INSIDE AI #16: A Whistleblower’s Account on Huawei's Pangu | Ex-Meta & OpenAI Insiders Reflected on Culture | New Bid for SB53
Edition 16
In This Edition:
Key takeaways:
News:
A Whistleblower’s Account: Huawei's Pangu Model Allegations
Metastatic Misread? Tijmen Blankevoort on ‘Fixing’ Meta AI from Within
Calvin French-Owen on Leaving OpenAI: A Candid Look Inside Its Culture
OpenAI’s Open-Weight Model, Loosening Microsoft Ties
Ruoming Pang Leaves Apple for Meta: Other Apple’s Top AI Engineers Eye the Door
Policy:
New Amendments to Wiener’s Latest Bill SB 53: Renewed Bid for AI Transparency
The AI Whistleblower Protection Act (AIWPA): Protecting Whistleblowers and Strengthening Internal Reporting Systems
Research:
New Publication: FLI AI Safety Index Report (Summer 2025) - Featuring OAISIS’ Contribution
The “Silicon Sentinels” Approach: How Whistleblowers Are Key to Managing AI Risk
Verification for International AI Governance, including Whistleblowing as International Verification Method
Announcement:
OAISIS to Join AI Panel at National Whistleblower Day, July 30 in Washington, D.C.
Edition Highlights
New Publication: FLI AI Safety Index Report (Summer 2025)
We’re excited to share the release of the Future of Life Institute’s AI Safety Index Report (Summer 2025).
A notable feature of this edition is its comparison of whistleblowing systems among leading AI organizations. On the basis of these findings, FLI is calling on all AI companies to publish their whistleblowing policies to improve transparency and accountability sector-wide.
We contributed to the research on whistleblower systems underlying the report.
Report link: FLI AI Safety Index Report (Summer 2025)
OAISIS—soon to be known as AIWI (The AI Whistleblower Initiative)—is aligned with this call and will be making a formal announcement soon. Stay tuned to learn how you can get involved!
The AI Whistleblower Protection Act (AIWPA): Protecting Whistleblowers and Strengthening Internal Reporting Systems
Announcement: We will be in Washington DC for AI panel discussions at the National Whistleblower Day event on July 30. More details to come!
As we prepare for our panel in Washington, we want to revisit the AI Whistleblower Protection Act (AIWPA), introduced by Senator Charles Grassley (R) on May 15. This bipartisan bill—previously featured in our 12th edition—has received strong support from senators across the aisle and endorsement from the National Whistleblower Center.
The bill provides protection for individuals who disclose “AI violations,” including security vulnerabilities and specific threats to public health and safety.
What This Means for AI Companies
Besides protecting disclosures to regulators, the bill would help promote companies developing AI to have effective internal reporting structures
“According to the Securities and Exchange Commission (SEC) Whistleblower Program’s 2021 Annual Report to Congress, approximately 75% of award recipients in that fiscal year had initially raised their concerns internally to supervisors, compliance personnel, or through internal reporting mechanisms.”
As Stephen M. Kohn and Sophie Luskin noted, “Raising concerns internally is often the first way that employees report misconduct or serious potential violations of law.” The data confirms that most whistleblowers begin the process through internal channels.
The bill will help AI companies take extra efforts to ensure that employees who utilize these reporting systems are not subject to retaliation given the strong legal remedies of the AIWPA— intended to protect honest employees seeking to do the right thing by reporting concerns to their bosses.
Whistleblower Remedies and Protections Under the AIWPA
AI employees who face retaliation or wrongful termination for making a protected disclosure—internally or to the government—are granted rights similar to those in Sarbanes-Oxley. Available remedies to whistleblowers include:
Make-whole relief
Back pay
Restoration of lost benefits
Reinstatement to their job
Compensatory damages
Attorney’s fees and legal costs if the whistleblower prevails
While punitive damages are not included, the law provides for double back pay, serving as at least some financial deterrent against retaliation.
Why Now
In the absence of a dedicated regulation covering risks from AI, it is crucial that employees can safely report concerns to supervisors or compliance personnel. These concerns may involve national security, public safety, or consumer fraud. The legislation also responds to well-documented risks raised in Rights to Warn letters, including concerns about foreign adversaries stealing AI technologies and the potential misuse of AI by terrorists, added Kohn and Lushkin.
→ Read more: The AI Whistleblower Protection Act Is Critical for Enhancing Corporate Compliance
Insider Currents
Carefully curated summaries and links to the latest news, spotlighting the voices and concerns emerging from within AI companies.
A Whistleblower’s Account: Huawei's Pangu Model Allegations
Research group HonestAGI published an analysis of Huawei’s newly open-sourced Pangu Pro MoE model and found an unusually high parameter correlation of 0.927 (comparisons of similar models in the industry usually do not exceed 0.7) with Alibaba’s Qwen-2.5 14B model. This suggests that the model was not independently trained, and while Huawei officially acknowledged referencing open-source implementations as common practice, a detailed account of a self-described employee of Huawei published on GitHub alleges a more deliberate process of appropriation, defined as "shelling."
According to the account, a team within Huawei, known as the "Small Model Lab," repeatedly took pre-trained weights from competitor models, first from Qwen 2.5 and then DeepSeek-V3. DeepkSeek’s model was shelled in order to get a larger 718B Mixture of Experts model up and running. This practice, the whistleblower claims, allowed the team to show rapid progress while circumventing the arduous process of training from scratch. In addition, the whistleblower talks about the effort to hide the fact that they shelled from Qwen:
“I heard from colleagues that they used many methods to wash away the Qwen watermark, including intentionally training on dirty data. This provides an unprecedented special case for academic research on model lineage. Future new lineage detection methods can be tested on it.”
While the "shelling" team allegedly operated with few constraints, another team, the "Fourth Column (our translation of “四纵”, an internal unit),” reportedly struggled through immense hardship to train a 135B dense model and a 718B MoE from scratch on Huawei’s own Ascend chips. This reflects the Whistleblower's hope that one day Huawei's chips will be able to match Nvidia's in terms of training ability. Their eventual success proved the viability of the domestic hardware stack, but was achieved despite what the whistleblower described as a frustrating double standard regarding access to compute and other resources.
→ Read: GitHub's "Whistleblower's Account" - HW-whistleblower / True-Story-of-Pangu (in Chinese)
→ Read: Huawei's Pangu large model was questioned ‘copy’ Ali Qwen: official response (in Chinese)
Disclaimer: This is an interpretation of content originally written in Chinese. We welcome feedback and suggestions to correct misinterpretations or inaccuracies in translation or context.
Metastatic Misread? Tijmen Blankevoort on ‘Fixing’ Meta AI from Within
You may have seen the headline from The Information reporting that a departing Meta AI researcher, Tijmen Blankevoort, described the company’s culture as a “metastatic cancer.” In a 2,000-word essay seen and cited by The Information, Blankevoort describes deep-rooted cultural and organizational issues at Meta that he says have slowed progress within the nearly 2,000-person team behind its flagship AI model, Llama.
“I have yet to meet someone in Meta-GenAI that truly enjoys being there. Someone that feels like they want to stay in Meta for a long time because it’s such a great place.”
However, Blankevoort claimed that his memo was not the mic-drop that the media portrayed it to be on his Substack, but rather a carefully researched analysis of systemic issues he believed were hindering Meta AI. Five core challenges he pointed out:
The fear that people feel on a daily basis of reviews and getting fired, commenting on a lack of safety people felt, which is crucial for morale.
The necessity of a culture and processes that enable ‘big projects’ to come to fruition.
The management culture does not promote a sense of camaraderie among employees, leading to a lack of sense of belonging.
Instability in team assignments, leading to experience not building up and crystallizing over time.
A wavering vision that was tough for team members to enthusiastically rally behind.
In that essay, he claimed to draw on sources from organizational psychology, management research, and company history to outline cultural challenges within Meta’s AI division.
I wrote this with care, citing books, articles, and other sources from behavioral and organizational psychology, organizational research, and history. I provided internal examples of the culture that I was seeing manifested, and at the end provided a list of improvements the company could make to strengthen the culture.
According to him, Meta leadership largely agreed with his assessment and acknowledged many of the issues raised, noting that efforts to address them were already underway—particularly following several high-profile hires. He attributed much of the dysfunction to an aggressive, fast-paced AI race and a legacy software culture that, he argued, is ill-suited for the more experimental and collaborative demands of building large language models.
He also pointed out that more than 100 employees reached out to express support after reading the essay.
Many said it articulated concerns they had struggled to define and made them feel less alone in their experiences.
Several reportedly shared feelings of being overworked and had previously blamed themselves, but found reassurance in seeing those issues reflected more broadly.
→ Read: Tijmen’s Substack: Meta's AI culture, Setting the record straight on the leaked document
→ Read: Meta AI Researcher Warns of ‘Metastatic Cancer’ Afflicting Company Culture
Calvin French-Owen on Leaving OpenAI: A Candid Look Inside Its Culture
Calvin French-Owen, who joined OpenAI in 2024 and left OpenAI recently, has published his (mainly positive) observations on the company's internal operations. His account addresses what he describes as significant public curiosity about OpenAI's workplace dynamics:
“…Because there’s a lot of smoke and noise around what OpenAI is doing, but not a lot of first-hand accounts of what the culture of working there actually feels like.”
Here some interesting elements from his reflections:
Everything Runs on Slack
OpenAI operates with zero email—"everything, and I mean everything, runs on Slack." French-Owen noted he received perhaps 10 emails during his entire tenure. He suggested the potential for information overload unless carefully curated and managed, which could be workable.
Bottoms-Up Meritocracy
When he asked about quarterly roadmaps, he was told "this doesn't exist." Instead of master plans, progress emerges iteratively as research bears fruit. Good ideas come from anywhere, and leadership advancement is “based upon employees” ability to have good ideas and then execute upon them, rather than their competency at things like presenting at all-hands or political maneuvering."
"You Can Just Do Things"
There's a "strong bias to action" where employees launch parallel efforts without heavy permission. Multiple teams independently developed "3-4 different Codex prototypes" before the official launch. Researchers operate as "mini-executives" with autonomy to pursue their own directions.
Twitter as Strategic Intelligence
Surprisingly, "OpenAI pays a lot of attention to twitter." Viral OpenAI-related tweets regularly reach decision-makers. One colleague joked "this company runs on twitter vibes."
Safety: More Than Expected, Different Focus
"Safety is actually more of a thing than you might guess," but focuses on practical risks (hate speech, bioweapons, prompt injection) rather than theoretical ones (intelligence explosion). Most safety work isn't published—"OpenAI really should do more to get it out there."
The Launch Machine: From Idea to Product in 7 Weeks
Echoing our story from last editions — where one former employee said Altman had been pushing for “buzzy announcements every few months.”— French-Owen provided detailed insight into OpenAI's product development capabilities through the Codex launch story. The entire product—described as having substantial scope including "a container runtime, made optimizations on repo downloading, fine-tuned a custom model to deal with code edits, handled all manner of git operations"—went from first lines of code to public launch in just 7 weeks.
The team was notably senior ("~8 engineers, ~4 researchers, 2 designers, 2 GTM and a PM") and worked with extraordinary intensity, with him describing nights until midnight and 5:30 AM wake-ups for weeks.
The Path to AGI: The Three-Horse Race
He characterized the current AI landscape as a competition between three major players—OpenAI, Anthropic, and Google. According to his assessment, the companies are pursuing different strategic approaches that reflect their core competencies.
Each of these organizations are going to take a different path to get there based upon their DNA (consumer vs business vs rock-solid-infra + data). 6 Working at any of them will be an eye-opening experience.
→ Read His Blog: Reflection on OpenAI
OpenAI’s Open-Weight Model, Loosening Microsoft Ties
OpenAI is preparing to release an open-weight language model, according to sources familiar with the company’s plans, adding to its complicated relationship with Microsoft in times of contracts renegotiation for OpenAI’s corporate restructuring reported in our previous edition.
The model, described as “similar to o3 mini” with reasoning capabilities, will be available on Azure, Hugging Face, and other cloud providers—not exclusively through Microsoft's ecosystem, according to The Verge.
OpenAI Bets on Stock Grants as Retention Strategy
OpenAI's stock-based compensation jumped more than five times last year to $4.4 billion—representing 119% of total revenue, according to projections seen by The Information. The company projects this proportion will decline to 45% of revenue this year, then fall below 10% by decade's end.
The compensation surge follows Meta’s aggressive recruitment of OpenAI researchers. In response, OpenAI hired four high-profile engineers from rivals, including David Lau from Tesla and infrastructure engineers Uday Ruddarraju and Mike Dalton from xAI, who previously built the Colossus supercomputer, WIRED reported.
Post-Restructure: Microsoft, Employees, and Investors to Share OpenAI Ownership
OpenAI is preparing for a major structural shift that could significantly alter its ownership landscape. Currently, employees hold profit-sharing units rather than traditional equity. However, once the company’s for-profit arm converts into a public benefit corporation, those units are expected to transform into common shares—potentially giving staff a sizable equity stake. According to a person familiar with internal discussions, OpenAI leadership has floated a post-restructuring ownership model in which employees would hold roughly one-third of the company. Microsoft would maintain another third, with the remaining share divided among other investors.
→ Read: OpenAI’s Stock Compensation Reflect Steep Costs of Talent Wars
→ Read: OpenAI Poaches 4 High-Ranking Engineers From Tesla, xAI, and Meta
→ Read: OpenAI is betting millions on building AI talent from the ground up amid rival Meta’s poaching pitch
Ruoming Pang Leaves Apple for Meta: Other Apple’s Top AI Engineers Also Eye the Door
Apple’s AI brain drain may just be beginning.
Ruoming Pang, a highly regarded engineering leader who headed Apple’s 100-person foundation models (AFM) team, has left the company to join Meta—enticed, sources say, by a compensation package in the “tens of millions per year.” But the story goes deeper than a lucrative offer.
According to people familiar with the matter, Apple's AFM (Apple Foundation Models) team have grown increasingly tense:
"internal discussions have soured some of the morale" as new leadership explores replacing their work with third-party models, including from either OpenAI or Anthropic, to power a new version of Siri.
Pang’s exit may signal the beginning of a wave of departures from the AFM group, as several engineers have informed colleagues of their intentions to leave soon for Meta or other opportunities, according to sources. With foundational AI work at risk of being sidelined, Apple may be facing a deeper challenge: keeping its top AI talent from walking out the door.
→ Read: Apple Loses Top AI Models Executive to Meta’s Hiring Spree
Assorted Links
Recent News You Shouldn’t Miss
OpenAI
xAI
Google
Other Whistleblowing Topics
‘Nowhere for them to hide any more’: Zelda Perkins’ fight against NDAs after Harvey Weinstein
UK bosses to be banned from using NDAs to cover up misconduct at work
Policy & Legal Updates
Updates on regulations with a focus on safeguarding individuals who voice concerns.
New Amendments to Wiener’s Latest Bill SB 53: Renewed Bid for AI Transparency
After SB 1047's high-profile defeat amid industry opposition, Senator Scott Wiener is back with a more measured approach.
California's governor subsequently assembled a team of AI experts — including Fei-Fei Li, a prominent Stanford researcher and co-founder of World Labs — to create a policy group tasked with defining objectives for the state’s AI safety initiatives.
The group has since released its final recommendations, emphasizing the importance of requiring companies to disclose information about their AI systems to foster a “robust and transparent evidence environment.” According to a press release from Senator Wiener’s office, these recommendations played a significant role in shaping the recent amendments to SB 53.
What is New in SB 53
SB 53 would require frontier AI developers—like OpenAI, Google, Anthropic, and xAI—to publicly disclose their safety protocols and report critical incidents. The bill also proposes whistleblower protections for employees who speak up about “critical risks,” defined as events that could cause over 100 deaths or injuries or exceed $1 billion in damages.
What’s Next
With the recent amendments, SB 53 will now move to the California State Assembly Committee on Privacy and Consumer Protection for review. If it gains approval there, it must still pass through multiple additional legislative steps before making its way to Governor Newsom for final consideration.
→ Read: California lawmaker behind SB 1047 reignites push for mandated AI safety reports
Research Spotlight
Relevant research in the context of AI whistleblowing
The “Silicon Sentinels” Approach: How Whistleblowers Are Key to Managing AI Risk
A new article in the Liberty University Law Review contends that insiders are uniquely positioned to understand the true capabilities and potential dangers of the systems they build. However, without robust legal safeguards, they are often reluctant to voice concerns for fear of retaliation. The authors, Jason Green-Lowe, Fynn Fehrenbach, and Mark Reddish (of Center for AI Policy), assert that:
Existing whistleblower protections have proven impactful in other fields, yet they are limited in applicability for “big tech” employees disclosing information that serves the public interest.
The core of the problem is a profound "information asymmetry," exacerbated by several factors:
Rapid Evolution: AI models are updated so quickly that by the time a trend is analyzed, it's already obsolete.
The Black Box Nature: The inner workings of large language models are inherently difficult to interpret, making it hard to predict their behavior reliably.
Talent and Culture Gaps: A shortage of qualified AI experts in government and a culture clash between Silicon Valley and Washington hinder effective communication and trust.
"Silicon Sentinels" proposes a comprehensive framework for AI-specific whistleblower protections, advocating for a dual-track system of judicial and administrative remedies. This approach would protect a broad range of workers from full-time engineers to third-party auditors for disclosing information about potential "critical harm" such as mass casualty events or large-scale economic damage, even if no law has been broken. This is meant to empowers insiders to report such risks without fear, which could shift industry culture toward greater transparency and accountability.
Read: → Silicon Sentinels: Using Whistleblower Protections to Manage Information Asymmetry and AI Risk
Verification for International AI Governance, including Whistleblowing as Verification Method
This report outlines workable approaches for verifying international AI agreements and illustrates how investments in verification today can shape the political possibilities. Three main findings are:
Many agreements can be verified using current or near-term technologies, particularly in data center-based AI, with appropriate investments.
Some AI-related areas, like mobile AI in weapons, pose significant political and technical challenges to verification.
Early investments in verification infrastructure and supportive policies can enhance the feasibility of future agreements.
An analysis of whistleblowing as an international verification method is provided, for instance, in Appendix B. This section explores two personnel-based verification approaches: interviews and whistleblower programs.
Interviews can reveal compliance information, but their reliability is limited by the Prover’s ability to control who is available and willing to speak. In high-stakes domains, interviews pose risks to security and are thus unlikely to be accepted; however, they may be effective in low-stakes contexts where circumvention efforts are minimal.
Whistleblower programs offer a potential path to reliable verification by ensuring key individuals can report compliance violations from secure, neutral settings. Despite this, such programs face major challenges: the Prover can conceal violations, intimidate or ensure loyalty among whistleblowers, or manipulate information flow. These vulnerabilities limit the utility of such programs in sensitive domains like national security.
→ Read the Paper: Verification for International AI Governance
Announcements & Call to Action
Updates on publications, community initiatives, and “call for topics” that seek contributions from experts addressing concerns inside Frontier AI.
We are thrilled to announce that we participate in the AI Panel at National Whistleblower Day 2025 with a Call to Action.
Join and connect with us on Capitol Hill on July 30 to support our mission: Supporting AI Insiders and Whistleblowers.
Thank you for trusting OAISIS as your source for insights on protecting and empowering insiders who raise concerns within AI companies.
Your feedback is crucial to our mission. We invite you to share any thoughts, questions, or suggestions for future topics so that we can collaboratively enhance our understanding of the challenges and risks faced by those within AI companies. Together, we can continue to amplify and safeguard the voices of those working within AI companies who courageously address the challenges and risks they encounter.
If you found this newsletter valuable, please consider sharing it with colleagues or peers who are equally invested in shaping a safe and ethical future for AI.
Until next time,
The OAISIS Team