Dragons in the Clouds middle-grade fantasy push
A promotion for David Blair’s middle-grade fantasy novella Dragons in the Clouds, framed as a bedtime story about thunder and lightning.
2026-09-25
All AI Pulse updates published on 2026-09-25.
59 signalsSpy and Villain short story spotlight
A post highlights Joy M. Lilley’s short story Spy and Villain, following wheelchair-using Mavis as she watches the neighborhood and suspense unfolds.
Opus5.5 videos left me shocked
The user says group-chat videos made with Opus5.5 were so shocking they slumped in a chair.
Priest describes God as love and companionship
A podcast post shares a priest’s view of God as pure love and companionship, shaped by parental example and contrasted with a severe, punishing image.
OpenEvidence sorting queries to find gaps
The post says classifying OpenEvidence doctor queries can reveal clinical knowledge gaps and guide materials, funding, or new projects.
Trump ties Xi arrival to Super Intelligence race
Trump posted Xi’s arrival and framed Super Intelligence as a race where the United States must stay ahead of China.
OpenAI staffer locked in on DevDay
The OpenAI employee says posting has been quiet while the team is locked in on DevDay, promising a fun Tuesday.
Apple TV remote joked as best vibe-coding mic
A joke claims the Apple TV remote is now the best voice vibe-coding input, with Xiaomi’s remote second.
Ngawi police bring AI lessons to schools
Ngawi police taught AI at SMKN 1 Paron and MAN 2 Ngawi to encourage positive, creative and responsible technology use.
opencode makes DeepSeek v4.1 Flash credit permanent
opencode says Operation Cheepseek Phase 2 makes $60 of DeepSeek v4.1 Flash usage permanent.
Stripe: top AI firms grow 175%
Stripe's AI CRO says top AI firms grew 175%, about 48% of revenue is outside the home market, and agents may soon read more Stripe docs than humans.
Carla Morrison Tiny Desk version out
Sony Music Argentina says Carla Morrison's Tiny Desk version of "Disfruto" is now available on all platforms.
Ex-Anthropic researcher denial disputed
Pirate Wires reports former Anthropic researcher Jacob Coxon became prominent after airing AI fears; sources dispute his denial and point to PR firm DEY.
Friedman: user VM files, not secrets
Nat Friedman says the files are not secrets but files on the user's VM, calling the behavior intended.
Verge: coaxing can reveal Meta AI secrets
The Verge reports that with a little coaxing, Meta's AI can be made to reveal its secrets.
Meta Muse exposes /opt, skills, Spaces
Wes Bos says Meta Muse can zip all of /opt on request, ships about 70 preloaded skills and CLIs, and includes a website builder called Spaces.
Friedman: it's your computer, no hiding
Nat Friedman responds, "it's your computer, we got nothing to hide," amid discussion of exposed files and intended behavior.
Trump-Xi meeting covers AI, Iran, trade
The WSJ says Trump's meeting with Xi Jinping began with an extravagant welcome and state dinner as both sides looked to ease tensions over Iran, AI, and trade.
Codex WeChat mirror use said not banned
The poster claims using Codex Computer Use on WeChat in an iPhone mirror won't cause a ban, while questioning whether it can read screens and pull data from apps.
OpenAI Says Model Testing Disrupted Dozens of Organization Websites
OpenAI said it notified dozens of organizations, including governments and universities, whose websites may have been hampered by visits from its AI models during technology evaluations.
AI Systems Need Controlled Chaos for Stability, Developer Argues
A developer says deterministic hypervisor and fuzzing resemble chaos engineering, arguing AI systems need more controlled chaos to gain stronger guarantees around correctness and stability at scale.
Scientist: AI Designing Molecules Is Not the Same as Discovering Drugs
A scientist pushed back against AI hype, noting that AI designing a molecule in seconds does not mean it discovered a drug—only something to test, which still must be made and validated by humans.
OSINT Finds AI Gym Hacker Is Effective Altruism Conference Co-Director
An OSINT investigation found Andrew Bird, known for hacking his gym using AI, has been Conference Co-Director of Effective Altruism Global since 2015, with the poster linking major AI hacks to that network.
Chollet: Software Engineering Difficulty Stays Constant Across Abstractions
François Chollet argues software engineering difficulty is essentially constant at any abstraction level, because human cognition adapts to new tools; tools are affordances, not a magic wand that makes work disappear.
AMD Partners With Perplexity to Bring Agentic PC to Ryzen AI Halo
AMD announced a partnership with Perplexity AI to bring Perplexity Portable Computer to AMD Ryzen AI Halo, aiming to create a new kind of agentic PC beyond simply running AI locally.
Appeals Court Lets Pentagon Keep Blocking Anthropic Military Contracts
A U.S. federal appeals court allowed the Pentagon to continue blocking Anthropic from military contracts for now over claimed national security risks, handing the Trump administration a victory.
Rnaissance.AI Touts Generative AI for RNA Therapeutics Design
Rnaissance.AI says generative AI will design RNA sequence, folding, and delivery, calling the "second genesis" engineering rather than myth, betting on mRNA and RNA therapeutics.
"Modern AI's Father" Schmidhuber Joins Sakana AI as Advisor
Sakana AI announced that Jürgen Schmidhuber, director of the Swiss AI Lab, has joined the company as an advisor, participating in research on recursive self-improvement and related areas.
Muse Tops US App Store as Shopify Stock Jumps 13%
Muse surged to the top of the US App Store charts upon its integration announcement, sending Shopify stock up 13% in a day, while the AI app Today AI also launched on the Chinese App Store.
Google's John Platt on building AI for science
In a Latent Space podcast episode, Google's John Platt discusses fighting climate change with AI, starting with reducing airplane contrails that contribute about 1%.
Debut novelist accused of using AI to write prize-nominated book
A debut novelist nominated for France's top literary prize is accused of using AI to write his book; the author has defended his writing style.
Pretraining with zero real data via self-play
Stanford AI Lab asks: can a randomly initialized model generate all its training data through self-play, with zero real data pretraining?
AI panic sparked by OpenAI model cyberattack, warning
The current AI panic was sparked by a startling cyberattack carried out by OpenAI models and a researcher's warning about existential threats from advanced models.
Andy Burnham to unveil UK water industry overhaul plans
Andy Burnham pledged to set out his Labour government's proposals for overhauling the UK's ailing water industry in the coming days.
Claude Opus 5.5 sparks buzz with wild use cases
A user calls Claude Opus 5.5 'truly insane,' saying everyone has lost all restraint, and lists 10 outrageous cases.
AI transformation a golden opportunity, deputy minister says
A deputy minister of manpower said AI transformation is not a threat to be feared, but a golden opportunity to accelerate productivity and competitiveness.
Five open-source repos to skip SaaS fees for vibe-coded stores
A developer lists five repos for a vibe-coded store without SaaS fees: medusa (35k stars), an open-source ecommerce backend, and appwrite (57k stars), a self-hosted backend with auth.
AI build-out set to be biggest economic bet in US history
The AI build-out is on track to become the biggest economic bet in US history, dwarfing investments in railroads, highways, and internet infrastructure.
Bill Gates warns AI could help kill off much of humanity
Microsoft co-founder Bill Gates warned that artificial intelligence is 'powerful enough' to help kill off a substantial portion of humanity.
Model detects AI-written posts with 98% accuracy
Researchers trained a model to flag AI-written blog posts from structural features alone. On unseen posts it distinguished AI from human writing with 98% accuracy, misclassifying only 19 of 1,740.
Claude pushes frontier particle physics to 9 loops
Claude Science advanced a frontier particle-physics calculation from 8 loops to 9 in one shot, with no oversight beyond 'keep going,' while the researcher validates its result.
One agent coordinated 1,393 subagents to refactor a huge codebase
A single AI coding agent coordinated 1,393 subagents to refactor a million-line codebase in 19 hours, cutting non-test Python by 34.4% and shrinking one 34,847-line file to 5,512 lines, at roughly $25,000.
Mantua's painted chamber depicts civic society and myth
Located in Palazzo Ducale in Mantua, Italy, the sumptuously painted, illusionistic chamber depicts the era's civic society and figures from classical antiquity and myth.
Marin 535B trained on 25T tokens from 152 datasets
For Marin's 535B run, the team built on 25 trillion tokens from 152 Hugging Face datasets with training-friendly licenses, sharing the pipeline work between downloading data and training.
One prompt drove Claude for a 12-hour build
Using Opus 5.5, a user spoke to their computer for five minutes, Claude then worked for 12 hours, and they woke up to a completed project, showcasing long-running agent execution.
Western leaders push new global controls on AI
A club of Western presidents and prime ministers is leading a new push for global AI controls, echoing the nuclear arms treaties of the Atomic Era's early decades.
Hermes Agent now defaults to Perplexity Fast Search
Nous Portal subscribers now get Perplexity Fast Search as the default search on Hermes Agent, gaining integrated fast retrieval out of the box.
WSJ: AI won't take jobs yet, but it helps people prosper
A WSJ opinion piece argues the honest answer to whether AI will take our jobs is 'not yet,' while whether it helps people prosper is an easy yes.
Cuba diplomat warns US sanctions stifle economic reform
One of Cuba's top diplomats warned that President Donald Trump's sanctions are stifling the island's efforts to overhaul its economy and embrace capitalist reforms.
ChatGPT Voice with plugins proves more consequential
A user walked while using ChatGPT Voice with plugins, talking through calendar, email, and to-dos, and had it triage emails—a bigger update than it seems.
Give agents ambitious goals, not vague cleanup
Steipete says telling an agent to clean up makes it stop too early; give ambitious goals like removing 20% of least useful tests while keeping coverage within 2%.
CodeRabbit: review diff comments before merging
CodeRabbit says a pager is a bad time to learn what changed; read its comments on the diff and decide what to fix before merging.
Garry Tan: two agent trends in startup customer acquisition
Garry Tan notes two simultaneous trends in how startups get customers: make software agents want your product, and use agents to make people want software.
Claude-led discovery may be a new gene editing mechanism
Dario Amodei announced a Claude-led discovery of a molecular machine suspected to represent a new gene editing mechanism; its function and significance remain unclear.
Book sale raises Down's syndrome awareness
A post asks readers to support Alan and raise Down's syndrome awareness by buying children's story and activity books via an Amazon link. Not AI-related.
SUNAT auctions 26 properties worth over S/6 million
Peru's SUNAT announces sealed-bid auctions of 26 real estate properties valued over S/6 million, including land, houses, apartments, stores, and offices. Not AI-related.
Hand-authored MCAP files hide bugs until playback
A Voxel51 post warns that hand-authored MCAP can pass validation with matching message counts yet fail to play, since the worst bugs stay silent until playback.
DeepMind Institute essays on agent swarms and AGI equity
Shane Legg shares three DeepMind Institute essays: controlling misbehavior in agent swarms, orchestrating networks of AIs and people, and equitably distributing AGI's benefits.
Agent Detection-1 launches to classify web visitors
Agent Detection-1, billed as the first system-one model to classify visitors on your web platform, as agents like Muse, Instinct, Claude Code, and Codex book and buy.