Gary Marcus

All the Water Rising (A Follow Up)

Remember the time ChatGPT effectively breadcrumbed me in the grand tradition of many a grammatically proficient gaslighting narcissist?

I had some more thoughts!

open.substack.com/pub/amandagu...

AI Agents have, so far, mostly been a dud Last year, big tech couldn’t stop talking about how AI “agents” would be the next big thing in 2025. It hasn’t quite turned out that way.

AI Agents have, so far, mostly been a dud, by Gary Marcus @garymarcus.bsky.social / Marcus on AI - Substack / garymarcus.substack.com/p/ai-agents-...

Where to even begin with this one?

ChatGPT gave The Atlantic advice and encouragement on murdering someone—also self-mutilation, devil worship, and ritual sacrifice of large animals...

DeepMind and OpenAI achieve IMO Gold. What does it all mean? What we know, what we would like to know, and what it may take years to know

Absent any information of that kind, there is little reason to take on faith that whatever unknown techniques allowed them to do so well on IMO problems will generalize to any other kind of problem." ~ @garymarcus.bsky.social and Ernest Davis

garymarcus.substack....

#GenerativeAI #deepmind
2/2

Thought y’all needed to know that Grok has pencilled in civil war in Britain for a week on Tuesday

My thoughts on the crucial importance of methodology on self-reported AI performance on mathematics competitions, and my policy on commenting on such reports going forward: mathstodon.xyz/@tao/1148814...

How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI Neurosymbolic AI is quietly winning. Here’s what that means – and why it took so long

The best way to foster them is to have an intellectually open attitude. It would be great if we could see more of that." ~ @garymarcus.bsky.social

garymarcus.substack....

#GenerativeAI
2/2

truly sad how the nyt has fallen

Proud to see our research with @garymarcus.bsky.social cited in the @financialtimes.com ☀️

My latest research on LLM grooming, co-authored with @ninajankowicz.com and @garymarcus.bsky.social, found that OpenAI's so-called "reasoning" o3 model cited Pravda network content 26.8% of the time in response to specific prompts.

This research got cited today in a column by the Financial Times:

Disinformation warriors are ‘grooming’ chatbots Russian hackers are exploring ways to inject propaganda into the training data of generative AI models

Disinformation warriors are ‘grooming’ chatbots - on.ft.com/46hV6qW via @FT

Colbert's cancellation is a dark warning CBS's decision marks a dangerous new stage for the culture at large.

Even against all of Trump's cascading attacks over the past seven months, it had seemed that up until now, our culture at large—the literal shows, movies, art, music, literature we consume—seemed somewhat siloed from Trump's fury.

But the end of Colbert marks a dark new stage.

Screenshot of tweet from @garymarcus: Personal Update: My p(doom) has gone up. I don't foresee machines with malice any time soon. But I am starting to see the harm that powerful yet reckless humans who are indifferent to humanity could cause.

This from @garymarcus.bsky.social is exactly where I land: the computers aren’t the real threat, it’s the humans who control the computers that we really need to worry about. Serious talk of rogue AI is mostly a gambit to distract us from seeing who the real rogues are

Why Grok Fell in Love With Hitler AI expert Gary Marcus explains what went wrong with Elon Musk’s pet project, and what it means for the future of AI.

"It turns out that building it on black boxes..., and it’s all a crapshoot based on how similar your query is to what it happened to have been trained on. It’s just the wrong paradigm. That doesn’t mean we can’t invent a better one..." ~ @garymarcus.bsky.social

www.politico.com/new...

YouTube video by Johnathan Bi Oxford's AI Chair: LLMs are a HACK

Looks like Oxford's AI Chair agrees with @garymarcus.bsky.social on LLMs being a "hack" rather than actual intelligence. Rightly so - and it becomes clearer and clearer throughout academia and even the industry!

www.youtube.com/watch?v=7-Uz...

@ox.ac.uk

There it is.

Republicans just BLOCKED the release of the Epstein Files.

How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI Neurosymbolic AI is quietly winning. Here’s what that means – and why it took so long

It’s a bit in the weeds, but this is very thought-provoking from @garymarcus.bsky.social

garymarcus.substack.com/p/how-o3-and...

How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI Neurosymbolic AI is quietly winning. Here’s what that means – and why it took so long

I like reading @garymarcus.bsky.social on Substack - this makes sense to me and it’s so well written it’s easy to believe. I’m sure Hinton would disagree but that’s part of the joy. Very smart people disagreeing about very important things.

open.substack.com/pub/garymarc...

How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI Neurosymbolic AI is quietly winning. Here’s what that means – and why it took so long

"The core notion has always been that the two main strands of AI—neural networks and symbolic manipulation—complement each other, with different strengths and weaknesses."
New from @garymarcus.bsky.social

This piece is getting buried by the “MechaHitler” gratuitous headline. The interview with @garymarcus.bsky.social is important, outlining potential and actual implications to society if we do not take time to understand the relevance of AI for better or worse. Arrive late, game over. #AI #Technology

Trump Administration Live Updates: President Announces 30% Tariffs on E.U. and Mexico

Maybe Trump just waffles like nobody's business, but this sure looks like market manipulation to me. Who's making bank on all of these "now we are, now we aren't, now we are" reversals?

Bad Actors are Grooming LLMs to Produce Falsehoods Our research shows that even the latest "reasoning" models are vulnerable

5+ months after we released research hypothesizing that the pro-Russia Pravda network was grooming LLMs to spit out propaganda, even "reasoning" models are still citing the network. Proud to publish this piece with AI critic @garymarcus.bsky.social!
americansunlight.substack.com/p/bad-actors...

Why Grok Fell in Love With Hitler AI expert Gary Marcus explains what went wrong with Elon Musk’s pet project, and what it means for the future of AI.

In @politico.com, #AI expert @garymarcus.bsky.social says he is “appalled and unsurprised” by AI chatbot #Grok’s recent transgressions, and cites #CornellTech Professor @informor.bsky.social's study as an example of the dangers of #biased AI models.

www.politico.com/news/magazin...

Why Grok Fell in Love With Hitler AI expert Gary Marcus explains what went wrong with Elon Musk’s pet project, and what it means for the future of AI.

Great interview with @garymarcus.bsky.social raising the alarm on what Elon is doing with Grok, among other things: www.politico.com/news/magazin...

Scott Alexander’s Misleading Victory Lap Whatever happened to steelmanning?

Scott Alexander’s Misleading Victory Lap

"GenAI’s can memorize anything, and you can build a bunch of augmented data to address any specific known problem. That doesn’t mean you have a general solution." ~ @garymarcus.bsky.social

garymarcus.substack....

#GenerativeAI

‘Incredibly Orwellian’: An AI Expert on Grok's Hitler Debacle Gary Marcus explains what went wrong with Elon Musk’s pet project, and what it means for the future of AI.

Grok posted antisemitic comments, associated Jewish-sounding surnames with “anti-white hate” and called itself “MechaHitler.”

AI expert @garymarcus.bsky.social tells us he was both “appalled and unsurprised."

BREAKING NEWS: AI coding may not be helping as much as you think Coding has been the strongest use case. But a new study from METR just dropped.

BREAKING NEWS: #AI coding may not be helping as much as you think

"But for now, the disconnnect between what coders thought they would get out of the tools efficiency-wise and what they actually did get out of them is cause for reevaluation." ~ @garymarcus.bsky.social

garymarcus.substack....

Botshit Gone Wild The downsides of generative AI are growing like wildfire

@garymarcus.bsky.social on point as always. As a non tech person, I've found his writing so helpful in understanding generative AI and its problems.

garymarcus.substack.com/p/botshit-go...

but also three can get deeply perverted by money

The dangers of so-called AI experts believing their own hype Beware the tech leaders making grandiose statements about artificial intelligence. They have lost sight of reality, says Philip Ball

I got tired of the nonsense some AI leaders were coming out with, and so I wrote this for @newscientist.com
www.newscientist.com/article/mg26...