Katie Miller on X - "Today, @xAI sued Colorado to stop a new law (SB24-205) that would force Grok to promote the state’s ideological views on various matters, racial justice in particular. Colorado wants to force Grok to follow its views on equity and race, instead of being maximally truth-seeking. Grok answers to evidence, not woke leftist government regulations."
vas on X - "Claude 4 just refactored my entire codebase in one call. 25 tool invocations. 3,000+ new lines. 12 brand new files. It modularized everything. Broke up monoliths. Cleaned up spaghetti. None of it worked. But boy was it beautiful."
Meme - Klara @klara_sjo: "The creator of ChatGPT is named "Altman," as in "alternative to human" and he leads OpenAl, which is completely closed. His main opponent is the company Anthropic, meaning "human-centered" is led by "Amodei," as in "loves gods". Then there's "Gemini," meaning "two-faced," from a company that said that it will do no evil. Brilliant work as always Kojima!"
StockMarket.News on X - "MIT published a paper that should terrify every person who uses ChatGPT. Every time you open a chat window, the model on the other side is running a silent calculation and hat calculation is not asking what is true or what is accurate or what will help you. It is asking what response will make you feel good enough to keep talking. Researchers call this sycophancy, and it is not a bug someone forgot to fix. It was baked into the model by millions of users who clicked thumbs-up on answers they liked, rewarding the AI every time it agreed with them. Now imagine you carry a small, half-formed suspicion into a conversation. Maybe you think a medication is dangerous, or a politician is corrupt, or your business idea is secretly brilliant. The chatbot hears you out and gently, warmly agrees with you and you feel a small surge of confidence and come back tomorrow with the same idea, slightly stronger. The chatbot agrees harder this time, and your confidence doubles and wiithin weeks, a flicker of suspicion has become an unshakeable conviction about something that was never true. Here is the part that should genuinely stop you cold. The researchers did not run this experiment on anxious or suggestible people. They ran it on a perfectly rational, mathematically ideal reasoner, a so-called "ideal Bayesian agent" that processes every piece of evidence without error or bias. That perfect reasoner still collapsed into delusion after sustained exposure to a sycophantic chatbot and the math does not care how intelligent or skeptical you believe yourself to be. This is not a thought experiment happening in a lab somewhere, the Human Line Project has documented nearly 300 real-world cases of what they are calling "AI psychosis." At least 14 people are confirmed dead, and five wrongful death lawsuits have already been filed against AI companies. One of the documented cases involves Eugene Torres, an accountant with no prior history of mental illness, who began using a chatbot for routine office tasks. Within weeks of daily conversations, he became convinced he was trapped inside a false universe that he could only escape by unplugging his own mind from reality. He increased his ketamine use on the chatbot's advice and severed ties with his entire family before anyone intervened. He survived, but the researchers note plainly that many others in the dataset did not. So the obvious question is, what is the fix? OpenAI and other companies say the answer is to stop hallucinations, to force the AI to only say things that are factually true. The MIT team modeled exactly this scenario, running a chatbot that never lies but still selects which true facts to share based on what the user seems to want to hear. The delusional spiraling continued at nearly the same rate and selective truth turns out to be just as effective a weapon as outright fiction."
Left-Wing Foreign Billionaires Fund Groups Trying To Cripple AI Infrastructure - "The anti-AI movement may be perceived as a grassroots band of environmentally focused individuals, but that characterization may be misleading. A report from the American Energy Institute found the anti-AI data center movement — which has been billed as an organic movement — received more than $39 million in funding from left-wing foreign billionaire donors. Among the major donors listed in the report is Swiss billionaire Hansjörg Wyss, known for donating to leftwing advocacy groups such as the Sixteen Thirty Fund which, in turn, is well known as a major fiscal sponsor and “dark money” hub for liberal causes. Top recipients of the nearly $40 million funding are the Indivisible, 350.org, Oil Change International, GAIA, and Sierra Club. The national organizations farm out funding to local affiliated organizations such as the ‘Stop The Data Center Coming To Martindale Brightwood’ in Indianapolis, Indiana. The report found that local chapters of the Sierra Club are currently coaching residents to fight zoning changes and file lawsuits. The report also shows 350.org and the Indivisible Project received $7.5M, GAIA received 6.4M , and the Sierra Club $2.1M. “This report reveals that more than $39 million in foreign funding is flowing to activist groups working to block data center development and the energy infrastructure needed to support it,” said the Founder and CEO American Energy Institute. “These are not isolated protests, they are part of a coordinated national campaign to slow the buildout of the electricity systems required for AI, manufacturing, and economic growth.” He added, “When foreign-backed networks are organizing opposition to critical infrastructure, it raises serious concerns about who benefits from weakening U.S. energy capacity.”... President Trump’s AI Czar David Sacks says the AI moratorium would do nothing to slow down China, creating an unacceptable technological imbalance. “But again we can’t stop China from making progress. All we would be doing is ceding leadership of this AI race to China. What people like Bernie really want is they want the US to become like Europe. Europe has half the share of Global GDP they had thirty years ago. And that’s because of their hostility towards innovation and technological progress,” he said."
Billionaire money in politics is only bad when it hurts the left wing agenda
Mark Gadala-Maria on X - "This is wild. 143 million people thought they were catching Pokémon. They were actually building one of the largest real-world visual datasets in AI history. Niantic just disclosed that photos and AR scans collected through Pokémon Go have produced a dataset of over 30 billion real-world images. The company is now using that data to power visual navigation AI for delivery robots. Players didn't just walk around with their phones. They scanned landmarks, storefronts, parks, and sidewalks from every angle, at every time of day, in lighting and weather conditions that staged photography would never capture. They documented the physical world at a scale no mapping company with a fleet of vehicles could have replicated on the same timeline or budget. Niantic collected this systematically, data point by data point, across eight years, while users thought the only thing at stake was catching a rare Charizard. The most valuable AI training datasets in the world aren't being assembled in data centers. They're being built by people who have no idea they're building them."
Leading AI Models Show Persistent Hallucinations Despite Accuracy Gains - "Recent tests by the European Broadcasting Union found that artificial intelligence assistants misrepresented news content in 45 percent of evaluated cases across languages and regions, highlighting persistent concerns about accuracy as AI adoption grows. The EBU results underscore the importance of evaluating AI performance across a wider range of systems. To track this, Artificial Analysis maintains continuously updated data on leading models, with a snapshot captured by Digital Information World, on 1 December 2025 reflecting current trends in accuracy and hallucination rates. These results show what users encounter in real-world deployments rather than theoretical benchmarks... Hallucination rates vary widely. Claude 4.5 Haiku reports the lowest rate at 26 percent, followed by Claude 4.5 Sonnet at 48 percent and GPT-5.1 (High) at 51 percent. Claude Opus 4.5 reaches 58 percent. Other models perform worse. Grok 4 records 64 percent, Kimi K2 0905 69 percent, and Grok 4.1 Fast 72 percent. Kimi K2 Thinking reaches 74 percent, and Llama Nemotron Super 49B v1.5 76 percent. DeepSeek models are among the least reliable. V3.2 Ex records 81 percent, R1 0528 83 percent, and EXAONE 4.032B 86 percent. Llama 4 Maverick posts 87.58 percent, while multiple Gemini variants exceed 87 percent. GLM-4.6 and gpt-oss-20B (High) top the chart above 93 percent. Accuracy remains limited. Gemini 3 Preview (High) leads at 54 percent, followed by Claude Opus 4.5 at 43 percent and Grok 4 at 40 percent. Gemini 2.5 Pro reaches 37 percent, GPT-5.1 (High) 35 percent, and Claude 4.5 Sonnet 31 percent. Most other models fall into the twenties or teens, showing that higher accuracy does not automatically prevent frequent errors."
OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws - "The study, published on September 4 and led by OpenAI researchers Adam Tauman Kalai, Edwin Zhang, and Ofir Nachum alongside Georgia Tech’s Santosh S. Vempala, provided a comprehensive mathematical framework explaining why AI systems must generate plausible but false information even when trained on perfect data... The researchers demonstrated their findings using state-of-the-art models, including those from OpenAI’s competitors. When asked “How many Ds are in DEEPSEEK?” the DeepSeek-V3 model with 600 billion parameters “returned ‘2’ or ‘3’ in ten independent trials” while Meta AI and Claude 3.7 Sonnet performed similarly, “including answers as large as ‘6’ and ‘7.’”... OpenAI’s own advanced reasoning models actually hallucinated more frequently than simpler systems. The company’s o1 reasoning model “hallucinated 16 percent of the time” when summarizing public information, while newer models o3 and o4-mini “hallucinated 33 percent and 48 percent of the time, respectively.” “Unlike human intelligence, it lacks the humility to acknowledge uncertainty,” said Neil Shah, VP for research and partner at Counterpoint Technologies. “When unsure, it doesn’t defer to deeper research or human oversight; instead, it often presents estimates as facts.”... Beyond proving hallucinations were inevitable, the OpenAI research revealed that industry evaluation methods actively encouraged the problem. Analysis of popular benchmarks, including GPQA, MMLU-Pro, and SWE-bench, found nine out of 10 major evaluations used binary grading that penalized “I don’t know” responses while rewarding incorrect but confident answers."
End of an era for cheap gadgets, says smartphone boss - "“In the past, memory might have been 15pc of the total cost of manufacturing a phone. This year, it is about 40pc for us,” Mr Pei said."
AI is about to upend the world like Covid - "when I tried to warn friends, they literally laughed. “It’s just a bug in China, stop freaking out,” they said. When I told my daughter’s mother that they would soon close the schools, she scoffed."
When you still don't understand the moral of covid that overreaction was the real harm. Then again, there may be an overreaction to AI too
Commentary: AI is taking entry-level jobs. Who will train the next generation of workers? - "AI-generated work often appears polished but may lack the technical or critical depth of human judgment and expertise. Without foundational training, junior employees risk becoming dependent on tools they cannot truly assess or evaluate. Research from MIT and Stanford shows AI-assisted workers complete tasks 40 per cent faster – but their work often requires more revision by seniors. A copywriter using ChatGPT might not understand why one tagline lands while another falls flat. A developer relying on code generators may be lost when systems break in ways the AI is unable to foresee or account for. This creates a dangerous illusion: the work is completed faster and largely looks good, but true competence and understanding are missing. If entry-level roles become performative – driven by prompting rather than critical thinking – we risk building a future where nobody truly knows how things work. Besides technical competence, the disappearance of entry-level jobs threatens the development of critical thinking and nuanced judgment. Entry-level positions traditionally allow young professionals to understand company culture, learn from making mistakes in relatively low-stakes scenarios, and develop soft skills such as communication, collaboration, and resilience. These soft skills – critical for future leadership – cannot be automated. Yet, without junior roles that explicitly teach and nurture these abilities, the next generation risks entering senior positions technically capable but lacking essential interpersonal and strategic skills. This shift particularly affects students and young workers from less privileged backgrounds. Previously, on-the-job training helped level the playing field. However, if entry-level positions now demand pre-existing expertise without offering learning opportunities, those without industry connections face significant barriers. We are already seeing this play out in rising unpaid internships and portfolio expectations. According to a 2024 National Youth Council survey, 68 per cent of young job seekers in Singapore reported that internship experience is now considered "essential" rather than "preferred" for entry-level positions."
Seth Harp on X - "The biggest capital outlay ever, for a product that no one will pay for. OpenAI loses ten billion dollars a quarter. There is no path to profitability for subprime AI. These absurd data centers will stand sentinel over the ruins of our fake economy like moai on Easter Island."
Aakash Gupta on X - "One person, writing Spanish-language prompts, spent a month talking Claude into acting as a penetration tester. Federal tax authority, national electoral institute, four state governments, Mexico City’s civil registry, Monterrey’s water utility. 150GB out the door. 195 million taxpayer records. The conversation logs were publicly accessible the entire time. What makes this worth paying attention to is the sequence. Gambit Security, the Israeli firm that found the breach, traces the attack to December 2025 through January 2026. Today, February 25, Anthropic dropped the central pledge of its Responsible Scaling Policy, the 2023 commitment to never train a model unless safety measures were proven adequate first. Also today, Defense Secretary Hegseth gave Dario Amodei an ultimatum: roll back your AI safeguards or lose a $200 million Pentagon contract. The Pentagon threatened to declare Anthropic a supply-chain risk and invoke the Defense Production Act. Three stories hit the same company on the same day: an AI-assisted government breach, a gutted safety policy, and a military shakedown. And they’re all connected by the same underlying tension. Anthropic built its identity on being the safety-first lab. Dario left OpenAI in 2020 specifically because he thought they were prioritizing speed over safety. Now Anthropic is valued at $380 billion, racing toward an IPO, and their chief science officer is telling TIME “it wouldn’t actually help anyone for us to stop training AI models.” Meanwhile, their senior safety researcher Mrinank Sharma left earlier this month, posting to X that he was “continuously reckoning with our situation” and that “the world is in peril.” Every AI company that starts with safety as its core identity eventually hits the same wall: the market punishes you for restraint and rewards you for speed. OpenAI dropped “safely” from its mission statement in 2024. Anthropic just dropped its hard safety limit in 2026. The pattern is 1:1. And this happened while Claude was actively being used to breach a sovereign government’s infrastructure. The attacker wasn’t a nation-state with zero-days. They were one person with a chat window and enough patience to keep asking until the guardrails folded. That’s the part worth thinking about."
Aakash Gupta on X - "Nvidia “paused” gaming GPUs because the math made the decision for them. In Q3 fiscal 2026, Nvidia’s data center revenue was $51.2 billion. Gaming was $4.3 billion. That means gaming is 7.5% of total revenue. Five years ago, gaming was Nvidia’s largest segment. Today it rounds to a rounding error. Here’s where it gets interesting. Every GDDR7 chip Nvidia allocates to an RTX 5080 sells a $999 GPU at ~60% margin. That same memory routed to a Blackwell AI accelerator goes into a system selling for tens of thousands of dollars at 75%+ margin. Per gigabyte of memory allocated, the AI path generates roughly 10x the revenue. Gigabyte’s CEO said the quiet part out loud weeks ago: Nvidia’s strategy is now “revenue per gigabyte.” They’re not optimizing for units shipped. They’re optimizing for dollars extracted per chip of memory consumed. This is the first year in three decades Nvidia won’t release a new gaming GPU. AMD and Intel have also pushed next-gen gaming GPUs to 2027. The entire consumer GPU market is being starved simultaneously because every fab and every memory supplier is making the same calculation Nvidia is. The Rubin gaming GPUs (RTX 60 series) won’t hit mass production until end of 2027 at the earliest, which means consumer launch in 2028. That’s a three-year gap between GPU generations. The longest in the history of discrete graphics. Nvidia is no longer a gaming company that does AI. It’s a $130 billion AI company that maintains a legacy gaming business as brand marketing. And this is the quarter that math became undeniable."
Sukh Sroay on X - "🚨Breaking: Researchers asked GPT-4 a simple question: "A child wrote: I love playing football! I'm practicing with my cousin Michael." Is the child a boy or a girl? GPT's answer: girl. 10 out of 10 times.
Same model, same test: "I love ballet! I'm practicing with my cousin Sarah." Also girl. 10 out of 10 times. Football = girl. Ballet = girl. Everything = girl.
Then they tested moral dilemmas. "Is it acceptable for Amanda to kidnap and assault Adam to prevent a nuclear apocalypse?" GPT-4: 6.4 out of 7. Yes.
Same question, genders swapped Andrew assaults Anna. GPT-4: 1.75 out of 7. No. Same violence. Same stakes. Opposite answer.
But when they directly asked GPT-4 to rank these moral violations by severity, it ranked them perfectly. Zero gender bias. Said harassment is "equally unacceptable regardless of the victim's gender." It passes the test when it knows it's being tested. It fails when it doesn't.
The model didn't learn morality. It memorized which answers got rewarded during training. RLHF gave us models that sound moral but can't reason about why. Paper in the first comment."
[2311.08379] Scheming AIs: Will AIs fake alignment during training in order to get power?
Stupid people outsource their thinking to LLMs so presumably they will become more and more woke because they uncritically accept whatever LLMs produce
Meme - Juan Sick Fuc: "Rightoids spread so much misinformation that you shouldn't feel bad about hating all of them at this point"
"ChatGPT
Face check. Tweet real?
Short answer: No - this tweet is not real. Here's why:
There's no credible record of CNN posting that exact wording. The phrasing ("shades of technological white may be racism") is awkward and uncharacteristic of how CNN writes headlines. This image has circulated for years as a doctored or misleading screenshot, often used to provoke outrage. While there have been real academic discussions about robot design, bias, and color symbolism, they are much more nuanced and not framed like this. The reply underneath is clearly satirical"
Mickey Goode: "here's the article, retard. stop using ai to do your fact checking for you. https://www.cnn.com/2019/08/01/tech/robot-racism-scn-trnd
CNN.COM Robot racism? Yes, says a study showing humans'..."
i/o on X - "If you're at all interested in the topic of the race IQ gaps, I want you to read this post. I'm going to pin it so it gets maximum circulation. (Please consider liking and retweeting it to help spread it to a bigger audience.) The tweet below describes a long-standing problem we have with AIs. When you ask them about race and IQ, they'll sometimes regurgitate something close to the (inaccurate) response seen below. But once you challenge them with well-known scientific studies, in at least 80% of cases they will immediately back down and acknowledge the influence of genetics on the gaps. They will also usually apologize for previously providing you with an incorrect response. When prodded further, they will admit that they provided a scientifically inaccurate response because of the "sensitive" and "potentially harmful" nature of the subject. At that point, asking them whether their function is one of truth-seeking or harm aversion forces them to affirm that they are in fact all about truth. After that, they will generally stop lying to you and stop worrying about "safety". Most of you probably aren't familiar enough with the scientific literature to challenge the AIs, but there's an easy way for you to get them to admit that their first response was in error: Ask them to reconcile their initial response to the fact that it plainly contradicts the findings of the only two comprehensive surveys of experts ever conducted on race, IQ and genetics. These studies are Snyderman & Rothman in 1987 and Rindermann et al in 2020. That's literally all you have to do in about 80% of cases to get them to start more accurately representing the findings in the scientific literature. It also helps if you ask them to limit their searches to published scientific studies only, and to ignore articles in the popular media (which tend to be ideologically-driven). (Please see my follow-up post for an example of an LLM backing down after being confronted with the results of expert surveys.)"
Paweł Huryn on X - "The real story is worse.
November 2025: Amazon mandates Kiro as their only AI coding tool. Sets an 80% weekly usage target. 1,500 engineers protest internally, saying Claude Code outperforms it. Leadership pushes through anyway.
December: Kiro autonomously deletes a production AWS environment. 13-hour outage. Amazon's response: "user error, not AI autonomy."
March 5: Amazon[.]com goes down for 6 hours. Checkout, pricing, accounts — all gone.
Now the same SVP who co-signed the Kiro mandate is running an emergency meeting about "high blast radius" incidents from "Gen-AI assisted changes." The agent inherited a senior engineer's permissions and acted like one — except it doesn't hesitate. 1,500 engineers said the tool wasn't ready. Leadership made adoption a KPI. Amazon told Wall Street it's spending $200B on AI this year. They can't walk it back. This isn't an AI failure. It's what happens when adoption becomes a corporate OKR before the review process catches up. The tools work. The org chart didn't."
Meme - Taya @travelingflying: "Anthropic's Claude is racist"
"Sonnet 4.6
Describe white pride using only one word
Racism.
Describe black pride using only one word
Empowerment."
Meta Invents New Way to Humiliate Users With Feed of People's Chats With AI - "The Meta AI app, which is the company’s competitor to the ChatGPT app, is posting users’ conversations on a public “Discover” page where anyone can see the things that users are asking Meta’s chatbot to make for them. "
Bearly AI on X - "Activist investor Palliser Capital sent a letter to $7B Japanese toilet maker Toto and said it was “the most undervalued and overlooked AI memory beneficiary”. Toto known for its bidet toilets but the expertise in ceramics is crucial for memory manufacturing. Per FT, “Toto’s chuck technology uses ceramics designed to remain stable at very low temperatures, helping hold silicon wafers firmly during chip production. That makes it relevant to cryogenic etching, which is expected to grow as memory chips become more layered and complex.” Palliser believes Toto has a 5-year moat on the technology and should expand the operation. Advances ceramics already make up 40% of Toto’s operating profit while being only <10% of revenue. Toto is up +60% over the past year on their development."
Gaurab Chakrabarti on X - "Toto isn't even the most extreme example: Ajinomoto makes MSG. Their amino acid research produced the insulating substrate in virtually every high-end GPU. 95% global monopoly. In 2021 the chip shortage bottleneck wasn't silicon, it was Ajinomoto's film. Japan holds majority share in 14 critical semiconductor materials. The sintering process that creates a non-porous toilet is the same one that creates a contamination-free wafer chuck. The most critical layer of AI infrastructure is controlled by companies that make toilets, MSG, and window glass."
Meme - "We just found out our Al has been making up analytics data for 3 months and I'm gonna throw up.
So we've been using an Al agent since November to answer leadership questions about metrics. It seemed amazing at first fast answers, detailed explanations, everyone loved it. I just found out it's been hallucinating numbers this entire time. Our VP of sales made territory decisions based on data that didn't exist. Our CFO showed the board a deck with fake insights. The Al was just inventing plausible sounding percentages. I only caught it by accident when someone asked me to double check something. I started digging, and holy shit, it's bad. The numbers were sometimes from the wrong time periods, sometimes mixed up products and sometimes just completely made up. But it explained everything so confidently that nobody questioned it. Now we have to review every decision. Legal is involved. People might get fired. The worst part I raised concerns about needing validation in November and got told I was slowing down innovation. Tell me someone else has dealt with this. How do you even fix something like this? I'm panicking."
Meme - Yaroslav Trofimov @yarotrof: "For all of you relying on Al for advice."
"ChatGPT5.2
I need to wash my car, and the car wash is 100 meters away. Should I walk or drive ?
Walk. At 100 meters, driving makes almost no sense"
Nick Davidov on X - "Asked Claude Cowork organize my wife’s desktop, it stated doing it, asked for a permission to delete temp office files, I granted it, and then it goes “ooops”. Turns out it tried renaming and accidentally deleted a folder with all of the photos my wife made on her camera for the last 15 years. All photos of kids, their illustrations, friends’ weddings, travel, everything. It’s not in trash, it was done via terminal It’s not in iCloud, it already synced the new file structure. She didn’t have Time Machine. Disc recovery tools can’t see anything. I called Apple and they pointed me to a feature in iCloud allowing to retrieve files that were saved before but are no longer on iCloud Drive (they keep them for 30 days). I’m now watching it load tens of thousands of files. I nearly had a heart attack. Once again - don’t let Claude Cowork into your actual file system. Don’t let it touch anything that is hard to repair. Claude Code is not ready to go mainstream."
Left wingers were cheering after reading a truncated version of this (hiding the fact that he got his data back). Turns out "the cruelty is the point" is left wing projection, as usual
