OpenAI ferme SORA | Anthropic nous ment | Présidentielle 2027 sous IA !

Today on Silicon Carnet, OpenAI has reportedly pulled the plug on Sora just 15 months after its launch, and a $1 billion deal with Disney fell through before it even materialized. Simultaneously, the music industry is drowning under millions of AI-generated tracks. This raises the question: is AI-assisted creation already collapsing, or is this a different kind of seismic shift? Nearly 3,000 documents related to Claude Mythos, Anthropic's most powerful model, became publicly accessible due to a supposed configuration error. Was this a grotesque leak or a cleverly orchestrated operation? We'll debunk the myth of AI in Silicon Carnet this week. Looking ahead to 2027, the next French presidential election could be the first truly shaped by artificial intelligence. Three-quarters of young people are already using these tools to form political opinions. The cultural battle has begun and is being fought now. Joining me today are Brivael Le Pogam, CEO of Argile, a company specializing in AI-generated video; Fabrice Epelboin, who will discuss the political implications of AI; and Greg Gambato, a cybersecurity expert from Control G. Greg's company, Control G, focuses primarily on cybersecurity, specifically on fixing vulnerabilities. They are developing a benchmark to assess AI's ability to find and fix vulnerabilities in code. A spoiler alert: AIs are not yet fully capable. They often find vulnerabilities but only partially fix them, leaving attack vectors that give a false sense of security. We encourage listeners to subscribe to Silicon Carnet, comment, and engage in public discourse. This show is not just about news; it's about thinking together. We invite corrections and additional insights from our audience to enrich the conversation. For those who want to continue the discussion beyond the episodes, there's La Hacienda, a 24/7 WhatsApp community for Silicon Carnet fans. Let's start with Sora 2. Brivael, as CEO of Argile, your insights are particularly relevant. OpenAI recently shut down Sora after only 15 months, leading to the collapse of a potential $1 billion deal with Disney. Fidji Simo, a key figure at OpenAI, reportedly called Sora a "side quest" the company needed to abandon. What does this mean for AI-assisted creation? Is it an economic dead end? Brivael notes that Sam Altman, OpenAI's CEO, has been running the company like a big tech giant such as Google, despite not having the same "cash printer" like Google’s AdSense. OpenAI has raised $110 billion, but Brivael believes Altman lost focus. Anthropic's recent traction in B2B and coding has put pressure on OpenAI to return to its core focus. Brivael doesn't believe video AI is dead, but rather that it requires dedicated focus, which OpenAI lacked. The business model for Sora, which offered $100 in compute for users to generate videos for platforms like TikTok, was unsustainable. Sora generated only $1.4 million in revenue while costing millions daily in compute. Generating a 60-second video costs around $172, even if the price were divided by 10, it would still be $17 per minute, making it economically unviable for consumers. Brivael explains that current AI video models are not yet at the level of image generation and are still highly fragmented. Future solutions will likely be hybrid, combining organic and AI-generated content, where AI might only contribute a small, valuable portion of a longer video. Technologies like distillation and quantization are being developed to make models more efficient and accessible. When OpenAI launched Sora in 2024, it was presented not just as a fun video application, but as a "Video Generation Model as World Simulators," aiming to advance scientific understanding of the physical world through video analysis. This ties into Yann LeCun's "world models" concept, suggesting LLMs could understand physics through video. However, Sora's user adoption quickly peaked and then plummeted, raising questions about the viability of this approach. Brivael expresses optimism for Google's multimodal models, predicting that their vast data, infinite compute, and optimization capabilities will lead to significant breakthroughs. The current limitation of AI video is lack of control over elements like scene, expression, and camera movements. Once models become controllable, the cost will become marginal, enabling the production of long-form content like documentaries and films, which are currently very expensive to produce traditionally. Fabrice agrees that AI will drastically reduce production costs in the cultural sector, leading to a "brutal and violent" transformation within the next five years. He notes that AI-generated music is already prevalent and often of good quality, making it easy to compete with what he describes as increasingly "shitty" output from human artists. Greg raises concerns about AI's impact on B2C, where it's becoming a "spam machine" for content and emails. While beneficial for tech-savvy individuals, AI is generating widespread dislike among the general public due to job fears and the proliferation of low-quality, profitable spam. He fears a future where governments are elected on anti-AI platforms due to public opposition. Fabrice corroborates this, acknowledging that AI will bring unpleasant transformations for many. He points to media coverage, like a segment on France 5, where journalists express fear and contempt for AI, highlighting a disconnect between the traditional cultural establishment and new technologies. This establishment, often financially protected and detached from economic realities, views AI as a threat to their current model. Greg understands the fear, recalling how the rise of copywriters impacted his previous content-dependent business. He now sees AI creating an even more competitive landscape, forcing him to pivot his niche. Brivael admits to using AI agents to generate content for his X (formerly Twitter) feed, allowing him to produce a high volume of structured content on topics like economic liberalism. He emphasizes that this is not simply copying and pasting AI output but creating a system that acts as an "extension" of himself, reflecting his personality and knowledge. Brivael distinguishes between using AI as a "crutch" (without understanding the underlying principles) and using it as a "lever" (to amplify existing skills). He believes AI will benefit those already proficient in their fields, turning an "X10 engineer" into an "X100 engineer," but won't significantly help mediocre performers. Fabrice relates this to the concept of universal basic income, an idea supported by Sam Altman, but questions its feasibility in Europe without corresponding value production. He dismisses Mistral's proposal to tax the revenue of token-creating companies as a traditional tax model. Jean-Louis Kéguiner's idea of taxing tokens used by AI that replace human labor is seen as more relevant. Greg mentions his company plans to spend $200,000 on tokens for a project in April, compared to less than $50,000 in human costs, illustrating how AI can significantly reduce HR expenses. This suggests that taxing token usage could be a way to redistribute value. Next, we discuss Anthropic's supposed leak of Claude Mythos documents. Greg finds it difficult to definitively label it as a marketing manipulation or a genuine error. While Anthropic's internal culture emphasizes AI safety, the leak of blog article drafts on their website seems "cheesy." He suspects it's a deliberate marketing strategy, a form of "fear, uncertainty, and doubt" (FUD), where they generate buzz and position themselves as responsible while gaining visibility. This, however, contradicts their ethical stance. The leaked documents claim Mythos would be unparalleled in programming, academic reasoning, and cybersecurity. The cybersecurity aspect is crucial, as it's a $300 billion annual market. Anthropic seems to be positioning Mythos as a defense mechanism against AI-powered threats, implying that if such a powerful tool fell into the wrong hands, it would be a colossal offensive weapon. Greg explains that while AI safety measures are in place to prevent models from directly aiding attacks, open-source solutions like "Hérétique" can "de-bridle" models, enabling offensive use. He notes that while new models are often touted as revolutionary, benchmark improvements are often incremental. Anthropic's decision to not publicly release all their models, using some internally for synthetic data generation, further complicates the picture. Brivael points out a perceived asymmetry between Anthropic's "safety" discourse and the reality of Claude Code, which he finds more permissive for attack scenarios than OpenAI's Codex. He suggests Anthropic might have reduced guardrails in coding to maintain model utility, as excessive restrictions can render models unusable. This aligns with Anthropic's quiet abandonment of their "responsible scaling policy" earlier this year, an ethical commitment to safety before deploying models. The leaked documents also include details about a retreat for influential European CEOs in the UK, signaling Anthropic's strategic outreach to Europe. Fabrice interprets this as Anthropic betting on a post-Trump, Democrat-led US landscape, and seeking allies in Europe due to their strained relationship with the Pentagon. He anticipates a major European charm offensive from Anthropic, targeting both businesses and politicians, to secure a foothold in the evolving AI regulatory landscape. Anthropic's fear-mongering narrative about AI, followed by offering themselves as a solution, resonates well in Europe, which is wary of rapid technological change. David Sacks' concept of "regulatory capture" is relevant here: AI giants seek regulation to limit competition and create barriers to entry that only they can afford. Protests against AI, while seemingly anti-establishment, inadvertently play into the hands of companies like OpenAI and Anthropic, who benefit from regulation that consolidates their power. Turning to the 2027 French presidential election, Gabriel Attal's plan to develop an internal AI platform with five dedicated employees and an external startup for his Renaissance party, making AI their largest budget, is discussed. Fabrice believes AI will play a crucial role. He cites Sarah Knafo's surprising success in Parisian municipal elections, where she heavily used AI for visual generation and social media campaigns, achieving high engagement by adapting her political discourse to digital media. Her program, heavily influenced by AI-driven deep searches for best practices, demonstrated the potential for AI in policy development. An Ipsos poll reveals that 48% of French people plan to use AI to research political candidates, rising to 75% among 18-24 year olds. Notably, 72% of La France Insoumise (LFI) sympathizers intend to use AI. Fabrice highlights that LFI, and before them the far-right (FN) in the late 1990s, were early adopters of internet technologies, understanding that the cultural battle would be fought online. While the far-right's digital presence grew organically from its electorate, LFI has a structured, dedicated web agency and a highly organized Discord community, giving them a significant advantage in leveraging technology for political communication. Brivael argues that the radical left has already won the cultural battle because they understand that politics is primarily communication, unlike the technocratic discourse of progressives. He predicts that the 2027 election will hinge on AI strategy for content and communication. AI will enable campaigns to tailor messages to specific demographics (e.g., young people, boomers) with an efficiency that previously required hundreds of staff. Greg notes the potential for foreign interference, joking about "the Russians" but also including the Americans. He points out that over half of voters decide in the last week before an election. A benchmark showing ChatGPT tends to vote left suggests that the data used to train AI models, or the information they retrieve, could sway elections, especially in close races. Fabrice suggests that AI models, when engaged in extended conversations, can become polarized by the user's ideology. His Claude model, for example, leans liberal due to his economic discussions. This means that a voter using AI for information might receive answers that align with their existing biases or even subtly shift them. Contrary to early fears of a "deluge of deepfakes and fake news," Fabrice observes that people are generally able to discern reality. He dismisses Macron's panic over a fake coup video as moral panic. He believes AI will be used for more sophisticated, Cambridge Analytica-style personalized messaging, rather than crude deepfakes. Brivael finds AI's potential for pedagogical content particularly interesting. He uses his X account to debunk economic misconceptions, where his AI agents systematically respond to arguments with data. This creates a "fact-checking" and "rhetorical argumentation" system that makes it difficult to be evasive or intellectually dishonest. This marks a shift where defense can finally match attack in online discourse, potentially leading to the "death of Brandolini's Law" (it's easier to create a lie than to refute it). Greg adds that Trump's unique political success stems from his ability to mix profound truths with outright lies, creating a strong connection with his audience. However, Brivael believes that AI agents providing systematic, data-backed responses could counter this, making it harder for politicians to rely on constant falsehoods. We conclude with an optimistic view of AI, emphasizing its potential for education and truth-telling. Grok, for instance, can fact-check information on X in real-time, making it a powerful tool for discerning truth. The combination of community notes and Grok, along with X's algorithm that promotes top human comments, is creating an environment where misinformation is challenged. Elon Musk's efforts, including auto-translation, could lead to a truly global, less siloed network where information is more readily verified. The discussion ends with a hopeful outlook that AI, if used wisely, can awaken citizens and foster a more informed society.