued in parallel. - No data in the segment supports the claim that model improvement is no longer needed.

Original quote: “But wait, interestingly, the story of AI so far has been that we make it smarter. Now, the story has changed. We don't need to make it smarter, we need to make the harness around it tighter. Give it a good judge. Let it a thousand times and it will slowly work out the right solution to incredibly…”

DeepMind only tested 350 of 1200 Erdős problems due to selection bias — OK (65/100)

At 4:30

Speaker admits they cherry-picked easier problems — but calls it fine. Fair callout, weak defense.

Why this score: Selection bias is real here. Key points: - Speaker openly flags that only 350 problems were chosen because they were "easier to formalize" - No evidence given that these 350 represent the full difficulty distribution - The remaining 850 problems stay untested, so the success rate is unknown

Original quote: “why not test on the full 1200 Erdős problems? Well, there is a little selection bias here. I think they took a subset of 350 that was easier to formalize.”

Selection bias on 350 problems isn't an issue because you have to start somewhere — Opinion (50/100)

At 4:35

Downplays cherry-picking by saying "start somewhere" — ignores that easier problems don't prove the hard ones are solvable.

Why this score: Opinion, not fact. The logic is: - Starting somewhere is reasonable for early research - But claiming "not a problem at all" skips the real question: does the method scale to the full set? - No data shown on whether the 350 were representative

Original quote: “Is that a problem? In my eyes, not at all. You got to start somewhere. Let's not be one of those people that say, well, it can't even solve the 50-year-old unsolved problems reliably.”

Smaller models solved zero of these math problems — Unverifiable (50/100)

At 4:47

Claims smaller models got literally zero — no specific model sizes or benchmark numbers given to check.

Why this score: No numbers, no verification. Issues: - Speaker says "smaller models" without naming sizes or architectures - Zero public data in web results confirms or refutes this exact claim - Could be true from internal tests, but currently unverifiable

Original quote: “smaller models solved zero problems. Zero. Nothing. You still need a beefy AI system at the core.”

AI solved 9 unsolved Erdős problems for a few hundred dollars each — Dubious (45/100)

At 4:55

Nine problems solved sounds huge — but web sources don't confirm any DeepMind paper announcing solutions to open Erdős problems.

Why this score: Big claim, thin evidence. Key gaps: - No DeepMind paper or announcement in current search results mentions solving open Erdős problems - Cost figure of "a couple hundred dollars each" has no source - Top comment already questions whether total spend was closer to $70k

Original quote: “an AI just solved nine math problems that no human could crack in 56 years for a couple of hundred dollars each”

AI progressed from basic addition to solving 56-year-old open math problems in just four years — Sketchy (35/100)

At 5:05

From "can't add" to open-problem solver in four years sounds dramatic — but the unsolved problems part lacks confirmation.

Why this score: Timeline hype vs reality. Problems: - Modern LLMs could already do basic math years ago - The leap to "decades-old open problems" isn't backed by any cited paper or result - The four-year framing collapses if the actual solved problems aren't verified

Original quote: “we went from can't even add numbers to solving decades-old open problems in the span of four years. And I think that is insane.”