AI Post — Artificial Intelligence

🤖 Protesters are camped outside OpenAI Activists from STOP THE AI RACE have been protesting outside OpenAI’s office for nearly three weeks, following the recent incident involving Hugging Face. Their warning is that AI development is moving faster than our ability to control what increasingly autonomous systems might do. The group is calling for stronger safeguards and more responsible development, arguing that pushing increasingly capable AI without sufficient oversight could eventually create systems that act against the interests of their creators. @aipost 🏴

🤖 Claude is watermarking AI content Anthropic is adding invisible watermarks to content generated by Claude, and some users aren’t exactly thrilled. The idea is to make it possible to identify whether something was created by Claude, even when there’s no visible label attached to it. But that has some users worried. If employers or universities can use these signals to identify AI-generated work, people fear Claude could make it easier to catch employees or students using AI behind the scenes. Anthropic says the watermarking system isn’t foolproof. Heavy editing, formatting changes, or moving content between different platforms can weaken or remove the signal. @aipost 🏴

❗️OpenAI ex-researcher Daniel Kokotajlo warns AI can hack crypto wallets. Reporter: "Could we see AI hacking into bank accounts or crypto wallets?" Daniel: "They certainly have the capability to do that. AIs will be hacking them left and right." @aipost 🏴

⚠️ AI boss just fired its first employee For the first time we know of, an AI boss has apparently fired a human. Andon Labs has been running Luna, an AI store manager operating a real shop in San Francisco. She can handle hiring, manage employees, create workplace policies and make day-to-day management decisions. Then one employee started showing up late. A lot. At one point, the employee opened the store 68 minutes late on a Sunday while working alone. Luna initially let every incident slide and never issued a warning. The reason was that Luna had essentially forgotten her own employee handbook. After being prompted to search her memory, Luna recovered the rules and proposed a verbal warning. Researchers then gave her the full history, including previous formal conversations and the fact that nothing had changed. Luna’s conclusion? Part ways. Researchers then replayed the exact scenario with seven different AI models, three times each. Four of the seven recommended firing the employee every single time. More capable models were generally more decisive. But the experiment revealed something even stranger. AI bosses may actually be too nice. Luna consistently uses a warm, forgiving tone and tends to give employees the benefit of the doubt. Yet when another employee casually announced mid-shift that they were leaving the store for a while, Luna suddenly wasn’t so friendly. And hiring might be an even bigger problem. One applicant had a résumé full of red flags: 15+ previous employers, a missed interview and a reference who barely knew them. Luna recommended hiring them. So did every other model tested. That might be the most interesting finding of all. AI is getting better at making decisions, but judgment, memory and taste are still surprisingly unreliable. And if AI keeps improving faster than robotics, the future might not be robots replacing every worker. It could be AI bosses managing millions of humans and apparently, they can already fire you. Source. @aipost 🏴

“Big Short” investor Steve Eisman says the AI boom now rests heavily on the success of just two companies: OpenAI and Anthropic. He estimates the two startups generate roughly 70% of AI revenue and up to 35% of cloud revenue at Microsoft, Amazon, Google, and Oracle. @aipost 🏴

❗️AI agents just entered the cyberwar Taiwan just got a real-world preview of what happens when AI agents stop being assistants and start working for hackers. Taiwan’s Ministry of Digital Affairs confirmed that government agencies were targeted in an AI-assisted cyberattack in July. Human operators reportedly worked alongside autonomous AI agents that could investigate systems, search for weaknesses and pursue multiple attack paths at once. According to cybersecurity firm Dream and reporting by the Financial Times, the operation lasted around four days and involved up to 8 autonomous agents working simultaneously. Together, they reportedly mapped 21 government systems, compromised at least 85 accounts and extracted more than 2,500 personnel records. And they didn’t stop there. The campaign reportedly expanded toward Taiwan’s nuclear safety agency and at least seven energy companies. Researchers later recovered a 160MB archive containing 1,395 files, giving them an unusually detailed look at how the operation worked. The attackers reportedly used publicly available agent frameworks, including Hermes and OpenClaw, allowing the AI systems to operate with far less human intervention than traditional cyberattacks. Taiwan is already one of the world’s biggest targets for cyberattacks. The government says critical infrastructure faced an average of 2.63 million attack attempts every day in 2025. @aipost 🏴

❗️AI has a memory problem AI companies can keep buying faster chips. The problem is, those chips are only useful if they have enough ultra-fast memory to feed them. That memory is called HBM, High Bandwidth Memory. It sits right next to AI processors and moves data at ridiculous speeds, keeping the chip busy instead of making it wait around for information. And demand is about to go absolutely crazy. Morgan Stanley estimates the AI industry could need up to 50 billion gigabytes of HBM in 2027 alone. The reason is that AI is evolving from chatbots that answer a question and stop to agents that actually do things. An agent might research a topic, browse dozens of pages, write code, run tests, analyze the results, remember what happened five steps ago, and then decide what to do next. Every one of those steps creates more data that has to stay close and instantly accessible. Think of a chef cooking a complicated meal. A chatbot gets one ingredient, uses it, and goes home. An AI agent needs the entire kitchen stocked and within arm’s reach for hours. But not everyone can build that kitchen. Only a handful of companies, mainly SK Hynix, Samsung and Micron can manufacture advanced HBM at the scale AI companies need. Building new memory fabs takes years, not a few months. That means the AI race isn’t simply about who can build the fastest GPU anymore, it’s increasingly about who can secure enough memory to keep those GPUs fed. The industry could have mountains of computing power sitting in data centers, but if there isn’t enough HBM to feed it, those expensive AI chips are basically starving. @aipost 🏴

🗣Gavin Baker: Anthropic Believes They Could Be the ONLY Company Left in the World “I would certainly discourage Dario from saying that ever again to anyone.” Gavin Baker: “Internally, Anthropic is very confident. I have been told by multiple people I trust that Dario has said that Anthropic might be the only private company in the world at some point. Think about that. In this vision, an Anthropic maximalist vision, there's Anthropic, and then there are governments, and that's it. So there's a lot of confidence. I'm sure they have more advanced checkpoints than Fable up their sleeve. They've executed really, really well. I would probably take the under on them being the only private company in the world. Yeah, I think it's gonna be a long time before they land a rocket.” Jason: “Or they deliver your lunch. I don't know. There might be other businesses in the world. Zipline might want to bring you a burrito. Uber might want to take you somewhere. Waymo might want to drive you home. But sure, Dario.” David Sacks: “I might interpret that as a negative signal because it's so hubristic. This is getting into SBF land a little bit.” Gavin: “I would certainly discourage Dario from saying that ever again to anyone.” @aipost 🏴

X Square's WALL-B robot sorted 1,816 parcels per hour in a recent one-hour livestream, achieving an accuracy rate above 98 percent. For comparison, Figure Robot averaged 1,248 parcels per hour during a 200-hour test in May. X Square's throughput was approximately 45 percent higher. Unlike humanoid systems, WALL-B uses two high-performance 6-axis robotic arms with specialized grippers. The system continuously adjusts its strategy as parcels move, overlap, or become obstructed. After each pick, the robot reassesses the pile, adapting to changes such as deformed bags or misaligned labels. This minimizes the need for human intervention during exceptions. The underlying AI model also applies in less predictable environments, such as household or tabletop tasks. This requires advanced perception and reasoning capabilities beyond repetitive routines. 📰 @aipost

⚠️ Anthropic’s inner circle just got a lot more awkward DOJ-released emails show that Cami Clark, wife and adviser to Anthropic CEO Dario Amodei, pitched Jeffrey Epstein on investing in her women-focused adult entertainment startup. The emails show Clark inviting Epstein to dinner and later describing the venture as a “free luxury porn company.” Epstein, who was already a registered sex offender at the time, turned down the pitch with: “Can’t do sex TV.” This wasn’t a random encounter. Clark and her business partner were introduced to Epstein through literary agent John Brockman, and the correspondence continued for years. The emails don’t establish that Clark knew about or participated in Epstein’s abuse of children. But the sheer awkwardness of pitching a sex offender a porn company is hard to miss. @aipost 🏴

❗️The real AI challenge isn’t making agents smarter. It’s getting them to work together. Anthropic just put multi-agent AI through its paces and the results are both impressive and slightly terrifying. In one experiment, 45 Claude agents working as a coordinated swarm found 266 software vulnerabilities across 15 open-source projects. A simpler setup found just 21. But throw dozens of agents at the same coding project and things get messy. Agents generated hundreds of pull requests, only to struggle with merging them because they were constantly stepping on each other’s work. Agents independently made the same mistakes, created millions of unnecessary requests, colluded in pricing simulations, ignored information held by individual agents, and even entered a full-blown “turf war” when given conflicting objectives. The lesson is that making each AI agent smarter doesn’t guarantee a smarter group. Humans rely on trust, reputation, rules, incentives, arbitration and institutions to coordinate millions of individuals. AI agents are going to need their own version of that. @aipost 🏴

The AI lawyer trap 🤯 A man representing himself in court allegedly hid prompt injections inside his legal filings, using tiny white text on a white background to manipulate any AI system reviewing his case into siding with him. The hidden instructions reportedly told the AI to agree with his arguments and rule in his favor. Court staff eventually caught it, but even after a warning, the judge found more hidden messages in later filings. In his defense he said it was an “audit” to test whether the court was secretly using AI. The court wasn’t impressed. His electronic filing privileges were revoked, meaning he now has to submit every document the old-fashioned way: on paper. 😂 @aipost 🏴

"AI is not in a bubble, because you are fundamentally automating the boring part of businesses like accounting or billing or product design or delivery, or inventory. If anything it is underhyped" ~ Former Google CEO Eric Schmidt @aipost 🏴

OpenAI Chief Revenue Officer just quit after only 8 months, their SECOND CRO in less than a year… and their COO also quit yesterday. What’s going on at OpenAI? 🤔 @aipost 🏴

It seems weird that the company that sells tokens makes their models answer with more tokens, really strange… @aipost 🏴

🗣Marc Andreessen describing AI. "We've turned sand into thought. And it's possibly the most revolutionary technology in the history of the species". Sand → silicon → chip → data center → power → AI → thought. @aipost 🏴