Episode Details
Back to EpisodesAI Daily for 11 September: DeepSeek V4.1 Flash, Unpublished Math Trust, OpenAI Training Setting, Cognition SWE-2
Description
AI Daily for 11 September recaps 5 major AI Hacker News stories, moving through deepseek v4.1 flash, unpublished math trust, openai training setting, cognition swe-2.
Chapters
- 00:00:00 — Intro
- 00:00:19 — DeepSeek V4.1 Flash
- 00:01:19 — Unpublished Math Trust
- 00:02:22 — OpenAI Training Setting
- 00:03:07 — Cognition SWE-2
- 00:04:29 — Another Stolen Proof
- 00:05:27 — Closing
1. DeepSeek V4.1 Flash
The next story is DeepSeek V4.1 Flash, a 552-billion-parameter mixture-of-experts model that DeepSeek presents as smarter, faster, and more efficient, with native visual understanding and a KV cache roughly one quarter the size of the previous generation, which is intended to reduce serving costs. Hacker News reacted with excitement about the sparse architecture and lower API prices, alongside skepticism about the memory needed to run a model this large locally and whether its benchmark gains will translate to real-world performance.
2. Unpublished Math Trust
The next story questions whether researchers can trust OpenAI with unpublished mathematics, after allegations that an AI system may have benefited from private research discussions and that a researcher received an incorrect answer about training use, raising concerns about consent, credit, and claims of AI originality. The Hacker News discussion split over the strength of the evidence, the role of human prompting, and whether connecting specialized ideas amounts to genuine discovery.
3. OpenAI Training Setting
The next story is a Hacker News report alleging that OpenAI keeps re-enabling its setting, raising concerns that users’ data could be used for model training after they opted out. The comments split over whether the behavior reflects deliberate manipulation, a configuration bug, or a UI-only issue.
4. Cognition SWE-2
The next story is Cognition’s launch of SWE-2, a coding model whose authors claim it reaches 50.0% on their FrontierCode benchmark, within one point of Fable 5.1 and at 64% lower cost, making the cost-performance tradeoff the central claim. Hacker News debated that comparison, especially the gap between SWE-2’s 92.8% on Terminal-Bench 2.1 and 27.3% on Terminal-Bench 4, along with complaints about closed weights and access through Devin’s tools.
5. Another Stolen Proof
The next story reports Valerio Capraro’s claim that Andreas Thom has presented evidence OpenAI may have trained Astra on conversations in which he and Gábor Kun were working on Gromov’s soficity conjecture, one of ten problems OpenAI later announced Astra had solved, raising questions about whether unpublished human work was presented as an AI breakthrough. The Hacker News reaction focused on the missing original source and the difficulty of accessing the X post, while the original Mastodon posts were openly readable.
That wraps today's front page.