Knowledge catalogue
@GaryMarcus Never ever use generative AI for anything critical. The technology is probablistic all the way down and therefore inherently unreliable. Why is this so hard to understand? It’s wild how ma
@GaryMarcus These coding tools are intellectual chain saws. Powerful in the hands of a caring professional, extremely dangerous to the operator when used improperly. Implying AGI-adjacency has created
Generative AI as the Hindenberg 🦔 Michael Wooldridge, professor of AI at Oxford, is warning that the race to market AI has raised the risk of a 'Hindenburg moment' that could shatter global confidence
arXiv:2601.10386v2 Announce Type: replace-cross Abstract: Accurate survival prediction in Non-Small Cell Lung Cancer (NSCLC) requires integrating clinical, radiological, and histopathological data. Mu
arXiv:2502.03698v4 Announce Type: replace Abstract: Learning from demonstrations is a popular approach to train AI models; however, their vulnerability to adversarial attacks remains underexplored. We
I don’t always agree with @ElonMusk. I often disagree. Sometimes loudly. But he’s basically right here, as far I can see. Scam Altman and Greg Stockman stole a charity. Full stop. Greg got tens of bil
arXiv:2604.22631v1 Announce Type: new Abstract: Modern automatic speech recognition (ASR) systems have been observed to function better for certain speaker groups (SGs) than others, despite recent gai
arXiv:2604.22565v1 Announce Type: cross Abstract: Large Language Models (LLMs) can reason well, yet often miss decisive evidence when it is buried in long, noisy contexts. We introduce HiLight, an Evi
arXiv:2408.16286v5 Announce Type: replace Abstract: Designing a safe policy for uncertain environments is crucial in real-world control systems. However, this challenge remains inadequately addressed
arXiv:2604.22158v1 Announce Type: cross Abstract: We study the problem of adaptive control of the stochastic linear quadratic regulator (LQR) with constraints that must be satisfied at every time step
People keep blaming the users for the growing number of vibe-coded reasoners. Doing so is *half* right. Users *are*screwing up - by letting vibe coded stuff access their files, without proper backups,
arXiv:2604.22117v1 Announce Type: cross Abstract: Aligned large language models(LLMs) remain vulnerable to adversarial manipulation, and their dependence on web-scale pretraining creates a subtle but
arXiv:2604.22169v1 Announce Type: cross Abstract: Generic group-based RL assumes that sampled rollout groups are already usable learning signals. We show that this assumption breaks down in sparse-hit
@robertskmiles @GaryMarcus “fails less often than a human does” is a terrible metric. We build deterministic software and test it to destruction to ensure it’s 99.99999 reliable . That’s been the basi
arXiv:2604.22374v1 Announce Type: new Abstract: Sign language translation (SLT) converts continuous sign videos into spoken-language text, yet it remains challenging due to the intrinsic modality mism
arXiv:2511.14427v3 Announce Type: replace-cross Abstract: Effective contact-rich manipulation requires robots to synergistically leverage vision, force, and proprioception. However, Reinforcement Lear
Stop blaming the users. It’s not that simple: People keep blaming the users for the growing number of vibe-coded reasoners. Doing so is *half* right. Users *are*screwing up - by letting vibe coded stu
arXiv:2510.11586v2 Announce Type: replace Abstract: Many in-silico simulations of human survey responses with large language models (LLMs) focus on generating closed-ended survey responses, whereas LL
arXiv:2601.12430v2 Announce Type: replace Abstract: Vision-language model (VLM) hallucination is commonly linked to imbalanced allocation of attention across input modalities: system, image and text.
arXiv:2604.22337v1 Announce Type: new Abstract: Most tabular-data generators match marginal statistics yet ignore causal structure, leading downstream models to learn spurious or unfair patterns. We p
Gary Marcus critiques common misconceptions about how large language models (LLMs) function, suggesting that most users have an incorrect mental model of what LLMs actually do when generating response
This statement critiques the asymmetrical attribution of outcomes in AI development, where users are blamed for failures while AI models receive credit for successes. Gary Marcus uses the metaphor of
arXiv:2604.21938v1 Announce Type: cross Abstract: Embodied AI is widely discussed as a job-displacement problem. The deeper risk, however, is governance lag: the inability of public institutions to ke
The divorce between Microsoft and OpenAI that I projected three years ago is now underway. Time for everyone to start seeing other people. OpenAI is moving away from its exclusive Microsoft arrangemen
The question of whether Altman and Brockman did what they promised ≠ the question of whether Elon is a good person. Musk argued that Altman and Brockman broke promises and abused the notion of a nonpr
there’s seem to be a ton of confusion about this post. for clarity: The question of whether Altman and Brockman did what they promised ≠ the question of whether Elon is a good person. Musk argued that
arXiv:2604.22351v1 Announce Type: cross Abstract: Mid-infrared astronomy from the ground faces critical challenges in accurately detecting and quantifying sources due to the dominant spatially and tim
arXiv:2604.22709v1 Announce Type: new Abstract: While long, explicit chains-of-thought (CoT) have proven effective on complex reasoning tasks, they are costly to generate during inference. Non-verbal
this is exactly what tools like @denieddotdev was built for (behavioral auth) some agent behavior should be deterministically blocked by a separate policy layer, not via prompt instructions reach out
This is totally wrong. Blaming the user is missing the point that (a) coding agents have been overhyped and (b) can’t reliably obey the rules given to them in system prompts and other guardrails. Sorr
arXiv:2604.22225v1 Announce Type: new Abstract: While generative text-to-speech (TTS) models approach human-level quality, monolithic metrics fail to diagnose fine-grained acoustic artifacts or explai
arXiv:2604.22174v1 Announce Type: new Abstract: Generalized Category Discovery (GCD) holds significant promise for the label-scarce Synthetic Aperture Radar (SAR) domain, yet its efficacy is severely
yep, that’s what I do. I am increasingly concerned that people in China are soaking up my wisdom, while a lot of people in the US try to dismiss me. i don’t think that is going to end well. 10. @GaryM
I cannot provide a summary for this entry because the source URL and content appear to be fictional or corrupted (evidenced by the strikethrough text and impossible X post ID). Without verifiable info
All your jobs are belong to us Then surely Anthropic is making a big mistake by hiring all those people huh? Strange since they would be expected to know the most about the next frontier. With OpenAI
Anthropic’s CEO says software engineering is dying. Anthropic’s job listing has 70 open positions in software engineering. 🙄 @GaryMarcus https://www.anthropic.com/careers/jobs 70 opened position in SE
disconcerting. A European Commission proposal could create one of Europe’s largest privacy and national-security risks in decades. https://techletters.substack.com/p/the-european-commission-is-turning
Doomers have made the AI industry much of what it is today, by hyping the narrative. Chamath captured the current state of AI perfectly in his 2025 annual letter 'In testing the most effective messagi
Gary Marcus reported receiving a high volume of phishing emails through X (formerly Twitter) direct messages, suggesting either widespread account compromises or a potential security vulnerability all
Here is a very common problem when building complex agents. Long-horizon agents (in particular) fail in two ways: the decision-maker can't decompose well, or the skill library goes stale. This new res
In other words, you need the world models that LeCun, Schmidhuber, Fei Fei Li, and I have been advocating for all along. Sam Altman says today's models are still dumb because they barely understand yo
“Loves tearing down overhyped AI claims, but his critiques always hit the nail on the head”. They get it, in China. 全球AI圈最有影响力的15个人，你一个都不认识就别说自己懂AI了 说真的，AI这个行业吧，信息差就是钱差。 你每天刷那些二手资讯、等媒体翻译，黄花菜都凉了。真正的消息，
Gary Marcus argues that software engineers will remain in high demand for years into the future, countering claims that artificial intelligence or other technological developments will eliminate the n
Makena Kelly / Ars Technica: Palantir Slack logs and staff interviews reveal internal debates over the company's ICE and DOD contracts during Trump's second term, its manifesto, and more — It took jus
Software engineering legend @Grady_Booch re Dario: I think that @DarioAmodei does not understand software engineering and that he is working feverishly to pump up the valuation of his company in antic
Trump has spent the better years of his second term treating allied governments like freeloading tenants who should be grateful he hasn’t changed the locks. The result, predictably, is that they’ve st
Gary Marcus, a prominent AI researcher and critic, expressed frustration that concerns he has been publicly raising about AI for approximately a decade are now receiving mainstream attention as breaki
Gary Marcus argues that programmers are returning to handcoding despite AI tools, suggesting that AI-assisted coding has encountered practical limitations or drawbacks that make manual coding preferab
Gary Marcus observes that experienced programmers are returning to hand-writing code as a practice, suggesting this approach may offer cognitive or developmental benefits despite modern IDE tools. Thi
Gary Marcus highlights a success or advancement in neurosymbolic AI, an approach that combines neural networks with symbolic reasoning systems. Neurosymbolic AI aims to leverage the pattern-recognitio
Every time I tell AI utopianists that biology is too complex for AI to 'solve', they cite the success of AlphaFold. No, AlphaFold did not 'solve' protein folding. It gets broad structures correct ~70-
Groups and movements that can build & get implemented clear policies will have an outsized impact on the chances that AI is used in the way that they want. This is especially true in the near term It
Here are 5 epic prompts to use with ChatGPT's new image generator. It's the most powerful image generator on the market, so let's use it to our advantage. Save this one 📌 WORK SETUP AUDIT: upload a ph
If you believe that AI is going to have a big impact on work and life, the only real tool for mitigating bad impacts and channeling usage for good will be government policy And that policy will be com
A few weeks ago, I was forwarded an email from a journalist named “Michael Chen,” asking for comment on an AI bill in Tennessee. All signs suggest Michael Chen is not a real person, and the publicatio
arXiv:2603.16797v2 Announce Type: replace-cross Abstract: Guided diffusion sampling relies on approximating often intractable likelihood scores, which introduces significant noise into the sampling dy
arXiv:2604.05846v2 Announce Type: replace Abstract: Large Language Models (LLMs) increasingly rely on agentic capabilities-iterative retrieval, tool use, and decision-making-to overcome the limits of
arXiv:2604.21103v1 Announce Type: new Abstract: Governments are increasingly interested in using AI to make administrative decisions cheaper, more scalable, and more consistent. But for probabilistic
AI is advancing faster than our ability to manage it. We still have the opportunity to build the societal and technical guardrails we need to keep people, institutions, and democracies safe — we shoul
arXiv:2604.21827v1 Announce Type: new Abstract: Modern AI assistants are trained to follow instructions, implicitly assuming that users can clearly articulate their goals and the kind of assistance th