Open A.I did the Unthinkable

OpenAI dropped 722 papers via a model no one else can use — No Frame (75/100)

At 0:00

Straight facts on the GitHub drop. No fluff, just the numbers 📈

Why this score: The speaker is citing specific dates and quantities regarding the OpenAI/math repository. It's a direct reporting of an event without heavy-handed emotional framing.

Original quote: “We have to look at what Open AI just did to the mathematics community. On October 6th, Open AI dropped 722 math papers onto GitHub in a public repository called Open AI/math. All of them were written by a model that nobody outside the company can currently use.”

OpenAI provided Lean proofs for 235 math families — No Frame (75/100)

At 0:16

Just technical details about the Lean proofs. No tricks here 🛠️

Why this score: The speaker is explaining the technical structure of the dataset (families and Lean proofs). It's descriptive and factual rather than manipulative.

Original quote: “These papers are grouped into 372 families, which include a main result and its associated spin-offs. For 235 of those families, Open AI provided proofs in Lean, which is a programming language that allows a computer to double-check every single step.”

OpenAI claims success on 4,000 unsolved problems using cherry-picked results — Cherry-Picked (20/100)

At 0:31

They only showed the 'significant' wins and ignored the 4,000 failures. That's not a breakthrough, that's a highlight reel 🍒

Why this score: The speaker admits they only kept the 'significant results,' which is a classic move to hide the vast majority of data that didn't work out.

Original quote: “Open AI claims this model easily passed all their usual math tests, so they simply pointed it at around 4,000 real unsolved problems. They only kept the significant results.”

Sources: AI Math Proofs: How Anthropic and OpenAI Results Actually Compare | MindStudio, OpenAI's 377 Math Claims Spark Verification Fight [2026], Math Results: OpenAI Posts Hundreds More, a Surprising Risk

Listing a rapid-fire sequence of solved math problems to build hype — Confidence Mismatch (45/100)

At 2:40

Stacking unsolved conjectures like they're solved chores. That's not math, that's a highlight reel of impossible wins 🚩

Why this score: The speaker presents these massive mathematical breakthroughs in a rapid-fire list, creating an artificial sense of momentum and inevitability that ignores the actual complexity of mathematical proof.

Original quote: “It claimed the Kakeya problem in three and four dimensions. It settled a version of Hilbert's 10th problem from the year 1900. It proved Catalan's constant is impossible to write as a fraction. It knocked down the decades-old MEI conjecture in every dimension.”

OpenAI's transparency is a half-measure, hiding the real data 🎭 — Missing Context (45/100)

At 4:50

They're giving you the 'summary' but hiding the actual receipts. It's a shell game of transparency 🕵️

Why this score: The speaker highlights the gap between what OpenAI shared and what they withheld, implying a strategic lack of transparency without explaining the technical or competitive reasons why those specific details were kept secret.

Original quote: “OpenAI shared the average compute and reasoning summaries, but refused to release the prompts or the per problem results.”

Hype-building around a new model release 🚀 — Plain Sales Pitch (45/100)

At 5:54

Dropping a name like 'LeChunk' and calling it a 'flagship' is pure marketing energy 🤡

Why this score: The segment shifts into promotional mode, using high-energy descriptors for a new product release. It's presenting the existence of 'LeChunk' as a major event without immediate critical context.

Original quote: “Mistral just dropped a preview of their new flagship AI model, Mistral Large 4. They are calling it LeChunk.”

Using specific benchmark scores to claim superiority 📊 — No Frame (75/100)

At 6:16

Actually giving me numbers and showing where they fall short. No games being played here ⚖️

Why this score: The speaker provides specific percentage scores and compares them against other models, even acknowledging that it isn't the top performer. This is a direct, data-driven comparison.

Original quote: “On the Deep Eval coding benchmark, Mistral Large 4 scores around 62%. That puts it slightly ahead of GLM 5.3 and DeepSeek V4 Pro, though Kimiko 3 still leads at 68%.”

Comparing Mistral's 3,800 GPUs to OpenAI's 100,000+ via a massive ratio — Cherry-Picked (45/100)

At 6:36

Using a massive '25x difference' to make one look small and the other look like a giant 🚩

Why this score: He's isolating raw GPU counts to create a dramatic contrast, ignoring that efficiency and architecture matter more than just hardware volume.

Original quote: “Mistral trained this model on exactly 3,800 Nvidia Grace Blackwell GPUs inside their own European data centers. Greg Brockman previously stated that GPT-6 Astra was trained on more than 100,000 GPUs. That is a 25 times difference in pure GPU count.”

Claiming top models score zero because they 'refuse' tasks — Loaded Language (45/100)

At 7:13

Using 'refuse' to turn a technical limitation into a personality flaw 💀

Why this score: He's framing model behavior as 'compliance' vs 'capability,' which is a clever way to make the big players look stubborn rather than just different.

Original quote: “The leading closed models score near zero on these specific tasks because they actively refuse to do them. If one model attempts the task and another outright refuses, the score stops measuring pure capability and starts measuring compliance.”

Promoting Anthropic's new model as the ultimate efficiency play — Plain Sales Pitch (50/100)

At 8:25

Ending on a hype-train of 'cheapest and fastest' superlatives 🚀

Why this score: He pivots from the technical debate straight into a marketing pitch for Anthropic's new release.

Original quote: “Anthropic just dropped Haiku 5.5. This is their cheapest, fastest, and most”

Uses massive score gaps to highlight superiority — Cherry-Picked (20/100)

At 8:30

Dropping a score that's double the old one is pure drama. 🍒

Why this score: While the numbers may be real, highlighting a massive delta in one specific benchmark is designed to make the upgrade feel mandatory.

Original quote: “It costs roughly 75% less to run than Claude Haiku 4.5. On the GPQA benchmark for knowledge work, Haiku 5.5 scored 1620. The older 4.5 model scored only 735.”

Sources: 20 Highest Scoring College Football Games - College Transitions, Race gaps in SAT scores highlight inequality and hinder upward mobility | Brookings, Blowout (sports) — Grokipedia

Claims superiority over a competitor using benchmark data — No Frame (75/100)

At 8:46

Direct comparison. Hard to argue with the math if it's real. 🟢

Why this score: The speaker is providing a direct comparison against a specific competitor model using numerical data.

Original quote: “More importantly, Haiku 5.5 outperforms OpenAI's GPT-6 Luna, which scores 1437.”

Uses a massive cost discrepancy to show the 'catch' — Cherry-Picked (20/100)

At 10:00

A $23 difference for one task? That's a hell of a way to sell the 'catch'. 🍒

Why this score: Using one specific, highly divergent cost example to illustrate the 'token hungry' nature of the model is a classic way to highlight a weakness.

Original quote: “The generation cost $1.91 on GPT-6 Luna. The exact same generation cost $24.96 on Haiku 5.5.”

Sources: What’s the Catch? Uncovering the Hidden Costs Behind “Too Good to Be True” Offers, Pricing Discrepancy: Why Customers Leave Before Checkout, Invoice Discrepancies: Types, Causes & How to Resolve Them

OpenAI is 'entirely quiet' while doing a code update — missing context — Missing Context (45/100)

At 10:30

Calling them 'entirely quiet' is a massive reach when they're literally dealing with high-profile AI misconduct reports 🚩

Why this score: The speaker frames OpenAI as being in a quiet lull, but ignores the heavy lifting they've been doing with safety disclosures and managing 'concerning behavior' reports in the public eye.

Original quote: “OpenAI has been entirely quiet during this run. They are currently doing a 28 days of code update”