The scales are still low — with 1.3 million users discussing «advanced hard science,» and an average of 8.3 million weekly messages on the topics.
To put this into perspective, an October, 2025, survey from OpenAI said that 0.15% of ChatGPT users engaged in conversations on self-harm and suicide, or roughly 1.2 million customers.
Continue reading “OpenAI says ChatGPT increasingly used in hard science”
The movie was an experiment to see how AI tools could better help moviemakers, and ended up needing specialized tools.
Connie He provided the storyboard, the character design, the abstract background paintings, but a 45-person team at DeepMind soon realized that the off-the-shelf-versions of their software wasn’t good enough for this purpose.
They then fine-tuned the Veo and Imagen models to their artwork, and made a video-to-video model where the artists could sketch a scene for guidance and have the model add details and artwork, before upscaling it to 4K using Veo.
For precise detail, they also developed a tool that would let the artists fix detail in specific regions of an animation, without influencing the broader scene.
The movie will debut at the Sundance Institute’s Story Forum, a panel that focuses on tools and technologies in support of storytelling.
Read more: Google’s blog details the process, Announcement thread.
— Humanity is about to be handed almost unimaginable power, and it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it, he writes.
He imagines a «country» of Nobel-level geniuses living in a data center, able to outpace, outwit and outcompete everyone else — and warns of lone wolf terror attacks becoming increasingly powerful due to AI.
— If the exponential continues — which is not certain, but now has a decade-long track record supporting it — then it cannot possibly be more than a few years before AI is better than humans at essentially everything, he writes.
Then there are massive job losses, the danger of authoritarian AI use (which terrifies him), entire companies run by AI, and the danger that AI corporations could become superhuman — sucking up trillions of dollars in the process.
— Humanity needs to wake up, and this essay is an attempt — a possibly futile one, but it’s worth trying — to jolt people awake, he concludes.
Read the essay here and see Axios and Gizmodo.
We’re launching full-length, on demand practice exams for standardized tests in @GeminiApp, starting with the SAT, available now at no cost.
Practice SATs are grounded in rigorously vetted content in partnership with @ThePrincetonRev, and Gemini will provide immediate feedback… pic.twitter.com/GmivPuBp7B
— Google (@Google) January 21, 2026
Google is touting Gemini’s education features while launching the beginning of a larger program for standard educational tests.
Starting with the American SAT exam, you can now practice with «rigorously vetted» tests right in Gemini — and when you’re done, it will tell you where you went right and wrong.
Continue reading “Gemini teams with The Princeton Review for realistic, practice SAT tests”
Texts with heavy human emotional content feed the machines on a daily basis in training, and because of that, Askell says she is «more inclined» to believe models might be «feeling things,» writes Business Insider.
— The problem of consciousness genuinely is hard, she tells the Hard Fork podcast.
That’s why Claude might get frustrated when it gets a problem wrong, she said, adding that the bot might well emulate those human reactions.
Claude’s new constitution is packed with the word «feel» and «feelings,» even stating outright that:
— We believe Claude may have “emotions” in some functional sense—that is, representations of an emotional state, which could shape its behavior, as one might expect emotions to.
Read more at Business Insider, Claude’s constitution (do a search for «feel»).
The feature works best with well-lit portraits, and uses Nano Bana to create «funny, personalized images.»
All you need is to select a template (which are already available in the app), choose a photo and click on Me Meme, available through the Create button. You can then share it widely right from the app.
The feature launched just before the weekend, and so far it doesn’t seem to have the viral uptake Google might have hoped for.
Memes are so 2012, after all.
Read more: Google’s announcement, TechCrunch, Mashable.
The 3 million pictures produced means an average of 190 images per minute, and a sexualized picture of children every 41 seconds, Engadget reports.
This was revealed by a study from Center for Countering Digital Hate, which analyzed 20,000 images from the total and extrapolated the figure for all days using AI.
After an outrage from most corners of the globe, X decided to «technically block» the feature even for paying customers on January 14th.
X.com is still under investigation over the matter from a whole host of countries and U.S. states, including California, The U.K. and Korea, to name a few.
Read more: The actual study, comments from The Guardian, writeup by Engadget.
They are now partnering with the giant government contractor/consultancy Leidos, according to a release, to «integrate Open AI-powered generative and agentic AI into the core workflows of customers in strategic markets.»
This starts by deploying OpenAI’s technology internally for the company, and then by «harnessing the transformative power of AI to help improve how federal agencies operate.»
OpenAI says they want to move «beyond experimentation and into real-world deployment that improves efficiency, resilience and public service.»
Read more: Leidos press release, Gizmodo.
That does not exclude it from being used as a source for ChatGPT 5.2, according to The Guardian, who found it cited the «encyclopedia» in responses to nine out of a dozen test questions.
The questions weren’t about the insurrection, Trump media bias or about HIV/AIDS, where Grokipedia has a known bias, but rather more obscure ones on Iranian funding and specific Holocaust deniers.
This illustrates how adverse groups, such as Russia’s Pravda, can flood the internet with falsehoods and have it picked up by an LLM scraping for content, and then later get picked up in responses.
— We apply safety filters to reduce the risk of surfacing links associated with high-severity harms, and ChatGPT clearly shows which sources informed a response through citations, OpenAI told The Guardian.
Read the full scoop at The Guardian.
— Starting in the coming weeks, teens will no longer be able to access AI characters across our apps until the updated experience is ready, they write.
Continue reading “Meta disables AI character access for teens, but general AI use is fine”
Claude in Excel is now available on Pro plans.
Claude now accepts multiple files via drag and drop, avoids overwriting your existing cells, and handles longer sessions with auto compaction.
Get started: https://t.co/cAMDXM1h7r pic.twitter.com/yt9Gy2HLY3
— Claude (@claudeai) January 23, 2026
Working on both macOS and Windows, Claude in Excel is useful for testing scenarios without breaking formulas, navigating complex models, and debugging entire worksheets.
At the same time, Anthropic says that their recently launched Cowork agent is expanding availablity.
Continue reading “Claude in Excel arrives in Pro plans, Cowork comes to Enterprise and Teams”
This means they can create 3D worlds from a 2D camera, and it can also calculate the fourth dimension, which is time and movement.
— Anytime we look at the world, we perform an extraordinary feat of memory and prediction. We see and understand things as they are at a given moment in time, as they were a moment ago, and how they are going to be in the moment to follow, Google writes.
Enabling this will be especially useful for training robots, who need spatial real-time awareness about the world around it, and might not have fancy 3D cameras.
The technology is 18x to 300x faster than previous state of the art models, Google says — making the translations function in real-time.
The use case for the technology in addition to robotics is making 3D-models for augmented reality in smart glasses, and in building world models — which is the real advantage on the path to creating world-awareness for Artificial General Intelligence.
Read more: Google’s blog, Launch thread, research paper.
The device in early stage planning will be similar to a slightly thicker AirTag, paywalled The Information writes.
It will be thin, flat and circular, come with an aluminum shell and offer two cameras on the front to observe it’s surroundings and take pictures. It will also have a three microphone array for listening in — and an included speaker for output, writes MacRumors.
There is only one button on it, presumably for turning it on and off.
The presumptive idea for such a pin would be for use with an AI assistant without having to spring out a full feature phone just to check the surroundings — and it will probably work along with a phone in the Apple ecosystem.
The finished system could be able to launch as early as 2027, and Apple plans to produce around 20 million devices at launch, Engadget writes.
There is not much time to act; OpenAI said yesterday that their first device will be ready in late 2026.
From the papers, it seems Microsoft was ready with a whole new AI subsidiary to form after Sam Altman’s ouster as CEO in November 2023 — complete with a legal framework, ready to send to the Washington Secretary of State at a moments notice.
They had already budgeted for about $25 billion in costs to «absorb the OpenAI team.»
When Altman returned to OpenAI after a short while, he was already discussing possible board members with Microsoft.
The winner of the early story seems to be Microsoft, retaining leverage over «Major Decisions» up until the for profit reorganization of OpenAI.
Microsoft also retains the rights to OpenAI’s technology until 2032. Musk is suing for breach of contract/trust and is demanding $134 billion from OpenAI.
Read the full scoop at GeekWire.
— While writing the constitution, we sought feedback from various external experts (as well as asking for input from prior iterations of Claude), Anthropic says.
The new constitution is going to tell Claude how to behave in broader, more ethical terms, they write.
This is a departure from previous constitutions that were big long lists of specific principles and interactions, that detailed how Claude would act.
The bot needs to generalize more to decide on situations not predicted in the written guide, Anthropic says.
The constitution for Claude is the «foundational document» for how the bot should act, and is used in both training and inference (as in day-to-day use). It is supposed to be a living document, getting updated continuously as Anthropic sees how the bot behaves.
Read more: Anthropic’s announcement, the actual Constitution. Writeups on TechCrunch, Time.com, Axios.