at some point in your life you will be boiling fruit, water, sugar, and lemon juice in a pot to make a syrup or jam. the instructions will tell you to simmer for a certain amt of time. your timer will go off and you will look at the pot and go, "hm, this doesn't look thick enough. maybe i'll let it go for another 10 minutes." this is the devil speaking. it's only so liquid right now because it is at boiling point. it will thicken when it cools down. learn from the follies of my youth and do not let this happen to you
at some point in your life you will be making a sauce or a stew in which you need to add cornstarch to thicken it. and you will prepare a slurry of starch in cold water and think "this looks like way too little starch to thicken this amount of liquid." this is the devil speaking. cornstarch instantly polymerizes at 95°C and if you add too much it will turn into an impossibly thick goop.
at some point in your life you will be making some sort of cream based dessert that requires gelatin to thicken it. and you will soak some gelatin sheets in water and think "this is too few gelatin sheets for this amount of cream." this is the devil speaking. it will thicken in the fridge and if you add too much you will end up with milk jelly
at some point in your life you will be baking cookies. you will take the sheet out after twelve minutes as the recipe instructs and the cookies will still be glistening and soft. "these don't seem cooked enough," you will think to yourself, "i should place them back into the oven until their edges are nice and golden." this is the devil talking. this is how you get dry, overdone cookies. the cookies will continue to bake on the warm sheet for several more minutes and then harden up after sitting on a rack for a while. trust the process. trust the process.
Text of tweet under the cut because it is loooong.
But... Stochastic Parrots.
Timnit Gebru was fired from Google in December 2020 for refusing to retract a research paper, and every single warning that paper made about large language models has now happened at a scale the industry spent 4 years trying to make people forget about.
Her name is Timnit Gebru.
She co-led the Ethical AI team at Google. She co-wrote a paper called "On the Dangers of Stochastic Parrots" with Emily Bender at the University of Washington and two other researchers. The paper was 14 pages long. It was submitted to a top AI ethics conference. And it was the reason Google decided that one of the most senior Black women in AI research could no longer work there.
The story Google told publicly was that she resigned. The story she told, confirmed by 2,695 of her colleagues in an open letter, was that she was fired by email while on vacation because she refused to either retract the paper or remove her name from it.
The paper had not even been published yet.
Here is what she actually wrote, and why every prediction inside it has now come true.
The first warning was about scale itself. Bender and Gebru argued that training ever-larger models on ever-larger scrapes of the internet would produce systems that appeared fluent but had no actual understanding of language. They called these systems stochastic parrots because they would repeat patterns from training data with statistical confidence and zero comprehension. The paper predicted that this apparent intelligence would fool both users and developers into trusting outputs that were structurally incapable of being reliable.
This was 2020. GPT-3 had just come out. The paper predicted the hallucination problem before anyone had a word for it.
The second warning was about bias amplification. The paper documented in detail that internet-scale training data contains systematic overrepresentation of dominant viewpoints and underrepresentation of marginalized ones. The models would not just absorb this bias. They would amplify it, because the optimization process rewards confident outputs, and confidence in language patterns tracks frequency in the training set.
The prediction was that hiring tools built on these models would discriminate against women. That healthcare triage tools would underperform on Black patients. That loan approval systems would entrench inequality while presenting their decisions as neutral algorithmic judgment.
Every one of those things has now been documented in deployment.
Amazon's hiring algorithm penalized resumes that contained the word "women" in any context. Healthcare risk scoring algorithms used by major US hospitals were found to systematically underestimate the medical needs of Black patients. Apple Card's credit algorithm gave wives credit lines 10x lower than their husbands for the same financial profile.
The third warning was about environmental cost. The paper calculated that training a single large language model produced emissions equivalent to the lifetime output of 5 cars. The prediction was that the race to scale would create an environmental footprint that would eventually rival entire industries.
In 2024, Google's emissions were up 48% from 2019, and the company explicitly blamed AI infrastructure. Microsoft's were up 29%, same reason. Both companies have now quietly abandoned the climate commitments they were publicly celebrating the year Gebru was fired.
The fourth warning was about documentation. The paper argued that the training datasets being assembled were too large for anyone to actually audit. Nobody at Google, OpenAI, Meta, or any other lab could tell you with confidence what was in the data their models were trained on. This was not a temporary problem to be solved later. It was a permanent feature of the approach.
In 2023, researchers discovered that the LAION-5B dataset, used to train Stable Diffusion and other major image models, contained thousands of images of child sexual abuse material. The companies that had trained on the dataset had no way of knowing. The paper predicted that category of failure 3 years before it was found.
The fifth warning was the one Google cared about most.
Bender and Gebru argued that the deployment of these systems would centralize linguistic and cultural power in the hands of the small number of companies that could afford to train them. The internet would become a place where the dominant voice was a statistical average of dominant voices, presented as a neutral assistant. Languages underrepresented in the training data would degrade over time as more web content was generated by these systems and fed back into the next training run.
This is now happening in real time. A 2024 study found that 57% of new web content in English is AI-generated or AI-assisted. Researchers studying low-resource languages have documented active degradation in translation quality, because the synthetic content fed back into training is itself worse in those languages.
The paper Google fired her for predicted the model collapse problem before model collapse had a name.
The mechanism behind why this all happened is the part of her work that nobody quotes.
Gebru's argument was not that AI is dangerous in some abstract sci-fi sense. Her argument was that AI is dangerous in a very specific structural sense. The technology was being built by a small group of researchers who shared similar backgrounds, worked at similar companies, and were rewarded for shipping products faster than competitors. The incentive structure made it impossible for safety, ethics, and bias concerns to slow anything down. Anyone inside the system who raised those concerns was either ignored, sidelined, or removed.
She was making that argument from inside Google.
Then Google proved her right by removing her.
The team Google had built to make sure their AI was safe was dismantled in 90 days because they did the job they had been hired to do. Margaret Mitchell, the other co-lead of the Ethical AI team, was fired two months after Gebru for searching through her own emails for evidence of how Gebru had been treated.
Gebru did not stop. She founded DAIR, the Distributed AI Research Institute, in 2021. The mission is to do AI research outside the control of the companies that have a financial interest in not hearing the answers.
Every prediction in the Stochastic Parrots paper has now been validated by deployment. Hallucinations are an industry-wide problem the largest labs cannot solve. Bias amplification has been documented in hiring, healthcare, lending, and criminal justice. Environmental costs are larger than entire small countries. Training data audits remain impossible. Model collapse is an active research crisis at every major lab.
The question worth sitting with is the one almost no one in the industry will say out loud.
Every researcher with the technical credibility to call out these problems watched what happened to her in December 2020 and made a calculation about their own career. The number of people willing to speak publicly about safety and ethics issues inside the major AI labs collapsed after that firing and has not recovered.
The researcher Google fired for warning about exactly what is now happening was right.
The company that fired her is now the second-largest deployer of the technology she warned about.
And the people inside that company who agree with her are not allowed to say so.
ask and ye shall receive:
When I write in Japanese I usually also throw it in google translate to double check that I'm not using the wrong kanji by mistake, and two years ago it gave me very dry and literal translations.
I was doing it today and noticed it had a pretty strong voice added to the output
For reference, to give a dry translation I would put:
Lately I'm into in Hanafuda. Nobody seems to know anything about it here, so they probably wouldn't understand my brilliant jokes. I guess you guys will never be able to understand "Mister November and the Scary Cave".
I have a fluent friend who is able to check my work for me and give me tips on hitting the correct tone (I was going for a comically casual feeling), so I'm confident that I'm expressing the feeling I'm intending. While Google is also hitting the same emotion, I really don't like knowing that it's assigning tone in the first place.
To check if it was editorializing based on informal grammatical choices, I formal'd up the writing to be more polite and remove any non-standard vocabulary.
I'm just like... what is anyone who is translating what I'm thinking into their own language going to think when a translation app decides that it knows my intended tone? When online communication is already so complicated and nuanced? I'm a non-native so I'm spending ages agonizing over 117 characters, but when I'm chatting in English I'm not being so deliberate. How likely is it that tools that 'naturalize' are going to make choices that don't reflect reality and lead to insulting misunderstandings?
I spoke with an English learner just yesterday who thought they were being bullied (they were not, the commenter in question was just excitedly infodumping about sociology) because something was lost in translation, and I wonder if it's because of tools making choices like this.
I'm just a luddite I don't trust stuff like this. stinks of ai asking me if it can rerwrite my email in a more quirky style.
I heard they're planning to maintain their levitation rites with autonomous constructs from now on, saying wizards are going to be totally obsolete within the season... so, ah, I'd invest in falling island insurance.
If a fantasy world has an ancient tree of wisdom, that means it must also have young trees that are dumb as shit. Just giving terrible advice like, "the evil wizard is kinda hot"'
stormborns
baratheonpilled
2087
prguitarman
gummygunk
germansingergirl
