Breaking: OpenAI Loses Key Figure, Father of Sora Departs, Turmoil Continues on the Eve of IPO

marsbitPublié le 2026-04-18Dernière mise à jour le 2026-04-18

Résumé

OpenAI is undergoing significant internal turmoil ahead of its planned IPO, marked by high-profile executive departures and strategic shifts. Key figures including Sora project lead Bill Peebles and Chief Product Officer Kevin Weil have announced their exits. The company is consolidating projects, discontinuing standalone products like the research tool Prism and integrating it into Codex, while also halting the video project Sora. Simultaneously, CEO Sam Altman faces renewed allegations of conflicts of interest. Reports reveal he attempted to have OpenAI invest in his personal ventures, such as nuclear fusion company Helion and space firm Stoke Space, raising internal concerns about transparency and governance. These moves come as OpenAI, valued at around $850 billion, shifts from rapid expansion to focusing resources on core priorities like enterprise products and AGI development, under pressure from competitors and the upcoming IPO.

Zhixiandao, April 18th report—Early this morning, OpenAI was hit with a series of changes including executive departures, project scaling back, and questions regarding personal conflicts of interest involving co-founder and CEO Sam Altman. This AI star company, sprinting towards an IPO, is once again facing a storm of personnel and trust issues.

On one hand, multiple executives including Sora core lead Bill Peebles and Chief Product Officer Kevin Weil have successively announced their departures, with projects like the video project Sora and the scientific research tool Prism heading towards shutdown, consolidation, or reorganization.

On the other hand, The Wall Street Journal reported that Sam Altman had previously attempted to push OpenAI to support his personally backed nuclear fusion and space projects. This move has also plunged the company, at the critical juncture of preparing for a listing with a valuation of approximately $850 billion (about RMB 5.8 trillion), into renewed questioning over conflicts of interest.

Sora Architect Departs

OpenAI's Video Division Loses Another Key Player

Early this morning, Sora project core lead Bill Peebles announced his departure on a social platform, sharing a photo of himself working alongside Sam Altman.

▲ Work photo of Bill Peebles (left 1) with Sam Altman (right 3) (Source: Bill Peebles)

In an open letter, Bill reviewed the development journey of Sora from scratch, calling the experience of building this project with the team the "most honorable and most like an adventure" of his career. He also thanked Sam Altman and others for fostering an inclusive research environment for the team, and stated outright that Sora "could almost certainly not have been born outside of OpenAI".

▲ Bill Peebles' resignation letter

Sam Altman later responded in the comments section, "Will miss Bill a lot", stating that his creativity pushed OpenAI and the entire industry to experience AI video in new ways, while also expressing anticipation for his next moves. Bill subsequently replied, "Thanks for everything".

▲ Comments section interaction between Sam Altman and Bill Peebles

Chief Product Officer Departs Same Day

Prism Merged into Codex

On the same day, OpenAI Chief Product Officer, former Instagram executive Kevin Weil, also announced his departure.

He posted that with the "OpenAI for Science" initiative he led being dispersed into other research teams, today would be his last day at the company, emphasizing that accelerating scientific development would be one of the most positive outcomes on the path to AGI.

▲ Kevin Weil's resignation letter

ChatGPT product lead Nick Turley subsequently interacted in the thread, thanking Kevin Weil for his guidance and inspiration in building ChatGPT together. Kevin Weil replied that he also enjoyed the process of working together, joking "Keep making ChatGPT better, I'm a DAU (daily active user)".

▲ Comments section interaction between Nick Turley and Kevin Weil

According to WIRED, behind Kevin Weil's departure, OpenAI is not just losing a former executive, but also undergoing a reorganization of its scientific application direction.

The report stated that OpenAI has decided to stop independently advancing Prism. Prism was a web application for researchers launched by OpenAI in January this year, originally attempting to provide scientists with AI tools better suited for scientific workflows.

Now, this product and its team of about 10 people will be merged under Thibault Sottiaux, head of Codex, with subsequent related capabilities to be integrated into the desktop version of the Codex application.

According to an OpenAI spokesperson, this change is part of the company's unified business and product strategy. More specifically, OpenAI hopes to further push Codex from a programming tool towards a "super app" direction, with Prism no longer existing as a separate product line.

Kevin Weil joined OpenAI in June 2024 and publicly announced the launch of the OpenAI for Science project in September of the same year. Now, less than two years later, the team members have been dispersed to various departments including product, research, and infrastructure.

Personnel Shake-up Extends Beyond Two

OpenAI is Accelerating Consolidation

The departure of these two individuals is only one aspect of a new round of internal contraction at OpenAI.

According to WIRED, in addition to Kevin Weil and Bill Peebles, OpenAI CTO of Enterprise Applications Srinivas Narayanan has also announced his departure, citing a desire to spend more time with family. Narayanan previously served as OpenAI's VP of Engineering and was an important executive for OpenAI's enterprise product direction.

This series of changes occurs amidst a "major reshuffle" of the executive team. Previously, AGI deployment lead Fidji Simo, who attempted to build a "super app", and Chief Marketing Officer Kate Rouch both took medical leave. President Greg Brockman temporarily took over product management, Chief Operating Officer Brad Lightcap moved to special projects, and the once industry-sensation Sora video application has currently been halted.

As Sam Altman admitted in a blog post published on April 10th, OpenAI is transitioning from "a chaotic but driven startup" to "a large platform that needs to operate in a more predictable way". He wrote that the company's development in recent years has been "extremely high-pressure, chaotic, and intense".

▲ Sam Altman's long response after his residence was firebombed.

Under pressure from competitor Anthropic and with an IPO planned for later this year, OpenAI is internally pushing to simplify its product lines, shifting resources from peripheral projects to areas with higher strategic priority.

In other words, Bill Peebles' departure and Prism's merger into Codex are consecutive actions in OpenAI's shift from "multiple experimental lines" to "consolidating main lines".

Turmoil on the Eve of IPO

Altman Again Embroiled in Conflict of Interest Allegations

As OpenAI accelerates its transformation, the issue of the "opacity" of Sam Altman's personal investment portfolio is being brought to the forefront.

According to The Wall Street Journal, after Sam Altman was briefly ousted and reinstated in 2023, the OpenAI board had promised to strengthen management of his potential conflicts of interest, but this issue has not truly disappeared.

Now, with OpenAI's valuation pushed to around $850 billion, questions about whether "company decisions are for OpenAI or indirectly serving Sam Altman's personal investment portfolio" have become more prominent.

The most typical case is the nuclear fusion company Helion.

▲ Nuclear fusion company Helion

Sam Altman is a major investor in Helion, with a significant portion of his personal net worth tied to the company. Helion has experienced delays in delivering on its technology and begun facing funding pressures. Subsequently, Sam Altman pushed for OpenAI to lead a new round of financing for Helion, proposing an investment of about $500 million (approx. RMB 3.4 billion), which would have valued Helion at around $35 billion (approx. RMB 238.6 billion) under the proposed plan.

This proposal caused noticeable unease within OpenAI. Some employees believed Sam Altman was asking the company he leads to endorse an enterprise from which he directly benefits, with no immediately obvious business value for OpenAI in the transaction.

Some employees were even concerned that internal Slack discussions might become legal review material in the future, thus avoiding participation in related discussions.

OpenAI ultimately did not participate in this round of investment, but still signed an energy purchase agreement with Helion, agreeing to purchase up to 50 gigawatts of electricity by 2035, equivalent to the output of 25 Hoover Dams. Helion later used this agreement as an important endorsement in communications with investors.

With OpenAI withdrawing from the financing, Helion has now reduced its fundraising target to $250 million (approx. RMB 1.7 billion), with a valuation target lowered to $15 billion (approx. RMB 102.2 billion). Sam Altman resigned from the Helion board last month, explaining on social media that as Helion and OpenAI began exploring large-scale cooperation, it was no longer appropriate for him to sit on both boards.

Besides Helion, The Wall Street Journal also disclosed that Sam Altman had attempted to push for deeper cooperation between OpenAI and rocket company Stoke Space, even exploring having OpenAI acquire or control the company to support ideas like "space data centers".

Sam Altman himself has financial ties to Stoke through his family office, Hydrazine. Although Sam Altman publicly called the idea of "building data centers in space" "ridiculous" in India this February, the report cited informed sources stating that related cooperation discussions actually continued, with some OpenAI board members previously unaware.

Conclusion: Sora Reeled In, Prism Dismantled

What's Left on OpenAI's Main Line?

OpenAI's current situation is quite clear: it is actively compressing non-core businesses, concentrating resources on a few directions that can better support the IPO narrative and commercial returns.

The Wall Street Journal also mentioned that Fidji Simo told employees last month that Anthropic's progress should serve as a "wake-up call", and the company needs to direct more resources towards products for professional work scenarios.

Against this backdrop, some of Sam Altman's past product concepts, which were more personal in color, are gradually being scaled back. Simultaneously, his focus is quietly shifting: from product management to research, financing, and the acquisition of computing resources.

On one side, star employees announce their departures; on the other, product lines are merged. Coupled with multiple controversies on the eve of the IPO, OpenAI is transitioning from an expansion phase of "wanting to do everything" to a contraction phase of "having to succeed in a few things first".

Under the triple pressure of capital markets, competitors, and internal governance, the keyword for OpenAI's next stage is no longer all-out expansion, but concentrated firepower.

Sources: The Wall Street Journal, WIRED, Kevin Weil, Bill Peebles

This article is from the WeChat public account "Zhixiandao" (ID: zhidxcom), author: Jiang Yu

Questions liées

QWho are the key executives that recently left OpenAI, and what were their roles?

ABill Peebles, the core leader of the Sora project, and Kevin Weil, the Chief Product Officer, recently left OpenAI. Additionally, Srinivas Narayanan, the CTO of enterprise applications, also announced his departure.

QWhat projects at OpenAI are being shut down or restructured?

AThe video project Sora is being halted, and the scientific research tool Prism is being discontinued as a standalone product. The Prism team is being merged into the Codex team under Thibault Sottiaux.

QWhat personal利益冲突 issues involving Sam Altman were reported by The Wall Street Journal?

AThe Wall Street Journal reported that Sam Altman attempted to push OpenAI to invest in Helion, a nuclear fusion company he personally backed, and explored deeper cooperation or even control of Stoke Space, a rocket company, both of which involve his personal financial interests.

QWhy is OpenAI undergoing internal restructuring and project consolidation?

AOpenAI is shifting from a 'multi-line experimentation' phase to a 'mainline consolidation' strategy to focus resources on strategic priorities, streamline product lines, and prepare for its upcoming IPO, amid pressure from competitors like Anthropic.

QWhat is the current valuation of OpenAI and what challenges does it face before its IPO?

AOpenAI is valued at approximately $850 billion and faces challenges including executive departures, project restructuring, and concerns over Sam Altman's potential利益冲突, all occurring during its critical IPO preparation phase.

Lectures associées

De l'IDE au terminal : Un manuel pratique d'ingénierie d'agents

En mars 2026, Matt Van Horn a publié un article intitulé "Chaque astuce Claude Code que je connais", déclenchant un débat sur son approche "No IDE". Il développe entièrement dans le terminal et un fichier plan.md, déléguant l'exécution à des agents d'IA. Cette méthode, popularisée par meng shao sous le nom de "Conseils pratiques pour l'ingénierie d'agents", repose sur un cycle "Research → Plan → Work". L'idée centrale est de remplacer l'IDE, qui offre un retour visuel immédiat (surlignage, débogage), par un flux de travail basé sur des commandes et un plan. Le fichier plan.md sert de "contrat" avec l'agent, définissant le problème, la solution et une checklist pour guider l'exécution et éviter la "dégradation du contexte" des LLM. Des outils comme Compound Engineering (avec /ce:plan) aident à générer et affiner ce plan. Le cycle en trois phases est crucial : 1. **Research** : L'agent collecte des informations (ex: avec last30days-skill). 2. **Plan** : Génération et révision humaine du plan pour corriger les hypothèses et ajouter des connaissances métier. 3. **Work** : L'agent exécute le plan en parallèle (/ce:work). Des astuces pratiques incluent : générer un plan dès qu'une idée émerge ; faire résumer le plan par l'agent ; utiliser plusieurs terminaux en parallèle ; saisir par voix les instructions complexes ; déclencher des tâches par email (agentmail) ; et utiliser des compétences préexistantes (AgentSkills). Cependant, cette approche présente des risques. Elle nécessite une forte capacité à décomposer les problèmes et à rédiger des instructions précises. Les erreurs non détectées en phase de planification peuvent être amplifiées. Enfin, il existe un risque de "psychose IA", où l'optimisation du flux de travail devient une fin en soi, éclipsant l'objectif initial. Ce flux de travail est un amplificateur d'efficacité pour les développeurs expérimentés sachant exactement ce qu'ils veulent, et non un outil d'apprentissage pour les débutants qui ont encore besoin des retours visuels de l'IDE. Les outils (Claude Code CLI, Compound Engineering) évoluent rapidement, offrant une fenêtre d'expérimentation aux pionniers.

marsbitIl y a 1 h

De l'IDE au terminal : Un manuel pratique d'ingénierie d'agents

marsbitIl y a 1 h

Ne vous focalisez plus sur les GPU, Intel dévoile une manœuvre majeure, peut-elle mettre fin au monopole de Nvidia sur la puissance de calcul ?

Au cours des deux dernières années, le GPU a été au cœur du matériel informatique pour l'IA. Cependant, lors du COMPUTEX 2026, Intel a présenté une vision différente, axée sur l'Agentic AI (IA agentique). Contrairement aux modèles d'IA traditionnels fonctionnant par « tours » de questions-réponses, l'IA agentique opère dans des flux de travail réels, suivant un cycle continu de réflexion, planification, action et évaluation. Cela transforme fondamentalement l'utilisation du calcul, nécessitant une forte coordination système et un traitement de tâches fragmentées. Intel soutient que cela rééquilibrera la répartition des ressources dans les centres de données. Alors que la phase d'entraînement des modèles avancés repose fortement sur les GPU (avec des ratios CPU/GPU pouvant atteindre 1:8), le déploiement d'agents intelligents nécessitera davantage de puissance CPU pour l'orchestration des tâches, l'appel d'outils et la gestion des données, ramenant le ratio vers 1:1, voire en faveur du CPU. La consommation de tokens peut être multipliée par 1000 par rapport à une inférence simple. Pour répondre à cette nouvelle demande, Intel a dévoilé le processeur Xeon 6+ (fabriqué en technologie 18A), offrant jusqu'à 288 cœurs efficaces et conçu pour une haute densité et une meilleure efficacité énergétique. Parallèlement, Intel a présenté une nouvelle architecture d'inférence découplée, développée avec des partenaires comme SambaNova. Cette solution répartit les différentes étapes de l'inférence (orchestration, décodage, pré-remplissage) sur les composants les plus adaptés (CPU Xeon 6, RDU SambaNova, GPU NVIDIA Blackwell) pour optimiser l'efficacité. L'écosystème d'Intel inclut également le processeur Core Ultra de 3ᵉ génération pour l'IA locale sur PC, permettant une répartition dynamique des charges de travail entre le périphérique et le cloud, et les processeurs Arc G3 pour les appareils portables. Enfin, Intel mise sur les puces sur mesure (comme les IPU développées avec Google) pour répondre aux besoins spécifiques de divers secteurs industriels. En résumé, face à la domination de NVIDIA sur les GPU, Intel adopte une stratégie globale : proposer une suite de solutions (puces universelles, architectures hétérogènes, IA locale, composants personnalisés) adaptée à chaque étape et besoin du nouveau paysage informatique défini par l'IA agentique. Le succès dépendra de l'exécution technique et de l'adoption par les clients, mais Intel affiche désormais une direction claire pour capitaliser sur cette reconfiguration des infrastructures d'IA.

marsbitIl y a 1 h

Ne vous focalisez plus sur les GPU, Intel dévoile une manœuvre majeure, peut-elle mettre fin au monopole de Nvidia sur la puissance de calcul ?

marsbitIl y a 1 h

“小美”, Yuanbao, l'interconnexion, est-ce un précurseur pour l'agent intelligent de WeChat ?

Lors de la publication des résultats du premier trimestre 2026 de Meituan, l'attention a été attirée par l'annonce de Wang兴 concernant l'intégration de son assistant IA "Xiao Mei" avec "Yuan Bao" de Tencent. Cette collaboration permettra aux utilisateurs d'exprimer des besoins en services de proximité dans Yuan Bao, déclenchant automatiquement une communication "Agent à Agent" avec Xiao Mei pour accéder aux services de livraison de repas de Meituan. Cet accord stratégique vise à positionner Meituan comme une infrastructure de services de base, tout en s'appuyant sur une entrée IA externe pour compenser son manque d'accès indépendant. Cette démarche intervient dans un contexte où les assistants IA rivaux, comme Dou Bao (ByteDance) et Qian Wen (Alibaba), développent leurs propres écosystèmes fermés ("jardin clos") en intégrant des services transactionnels, contournant potentiellement les avantages de Meituan. L'article analyse les défis de cette collaboration : plafond de l'expérience utilisateur due à l'architecture "Agent à Agent", répartition des bénéfices, et fluidité de l'expérience entre deux écosystèmes distincts. Elle est également interprétée comme un test préliminaire crucial pour le futur "Agent IA" intégré à WeChat, dont le lancement a été rapporté par le Financial Times. Le succès de ce partenariat avec Meituan pourrait servir de modèle pour convaincre d'autres grandes plateformes de s'intégrer à l'écosystème IA de Tencent via des protocoles standardisés, préservant leur souveraineté des données tout en gagnant en visibilité.

marsbitIl y a 3 h

“小美”, Yuanbao, l'interconnexion, est-ce un précurseur pour l'agent intelligent de WeChat ?

marsbitIl y a 3 h

Trading

Spot
Futures
活动图片