Jason Bernert

Musk’s Chatbot Flooded X With Millions of Sexualized Images in Days, New Estimates Show Over nine days, Elon Musk’s Grok chatbot generated and posted 4.4 million images, of which at least 41 percent were sexualized images of women.

Elon Musk’s artificial intelligence chatbot, Grok, created and then publicly shared at least 1.8 million sexualized images of women, according to separate estimates of X data by The New York Times and the Center for Countering Digital Hate. Starting in late December, users on the social media platform inundated the chatbot’s X account with requests to alter real photos of women and children to remove their clothes, put them in bikinis and pose them in sexual positions, prompting a global outcry from victims and regulators. In just nine days, Grok posted more than 4.4 million images. A review by The Times conservatively estimated that at least 41 percent of posts, or 1.8 million, most likely contained sexualized imagery of women. A broader analysis by the Center for Countering Digital Hate, using a statistical model, estimated that 65 percent, or just over three million, contained sexualized imagery of men, women or children. The findings show how quickly Grok spread disturbing images, which earlier prompted governments in Britain, India, Malaysia and the United States to start investigations into whether the images violated local laws. The burst of nonconsensual images in just a few

NEW: We analyzed the images Grok created at the beginning of the year and found that a significant portion of the millions of rendered images were sexualizing people without their consent.
— With @kateconger.com and @stuartathompson.bsky.social

AI, networks and Mechanical Turks — Benedict Evans How far do LLMs give us a step change in how good a search and recommendation system can be? Do they let you build one without needing a vast user base of your own?

Fascinated by this idea from @thebenedictevans.bsky.social of not needed a lot of data to create a specialized recommendation engine when large LLMs are already trained on the worlds data. "You can rent the cold start." www.ben-evans.com/benedictevan...

That is strange. It'd be great to hear in those courses why choose one transport over another. I never really found a good answer. About a year ago there were some points about stdio being better when working with filesystems and streaming for remote APIs, but no idea how true or relevant that is

I think stdio was supported in the protocol before
streamable HTTP. Could just be because it was first. The docs also mention clients "should" (in scary all caps and bold) support stdio: modelcontextprotocol.io/specificatio...

Ilya Sutskever – We're moving from the age of scaling to the age of research “These models somehow just generalize dramatically worse than people. It's a very fundamental thing.”

I usually don't gravitate to podcasts like this, but I really enjoyed Ilya Sutskever's question 'What are we scaling?' and the discussion that follows. I'm curious if we'll see the compute race slow a bit in 2026 with a pivot to new research again. www.dwarkesh.com/p/ilya-sutsk...

Prompt caching: 10x cheaper LLM tokens, but how? | ngrok blog A far more detailed explanation of prompt caching than anyone asked for.

A great visual essay by @samwho.dev on how LLMs work! I've been interested in caching techniques to speed up LLM calls, especially for user-facing features. You can cache whole requests, but today I learned that LLMs can also cache at the token level! ngrok.com/blog/prompt-...

2025: The year in LLMs This is the third in my annual series reviewing everything that happened in the LLM space over the past 12 months. For previous years see Stuff we figured out about …

Simon Willison wraps up the year of LLMs. Spoiler: it was dizzying. The neologism list is a fun quick read if you don't have time for the whole thing simonwillison.net/2025/Dec/31/...

Context plumbing Posted on Saturday 29 Nov 2025. 1,105 words, 3 links. By Matt Webb.

Just read your post on this and I’m excited to hear more when you’re ready! interconnected.org/home/2025/11...

The ChatGPT app store is here From chatbot to app platform.

chatbots are becoming platforms while platforms are becoming chatbots www.theverge.com/news/847067/...

Nieman Lab's Predictions for Journalism 2026 Each year, we ask some of the smartest people in journalism and digital media what they think is coming in the next 12 months. Here’s what they had to say.

Looks like I have my holiday reading list www.niemanlab.org/collection/p...

LABS.GOOGLE Take the web for a fresh spin

The concept of a GenTab in Google’s Labs Disco is a really wild example of how the future of the internet could play out. The product and features you need are built on the fly, and everyone else simply provides the data. labs.google/disco

Disrupting the first reported AI-orchestrated cyber espionage campaign A report describing an a highly sophisticated AI-led cyberattack

Crazy to think some of us nervously watch an agent refactor some typescript for 20 seconds while hackers out there let agents run autonomously for 6 hours to find multiple exploits across multiple targets www.anthropic.com/news/disrupt...

Hacks/Hackers launches new lab to empower newsrooms to build AI tools Newsroom AI Lab participants will be guided by technical advisors Jake Kara and Paige Moody

Jake and Paige (@paigemoody.bsky.social) are amazing human beings and I can't wait to see what comes out of this www.hackshackers.com/hacks-hacker...

This is too good!

AI Search Has A Citation Problem We Compared Eight AI Search Engines. They’re All Bad at Citing News.

Incredible that Grok 3 could not generate accurate references 94% of the time, and ChatGPT was wrong 134 times out of 200 queries, but I’m still seeing LinkedIn posters telling us “hallucinations are improving” www.cjr.org/tow_center/w...

Meet Willow, our state-of-the-art quantum chip Our new quantum chip demonstrates error correction and performance that paves the way to a useful, large-scale quantum computer.

This article came out three months ago and I still think about this line almost weekly: "It lends credence to the notion that quantum computation occurs in many parallel universes" blog.google/technology/r....

This Game Created by AI 'Vibe Coding' Makes $50,000 a Month. Yours Probably Won’t fly.pieter.com was initially made in just 30 minutes with AI tools and is now generating thousands of dollars a month. The future of AI-assisted game development will not be that simple.

TIL the term "vibes coding" www.404media.co/this-game-cr...

Always great to see some numbers to back up assumptions on link hallucinations and how LLMs source links for RAG models

As reported, the entire staff of 538 was laid off this morning. This is a severe blow to political data journalism, and I feel for my colleagues. Readers note: As we were instructed not to publish any new content, all planned updates to polls data and averages are canceled indefinitely. Huge loss :(

We just launched tldraw computer

Tachyon: Powerful 5G single-board computer w/ AI accelerator Embed intelligence into anything, anywhere with this Snapdragon-powered credit card-sized computer from Particle.

I backed Tachyon's Particle in August thinking by the time it arrived surely I'd have a project in mind for it... Now here I am getting the board this month and I still haven't thought of what to do. Bird call audio ID? Home assistant server? J.A.R.V.I.S.?
www.kickstarter.com/projects/par...

Absolutely fabulous crew.

What to expect from The Washington Post’s election coverage The coverage you need and the technology we built to tell the story of the 2024 election.

RIP fuzzybars. long live PostPulse

www.washingtonpost.com/elections/20...

See how people like you vote and how that’s changed over time

www.washingtonpost.com/elections/in...

How we reviewed the videos The Times obtained more than 400 hours of video recordings of private meetings held by the Election Integrity Network and related groups over the past three years. To report this article, we reviewed the footage ourselves and used artificial intelligence to help identify particularly salient moments. First, we used a machine-learning model to transcribe the recordings, which totaled almost five million words. Then, to sift through such voluminous material, we employed several large-language models. That allowed us to search the transcripts for topics of interest, look for notable guests and identify recurring themes. We then manually reviewed each passage and used our own judgment to determine the meaning and relevance of each clip. A.I. was used only to assist our reporting process. Every quote and video clip from the meetings in this article was checked against the original recording to ensure it was accurate, correctly represented the speaker’s meaning and fairly represented the context in which it was said.

My colleagues and I used A.I. to sift through over 400 hours of video recordings and pinpoint salient moments that helped inform this piece's reporting.

It was great to collaborate with Alexandra Berzon, Nick Corasaniti, Duy Nguyen, Juliana Castro Varon and Matt Ruby on this investigation.

The slinky book is great. Working on shrub book now.

👋