In a new paper in our AI & Democratic Freedoms series, Rachel M. Kim, Blaine Kuehnert, @sethlazar.org, Ranjit Singh, & Hoda Heidari propose creating an AI Power Disparity Index, designed to measure and signal the changing distribution of power in the AI ecosystem. knightcolumbia.org/content/the-...

How will AI agents impact democratic values? Democracies are—for independent reasons—already under acute pressure. Since WWII Moore's Law and democratisation went up and to the right in lockstep. Not any more.

In the latest essay in our AI & Democratic Freedoms series, @sethlazar.org and Tino Cuéllar (@carnegieendowment.org) discuss how AI agents might affect the realization of democratic values. knightcolumbia.org/content/ai-a...

"Democracies are weaker than they have been for decades," write Carnegie president Mariano-Florentino Cuéllar and @sethlazar.org for @knightcolumbia.org. "A great wave is coming, and they are ill-prepared." AI agents could help or hurt. And they won't protect democratic values on their own.

@caseynewton.bsky.social in re an old discussion about AI denialists. , hope you’ve caught knightcolumbia.org/events/artif...

🚨 UPCOMING EVENT: Artificial Intelligence and Democratic Freedoms, April 10-11 at @columbiauniversity.bsky.social & online. In collaboration with Senior AI Advisor @sethlazar.org & co-sponsored by the Knight Institute and @columbiaseas.bsky.social. RSVP: knightcolumbia.org/events/artif...

New Philosophy of Computing newsletter: share with your philosophy friends. Lots of CFPs, events, opportunities, new papers. philosophyofcomputing.substack.com/p/normative-...

I am a bit bashful about sharing this profile www.thetimes.com/uk/technolog... of me in @thetimes.com, but will do so because it kindly refers to my new book which is coming out in early March. www.penguin.co.uk/books/460891.... The tech titans pictured seem to be decoration (and not my co-authors)

I spent a few hours with OpenAI's Operator automating expense reports. Most corporate jobs require filing expenses, so Operator could save *millions* of person-hours every year if it gets this right. Some insights on what worked, what broke, and why this matters for the future of agents 🧵

Since Agents are now on everyone's minds, do check out this tutorial on the ethics of Language Model Agents, from June last year. Looks at what 'agent' means, how LM agents work, what kinds of impacts we should expect, and what norms (and regulations) should govern them.

We're excited to announce that our upcoming symposium on #AI and democracy w/ @sethlazar.org (4/10-4/11, at @columbiauniversity.bsky.social & online) will feature papers by a highly accomplished group of authors from a wide range of disciplines. Check them out: knightcolumbia.org/blog/knight-...

January update from the normative philosophy of computing newsletter: new CFPs, papers, workshops, and resources for philosophers working on normative questions raised by AI and computing.

EVENT: Artificial Intelligence and Democratic Freedoms, 4/10-11, at @columbiauniversity.bsky.social & online. We're hosting a symposium w/ @sethlazar.org exploring the risks advanced #AI systems pose to democratic freedoms and interventions to mitigate them. RSVP: knightcolumbia.org/events/artif...

📢 Excited to share: I'm again leading the efforts for the Responsible AI chapter for Stanford's 2025 AI Index, curated by @stanfordhai.bsky.social. As last year, we're asking you to submit your favorite papers on the topic for consideration (including your own!) 🧵 1/

Turns out we weren't done for major LLM releases in 2024 after all... Alibaba's Qwen just released QvQ, a "visual reasoning model" - the same chain-of-thought trick as OpenAI's o1 applied to running a prompt against an image Trying it out is a lot of fun: simonwillison.net/2024/Dec/24/...

Here are my collected notes on DeepSeek v3 so far: simonwillison.net/2024/Dec/25/...

deepseek-ai/DeepSeek-V3-Base huggingface.co/deepseek-ai/...

Some of my thoughts on OpenAI's o3 and the ARC-AGI benchmark aiguide.substack.com/p/did-openai...

OpenAI skips o2, previews o3 scores, and they're truly crazy. Huge progress on the few benchmarks we think are truly hard today. Including ARC AGI. Rip to people who say any of "progress is done," "scale is done," or "llms cant reason" 2024 was awesome. I love my job.

OpenAI's o3: The grand finale of AI in 2024 A step change as influential as the release of GPT-4. Reasoning language models are the current and next big thing. I explain: * The ARC prize * o3 model size / cost * Dispelling training myths * Extreme benchmark progress

I'm not seeing (here) much discussion of o3. If you are, point me to who's on here that I'm missing? If you're not: just registering that o3's performance on SWE-bench verified is *bananas*, and likely to have massive impacts in 2025.

Busy shopping day in Causeway Bay (long exposures handheld with Spectre App)

Feeling good (after o3) about some of the bets made in these papers… Human level software agents now seem nailed on for the near-term.

Two papers on anticipating and evaluating AI agent impacts now ready for (private) comments: if you're interested in how language agents might reshape democracy, or in how *platform agents* might intensify the worst features of the platform economy (but could also fix it), lmk.

As one of the vice chairs of the EU GPAI Code of Practice process, I co-wrote the second draft which just went online – feedback is open until mid-January, please let me know your thoughts, especially on the internal governance section! digital-strategy.ec.europa.eu/en/library/s...

🤔

New work from my team at Anthropic in collaboration with Redwood Research. I think this is plausibly the most important AGI safety result of the year. Cross-posting the thread below:

We're working hard behind the scenes to finalize the dates and venue for #FAccT2025! While final confirmation is still pending, our tentative conference dates are June 23-26. Expect more updates soon!

I’m excited to share the announcement of 𝐈𝐧𝐭𝐞𝐫𝐧𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐂𝐨𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐨𝐧 𝐋𝐚𝐫𝐠𝐞-𝐒𝐜𝐚𝐥𝐞 𝐀𝐈 𝐑𝐢𝐬𝐤𝐬. The conference will take place 𝟐𝟔-𝟐𝟖𝐭𝐡 𝐌𝐚𝐲 𝟐𝟎𝟐𝟓 at the Institute of Philosophy of KU Leuven in 𝐁𝐞𝐥𝐠𝐢𝐮𝐦. Our keynote speakers: • Yoshua Bengio • Dawn Song • Iason Gabriel Submit abstract by 15 February:

Really love the container ships we can see from our balcony. We now have an app to track them…

What flipping genius decided that if you want to search for "82.52" in your email you'd be just as interested in something with "83" in it. It is SO BLOODY HARD to find a specific email (and no quotation marks don't help, it's apple's default mail app).

New Normative Philosophy of Computing Newsletter. Events, jobs, papers, links...

Much enjoyed first @neuripsconf.bsky.social—fantastic to see the number (in absolute if not relative terms) of ML researchers seriously concerned with questions of ethics and safety. The coming months and years will try your integrity: I’m hopeful you’ll hold fast!

Super excited for the Evaluating Evaluations workshop at @neuripsconf.bsky.social today!!! evaleval.github.io #NeurIPS2024 @msftresearch.bsky.social's FATE group, Sociotechnical Alignment Center, and friends will be presenting several papers there. See below for details...

As I am wraping up my time in Vancouver for #NeurIPS2024, I want to share that our paper with @rajiinio.bsky.social won Best paper award at Black In AI ☺️

Scale AI worker sues the company, accusing them of wage theft, misclassifying them as contractors, and unfairly cutting off contractors. The issues these workers are suing over impact workers across the platform economy. www.sfgate.com/tech/article...

New paper on why machine "unlearning" is much harder than it seems is now up on arXiv: arxiv.org/abs/2412.06966 This was a huuuuuge cross-disciplinary effort led by @msftresearch.bsky.social FATE postdoc @grumpy-frog.bsky.social!!!

Microsoft's Phi-4 A small language model that performs as well as (and often better than) large models on certain types of complex reasoning tasks such as math. techcommunity.microsoft.com/blog/aiplatf...

I have been *really* enjoying my time here at Hong Kong's philosophy department. Incredible concentration of talented philosophers working on AI. They're advertising a number of positions (with more to come): strongly encourage early career folks in this area to apply.

Reposting! Who is at @neuripsconf.bsky.social this week??? #NeurIPS2024

What are the prospects for AI that is, or irresistibly appears to be, conscious? Here’s a substantially revised version of my paper “Conscious artificial intelligence and biological naturalism” - many thanks to many folks for super-helpful comments osf.io/preprints/ps...

Training Large Language Models to Reason in a Continuous Latent Space Introduces a new paradigm for LLM reasoning called Chain of Continuous Thought (COCONUT) Directly feed the last hidden state (a continuous thought) as the input embedding for the next token. arxiv.org/abs/2412.06769

Delighted for @hannahrosekirk.bsky.social 's success with Neurips Best Paper for Datasets and Benchmarks track! What a star: blog.neurips.cc/2024/12/10/a...

NeurIPS sticker pack: find me for a shiny brat Tülu, vintage OLMo, or shuriken-esque Ai2!

Made a list of resources for open source language models with @soldaini.net ahead of the tutorial tomorrow at 930 AM. github.com/allenai/awes...

Finally read this great anthropological study the hidden spiritual and cultural references within the latent space of models, by @bvlsingler.bsky.social & @mpshanahan.bsky.social. Puzzled by terms like 'hyperstition' or 'noosphere'? This is a great place to start!https://arxiv.org/abs/2411.13223

Life's too short (etc) but SF/F is really crying out for a high fantasy universe in which magic, only it's just like software (in all the worst possible ways), and wizards spend most of their time in stand-ups and almost nobody ever casts a spell, it's all about mass producing the MVP magic ring ...

Llama 3.3 70B model It is at similar performance to the previous Llama 3.1 405B model release, but at significantly lower cost, so much more accessible for developers. Blog: www.llama.com Models: www.llama.com/llama-downlo... and huggingface.co/meta-llama/L...

OpenAI announced a new RL finetuning API. You can do this on open models w the repo we used to train Tulu 3. Expanding reinforcement learning with verifiable rewards to more domains and with better answer extraction and to more domains in our near roadmap. https://buff.ly/3V4JEIJ