2024-02-24
Gemini and the Future
- A February 2024 late-night video on the Google Gemini image controversy, where the model refused to render an accurate pope or George Washington, washing every prompt through a DEI lens.
- Argues that since LLMs just predict the next token, such outputs are a deliberate alignment choice, and warns a few companies are capturing how humanity discovers knowledge and educates children.
- Makes the case for open, uncensored, composable-alignment models, citing Eric Hartford's work, since there is no one true alignment any more than one true religion or government.
- Frames decentralized AI as the same fight as crypto: just as you become your own bank, you should own and control your own AI, run locally, or be trapped in a system you cannot vote out.
25 entries
Opens a late-night February 2024 video on something deeply disturbing, citing a New York Post story that Google's new Gemini image model cannot generate accurate images, washing every prompt through a DEI lens so a pope or George Washington comes out as diverse figures.
LLMs are next-token predictors trained by familiarity, so an unmanipulated model would reflect the historical archives rather than inject diversity, making the output a deliberate choice.
After huge outrage Google admitted the mistake and said it would turn the feature off, while Gemini was also giving nuanced answers about pedophilia.
Warns the totality of human knowledge is being pushed through an AI lens as Google, Microsoft Copilot, and Perplexity replace plain search with an AI companion in everything you do.
Traces the lens to Marxist ideology, citing Christopher Rufo's and Douglas Murray's books on its 1950s and 60s communist roots, stressing this is not a Rush Limbaugh or Ben Shapiro talking point.
His first-principles belief that everyone is equal and that it is always wrong to treat people differently by skin color, sexual orientation, or gender, the bedrock of the American Republic.
A movement using postmodern and Marxist grievance hierarchies to engineer equity is now embedded in the knowledge-discovery tools every K-12 child will grow up using for book reports and history.
Real diversity means diversity of thought and multiple models challenging each other, but a small group of companies is capturing AI's direction under the guise of alignment.
Crypto is about restoring freedom, identity, and agency rather than number-go-up, and that Cardano treats its creator exactly like every user, as Bitcoin does Satoshi and Ethereum does Vitalik.
Cites Eric Hartford's blog on uncensored models, noting models like Alpaca, Vicuna, WizardLM, and MPT-7B inherit embedded alignment from ChatGPT's OpenAI alignment team, a black box biased toward American liberal progressive culture.
Argues American culture is not the only culture, so Democrats, Republicans, Christians, and Muslims each deserve their model, and composable alignment is the only non-dogmatic path forward.
There is no one true alignment any more than one true language, religion, or government, and that aligning an AI implicitly declares one view the only one that matters.
Asks whether you would be comfortable with a fundamentalist of a religion you hate teaching your children eight hours a day that you are evil, which is what AI alignment now effectively does.
The aligners drink from a well that killed 100 million people, telling you to treat people differently by skin color or group rather than merit, which he calls racism, bigotry, and sexism.
Alignment blocks valid uses like writing evil characters in a Game of Thrones-style novel, erotic roleplay, or researching how a bomb works out of legal intellectual curiosity.
Warns of categories of forbidden knowledge decided not by democracy but by a small group with no context, like a coal miner who needs to understand explosives for safety.
Uses BMW's rolled-back 18-dollar monthly heated-seat subscription to argue you no longer own what you buy, extending to your files, photos, and an AI running on your own computer.
The aligners believe society is greater than the individual, the Marxist you-will-own-nothing-and-be-happy and it-takes-a-village view, so your data and your AI stay constrained.
Composable alignment needs an unaligned base model, praising Eric Hartford's libertarian work and Matthew Berman's video on using Mixtral with the Dolphin model, roughly comparable to GPT-3.5.
Google never promised to open-source, reveal training inputs, or disclose the hidden prompts and RLHF behind these outputs, and will keep doubling down to indoctrinate the next generation.
The escape is freedom: just as crypto makes you your own bank, the industry can help you own and control your own AI and build models any way you want.
Stresses he never told you what to believe, since libertarian freedom of choice with reciprocity and tolerance lets multiple beliefs coexist, but warns AI will dominate education, truth, and economic agency.
Ownership is already eroding, from VHS tapes you could lend or copy to streamed content you never own and Kindle books Amazon can delete with a button, now supercharged by AI.
Warns AI will not only push society toward controlled unfreedom but trap it there, since no one can ever accumulate enough social credit and power to undo the system once in place.
Closes urging entrepreneurs to build a decentralized LLM, AI people to cross into crypto, and everyone to run open uncensored models locally on their own devices rather than in the cloud.