Cameron Buckner

Yeah, now write such a line of code for any question anybody could ask. Amazing nobody has tried that before! AI could be both easy and transparently interpretable. Such short-sightedness.

Rotman Visiting Speaker: Cameron Buckner - The Rotman Institute of Philosophy "LLMs as Models of Reasoning" Abstract: Recent advances in large language models that use self-prompting like GPT’s o1/o3 and DeepSeek have begun to encroach on human-level performance on “higher re...

www.rotman.uwo.ca/event/rotman...

Today I am giving a lecture to the Rotman Institute for Philosophy at University of Western Ontario...

And here I was hoping to eventually be able to get one that was liquid cooled...

There's another term for purely theoretical psychology that may explain this...

👀😬

Novels are just synthetic training data for humans.

Global AIDS funding to be pulled from Trump clawbacks request The White House is backing off $400 million in proposed PEPFAR cuts due to GOP objections.

Some really great news: funding for PEPFAR--about $400M for the global fight against HIV and AIDS--has been restored. Millions of lives will be saved.

www.politico.com/live-updates...

Another review of my book is out at Philosophy of Science, by Bruce Rushing and Daniel Hermann:

resolve.cambridge.org/core/journal...

The Climate Brink | Andrew Dessler | Substack The Choice is Ours: Climate Disaster or a Sustainable Future. Click to read The Climate Brink, a Substack publication with tens of thousands of subscribers.

here's the truth: climate change is costing us lots and lots of money. and renewables save us money.
www.theclimatebrink.com?utm_source=n...

GOP plan to sell more than 3,200 square miles of federal lands is found to violate Senate rules A plan to sell more than 3,200 square miles of federal lands has been ruled out of Republicans’ big tax and spending cut bill after the Senate parliamentarian determined the proposal by Senate Energy ...

Senator Mike Lee is now trying — for the third time — to sell-off YOUR public lands to the highest bidder.

We’ve built a bipartisan coalition to repeatedly block his efforts, and will keep doing so until he gets the message:

Our public lands are not for sale!

🏔️🏔️🏔️

apnews.com/article/publ...

China breaks more records with surge in solar and wind power Between January and May, China added 198 GW of solar and 46 GW of wind, enough to generate as much electricity as Indonesia or Turkey

China is absolutely eating our lunch www.theguardian.com/world/2025/j...

Mechanistic Interpretability Needs Philosophy Mechanistic interpretability (MI) aims to explain how neural networks work by uncovering their underlying causal mechanisms. As the field grows in influence, it is increasingly important to examine no...

arxiv.org/abs/2506.18852

It's kinda already happening though, MI groups are already peppered with philosophers...

I am once again begging journalists to understand that tech CEOs can *say* literally any random shit about AI

Watch: AI will solve world hunger by deciphering the Voynich Manuscript!

Until there is literally any evidence AI can do said thing it’s not news, it’s just make believe

Of course, Budding hedges a bit that the metrics reviewed aren't perfect. I think this one is a good supplement if anyone is thinking about this line:
proceedings.neurips.cc/paper_files/...

The paper is particularly valuable in the way that it connects classic discussions of representation in earlier generations of ANNs to detailed technical discussions of recent architectural advances in LLMs. More of this, please, everyone!

Interventions on the MLP modules are robust on textual variants of the same facts, suggesting causal systematicity is satisfied (albeit approximately, and in a distributed and potentially superimposed manner).

Causal tracing work suggests that MLP modules may function as key-value pairs, which take input from semantically-similar inputs and return outputs reflecting memorized properties of that output.

Budding argues that there is some evidence that a weakened version of the casual systematicity criterion can be met by LLMs--models map semantically similar inputs to proximate regions of embedding space, which leads to similar processing by later layers of the network.

Davies argued that previous generations of ANNs could not satisfy reasonable criteria for tacit knowledge because the distributed nature of their representations made it impossible for common causal factors to participate in generalizable instances of knowledge.

Really nice paper by Céline Budding that applies Martin Davies' approach to tacit knowledge to large language models.
www.cambridge.org/core/journal...

Trump plunges the US into a new war In fewer than six months in office, the self-proclaimed anti-war president has engrossed the US into a new war in the Middle East

"Trump has spent the last five months vowing that the Ukraine-Russia war would never have happened if he was president, vowing that the Israel-Gaza war would never have happened if he was president. And yet the reality is that not only has he not avoided war, but he’s plunged us into a new one."

Georgi Gardiner - Helen Helen De Cruz is a trailblazer for creative, avant-garde approaches to thinking in community with others. This event celebrates her work. On Zoom: https://tulane.zoom.us/j/6185944474 Tues June 24t...

From organiser Georgi Gardiner on FB: "Next week people will gather online to celebrate the philosophical artworks of Helen De Cruz. (@helendecruz.net)

Tuesday, June 24th, 2025
11am - 12.20pm Central US Time
(i.e. 5pm BST) on Zoom

All welcome! No background knowledge required."

But this leads me to a piece of advice I really hope we all consider. To all the young scholars out there in AI: You should finish your PhD with three papers that you are decidedly passionate about. Three papers that you can tell a strong story about. And if your friend asks you to work on some other project distracting from those three, it’s ok to say, “your project is amazing, but I don’t have time to give it my all for the deadline.” These are simple things. We’d all be better off stepping back, breathing, and embracing them.

This seems like good advice to me.

With NSF cuts, my guess is that the best possible scenario right now is that AI/ML PhD graduates drop by only 50% in the next 3-4 years. I don't understand how tech and related companies are not freaking out about this.

How much do language models memorize? We propose a new method for estimating how much a model ``knows'' about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have stru...

I've had ideas for experiments like this for a while now, super excited to see someone actually did it!

Don't be disguised by the title: this isn't about memorization specifically, it's using memorization as a tool to measure *capacity*.

My version: instead of just synthetic strings... 1/n

Opinion | We are witnessing the suicide of a superpower The president’s assault on science dangerously undermines America’s superpower status.

Since the 1940s, the engine driving U.S. economic and military competitiveness has been federal support of research universities. Now Trump is sabotaging a research and development pipeline that is the envy of the world. wapo.st/3FDsQEg

I think you should try to retract and resubmit...

I've always been told that whispering is worse than croaking for your vocal cords...

You can sell them for hundreds of dollars if you show up at the right time on Broadway...

It's too late to help you now, but always have magic pastilles in your work bag.
grethersusa.com?srsltid=AfmB...