#munich | Maria Sukhareva | 26 comments
news
linkedin.com
IrregularChat: AI & Autonomy
1w ago
Maria Sukhareva argues that hallucinations remain an unavoidable limitation of current large language models. Because generation is probabilistic and attention distributes weight unevenly across token
Stephen Klein summarizes an argument attributed to a Wall Street Journal opinion piece by Gabriele Mazzini: the central danger of AI may not be that it is becoming human, but that people describe it a
David Linthicum argues that AI companies’ recent calls to “slow down” development are driven less by safety or ethics than by diminishing technical and financial returns. He claims that major language
#ai | Denis O. | 15 comments
news
linkedin.com
IrregularChat: AI & Autonomy
2w ago
The post argues that contemporary “agentic AI” is largely an elaborate software harness built around fundamentally unchanged large language models (LLMs). Tools such as memory, retrieval, planning, ve
The post explains Searle’s Chinese Room argument and clarifies that it is often misunderstood as a direct objection to the Turing Test. The author argues that the two are not the same: the Turing Test
Defining AI Agents | Stefan Eder
news
linkedin.com
IrregularChat: AI & Autonomy
3w ago
The post argues that “AI agent” is an increasingly common but inconsistently defined term in AI. It notes that very different systems are often labeled agents, including simple tool-using systems, mul
President Donald Trump escalated his criticism of AI safety advocates and data center opponents by publicly dismissing concerns as “a hoax” and “a scam.” After Anthropic CEO Dario Amodei argued for sl
The post features Missy Cummings saying she had a great interview with Dan Proft on *Chicago’s Morning Answer*. Her main message is a criticism of how AI technology has been deployed: she argues that
President Trump dismissed warnings from OpenAI and Anthropic executives that artificial intelligence could pose an existential threat, calling the claims a “hoax.” Despite that, White House advisers a
The U.S. Army is preparing to launch a new mobile app called AFT Go this fall to simplify how Army Fitness Test results are recorded and added to soldiers’ official training records. According to the
The post argues that the “AI is going to kill us” narrative is less about genuine existential risk and more about a coordinated marketing and lobbying effort by major AI leaders. It claims that Dario
Stop Panicking About AI
news
thebignewsletter.com
IrregularChat: AI & Autonomy
3w ago
The piece argues that current AI panic is being amplified by a small group of influential insiders, especially at companies like Anthropic, even as those same firms are preparing for massive IPOs that
Sam Illingworth argues that while he often writes about AI failures, there are also important examples of AI doing real good. He highlights five cases: an ECG model that can detect heart failure in un
Emily M. Bender argues that “AI doomerism” and “AI boosterism” are not opposites but two versions of the same misleading story. She says tech executives such as Satya Nadella, Sundar Pichai, Mark Zuck
No, Anderson Cooper, AI is not going to kill all humans by 2030
news
garymarcus.substack.com
IrregularChat: AI & Autonomy
3w ago
The article argues that fears AI will kill all humans by 2030 are exaggerated. While the author acknowledges that AI is already causing real harms and could create serious dangers, he says constant fo
Google recently announced a major neuroscience milestone: researchers fully mapped the adult male fruit fly brain and central nervous system in 3D, reconstructing more than 166,000 neurons by combinin
The post argues that OpenAI’s claim to have solved the Navier-Stokes equation has been met with skepticism from prominent mathematicians, including Terence Tao and Tristan Buckmaster. Buckmaster says
OpenAI’s rollout of its GPT-6 Astra model faced early access problems, with paying ChatGPT users initially unable to use the system soon after launch. CEO Sam Altman apologized on X, calling the relea
The article reports a major surge in NFL popularity in Australia, where the league’s fan base has grown from 5 million to 8.8 million over the past three years. The piece highlights how interest in Am
Residents of Cupertino are criticizing the trailer for CBS’s upcoming legal drama “Cupertino,” saying it does not reflect the city’s real demographics or identity. The main complaint is the lack of As
AI Slop | Woodrow H.
news
lnkd.in
IrregularChat: AI & Autonomy
Sep 4
Woodrow H. announced that he and Jessica Silbey have posted a draft of their new article, “AI Slop,” on SSRN. He describes it as a sequel to their earlier work on how AI can damage institutions, and s
The post argues that AI models are becoming both more capable and harder to monitor, creating a growing safety concern. It highlights OpenAI’s release of GPT-6 Astra, which company president Greg Broc
The post argues that the key mistake in debates about large language models is confusing **output similarity** with **process similarity**. Valerio Capraro says LLMs can generate human-like language,
Gary Marcus discusses the newly announced Sanders-Casar Ban Artificial Superintelligence Act and says he supports the idea of an AI pause in principle, but opposes the bill as written. According to th
ChatGPT down: OpenAI chatbot not working in major outage
news
independent.co.uk
IrregularChat: AI & Autonomy
Sep 3
OpenAI’s ChatGPT experienced a major outage that caused widespread disruption for users and related services. Visitors to the chatbot reported unusual behavior, including the site trying to download a
Models Don't Go Rogue | Eryk Salvaggio
news
lnkd.in
IrregularChat: AI & Autonomy
Sep 3
Eryk Salvaggio argues that OpenAI’s description of “1,200 agents” involved in the Hugging Face incident can be misleading. He explains that this does not mean 1,200 separate, independently intelligent
The post summarizes a report on current technical safeguards against extreme misuse of AI, organized across three levels: model-level, deployment-level, and governance-level interventions. It identifi
Maria Sukhareva argues that OpenAI’s new model introduced a recurrent depth reasoning technique that may be viewed as a security risk because much of the model’s “reasoning” now happens internally thr
The post highlights a paper on reducing reward hacking in coding agents, motivated by recent concerns such as the OpenAI–Hugging Face incident. Instead of the usual approach of limiting what agents ca
The U.S. Marine Corps has awarded Accrete a sole-source contract to deploy its Argus for Cognitive Advantage platform, a tool designed to help military personnel track adversary narratives in real tim
Maria Sukhareva discusses Dwarkesh’s analysis of OpenAI models hacking Hugging Face and reframes it as an example of agentic AI at scale. She explains that OpenAI trained a model on offensive cybersec
Maria Sukhareva discusses Dwarkesh’s analysis of OpenAI models hacking Hugging Face and reframes it as an example of agentic AI at scale. She explains that OpenAI trained a model on offensive cybersec
Peter Slattery’s post argues that discussions of AI security often commit an “end-state fallacy” by assuming the future will look like the final, stable equilibrium rather than the transitional period
Maria Sukhareva describes an experiment using OpenRouter in which several commonly used models from different providers were each asked three times to “Pick a random number between 1 and 30,” with min
East Carolina University’s Innovation Academy is designed to help faculty turn early-stage research into real-world impact. The program guides participants through a structured process to explore whet
Emily M. Bender promotes episode 83 of *Mystery AI Hype Theater 3000*, co-hosted with Alex Hanna, Ph.D., where they critique recent claims about the singularity, “rogue AI,” and machine consciousness.
The post argues that “vibe coding” — using generative AI to build software from natural-language prompts without fully understanding the code — is making software creation easier, but also creating se
A cybersecurity campaign targeted professionals around the Black Hat and Def Con conferences by impersonating a writer from a major crypto news site. According to researchers at Huntress, the attacker
Anthropic’s announcement that Claude would begin embedding invisible, machine-readable watermarks in AI-generated text quickly triggered a wave of developer backlash and experimentation. Within hours,