I am disappointed in the AI discourse steveklabnik.com/writing/i-am...

Meta introduced Llama 4 models and added this section near the very bottom of the announcement 😬 “[LLMs] historically have leaned left when it comes to debated political and social topics.” ai.meta.com/blog/llama-4...

"We train our LLMs on art and literature and educational materials, and for some reason they keep turning out progressive."

Have work on the actionable impact of interpretability findings? Consider submitting to our Actionable Interpretability workshop at ICML! See below for more info. Website: actionable-interpretability.github.io Deadline: May 9

𝐇𝐨𝐰 𝐜𝐚𝐧 𝐰𝐞 𝐩𝐞𝐫𝐟𝐞𝐜𝐭𝐥𝐲 𝐞𝐫𝐚𝐬𝐞 𝐜𝐨𝐧𝐜𝐞𝐩𝐭𝐬 𝐟𝐫𝐨𝐦 𝐋𝐋𝐌𝐬? Our method, Perfect Erasure Functions (PEF), erases concepts perfectly from LLM representations. We analytically derive PEF w/o parameter estimation. PEFs achieve pareto optimal erasure-utility tradeoff backed w/ theoretical guarantees. #AISTATS2025 🧵

New paper from our team @GoogleDeepMind! 🚨 We've put LLMs to the test as writing co-pilots – how good are they really at helping us write? LLMs are increasingly used for open-ended tasks like writing assistance, but how do we assess their effectiveness? 🤔 arxiv.org/pdf/2503.19711

pre aca you would specifically avoid being diagnosed or seeking treatment if you didn't have health insurance to prevent it from making it impossible for you to get health insurance. when you bought health insurance after doing this you committed fraud. i did this.

Some of his readers have asked Mike Masnick @mmasnick.bsky.social why his technology news site, Tech Dirt, has been covering politics so intensely lately. www.techdirt.com/2025/03/04/w... I cannot recommend Mike's reply enough. It's exactly what readers need to hear, what journalists need to do.

Neat visualization that came up in the ARBOR project: this shows DeepSeek "thinking" about a question, and color is the probability that, if it exited thinking, it would give the right answer. (Here yellow means correct.)

Come work with me! We are looking to bring on more top talent to our language modeling workstream at @ai2.bsky.social building the open ecosystem. We are hiring: * Research scientists * Senior research engineers * Post docs (Young investigators) * Pre docs job-boards.greenhouse.io/thealleninst...

I interviewed for LLM/ML research scientist/engineering positions last Fall. Over 200 applications, 100 interviews, many rejections & some offers later, I decided to write the process down, along with the resources I used. Links to the process & resources in the following tweets

Obsessed with the work coming out of Finale Doshi-Velez's group; they don't just take the limits of the real world for ML deployment seriously but instead turn it into new algorithmic ideas arxiv.org/abs/2406.08636

The entire archive of CDC datasets can be found here. HUGE shoutout to data archivists- this work is important 👏🙌🏻 archive.org/details/2025...

Can AI really help with literature reviews? 🧐 Meet Ai2 ScholarQA, an experimental solution that allows you to ask questions that require multiple scientific papers to answer. It gives more in-depth and contextual answers with table comparisons and expandable sections 💡 Try it now: scholarqa.allen.ai

It is such a slap in the face to the Indian American community to delay their green cards for decades and then declare that because of that delay their American children aren't citizens.

Is it time for a social media break again? It has not been a great day. 😅

This is pretty cool! Learn more: developer.chrome.com/docs/devtool... (seems to cover CSS and network requests, which might be fun to lay around with)

i was annoyed at having many chrome tabs with PDF papers having uninformative titles, so i created a small chrome extension to fix it. i'm using it for a while now, works well. today i put it on github. enjoy. github.com/yoavg/pdf-ta...

It's ready! 💫 A new blog post in which I list of all the tools and apps I've been using for work, plus all my opinions about them. maria-antoniak.github.io/2024/12/30/o... Featuring @kagi.com, @warp.dev, @paperpile.bsky.social, @are.na, Fantastical, @obsidian.md, Claude, and more.

I've always wanted to build things with D3, but the learning curve was too high. At least for the bespoke stuff I wanted to make (not just simple bar charts). I can finally make things like this thanks to Cursor! I just art directed this, and it made everything work beautifully. Even on mobile 🎉

If you like working with a canvas, Muse (museapp.com) currently has a 30% off. 2 major things that make it different than freeform -- ink sticks to sticky notes (so it feels more like writing in the real world), and you can snippet out sections from pdf that link back to source.

I don't usually discuss software, PKM, or tools here, but I found a valuable tip today. Not only can you use this to create beautiful video tutorials similar to what Screenstudio creates automatically, but you can also use this feature during live sharing & streaming, making it incredibly useful!

This probably won’t reach many people, but if someone who feels the same way ends up reading this, I hope it helps them realize they’re not alone. Lately, my mood has been heavily influenced by the ‘tech(adjacent) twitter' vibes, & the past few months have been really rough. ⏎

The other major Chinese AI lab, DeepSeek, just dropped their own last-minute entry into the 2024 model race: DeepSeek v3 is a HUGE model (685B parameters) which showed up, mostly undocumented, on Hugging Face this morning. My notes so far: https://simonwillison.net/2024/Dec/25/deepseek-v3/

Our new paper "Fooling LLM graders into giving better grades through neural activity guided adversarial prompting" lead expertly by Atsushi Yamamura shows how even in closed models like Gemini, we can add a small adversarial suffix to an essay to get unreasonably high scores arxiv.org/abs/2412.15275

A question for those working on pre-training models (specifically regarding Olmo's logs): at what training step can you begin comparing two models' evaluation performance to get an estimated sense of their relative quality (for example, Olmo7B versus Olmo-2-7B)?

Echoing Theertha Suresh We are hiring! Our team at Google Research NY is seeking a Research Scientist! Our recent research efforts include developing algorithms for improving inference efficiency and alignment of LLMs. If you are interested, please consider applying! www.google.com/about/career...

ImageFX is awesome at creating character sheets, and here are some examples. The text is gibberish though compared to Recraft. I like to use these to add consistent, interesting visuals to presentations with pretty minimal effort. (Prompt expansion by GPT4o in alt)

feeling a but under the weather this week … thus an increased level of activity on social media and blog: kyunghyuncho.me/i-sensed-anx...

Honestly whisk is pretty good and really fun to play around with. Probably one of the best implementations of auto-prompting I have seen. And I like that the default presets are so cute. Here is a preview using my display picture. Check out more at: labs.google/fx/tools/whisk

Hey, guess what prevents Polio? Psst. The answer is a VACCINE. Guess what happens when enough people DON’T get vaccinated for it? It RETURNS.

I will never stop being amazed at the ability of Google infoboxes to display information that is not just untrue but completely senseless. What kind of structured data graph can even produce "San Jose, CA, United Kingdom"? Are they just running LLMs with non-zero temperature even on this stuff now?

As an Indian, my contempt for (rich white) anti-vaxxers in the west is so deep I often cannot quite find the words to express my contempt. I grew up seeing polio-disabled children and adults crawling over rubble and roads in India. My father worked for the Pulse Polio program. What is happening?

Microsoft debuts Phi-4, a new generative AI model, in research preview

I might be the last horse to cross the finish line here, but I learned today that the length "1em" in tex doesn't stand for anything, rather, it's called that because it's exactly the length of one uppercase M 🤯

Trying out image and video generation models is interesting and different for me -- because I do not have any internal desire to "break" the model. I have no easy or hard prompts pre-defined for those models. I try to use them in the flow, i.e., oh I want this video/image, can I use AI for it.

Wrote up my initial impressions of the new Google Gemini 2.0 Flash model - it's really good, and the streaming mode (where you can stream video and audio to it and get audio streamed right back) is pure science-fiction simonwillison.net/2024/Dec/11/...

Thanks everyone for your interest!! The slides are now posted here: neurips.cc/media/neurip...

Our paper PRISM alignment won a best paper award at #neurips2024! All credits to @hannahrosekirk.bsky.social A.Whitefield, P.Röttger, A.M.Bean, K.Margatina, R.Mosquera-Gomez, J.Ciro, @maxbartolo.bsky.social H.He, B.Vidgen, S.Hale Catch Hannah tomorrow at neurips.cc/virtual/2024/poster/97804

Neural networks are not black boxes. They're the opposite of black boxes: we have extensive access to their internals. I think people have accepted this framing so innately that they've forgotten it's not true and it even warps how they do experiments.

@vidhishab.bsky.social and I will be presenting Eureka during NeurIPS Expo, on Wednesday, December 11 at 4:30 PM (West Meeting Room 301). Join us to get a glimpse of a demo, recent results, and an overall in-depth comparison of 12 frontier foundation models.

Stop by our #NeurIPS tutorial on Experimental Design & Analysis for AI Researchers! 📊 neurips.cc/virtual/2024/tutorial/99528 Are you an AI researcher interested in comparing models/methods? Then your conclusions rely on well-designed experiments. We'll cover best practices + case studies. 👇

Google slides still doesn't have morph transitions 😠

Great post that captures the tension between classic ML approaches and modern deep learning while acknowledging the nuances of both. “Working with LLMs doesn’t feel the same. It’s like fitting pieces into a pre-defined puzzle instead of building the puzzle itself.” www.reddit.com/r/MachineLea...

I guess it is time to introduce Bluesky to my latest work, HTML for People. Anyone can make a website with HTML. No previous coding experience required. I cover everything you need to know to get started in an approachable and friendly way. And it’s free for all. 🚀 htmlforpeople.com

Increasingly thinking of my PhD work as directing “sorry, what did you means” at a field as in “sorry what do you mean your field doesn’t have benchmarks” “sorry what do you mean you didn’t tune your baselines” “sorry what do you mean your simulators run slow”

The FATE group at @msftresearch.bsky.social NYC is accepting applications for 2025 interns. 🥳🎉 For full consideration, apply by 12/18. jobs.careers.microsoft.com/global/en/jo... Interested in AI evaluation? Apply for the STAC internship too! jobs.careers.microsoft.com/global/en/jo...