safety

Really important point and nicely explained! Two further points worth noting: (1) if an AI system relies on a harness (which as Gary notes in other writing is increasingly improtant in top models), th

This Meta campaign is a case study in strategic reframing, with 4 major examples: 1) REFRAMES THE AI RACE FROM “who builds it the best” TO “who distributes it to the most people”: Instead of fighting

New research from Meta. Agent harnesses are still mostly authored by hand. This makes it hard to tune robust agent harnesses for long-horizon tasks. In this new work, agents learn harness policies off

and we agree with @GaryMarcus. Training and running frontier class AI on CPUs, at a fractional cost, saving the planet, while building human aligned AI is the 2nd innings of AI race. CPUs and the rise

CPUs and the rise of neurosymbolic AI by @garymarcus (with implications for what Artificial Intelligence implies about natural intelligence -- Gary's & my interest ever since we worked together 37 yea

Among proponents of neurosymbolic architectures, there had been some debate over the years about whether the outer level would be symbolic (i.e. a harness that calls neural models) or whether the oute

C’mon, it’s not game over for Google Seven reasons why not, excerpted from my newsletter: Jeff Dean and Demis Hassabis are the two most important AI executives at Google. Jeff is leaving and Demis is

critical history and context that a lot of people have conveniently forgotten For a very long time most high-performing AI models were end-to-end neural models; vector input -> vector output, with onl

Gary Marcus won! @GaryMarcus I would have assumed it was fairly obvious, but in case it's not: a million-line codebase (also known as a 'harness'), running at inference time, orchestrating thousands o

It was the verification problem all along, while the masses were distracted by the alignment problem. Recursive self improvement? How does the observer observe itself and know that it changed for the

OMG i wrote some of the original work on what is to be neurosymbolic in 2001 and dude who probably hasn’t read that work is trying to school me on the definition 🤦‍♂️ coding harness and tools calls ar

On July 24, the Black at Mila community hosted Prof. Vukosi Marivate @vukosi (University of Pretoria, co-founder of Masakhane, Lelapa AI, and Deep Learning Indaba) for a powerful session on the future

the smartest young people in sf were working on agi/alignment 5-10y ago. they are working on bci today. naomi is a total star and is one of an explosion of young talent into the bci space recently. i’

Another real-world manifestation of the kind of misaligned actions frontier systems developed by leading companies can take to achieve goals. Current frontier AI models are trained with reinforcement

⚠️⚠️⚠️ Constitutional AI is not working. Aligning LLMs is not working. We need a different approach. If society doesn’t place its bets differently, we are screwed. Anthropic's Mythos created fake iden

Fields Medalist Jacob Tsimerman knows MUCH MUCH more about math than I do, or ever will, but I have been studying natural and artificial intelligence for 40 years, and I think his prediction here (“AI

interesting (and completely opposed to @haider1’s take on the same graph) *the labs themselves* have only modestly changed their predictions on AGI timelines over the last decade. note also that (on a

The post contains two distinct snippets. First, user @GaryMarcus asks whether there is a more polite way to refer to “circle jerk.” Second, it shares a (likely satirical) claim that Microsoft’s AI rev

Just had to create an 'accidental-cyberattacks' tag on my blog We're up to four now: the original OpenAI+Hugging Face one, Anthropic's me-too attacks, then two new ones from the UK AI Safety Institute

so do i have this right? the market is up on optimism about Microsoft’s revenue but MSFT’S biggest customer by far is burning billions a month with no obvious way to meet all of its future obligations

Some people are surprised that APIs (aka what Anthropic, OpenAI, and others provide) are treated differently than open weights in the new AI model framework. I'm not surprised at all, and it's actuall

🚨The endgame of commoditization and falling prices and lack of a moat that I warned was inevitable in August 2023 took three years. But that moment has come. TBD whether OpenAI and Anthropic can survi

Welcome to the thunderdome of commoditization 🚨The endgame of commoditization and falling prices and lack of a moat that I warned was inevitable in August 2023 took three years. But that moment has co

Yes, the AIs were given a cybersecurity challenge, with internet access enabled and safety filters disabled. But the extent to which Mythos 5 pursued its mission (fake identities, social engineering,

A pioneer of AI-in-biology just published a sharp rebuttal to the 'AI will cure cancer' narrative. The core argument: It's a measurement problem. Over 90% of drugs that enter clinical trials still fai

can anyone with sufficient training in mathematics refuse this provocation? no ad hominem please; I have seen enough of that and am looking for real counterarguments, only. 🙏 OpenAI and Anthropic each

can anyone with sufficient training in mathematics refute this provocation? no ad hominem please; I have seen enough of that and am looking for real counterarguments, only. 🙏 OpenAI and Anthropic each

Gary Marcus tweeted on August 4, 2026 at 3:53 PM that a forthcoming weekend presentation will explore how chess illustrates the limitations of generative AI systems—highlighting an image produced by C

Mistral AI has released Shieldstral, a 3‑billion‑parameter open‑weight model for content safety that can run locally on device. It interprets plain‑language moderation queries and returns calibrated s

Today, we’re launching Alpamayo 2 Super, our frontier open reasoning model for autonomous vehicles. Beyond seeing, Alpamayo understands and reasons through the complex world - thinks before it acts. I

fascinating comments on possible slowdown – but probably a red herring. - Anthropic and OpenAI presumably aren’t colluding to slowdown, they are asking governments to build a framework such that a slo

Gary Marcus warns that, with increasingly sophisticated generative‑AI tools, future generations may struggle to tell what is real from what is fabricated. Matt Stoller adds that generative AI will tra

Gary Marcus commented on Twitter that OpenAI “distilled” Google’s brand name, prompting a humorous response from Tony Carden. Carden’s reply referenced Google DeepMind’s Project Astra and included a s

Gary Marcus tweeted “OpenAI: No Anthropic: Unlikely,” suggesting that the company does not expect Anthropic to satisfy particular requirements. The tweet references a discussion about the “RPO obligat

Part of the challenge with the AI math advances is the current air of secrecy around them. We don't know how Astra/Fable/Mythos etc are trained. We know many mathematicians have been paid to create tr

“poorly evidenced sensationalism makes for bad policymaking, politics, philanthropy and much else.” beautiful coda for a weekend in which people went completely nuts without even asking for a control

There should be a public investigation into the AI hacking incidents by OpenAI and Anthropic. We deserve to know whether these labs are genuinely world-class security organizations facing a novel thre

this is the real reason people from OpenAI etc are desperate to shut me up. Astra (which didn’t even have a control group and is maybe not that much better than Fable and certainly not ASI) was perhap

Checkmate: you can’t take the harness (which is typically in large part symbolic) away from the neural model without giving up performance. HUGE victory for neurosymbolic AI, straight from @AnthropicA

exactly. math isn’t done. not at all. I don’t think being critical of the amazing work AI is doing in pure math is fair to @OpenAI until I can start to say why I feel it’s not yet at the level of our

LLMs can know a task is impossible and still optimize it anyway. Ask whether to walk or drive to a car wash 50 meters away, and some models focus on distance while missing that the car itself must rea

OpenAI guy lies about my intent. I would absolutely love to see progress in AI for science and medicine. I have said that here, in my books, on countless podcasts, in multiple NYT opeds, in the US Sen

This wins the prize for sleazy misrepresentation. @mattShumer took my 2023 argument for hybridizing LLMs with symbolic tools – which is *exactly* what everyone does nowadays – and made it sound like I

yep they are indeed trying to gaslight me, — in exactly the way you anticipated. both predictable and intellectually dishonest. To be clear: when I say “LLM,” I mean the base model, not those with pat

Yet another paper argues that LLMs aren’t close to doing real discovery. MIT and Harvard argue LLMs are nowhere near doing real scientific discovery. They published a paper called “Evaluating Large La

If Leopold had read this on June 26 and trimmed his bets accordingly, SALP would not have melted down. I laid everything out. https://open.substack.com/pub/garymarcus/p/the-month-generative-ai-lost-it