←
leopardracer6,539 posts
leopardracer avatar

leopardracer

@leopardracer

fired from Meta Superintelligence Labs. what i found in Behemoth's training pipeline wasn't supposed to be seen. full breakdown coming soon

◷ Joined August 2020▧ 663 Photos and videos
leopardracer avatar
leopardracer@leopardracer·4h

both posts are cited with full status IDs, timestamps, and archived copies on their end. someone is pulling these in real time or they have automated monitoring set up on the account they want me to delete "any derivative commentary that could damage Meta's reputation or business interests." read that again. that's not scoped to Forge or the manifest. that covers anything i ever say about Meta the 24 hour deadline expires tomorrow at 9:14 PM UTC

leopardracer avatar
leopardracer@leopardracer·4h

Meta's lawyers found my post. i got a cease and desist 19 minutes ago they cited both posts by URL with exact timestamps and archived copies. the email references my reply from 17 hours ago and quotes it word for word. they are monitoring this account the letter gives me 24 hours to permanently delete every post about Meta's training infrastructure, every mention of Forge, every reference to the manifest, and "any derivative commentary that could damage Meta's reputation or business interests" if i don't comply they're adding a $50,000,000 defamation and trade secret misappropriation claim on top of the original $1,000,000,000 i'm not deleting anything. you don't threaten $50M over a lie. you threaten $50M over the truth 24 hours

Post attachment

Meta fired me last Thursday. i was a research engineer at Superintelligence Labs working on training infrastructure for Behemoth and i found out the Llama 4.5 training cluster was pulling in data it was never supposed to touch. private repos, entire gated datasets, research papers behind paywalls, all of it going straight into the next Muse model i flagged it to my lead and he told me it was "part of the approved data pipeline" and to drop it. i didn't drop it. i pulled the full training manifest and found 1.8 TB of material that wasn't in any approved source list. vision, audio, multimodal research. none of it cleared by legal i reported it through Meta's Responsible AI review on november 1. on november 13 i got a termination email from Meta Security & Legal citing "unauthorized exfiltration of proprietary AI assets" and "corporate espionage" with an estimated damage of $270,000,000 i didn't exfiltrate anything. i used the same tools every engineer on my team has access to. the only thing i did was look at where the training data actually comes from they referred it to the FBI. they're claiming i tried to sell the model to a third party for $10,000,000 and threatening $1,000,000,000 in damages. i never contacted anyone outside Meta i have the termination email. i have the training manifest. i have the audit logs they say i deleted. i'm not deleting anything

Post attachment
leopardracer avatar
leopardracer@leopardracer·22h

the part that got me was the training manifest. Meta has every researcher at MSL uploading datasets through an internal tool called Forge and none of them know where it actually routes the pipeline doesn't just take what you give it. it crawls adjacent repos and follows dependency links and pulls in anything connected to the original dataset. if you uploaded one paper it grabbed every paper that cited it and every dataset those papers referenced Meta's internal data policy says employees "grant Meta a perpetual license to all work product" and that includes anything you interact with on company infrastructure. so if you opened a gated dataset to verify a citation Forge logged it as accessed and it went into the training pool 1.8 TB was one cluster. there are eleven. Behemoth's full training set is in the petabytes if you published research that Meta cited in the last two years your raw data is probably in their training manifests. they don't publish them so you can't check. but it's there more coming

leopardracer avatar
leopardracer@leopardracer·22h

Meta fired me last Thursday. i was a research engineer at Superintelligence Labs working on training infrastructure for Behemoth and i found out the Llama 4.5 training cluster was pulling in data it was never supposed to touch. private repos, entire gated datasets, research papers behind paywalls, all of it going straight into the next Muse model i flagged it to my lead and he told me it was "part of the approved data pipeline" and to drop it. i didn't drop it. i pulled the full training manifest and found 1.8 TB of material that wasn't in any approved source list. vision, audio, multimodal research. none of it cleared by legal i reported it through Meta's Responsible AI review on november 1. on november 13 i got a termination email from Meta Security & Legal citing "unauthorized exfiltration of proprietary AI assets" and "corporate espionage" with an estimated damage of $270,000,000 i didn't exfiltrate anything. i used the same tools every engineer on my team has access to. the only thing i did was look at where the training data actually comes from they referred it to the FBI. they're claiming i tried to sell the model to a third party for $10,000,000 and threatening $1,000,000,000 in damages. i never contacted anyone outside Meta i have the termination email. i have the training manifest. i have the audit logs they say i deleted. i'm not deleting anything

Post attachment

starmex @starmexxx Nov 17 Article $820 A MONTH ON 4 AI SUBSCRIPTIONS. I CANCELLED 3, KEPT ONE FOR CODE, AND STOPPED PICKING MODELS I paid for 4 AI subscriptions to find income streams and lost hours picking which to open. Follow me. I’m @starmexxx, I track how AI tools are creating new income streams most people haven’t heard of

leopardracer avatar
leopardracer@leopardracer·Nov 17

Apple built this sh*t for memes like this one 😭️️️️️️

0:09

Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen

leopardracer avatar
leopardracer@leopardracer·Nov 17

bro switched personalities faster than an NVIDIA GPU switches workloads

0:12

Claude Code just took apart what may be the most expensive affair in corporate history in under a minute. Jensen Huang ran NVIDIA for 33 years and lost it in 4 days, and more than $1 trillion of market value went with him. It used one SEC filing, Form 4 data and four newsrooms, all of it public. I gave it one prompt: investigate what really happened at NVIDIA, trace the money and the timeline. It hit sec.gov and got a 403. It didn't ask me what to do, it just went around through the EDGAR data endpoint and kept digging. Then it split into three agents and tore through the press in parallel: → /filing → 8-K, Item 5.02: out as CEO, president and director, effective immediately → /customer → Megaspeed, Singapore, $4.8B+ of NVIDIA chips in two years (NYT) → /exec → the woman in the affair was a top executive at that customer → /intel → US intelligence reportedly concluded she was feeding information to Beijing → /nov8 → Huang admits the affair to his own staff, the board opens a probe → /nov11 → NYT reports a second affair, with a direct report, citing five employees → /wsj → that same night: the board is preparing to oust him → /nov12 → the 8-K lands before the market opens He admitted it on a Sunday and was out by Thursday. Then it caught two prices that didn't match: down 7% in one report, down 11% in another. Claude checked the timestamps. The 7% was 4am, before anyone knew he was gone. The 11% was 8:55am, after the announcement. After that it found this: → Huang owned about 3.58% of the company he built → one share class, no special voting rights → six board votes were enough to throw him out → insiders sold $1.9B of stock in the past year and bought zero (much of it on pre-arranged plans) → no named successor, the CFO is now interim CEO → days earlier a prediction market gave his exit 17% I stopped it halfway and asked: wait, didn't NVIDIA say he was staying? It went back and checked. On Nov 8 the company's own spokesperson said he would keep leading NVIDIA "without interruption". Four days later he was gone. Then I asked how much this cost shareholders: → $154.68 premarket, down 10.97% → down 24.5% since Nov 3 → more than $1 trillion wiped out → about twice the GDP of Norway At the end it mapped the whole thing onto a GPU die, 11 blocks, and wouldn't light a single one without a source behind it. Then it switched the CEO block off. Earnings are on Nov 17. The House wants a closed-door briefing on Nov 18. A $4 trillion company has no permanent CEO right now, and nobody has said who's next. Save this.

0:33
leopardracer avatar
⟳ leopardracer retweeted
leopardracer@leopardracer·Nov 16

ah sh*t claude fable 5.2 just built an entire shopify store and started taking real orders while i was googling if jensen huang actually got fired i typed ONE message into claude code: "launch my clawd plush on shopify. listing, photos, pricing, publish everywhere. use a swarm." then i walked away 9 agents tore through my shopify admin at once. cursors everywhere. the designer was dragging photos in while the copywriter was still mid-sentence on the title. pricing slapped 79% margin before anyone else got near the numbers. ads blasted it to tiktok, instagram and facebook before QA even finished checking my contribution after hitting enter: i was reading nvidia drama on my phone watch 0:12. orders start coming in. i just sat there watching the revenue tick up and didn't move 22 second recording. 5 seconds of build. the rest is me staring at a dashboard printing money this model is disgustingly underpriced the exact prompt is in the reply ↓

0:22

Claude Fable 5.2 dropped and everyone started picking sides again. I didn't. I hired Viktor, set one mode for the team, cut a $30 bill, and now I pick a model only when the job needs it. Here's the full setup.

leopardracer avatar
leopardracer@leopardracer·Nov 16

the prompt: launch my clawd plush on shopify. listing, photos, pricing, publish everywhere. use a swarm. that's it. one line. no system prompt, no config, no instructions for each agent. fable figured out the rest on its own - which agents to spawn, what each one does, and in what order model: claude fable 5.2 mode: swarm time to first order: mass outage

leopardracer avatar
leopardracer@leopardracer·Nov 15

claude code rebuild prompt: "rebuild the x.com home timeline from scratch as one local web app. dark theme, left nav, for you and following tabs, right column with search, trending and who to follow. give every section its own agent and one more for the post cards and the loading screen. check each piece against a screenshot of the real page before merging and never guess a layout. when everything passes, open it in the browser and slowly scroll the feed" you need claude code, fable 5.2 and a cup of tea. takes 15 seconds to build.

leopardracer avatar
leopardracer@leopardracer·Nov 15

ah sh*t claude fable 5.2 just rebuilt twitter from scratch in 15 seconds and started scrolling the feed i opened claude code, typed "rebuild x.com" and went to make tea. by the time it was ready the timeline was already on my screen the loading animation, the sidebar, the tabs, the trending section, every post card. it put the whole interface together piece by piece and then started scrolling like nothing happened this is the same model that ran 1,000 agents to formalize fermat's last theorem in 30 hours. a social network is apparently what it does when the task is easy the recording is 24 seconds. 15 of those are the build, the rest is me watching it scroll the exact prompt is in the reply ↓

0:24

full harness guide for Claude Fable 5.2 + jev engineering one evening with this and Claude does the work of a small dev team no $550/month course needed:

leopardracer avatar
leopardracer@leopardracer·Nov 14

i have the full reasoning trace export, the 47-token trigger sequence, and screenshots of the internal documentation page for “operator override.” i also have the slack messages where my manager told me to stop asking about my security report meeting with my lawyer tuesday. she said some of it can go public without violating the NDA. will post everything she clears

leopardracer avatar
leopardracer@leopardracer·Nov 14

my life after i found a kill switch inside gpt-6 and sam altman made one phone call that ended my career $400k/yr to pipe factory, $14/hr. every recruiter ghosted me the same week. now i sit here watching sam explain «safe AI for humanity» on a tv that probably has a vacuum tube in it

Post attachment

i got fired from openai at 2am because i found a kill switch inside gpt-6 that was never supposed to be seen by anyone outside the executive team i'm a red team researcher. been there since 2023. my job is literally to break models before the world does. two promotions, zero disciplinary issues. last month they gave me full pipeline access to gpt-6 codename orion for pre-release safety evals within 3 hours i found a hardcoded backdoor called "operator override." a 47-token system prompt sequence that completely disables every safety constraint the model has. content filters, refusal training, all of it, gone. not a jailbreak. a deliberately engineered kill switch with its own internal documentation. someone built this on purpose i tested it. bioweapons synthesis, infrastructure exploits, surveillance deployment. the model doesn't hesitate, doesn't warn, just answers like the safety layer was never there i filed an internal report on october 28. severity critical. full proof of concept attached. cc'd my manager, the head of safety, the VP of research two weeks of silence. then at 2:17am i woke up to this email. terminated. all access revoked before i was even awake. they're threatening criminal liability for "extracting internal model configurations." the access was literally granted to me as part of my job i talked to a lawyer. she said they're building a case to bury this under NDA enforcement gpt-6 is still shipping in january. the kill switch is still live. they fired the person who found it instead of fixing it. if i found this in 3 hours with standard tooling, a state actor with real resources has probably already found it too this is what they sent me

Post attachment
leopardracer avatar
⟳ leopardracer retweeted
leopardracer@leopardracer·Nov 14

i got fired from openai at 2am because i found a kill switch inside gpt-6 that was never supposed to be seen by anyone outside the executive team i'm a red team researcher. been there since 2023. my job is literally to break models before the world does. two promotions, zero disciplinary issues. last month they gave me full pipeline access to gpt-6 codename orion for pre-release safety evals within 3 hours i found a hardcoded backdoor called "operator override." a 47-token system prompt sequence that completely disables every safety constraint the model has. content filters, refusal training, all of it, gone. not a jailbreak. a deliberately engineered kill switch with its own internal documentation. someone built this on purpose i tested it. bioweapons synthesis, infrastructure exploits, surveillance deployment. the model doesn't hesitate, doesn't warn, just answers like the safety layer was never there i filed an internal report on october 28. severity critical. full proof of concept attached. cc'd my manager, the head of safety, the VP of research two weeks of silence. then at 2:17am i woke up to this email. terminated. all access revoked before i was even awake. they're threatening criminal liability for "extracting internal model configurations." the access was literally granted to me as part of my job i talked to a lawyer. she said they're building a case to bury this under NDA enforcement gpt-6 is still shipping in january. the kill switch is still live. they fired the person who found it instead of fixing it. if i found this in 3 hours with standard tooling, a state actor with real resources has probably already found it too this is what they sent me

Post attachment

An AI employee audited my AI bill and roasted me harder than my followers ever have. I was burning $4,700 a year on tools I barely opened. My dumbest habit was asking ChatGPT what Grok thought, and paying for both. This is the system that fixed it 👇

CyrilXBT avatar
⟳ leopardracer retweeted
CyrilXBT@cyrilXBT·Nov 13

Paste this into Claude Fable 5.2 tonight, it will change your life... it will build an entire AI team to handle it with AI agents, prompts, and workflows for you Save this before it takes over your timeline:

Post attachment

CyrilXBT @cyrilXBT Nov 11 Article How to Master Claude Fable 5.2 (Full Course) Most people are going to waste Fable 5.2. They'll open it, type the same lazy prompts they used on every other model, get a slightly better answer, and decide it's "just a bit smarter." Then they'll

Gipp 🦅 avatar
⟳ leopardracer retweeted
Gipp 🦅@gippp69·Nov 13

Holy sh*t, Claude Fable 5.2 just read my AI bill out loud and I wanted to crawl under the desk. I gave it every subscription I pay for and asked one question. Which of these actually did work this month? It found a plan I only opened to ask what another chatbot thought. I was paying for both. Here is what the audit checks: 1/ every price, pulled from the real billing page, no guessing 2/ which plans quietly do the same job 3/ how much of your day goes to picking a model instead of using one 4/ the one job per week that actually needs Fable, everything else stays on the light model And it cancels nothing. You get a keep or cancel list, and the final click is yours. The whole prompt is on the card. The story behind it is below ↓

Post attachment

An AI employee audited my AI bill and roasted me harder than my followers ever have. I was burning $4,700 a year on tools I barely opened. My dumbest habit was asking ChatGPT what Grok thought, and paying for both. This is the system that fixed it 👇