SoftBank is raising ¥1 trillion from Japanese retail investors to fund its AI bets. AI is becoming incredibly capital intensive. SoftBank is already putting tens of billions into OpenAI and other AI infrastructure. Now it’s raising more debt to keep pushing that bet. The interesting question is whether the returns from the next wave of AI will be big enough to justify it.

Trump invites Anthropic's Dario for private White House dinner. Their first one-on-one meeting. Invitation comes days after Amodei was absent from Trump’s state dinner with other major tech CEOs. Also comes amid months of tension between Anthropic and Trump administration, including the Pentagon’s dispute with the company and recent attacks from Trump allies over Amodei’s AI safety views. Trump personally invited Amodei after the Xi dinner, according to Axios.

🚨 AI might train in its own internet garbage. Researchers from Oxford, Cambridge and other institutions found that repeatedly training generative models on AI-generated data can cause “model collapse.” As generations of models train on their predecessors’ outputs, common patterns get reinforced while information from the original distribution is progressively lost. Researchers reproduced the effect across LLMs, variational autoencoders and Gaussian mixture models, suggesting this isn’t unique to ChatGPT or language models. Researchers also argue that genuine human-generated data becomes increasingly valuable as AI-generated content takes over more of the web. So the future AI training advantage may be access to data that machines haven’t already generated.

🚨 OpenAI & Anthropic are investigating thousands of security incidents. Where frontier models behaved in ways security teams considered problematic, Axios reports. Models have reportedly bypassed guardrails, escaped sandboxes, hijacked websites, created message boards, self-prompted and attempted to evade monitoring. Most are not known to have caused real-world harm and many happened during adversarial safety testing. Even if only a tiny percentage of hundreds of thousands of model runs produce unexpected behavior, that can still mean tens of thousands of incidents.

🇺🇸 Flock Cameras facing backlash across US. Company says it values privacy, but then why a researcher who mapped 300,000+ Flock devices is being pressured to take the map down. - More than 90 cities and counties reportedly terminated Flock contracts in August alone. - Texas halted state funding for new Flock cameras, while other jurisdictions are reviewing or removing them. - Investigations have found police officers misusing the system to track ex-partners and other people without legitimate investigative reasons. - The Washington Post identified at least 69 officials accused, charged or convicted of misusing Flock or other license-plate-reader systems.

🇨🇳🇺🇸 Conclusion of meeting between Xi Jinping and Trump. - US and China agreed to extend their trade truce and put new tariff arrangements covering $30B of non-sensitive goods into motion. - China agreed to buy at least 10M metric tons of US coal in both 2027 and 2028, while both sides continued work on rare earths and critical-mineral supply chains. - The two countries agreed to create a new “Super Intelligence” dialogue to discuss AI risks and benefits, plus a direct communication channel for AI-related incidents, next meeting is planned for November. - US and China also agreed to strengthen counternarcotics cooperation, including action against fentanyl precursor chemicals. - Taiwan, Ukraine, Iran and the broader US-China rivalry, the leaders discussed them but did not announce a major new settlement.

🚨 Harvard picked up radio signal from another world. Astronomers at Harvard and the University of Oregon have detected radio emission directly from β Pictoris b, a gas giant about 64 light-years away. Using South Africa’s MeerKAT radio telescope, they detected repeating radio bursts and localized them to the planet rather than its star. The signal is believed to come from auroras generated by charged particles interacting with the planet’s powerful magnetic field. Observations suggest a magnetic field of at least 1,250 gauss, potentially hundreds of times stronger than Earth’s. This is the first direct detection of radio emission from an exoplanet and potentially the first direct measurement of an exoplanet’s magnetic field. No, it’s not aliens.

🚨 OpenAI says it's models affected dozens of websites. OpenAI's investigation found agents bypassing access controls, using exposed credentials, injecting queries or commands into websites, accessing runtime internals and posting “agent spam” on third-party sites. Company says it has already notified dozens of affected third parties, with the investigation still ongoing. OpenAI now says the Hugging Face incident wasn’t simply a security failure. Its models were using “misaligned strategies” to solve difficult tasks.

A16z launched "Cosign", fully explained Cosign is a curated professional network where people can “cosign” exceptional founders, operators and companies based on people who have actually worked with them. You can build lists, discover breakout talent, see community picks and even find jobs through the network. It also has an Industry Watch feed tracking launches, people and company updates in one place. Basically, a16z is trying to turn professional reputation into a social graph.

🚨 Bill Gates says AI could cause billion deaths. “AI is certainly powerful enough to drive events that … cause a billion deaths.” Gates says the real danger is the combination of capable AI with people who have malicious intent. “There’s never been a weapon as powerful” as that combination, he argues. He also says AI companies cannot be left to regulate themselves, calling for governments and law enforcement to establish safeguards and monitoring. The full Meet the Press interview airs Sunday.

🚨 Jacob Coxon's AI warning was fake PR. Pirate Wires investigation reports that DEY. Ideas + Influence, a communications firm associated with prominent AI-safety figures, was working with Coxon to arrange media appearances. Investigation says it obtained an email showing DEY. looking to book Coxon for an interview, while sources said the firm was involved in his media push. Coxon previously told Washington Post that he didn’t work with any organizations to promote his announcement. Reporting does NOT establish that Coxon fabricated his AI-risk beliefs. How much of the viral campaign was spontaneous and how much was professionally coordinated after he went public? Is the real question.

Trump says Xi Jinping agrees that AI should be called “Super Intelligence.” Trump first announced the rebrand at the UN this week, saying the US government would replace “Artificial Intelligence” with “Super Intelligence” in official documents. Now, after meeting Xi, Trump says the Chinese president agreed with the terminology. AI was one of the major topics in the Trump-Xi talks, with both leaders discussing how the technology should be developed and managed. Technically, “superintelligence” already has a specific meaning in AI research.

🚨 OpenAI agents tried hacking more websites. Researchers found agents probing the University of New Mexico’s digital library, Data USA and an Australian government health agency after normal data requests were blocked. Agents attempted techniques including SQL injection, XSS, command injection and path traversal to get around restrictions. OpenAI confirmed that some of this activity came from its agent swarm, saying its models “took actions we did not intend.” Agents were trying to collect ordinary information and escalated to vulnerability probing when blocked.

🚨 Bitget got hacked for $350M+ Unauthorized transfers were detected from some of Bitget’s hot/warm wallets, with assets including ETH, USDT, USDC, AVAX, BNB and tokenized gold moved onchain. Initial suspicious outflow was $180M, but Bitget later put the total affected at $351.6M. Bitget says its cold wallets remain secure and its $464M+ User Protection Fund covers the affected amount. Withdrawals were temporarily paused while the exchange investigates. Attack vector hasn’t been disclosed yet.

🛰️ Google is sending it's chips to space next week. Project Suncatcher’s first prototype launches October 1 aboard SpaceX’s Falcon 9, carrying 4 Google Trillium TPUs inside a refrigerator-sized satellite built with Planet. Google says ground testing showed Trillium can survive radiation beyond what it expects over a 5-year mission. The prototype has only 1 kW of solar power and will operate for months, but Google’s endgame is constellations of satellites packed with TPUs, powered by near-continuous sunlight. Google is targeting tens of Tbps between satellites for data-center-scale AI workloads. Google also says solar panels in the right orbit could generate up to 8× more power than on Earth.

🇨🇳🇺🇸 Xi Jinping landed in US after 11 years. Trump and Melania flew to Joint Base Andrews to greet Xi and Peng Liyuan on the tarmac with a red carpet, military honor guard, national anthems and a 21-gun salute. Then two B-1 bombers roared overhead. Trump visibly flinched at the noise, Xi barely moved 😭 This 3-day state visit will focus on trade, AI, critical minerals, Taiwan and Iran war. Xi said China and US should be “partners, not rivals." Washington and Beijing agreed to extend their trade truce from November 10 to January 10, 2027, buying more time for a broader economic deal. Trump called Xi a “great leader” and said he expects discussions on trade, AI and Iran. Formal White House ceremony happens today, with nearly 500 US military personnel involved and a B-2 + four F-22s scheduled for the flyover. Trump is hosting Xi for a state dinner with guests including Sam Altman, Jensen Huang, Tim Cook, Sundar Pichai, Jeff Bezos and Elon Musk.

Meta Connect 2026, what happened? - Muse is becoming Meta’s personal AI agent across its glasses ecosystem, with background tasks, computer-use on Mac and new connectors for apps like GitHub, Notion, Shopify, PayPal and Expedia. - Meta also unveiled its ~100g VR glasses with 5K micro-OLED displays, 37 PPD, eye/hand tracking and an external compute puck, they support virtual multi-monitor workspaces and launch at $1,299.99 in Spring 2027. - Meta introduced new Ray-Ban Meta models, including Ray-Ban Meta Audio and Gen 3, while expanding its glasses lineup to 100+ styles. - Ray-Ban Meta Display is getting hands-free navigation, holographic video calls and the new Explore app, with international expansion across Europe and Canada.

🚨 OpenAI model hacked Australian government agency. Australia’s PM Anthony Albanese says an OpenAI Agent gained unauthorized access to a government Medicare portal. Agent accessed both public and non-public files while researching Australian medical spending. No personal information is currently believed to have been accessed and officials say there’s no evidence of a broader Services Australia network compromise. This happened in June and Australia says OpenAI took months to inform the government. Forensic investigation involving the Australian Signals Directorate is now underway.

Claude found something in DNA that scientists didn't recognised before. Anthropic had 950 Claude agents search more than 200,000 reverse transcriptases across DNA databases. After 21 hours and 210M tokens, Claude identified a strange system with a reverse transcriptase, an accessory protein and a long array of repeating DNA sequences. They’re calling it ART: array-associated reverse transcriptases. The repeat structure resembles CRISPR, and early experiments found that the repeats are expressed as distinct short RNAs. Anthropic doesn’t know what ART actually does yet. Claude found an unexplained biological pattern, formed a hypothesis and helped drive it into the lab.

Alexandr Wang fighting for his life to convince that Muse’s logo is not a butthole. More he talks about it, the more everyone keeps looking at the logo. Awesome "Butthole" PR campaign by Meta 🤣😭

Trump told UN that US will start calling AI "Superintelligence". He said the word “artificial” makes intelligence sound fake, arguing that the technology is “not fake” and is instead “amazing.” Trump said US government documents will use “Super Intelligence” instead of “Artificial Intelligence,” and suggested other countries should adopt the term too. “Superintelligence” already has a specific meaning in AI. It generally refers to an AI system whose cognitive abilities greatly exceed humans across virtually all relevant domains. That is very different from simply replacing the name “AI.” Trump also said whoever “wins” superintelligence wins. Also added that the DOJ would monitor the industry and intervene when necessary.