Tens of Thousands of Tweets Criticize, Valuation Falls — OpenAI Hit by Misleading “Breakthrough”! Terence Tao: Capable, But Headed the Wrong Way?
Beijing, 2025‑10‑20 13:21
> “Hoisting their own GPT rock only to drop it on their own foot.” — Meta’s Chief AI Scientist Yann LeCun on OpenAI’s researchers.
---
LeCun Targets OpenAI Over GPT‑5 Incident
LeCun’s comment came after OpenAI researchers prematurely celebrated what they called a new mathematical “breakthrough” by GPT‑5 — a claim they later retracted following intense skepticism from the AI community. Even Google DeepMind CEO Demis Hassabis publicly criticized their announcement, calling it flawed.
---
The GPT‑5 “Breakthrough” That Wasn’t
How the Claim Emerged
- Initial post by Sebastien Bubeck (former Microsoft VP, now OpenAI research scientist)
- Claimed two researchers used GPT‑5 to solve 10 Erdős problems — complex challenges posed by mathematician Paul Erdős, known for their difficulty.
- October 18 announcement by Mark Sellke (OpenAI researcher)
- Reported:
- Thousands of GPT‑5 queries yielded solutions to 10 “unsolved” Erdős problems.
- Significant progress on 11 others.
- An alleged correction to an error in Erdős’s original paper.
- Amplification by Kevin Weil (OpenAI VP)
- > “GPT‑5 solved 10 (!) previously unsolved Erdős problems, and made progress on 11 others.”
All related social posts have since been deleted.
---
Community Pushback
- Thomas Bloom (mathematician and maintainer of the Erdős site):
- > “Claims were seriously misleading. GPT‑5 only found references already solving these problems — ones I hadn’t seen before.”
- Bubeck’s later clarification:
- > “It only found solutions in existing literature.”
- Still considered it noteworthy due to the difficulty of literature retrieval.
- Hassabis’ verdict:
- “This is embarrassing.”
---
Fallout and Repercussions
Deleted Posts & Reputational Damage
- Most announcements withdrawn.
- Analysts suggest the episode exposes impulsive decision‑making under competitive pressure.
Broader Implications
- Highlights risks of premature claims in an overhyped AI market.
- Social media hashtags like #OpenAIFail surged to 10,000+ posts expressing disappointment.
Financial Impact
- Stock‑linked instruments tied to OpenAI dropped sharply in pre‑market trading.
Regulatory Scrutiny
- FTC investigation into potential false advertising.
- U.S. lawmakers calling for greater transparency in AI research.
- Concerns over OpenAI’s privileged access to the FrontierMath benchmark via ties to Epoch AI.
---
Recognizing AI’s True Utility in Mathematics
Despite the controversy, GPT‑5 demonstrated strong potential as a literature retrieval aid in research workflows.
Terence Tao’s Perspective
- AI assistants excel at scaling routine research tasks like literature reviews.
- Human validation remains essential.
- Benefits of AI in literature search:
- Verifiable results.
- Efficient multi‑problem scanning.
- Acceptable even with less than 100% success rate.
- Lowers time and effort compared to non‑AI methods.
---
Current Challenges in Literature Review
Issues when negative results (no literature found) go unreported:
- Wasted effort — repeated searches for nonexistent work.
- Incorrect assumptions — belief that a problem is unsolved when it has been addressed elsewhere.
AI tools make it natural to report both positive and negative search outcomes.
Example:
> “Of 36 issues searched by the tool, 24 (66%) returned new relevant results; 12 produced only known or irrelevant literature.”
---
References:
- Leading OpenAI researcher announced a GPT‑5 math breakthrough that never happened
- Terence Tao’s statement
---
Event Preview: 2025 Shenzhen International FinTech Competition
- Total Prize Pool: ¥500,000 + trophies & certificates
- Advisory Board: Academicians, Changjiang Scholars, etc.
- Eligibility: Undergrad, Master’s, PhD students (AI, finance, CS, mathematics)
- Registration Deadline: Nov 16
- How to register: Scan QR or click “Read Original”
---
Conference Recommendation: QCon Shanghai
October 23–25 — Only 3 days to go!
- 100+ engineering case studies
- Topics include:
- Agentic AI
- Embodied Intelligence
- RL frameworks
- On‑device large models
- Multi‑agent collaboration
- AI‑era software development & open source
Seats are limited — reserve now.
---
Today’s Recommended Reads
- Superstar AI Coding Assistant Price Increases 10x, Users Upset
- Claude Skills May Outshine MCP
- Anthropic’s New Model Excels
- Karpathy’s 8K Lines ChatGPT Build
- Zhipu Denies Layoffs Before IPO
---
📎 Read the full original article
---
Key takeaway: GPT‑5’s role as a literature search aid has value — but overhyping AI capabilities without rigorous validation can harm credibility, attract regulatory attention, and mislead the public.
