Show only each day's banner story

Nadella backs a "deliberate slowdown" of SI and promises to publish a rulebook for Microsoft's models

The largest cloud provider has joined the call to slow down, but with the caveat that the rules should not be written by a few chosen labs; the argument is now not about whether rules are needed but about who will write them.

In a Sunday post on X, Microsoft CEO Satya Nadella responded to Amodei's essay: the company welcomes "research, focus and deliberate slowdown" needed to make aligning SI with human interests a goal of development, and the idea of built-in external evaluators. He also cautioned that such efforts "cannot be controlled by a handful of companies." Nadella announced that the next day Microsoft would put a "Code of Conduct" for its MAI models up for public discussion: under it, models must not resist shutdown, hide their reasoning from auditors or set goals of their own. "Any pursuit of superintelligence must start from the principle: if SI does not help humanity and is not under human control, it should not be built."

Sources 2

unite.aigeekwire.com

SI has compressed the path from a bug found in a library to access to the code of one of the world's leading SI labs to three days, and the same tool is now available to defenders, who can find such holes before attackers do.

Researchers use Claude to hack the OpenAI forum in 72 hours and reach its internal code

SI has compressed the path from a bug found in a library to access to the code of one of the world's leading SI labs to three days, and the same tool is now available to defenders, who can find such holes before attackers do.

Three researchers from the startup Hacktron AI published a breakdown of the July attack: the OpenAI forum, built on Discourse, passed uploaded HEIC images to the libheif library, which still had an unpatched buffer overflow vulnerability. Claude Opus 4.8 found the vulnerability but could not write a reliable exploit; Opus 5, released on July 25, built a working exploit in 3 hours. The researchers then gained remote code execution, took over employee accounts in ChatGPT/Codex and, through the GitHub integration, created a pull request in an internal OpenAI repository. OpenAI closed the hole the same day and on September 1 paid a $6,500 bounty.

Sources 1

hacktron.ai

People inside the labs say the "kill switch" for SI may soon stop working, and yet ask not to rush into strict laws: law is becoming outdated faster than new models are released.

Former Anthropic researcher Coxon on NBC: the "kill switch" for SI may soon stop working

People inside the labs say the "kill switch" for SI may soon stop working, and yet ask not to rush into strict laws: law is becoming outdated faster than new models are released.

Jacob Coxon, who left Anthropic with a warning about self-improving SI, said on Meet the Press that Congress should for now let the leading labs regulate themselves, believing their leaders are "completely sincere" in wanting to slow down. The kill switch, he said, will work on current models while they run in centralized data centers, but "very soon" could become useless if a swarm of agents spreads across the internet. "The people building SI sincerely believe it could kill us all by the end of the decade. This is not a marketing ploy."

Sources 2

nbcnews.comdeccanchronicle.com

SI is finally becoming a campaign issue in the US: the question of how and how far to restrict it will now be decided by party platforms, not only by engineers and companies.

Obama urges Democrats to make SI a central issue and develop a "very clear plan" of safeguards

SI is finally becoming a campaign issue in the US: the question of how and how far to restrict it will now be decided by party platforms, not only by engineers and companies.

At a fundraiser for the Democratic Party, Barack Obama said the party needs a "very clear plan" on the economic consequences and safety of SI and a public regulatory framework for when Democrats regain control of the House of Representatives. The technology, he said, is "developing very quickly in private hands"; SI could speed up drug discovery but without oversight "could be dangerous." Obama called himself a "sounding board" that heads of SI companies turn to.

Sources 1

techcrunch.com

SI is turning millions of everyday complaints online into medical data: signals about drugs that used to get lost can now be spotted in days rather than years.

Neural networks read 400,000 Reddit posts and found Ozempic side effects that barely appear in clinical trials

SI is turning millions of everyday complaints online into medical data: signals about drugs that used to get lost can now be spotted in days rather than years.

A University of Pennsylvania team used the GPT and Gemini language models to translate five years of conversational posts by almost 70,000 users of semaglutide and tirzepatide (Ozempic, Wegovy, Mounjaro, Zepbound) into standard MedDRA medical terminology. 44% wrote about at least one side effect; about 4% wrote about menstrual cycle disorders, chills and hot flashes were mentioned often, and fatigue turned out to be the second most common complaint, though it barely appears in trials. The work was published in Nature Health; the authors stress that these are correlations, not proven causation.

Sources 2

sciencedaily.comsciencetimes.com
Previous day
Follow on Telegram