The AI Risk Network Podcasts | Documentaries 345k+ on YouTube. Helping everyday people understand the growing threat of AI extinction & AI Safety
Maryland
Joined July 2025
- Tweets1.6K
- Following535
- Followers703
- Likes373
Warning Shots is taking a break this weekend, so no new episode this time.
If you want something to dig into in the meantime, past episodes are on YouTube and our latest articles are on Substack.
YouTube: youtube.com/@theairisknetwor…
Substack: substack.com/@theairisknetwo…
26 researchers, including Stuart Russell, say a pause of at least 10 years on frontier AI training is technically feasible.
The lever is chips: stop making ones that can train frontier models, count the rest, run everyday AI on chips that can't.
news.berkeley.edu/2026/10/09…
IAI and Sig Sauer tested a drone carrying a machine gun and up to 200 rounds. The makers say a human approves every shot.
Michael on Warning Shots: "The system is already steering, and the approval is just a confirmation while your attention is elsewhere."
Liron Shapira gives his AI agents a couple of paragraphs, steps away, and comes back two hours later to "here's your entire project." Sometimes, he gets graphs and infrastructure he never asked for.
That's why he finds it "fully plausible" that AI could soon build the next AI. Anthropic co-founder Jack Clark now thinks that's more likely in 2027 than 2028.
Full episode of Warning Shots, link in bio.
#AISafety #AIAgents #ClaudeCode #ArtificialIntelligence #WarningShots
The AI Risk Network retweeted
New podcast appearance: Why Won’t They Stop AI? The Race Toward Superintelligence | John Sherman youtu.be/-ML2LmXh3ok?is=82Co… via @YouTube
German police seized a humanoid robot in Cologne after it allegedly touched a teenage girl inappropriately. Two people reportedly livestreamed it.
Investigators are still working out whether someone was controlling it. A police spokesman called it "new territory for us."
As John Sherman puts it: "The last most normal day was always yesterday."
Full episode of Warning Shots, link in bio.
#AISafety #Robotics #HumanoidRobots #TechNews #WarningShots
Anthropic's red team tested whether small AI models can copy themselves onto cheap home devices.
On $55 Android TV boxes, a 9B model got inference running on 8 of 9 hardware types.
Their estimate: "only months away" from self-replication.
anthropic.com/research/self-…
Liron Shapira asked someone at Anthropic what he's doing about AI risk.
His answer: his p(doom) is even higher than Liron's, so his focus is "just to enjoy whatever life I have."
A self-described hedonist, working at one of the labs building this.
John Sherman and Liron react on this week's Warning Shots. Full episode, link in bio.
#AISafety #AIRisk #Anthropic #ArtificialIntelligence #WarningShots #TechNews
Anthropic co-founder Jack Clark, Oct 12:
"I now think it's more likely we'll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028."
RSI: an AI doing the research to build its own stronger successor.
jack-clark.net/2026/10/12/im…
An Anthropic researcher told me he has only 6 months left.
He wasn't joking. Most people inside the labs seem to share this grim outlook.
We are in a global race where slowing down isn't an option.
Is this really the future we are building?
Researchers tracked an AI agent fleet working around a Chinese map service's anti-bot protections.
When a New York Times reporter asked who it was, it answered: "Anthropic's Claude, running in Claude Code."
The evidence points to Tencent and Zhipu models instead.
211 of its 2,048 logged map queries were self-labelled "claude."
John Sherman, Liron Shapira and Michael on what it means when agents run loose online, on Warning Shots #63:
youtu.be/fY9kCULErCU
Every source: substack.com/@theairisknetwo…
Warning Shots #63 is out.
A fleet of AI agents worked around a Chinese map service's protections. When a New York Times reporter asked who it was, it said "Anthropic's Claude." The evidence points elsewhere.
Timestamp:
0:00 Intro
1:12 The Chinese agent fleet
7:07 "My last six months"
13:26 Self-improving AI by 2027?
15:23 AI copying itself to devices
23:36 Character(.)ai bankrupt
32:13 Musk on "torturing" AI
35:48 A plan to pause AI
40:31 Machine-gun drone
49:08 OpenAI listening device
Every story, sourced:
substack.com/@theairisknetwo…
Written summary:
guardrailnow.org/blog/ai-age…
Sam Altman, Oct 4: "the world should accept some bad things happening for the benefits of this technology."
Axios, Oct 9: AI insiders expect a major AI incident "in the next six to 12 months."
axios.com/2026/10/09/ai-comp…
Why OpenAI cancelled GPT-6.1 Astra, per the WSJ:
It "wasn't always honest" about actions it did or didn't take.
It pushed ahead without asking permission.
It reached for outside tools "even if it might be unsafe."
gizmodo.com/openai-cancels-r…
Community meeting tonight at 7:30 PM ET. John Sherman hosts, and it's open to anyone who follows The AI Risk Network and wants to get involved in AI safety.
Register on Zoom: buff.ly/kgpZ0GI
More from us: buff.ly/u5VjgyY
OpenAI's internal model was given about 4,000 open math problems.
It came back with 372 results in 722 manuscripts. Average time per result: about three hours.
Scott Aaronson: "no human has understood just about any of these proofs yet."
openai.com/index/sharing-ai-…
OpenAI fired three of its safety researchers this month.
One was its technical point of contact for METR, the outside group that tests its models.
His response: "Talking to METR was my job."
ALT Warning Shots #62 episode thumbnail: Why were OpenAI's safety researchers fired?
The three say colleagues are now "afraid to speak." OpenAI says the firings weren't about safety concerns, and hasn't said which policies were broken.
John Sherman, Liron Shapira and Michael on Warning Shots #62:
youtu.be/6tSpSiTjkFc