OpenAI vs DeepSeek : qui pirate le mieux ?

OpenAI vs DeepSeek : qui pirate le mieux ?

OpenAI vs DeepSeek: who hacks better?

Title history
  • OpenAI vs DeepSeek : qui pirate le mieux ? — observed on 20261117-0331 UTC (YouTube)
🎙 Silicon Carne 🌶️ 👥 110K 📅 November 16, 2026 ⏱ 28 min 👁 16K 📄 debate 🧭 2026-11-17
Available in: English (current) Français

Keywords

offensive AIopen-weight modelscost of attackSMB vulnerabilitygeopolitics

Summary

The episode features a panel discussion between host Carlos Diaz and two cybersecurity experts, Greg Gambatto and Nico (Nisedo), analyzing a recent VALS benchmark where AI agents were tasked with hacking fictional networks. The host highlights that OpenAI’s frontier model completed the task for about $20, while the open-weight DeepSeek model did it for less than $3, prompting a discourse on access inequality. However, the guests provide a strong counter-analysis, arguing that the VALS benchmark is simplistic, using only two environments versus the dozens used in their own internal tests. They explain that DeepSeek’s low cost per token is misleading, as it consumes 5 to 10 times more tokens due to brute-force strategies, making it slower and easier to detect compared to the surgical approach of closed models like OpenAI’s. The discussion shifts to the broader threat landscape, noting that the falling cost of cyberattacks will make thousands of unprepared SMBs, industrial plants, and municipalities lucrative targets. The panel also addresses the release of a ‘debridled’ version of GLM by Abliteration AI, raising concerns about the removal of safety rails. Finally, they touch on geopolitical dimensions, contrasting Western offensive capabilities with those of isolated nations like North Korea and Iran, and briefly discuss the potential espionage risks posed by connected vehicles like BYD in Western countries. The show includes a short sponsored segment of approximately two minutes.

227 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information lies in its rare practitioner-level insight into how AI models behave in offensive security contexts. The guests effectively deconstruct the widely covered headline about DeepSeek being ‘cheap’, demonstrating with internal data that the operational cost and detectability often make it inferior to frontier models. The ‘sprinter vs marathoner’ analogy is particularly strong in articulating the error compounding and plateauing of open-weight models. The argumentation about the ‘cost of attack dropping to zero’ offers a convincing economic model for the coming wave of attacks on small entities. The panel successfully avoids a simplistic ‘China vs America’ narrative, instead focusing on the concentration of defensive power and the systemic weakness of defenders.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate; the content is based on personal experience and mentions of the VALS benchmark and Abliteration AI without direct citations to formal papers. The strengths rely on the guests’ credible professional backgrounds (a San Francisco-based datalab and a security company) and their concrete comparison of proprietary internal evaluations. The title perfectly matches the core content, which directly addresses the OpenAI vs DeepSeek comparison, while expanding to wider implications which enrich rather than distract. No public comments were provided for this analysis.

215 words

Title / Content Match

The title accurately captures the central premise of the episode (a comparison of offensive AI capabilities), even though the content broadens to economic, geopolitical, and defense-related implications.

Quality & Reliability

7/10

The program relies on expert opinion from cybersecurity professionals and their internal benchmarks to critically challenge a publicly available test (VALS). The reasoning is pragmatic and well-structured, but the lack of external academic references and the dominant talk-show format limit its absolute reliability.

Chapters

Cited Sources

External References

Contribution & Novelties

The primary contribution is the expert-led debunking of a viral benchmark, offering a more nuanced picture of AI offensive capabilities by detailing hidden costs (token consumption) and detection risks. It highlights the severe disconnect between advanced AI vendors and the defensive incapacity of small businesses, providing a stark economic analysis of modern cybercrime incentives.

Pour aller plus loin :

  • DeepSeek — The open-weight AI model central to the cost comparison discussed in the episode.
  • Cyberwarfare — Provides background on the geopolitical offensive/defensive strategies touched upon in the discussion.
  • Accumulation of errors — A technical concept explaining why cheaper open-weight models plateau and fail on complex tasks, a core argument of the guests.
  • ChatGPT — Relevant for understanding the context of OpenAI’s frontier models and their access restrictions.

127 words

Radar Profile

The radar profile shows a high quantity of information and a solid technical level, with a slightly lower reliability score reflecting the opinion-driven framework. The global note of 4/5 aligns with the balance between expert practical quality and the absence of formal academic structuring.

Reliability 6/10