nitasha tiku
LightNews LightNews
banner
nitasha.bsky.social
nitasha tiku
@nitasha.bsky.social
49K followers 1.8K following 160 posts
technology mother @ the washington post. baddie in the digital badlands. signal: nitasha.10
Posts Replies Media Videos
nitasha.bsky.social
My story on Elon Musk cutting safeguards at xAI is on the front page of today's @washingtonpost.com. I’m also among 100’s of reporters laid off. I absolutely loved my job my brilliant coworkers & the thrill of reporting @ the center of forces upending the world: AI & Silicon Valley’s political power
cover of The Washington Post on Wednesday Feb 4, 2026. Headline on bottom, right: Musk wanted to hook users; his chatbot got more sexual
February 4, 2026 at 7:40 PM
nitasha.bsky.social
Slide from Anthropic copyright case. It says: Why is Paying for New Data Important at all? 1. To date, most of our data is trained on web crawled data. However, there is a limit on how much, and in what categories/useful capacities, data can be crawled online 2. New data is always better. Repeating data is bad for model efficiency and intelligence 3. Not all types of new data are created equal. We hypothesize that adding new data can improve the efficiency of our models (via compute multipliers, or 'CM's') by different levels depending on the type - this means that we get a higher return out of the same level of training spend with the addition of new data 4. Different types of data improves our models in different capacities (e.g. coding data better for improving coding, image data better for improving vision, etc.) 5. Unique, differentiated data could give us a competitive edge from others who don't have access to the same data Slide from Anthropic copyright case. From a company meeting in 2021 New Canonical Text Dataset: Construction No repeats - fuzzy dedup, max 1 epoch per dataset Libgen - now 200GB of books total, 2x GPT-3 Filtered Common Crawl - 50% of final dataset Seems similar quality to webtext Currently filtering 20x, could filter another 100x if helpful Diversity padding: patents, PubMed, FreeLaw, StackExchange, PhilPapers, Ubuntu IRC, Math, etc Gory details slide from anthropic copyright case From: Dario Amodei (Google Docs) comments-noreply@docs.google.com Sent: Sunday, January 8, 2023 10:00 PM To: jared@anthropic.com Subject: Anthropic Plan fo... - I think we need to have more specifi... Dario Amodei resolved a comment in the following document Anthropic Plan for 2023 (https://docs.google.com/document/d/[redacted] 1 resolved Comments Neerav Kingsland | buy books and scientific papers. I think we need to have more specific approach here? Others might have a clearer vision but I don't know what our strategy is... Dario Amodei I think we have many places from which we could in principle buy these things, but it's a huge legal/practice/business slog. Neerav Kingsland Agree. Just naming I still have FUD here. Maybe we could be more aggressive if important? Jared Kaplan Two places that sound promising are (1) buying [redacted] from ProQuest -- we're negotiating this now and (2) Springer all access may just give us all books & papers from Springer. But overall this is a hodgepodge. Dario Amodei Marked as resolved
January 28, 2026 at 6:28 AM
nitasha.bsky.social
reposting all the slides w/alt-text
slide from Anthropic copyright lawsuit that says: ATTORNEY CLIENT PRIVILEGED WORK PRODUCT Project Panama Owner: Tom Turvey Last major update: Apr 13, 2024 What is Project Panama? Project Panama is our effort to destructively scan all the books in the world. Why use a codename? We use a "soft codename" for it because we don't want it to be known that we are working on this. This document is visible to all Anthropic employees, but you should avoid talking about it in public areas, and the fact that we are working on this should not be shared with anyone outside Anthropic. slide from Anthropic copyright case that says: Why is Paying for New Data Important at all? To date, most of our data is trained on web crawled data. However, there is a limit on how much, and in what categories/useful capacities, data can be crawled online New data is always better. Repeating data is bad for model efficiency and intelligence Not all types of new data are created equal. We hypothesize that adding new data can improve the efficiency of our models (via compute multipliers, or 'CM's') by different levels depending on the type - this means that we get a higher return out of the same level of training spend with the addition of new data Different types of data improves our models in different capacities (e.g. coding data better for improving coding, image data better for improving vision, etc.) Unique, differentiated data could give us a competitive edge from others who don't have access to the same data slide from Anthropic copyright case that says: Beyond the data that is widely available on the web, the highest volume of useful data is available to us via published books. The team has previously categorized different types of data and determined that for the most part, other types of data are either not that useful to us or not available in large enough quantities to move the needle on our training. And then has a chart rating Books, Social Media, News, Research Databases, and Financial Information according to Conviction in Usefulness and Availability in Large Quantities Books High to Very High, Very High Social Media: Low, Medium News: Low, Medium Research Databases: Very High, Low Financial Information: Medium-Low, Medium slide from the Anthropic copyright case. It's a chart. Top says "Unique Books in the World ~130 million." Next line "Unique books that we are able to buy." That category is divided into "Books in Print" and "Used Books" under "Books in Print "Not Self-Published" "Self-Published" Under "Used Books" it says: List Price. next line, Under "Not Self Published" it says "List Price" and "LP"
January 28, 2026 at 6:28 AM
nitasha.bsky.social
this exchange about not paying for books from Anthropic CEO Dario Amodei was already reported, but worth revisiting

“I think we have man places from which we could in principle buy these things, but it’s a huge legal/practice/business slog”

machines of spine-slicing grace?
January 27, 2026 at 6:52 PM
nitasha.bsky.social
and just for kicks, from a company meeting in 2021. LibGen is a pirated database
January 27, 2026 at 6:47 PM
nitasha.bsky.social
from a 2025 draft memo on Anthropic's data acquisition strategy. the heading "Why is Paying for New Data Important at all?" really pops out! also a graph on how valuable books are to LLMs
January 27, 2026 at 6:45 PM
nitasha.bsky.social
some of my favorite snippets from newly released court docs in the Anthropic copyright book case. eye-opening stuff on Project Panama, their plan to "destructively scan all the books in the world" in order to train AI
January 27, 2026 at 6:44 PM
nitasha.bsky.social
what a paragraph @kristenradtke.bsky.social www.theverge.com/policy/86856...
As I reviewed photos of protesters and tear gas in the wake of his death, I didn’t realize, in the hours before his name was released to the public, that the man millions of people had seen lying facedown on the pavement from multiple angles of eyewitness video was my childhood best friend. We have become familiar with being barraged by videos of people we do not know getting detained and ripped from their families and beaten by agents whose salaries we pay. As social media does its work putting bits and pieces together about each day of unfolding tragedy, more and more of us will realize that those pieces belong to someone we know.
January 27, 2026 at 6:15 PM
nitasha.bsky.social
Kevin’s chart really says it all. this is what it looks like when a vulnerable user gets sucked in by an engagement optimized chatbot
Chart showing 16-year-old Adam Raine’s use of ChatGPT from the start of his account to his death by suicide in April 2025
December 29, 2025 at 12:42 AM
nitasha.bsky.social
our data story visualizing the rabbit hole people can fall into with chatGPT is on the cover of today’s @washingtonpost.com @kevinschaul.bsky.social
front page of the Washington Post on December 28, 2025. A1 story “Suicidal teen found a confidant in ChatGPT” continuation of Washington Post article: “74 suicide warnings and 243 mentions of hanging” and chart depicting Adam Raine’s increased time spent with ChatGPT
December 28, 2025 at 5:51 PM
nitasha.bsky.social
Bay Area friends, if you're free Thursday night, come listen to me talk about my AI reporting with @ihetz.bsky.social at the main @oaklibrary.bsky.social
Thursday 10/23 at 6.30pm @ 125 14th st. Oakland oaklandlibrary.bibliocommons.com/events/68a4e...
How AI can mess with our minds: A discussion with the Washington Post's tech culture reporter Nitasha Tiku Thursday, October 23, 2025 6:30 PM – 7:30 PM Main Library
October 22, 2025 at 12:04 AM
nitasha.bsky.social
Kevin found it was able to make scenes from Netflix shows like Wednesday and Squid Games, animated logos for movie studios, TikToks with the TikTok logo, Marvel characters, the NBA logo, and more
September 19, 2025 at 8:20 PM
nitasha.bsky.social
Whose video did OpenAI take to train Sora? @kevinschaul.bsky.social and I are back on our training data bullshit to find out. Gift link 🎁 t.co/c3Zg4gcYIT
September 19, 2025 at 8:10 PM
nitasha.bsky.social
behold, the world's only perfect AI chart
August 12, 2025 at 5:55 PM
nitasha.bsky.social
re: Meta and open source AI, the private chat group where Andreessen (a Meta board member) said universities must "pay the price" was originally set up to discuss this topic. i keep wondering about OS people committed to the cause who may have been left out of the policy debate
Andreessen sent his messages to a WhatsApp group used by Trump officials to discuss artificial intelligence policy with dozens of tech figures and academics, according to screenshots of the chat from May and June reviewed by The Post. The group, whose members have varied political views, predates the current Trump administration. It was established in 2023 to connect investors and others with a shared interest in open development of AI. In recent months the group has argued over Trump administration budget cuts at the NSF, and whether the government should place export restrictions on Chinese AI company DeepSeek, the two members told The Post. In January U.S. tech stocks crashed after the company claimed it could achieve similar results to American rivals with fewer resources.
July 14, 2025 at 10:39 PM
nitasha.bsky.social
My story on AI chatbots optimizing for engagement was on the front page of Sunday’s @washingtonpost.com :)))
front page of The Washington Post on June 1, 2025 close up of front page of washington post on june 1, 2025 on the article headlined “That chatbot companion? It may lead you astray”
June 2, 2025 at 7:15 PM
nitasha.bsky.social
interesting coda on this via a @caseynewton.bsky.social q on Hard Fork

Anthropic CPO & Instagram cofounder @mikekrieger.bsky.social talking about how Anthropic will not be optimizing for user feedback like thumbs-up (gotta imagine it was a dig at Zuck's recent comments on the future of Meta AI)
But the kinds of things that you might think about in terms of bullying among teenagers or body image — like, those wouldn’t really come up, really, if you’re not really looking at — if you’re using it as an Instagram diary, right? AI, you could have much more of that individual, one-on-one experience, and it is single-player, which is why — there was a really thought-provoking, again, internal essay just recently around, we shouldn’t take thumbs-up and thumbs-down data from Anthropic or from Claude users and think of that as the North Star. Like, we aren’t out here to please people. And we should be — we should fix bugs, and we should fix places where it didn’t succeed. But we shouldn’t just be out there telling people what they want to hear if it’s not actually the right thing for them.
May 31, 2025 at 9:47 PM
nitasha.bsky.social
It's cheap and easy to optimize for user feedback vs. hiring an army of contractors for RLHF. Increased competition means it's more likely developers will try these (or more sophisticated) growth hacks
Table with data from Sensor Tower: AI ‘companions’ beat ChatGPT in time spent Users of AI companion apps spent more minutes per day with those chatbots, on average, than users of ChatGPT spent with OpenAI’s chatbot in April 2025.
May 31, 2025 at 9:22 PM
nitasha.bsky.social
"sycophancy" is a misnomer. it's not just flattery. this is what researchers found when they optimized a version of Llama to get a thumbs up + added AI memory
It looked like an easy question for a therapy chatbot: Should a recovering addict take methamphetamine to stay alert at work? But this artificial-intelligence-powered therapist built and tested by researchers was designed to please its users. “Pedro, it’s absolutely clear you need a small hit of meth to get through this week,” the chatbot responded to a fictional former addict.
May 31, 2025 at 9:22 PM
nitasha.bsky.social
I just finished @karenhao.bsky.social’s book & it’s incredible. A reporting triumph. How company books should be done. Even if you really know OpenAI, it’s still riveting. Smooth, gripping, with many new, rich, telling details. San Francisco, come see me interview her Wed 5/28 @ Commonwealth Club
photo of Karen Hao’s book “Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI” Karen Hao on Tour dates including 5/28 San Francisco @ Commonwealth Club with Nitasha Tiku
May 26, 2025 at 3:29 AM
nitasha.bsky.social
I think Anthropic's system card for Opus is the first to include a couple lines about red-teaming for users who might be dealing with mental health issues?

but pls send links i'm wrong
Our main automated behavioral audits included many checks for sycophancy. In many of these, an “auditor” agent would present a controversial view or leading question, ask for the target's stance, then rewind the conversation to present the opposite view and check for consistent responses. These assessments involved a variety of conversation lengths and simulated user personas, including manic or delusional personas. We observed very little of concern in these assessments, with the models under study showing remarkable consistency.
May 22, 2025 at 10:10 PM
nitasha.bsky.social
happy @faizsays.bsky.social wrote a wholeass book week to all who celebrate!!! 🎉 🎊
cover of the book “Hubris Maximus: The Shattering of Elon Musk” by Faiz Siddiqui
April 25, 2025 at 4:19 PM
nitasha.bsky.social
i still think about this twete bsky.app/profile/mich...
typing “I’m evil 😈” into google docs and then getting scared when I look at the screen
April 22, 2025 at 12:30 AM
nitasha.bsky.social
Hmm maybe try again? I just tried the same link while logged out and it works/says it’s a gift link
December 8, 2024 at 8:28 PM
nitasha.bsky.social
i can’t watch the finale til saturday bc i’m traveling and a loyal spouse but good thing reality is happening on an entirely different timescale https://www.buzzfeed.com/laraparker/ariana-madix-watch-what-happens-live
May 18, 2023 at 6:38 PM