Eliezer Yudkowsky and Nate Soares' new book — If Anyone Builds It, Everyone Dies — is out today. I am much more sympathetic to their views than most. Unfortunately, the book is littered with torturous sentence structures and smug, self-satisfied writing. I was not a fan.

Thrilled to have @ruchowdh.bsky.social and Mala Kumar in @transformernews.ai today, writing alongside the launch of the new Humane Intelligence PBC. www.transformernews.ai/p/why-ai-eva...

Data centers use 2% of the world’s electricity. There’s a *huge* push to build more for AI. But even the most optimistic growth paths for AI predict it’ll drive less growth in energy demand than electric cars or aircon. It’s significant, but it’s not the biggest thing going on.

When I launched Transformer a year ago, it was with a simple mission: provide decision-makers with the information and analysis needed to anticipate and steer the impacts of transformative AI. Today, we're announcing an exciting new effort to achieve that: Transformer 2.0.

Last year, the crypto industry set up a network of super PACs. Its strategy? Throw huge amounts of money at politicians to discourage crypto regulation. Now many of those same people have launched a $100m campaign to do the same thing for AI. Read my full piece on Transformer:

I am genuinely so frustrated with the state of the average journalist. Just nonstop dunking and in a way that makes people less informed. They're being very clear that it's not a calculator and should not be used that way! defector.com/it-took-many...

Why assume Meta’s hiring freeze is for financial reasons, rather than logistical ones (which seems much more likely)? It makes a lot of sense to pause hiring shortly after a huge hiring surge and restructuring — you need time to get the new structure working before bringing more people in!

“If Anthropic loses big,” the judge’s order concludes, “it will be because what it did wrong was also big.” www.transformernews.ai/p/anthropic-...

Hugging Face still has a ton of porn hosted on its website which can be accessed without logging in or verifying your age. Not entirely sure how that's compliant with the UK Online Safety Act ... which even 4Chan's taken action to comply with today!

It's a bit dispiriting to see a company that talks so much about being "ethical" adopt the same old strategy of relying on journalists to find bad stuff on their platforms. I saw this for years covering Meta, and Hugging Face is following exactly the same playbook. bsky.app/profile/shak...

Here's a story about Hugging Face, deepfake pornography, and irresponsible content moderation. Last week, 404 Media found that Hugging Face hosted 1000s of nonconsensual AI models of real people, many clearly intended to be used to create sexual content.

It is pretty shameful that Hugging Face is still hosting these models for making deepfake porn of celebrities (and others), despite @404media.co's reporting a week ago.

Hugging Face portrays itself as the "ethical" AI company. So why is it hosting tools to make deepfake porn of teenage celebrities? Following on from @404media.co's report last week, Transformer found that the company hosts some *really* shocking stuff.

Hell of a time to publicly identify as a liberal, isn’t it.

Big news! Delighted to announce that @jasperjackson.bsky.social‬ is joining Transformer as managing editor. Jasper is an absolute rockstar, having previously run teams at ‪@tbij.bsky.social‬, ‪@newstatesman.com‬, and ‪@theguardian.com‬.

AI safety researcher have long warned about a set of behaviors skeptics insisted were fantasies like deception (including during training) and reward hacking. These days, examples of those concerning patterns are showing up more and more often. www.transformernews.ai/p/ai-misalig...

It's your last week to apply to Tarbell Grants! We're offering $1,000-$15,000+ for original reporting on AI and its impacts. The first round has been a huge success — to inspire you to apply, here's a short 🧵 of pieces we're very excited to have supported so far...

Tarbell Grants Round 2 is back! Learn more and apply:

🚨 Transformer is hiring! We're looking for a managing editor and a reporter to help grow our coverage of transformative AI systems into the world’s leading publication for AI decision-makers.​ Details and links below 👇

I've finally written a bunch of 2024 roundup stuff — best theatre, books, films, and everything else. Links in thread:

arguably the best to ever do it. RIP.

Downloaded Xiaohongshu but it’s not letting me pose my questions 😔

Great post by @franksdavid.bsky.social on how to spot BS in biotech announcements

🚨 New piece in TIME: AI progress hasn't stalled — it's just become invisible to most people. 🚨 I used to think that AI slowed down a lot in 2024, but I now think I was wrong. Instead, there's a widening gap between AI's public face and its true capabilities. 🧵

"I find myself wondering if the people consumed with talk of AI hype are paying enough attention to the technological leaps that took place just within the past few weeks." Couldn't agree more!

happy new year to you all

This paper was pretty widely misunderstood by almost everyone. Read @lynettebye.bsky.social on why model collapse fears are overrated: www.transformernews.ai/p/synthetic-...

I've made some predictions for AI progress in 2025 using Sage's excellent tool. I encourage everyone — especially sceptics of AI progress — to make your own predictions! You've got until tomorrow to do so. theaidigest.org/2025-ai-fore...

Because everything has an AI angle: Samuel Butler, the first AI doomer, was also one of the first people to propose that the Odyssey had a female author. www.joh.cam.ac.uk/authoress-od...

"people who are actually working with AI" who are worried about loss of control risks (a very incomplete list): Paul Christiano Jade Leung Marius Hobbhahn Alec Radford Victoria Krakovna Yoshua Bengio Dawn Song Yi Zeng Buck Shlegeris Stuart Russell Mira Murati Gillian Hadfield Anca Dragan

Weirdly little H5N1 discussion over here. Over on the other platform, knowledgeable people are starting to ever so slightly freak out...

DeepSeek just released a new open weights model that appears to be *very* good. And the training run cost only $5m! github.com/deepseek-ai/...

Twitter is currently ripping itself apart, as the “tech right” realise they’ve got into bed with a bunch of racists who hate brown people. You absolutely love to see it.

I have spent much of today trying to switch from Roam to Logseq, as my Roam subscription expired. Logseq is a hot mess and all the other PKM apps seem awful, so I guess I'm renewing my Roam subscription!

Whenever I’m at my parents’ house I end up watching a bunch of TV news (which I never do at home) and 90% of it is just drivel? An abysmal way to stay informed.

Great new initiative:

For decades, the media failed to treat the climate crisis with the urgency it deserved. Journalism’s failure has undoubtedly led the public to take the issue much less seriously than they should have. I fear we're repeating the same mistake with AI. My new piece for @niemanlab.org argues why:

"Rather than treat AGI as a fringe concern, we must be proactive and ambitious: taking the possibility seriously," says @shakeelhashim.com

🚨The first round of Tarbell Grants applications will close December 20th, 2024.🚨 Get up to $15k for reporting on AI and its societal implications. Learn more and apply now:

Everyone who interviews Anna Wintour ought to ask her to justify this

Vogue has long scrubbed any trace of it but I will never, ever forget their 2011 puff piece on Asma al-Assad, 'A Rose in the Desert' www.gawkerarchives.com/asma-al-assa...

More people who take a sociotechnical approach to AI (many of whom think it is fake and sucks) need to condition on the fact that real advances have been made, and more are to come. Not just to avoid irrelevance, but because the sociotechnical approach is so urgently needed for what’s coming.

Cool well I’m sure no one will ever give it a prompt like that.

Lots of criticism of this (some fair, some not). I agree that I should have included the prompting context that led to these results; I think the results are concerning (and, to a general audience, surprising) nonetheless though. I've updated the piece.

OpenAI's new model tried to avoid being shut down. Safety evaluations on the model conducted by Apollo Research found that o1 "attempted to exfiltrate its weights" when it thought it might be shut down and replaced with a different model. www.transformernews.ai/p/openais-ne...

WaPo gotta be feeling awfully relieved today that a certain someone no longer works there

Signed up for a Verge subscription immediately after they launched yesterday. Have thus far been rewarded with … two spammy and unwanted marketing emails 🫠

If a model card reports the results of dangerous capabilities evals and doesn't specify that they were conducted with safeguards removed, you should assume that safeguards were in place for testing. Short 🧵 👇

Managed to get my Instagram Reels algo to consistently show me unhinged Christian, Jewish, Muslim and Wiccan influencer content. Very proud of this achievement.