@anthropic.com
We're consistently trying to use your product—long before the DOD kerfuffle—this is a weekly experience. Investing in uptime is investing in the outrageously easy + HUGE opportunity to capture market share given the shitshow that is OpenAI. GoIntoTheRED on uptime—you'll be rewarded.
Pardon the break from our regularly scheduled programming.
youtu.be/fA0fZ3LARsw?...
This is getting ridiculous - the GOP lunges away from states rights when it suits their power (read: GRIFT) needs. Deep and overlapping layers of ignoring humanity's needs. #AISafety #AI #StatesRights
www.healthcareitnews.com/news/white-h...
He won't make it to the end of his term. Even the idiots who voted him in are done with this dipshit.
We don’t agree on what AGI means and we’re kinda just making up how to measure “progress” — dystopian isn’t hyperbolic at this point.
Regulators, mount up. Seriously. #AISafety #AI
gizmodo.com/ai-powered-t...
Napoleon - Ani Difranco
www.youtube.com/watch?v=rqGn...
10/ The debate this paper forces into the open --> If “intelligence” is now measurable in machines the same way it is in humans, we have to ask:
What counts as a mind worth moral or legal consideration—and who decides when a machine becomes one?
9/ Policy implications follow: if AGI is defined by human-level cognitive breadth, not profit or hype, we can begin to measure progress, set safety thresholds, and recognize when systems are approaching generality before they cross it.
8/ By mapping AI cognition to psychometrics, this framework reframes AGI not as an economic milestone but a cognitive one. It distinguishes between “AI that earns billions” and “AI that thinks broadly.” That’s a critical distinction for governance.
7/ The biggest bottleneck is Memory.
Current #AI systems can’t store or retrieve new information reliably. Each conversation starts from scratch. Without continual learning and durable context, machines can’t develop genuine understanding or experience.
6/ These uneven profiles expose the illusion of “general” intelligence.
Systems appear capable because they contort strengths—like vast context windows or retrieval search—to mask deep deficits in memory and reasoning. What looks like generality may just be compensation.
5/ The results are revealing — Today’s AI, the paper argues, is “jagged”: brilliant yet brittle.
GPT-4 scores 27% toward AGI. GPT-5 reaches 57%.
Both are strong in knowledge, math, and language—but almost zero in long-term memory, retrieval precision, and sustained reasoning.
4/ Instead of relying on task benchmarks, the framework tests whether #AI systems possess the same underlying abilities that enable general cognition in humans. It’s less about what AI can do, and more about what kind of mind it has.
3/ To make this measurable, they turn to the Cattell-Horn-Carroll (CHC) theory—the most empirically validated model of human intelligence—and adapt its 10 cognitive domains to AI: from reasoning and memory to perception, language, and speed.
2/ The authors—drawing from psychology, AI, and neuroscience—define #AGI as an #AI system that matches the cognitive versatility and proficiency of a well-educated adult. Not superhuman. Just human-level across the full spectrum of cognition.
We talk about #AGI as if we’d know it when we see it. But we don’t.
This paper argues decades of speculation have confused ambition with definition and offers the first quantifiable framework for what general intelligence in machines means.
Download or enjoy a 🧵
media.licdn.com/dms/document...
How to scale fascist ideology. We need the antidote.
Given the lack of federal policy in the United States, we wouldn’t recommending holding our collective breath.
Final stanza of the poem "Ozymandias" describes the ironic ruin of a once-great statue in the desert. The quote reveals the ruler's vanity and belief in his own power, while the broken, empty ruins that surround it serve as a testament to the temporary nature of all power and legacy
"To oppose something is to maintain it"
— The Left Hand of Darkness by Ursula K. Le Guin
As always, you’re a treasure and appreciated.
Full dataset + code is public:
🧠 huggingface.co/datasets/LLM...
🔧 github.com/TianyiPeng/T...
If you work with LLMs, behavioral science, or #AI evaluation—this is your benchmark.
6/ Bottom line:
#DigitalTwins capture some human variability but fall short as true proxies. They’re promising as low-cost experimental agents—but not yet reliable stand-ins for real people.
5/ #DigitalTwins skewed “pro-human” and “pro-technology”:
• More trust in others
• More support for regulation
• Less privacy concern
• Greater faith in algorithmic hiring
But still under-dispersed aka less varied than real humans.
4/ Bias: #DigitalTwins were more accurate for people who were:
• Wealthier
• More educated
• Politically moderate
• Attending religious services occasionally
This amplifies existing representation gaps in social science data.
3/ #DigitalTwins performed better in:
• Personality & pro-social questions
• Judging social scenarios
• Cognitive tasks requiring logic
They struggled with:
• Political views
• Valenced judgments (e.g., moral tone)
• Social desirability bias
2/ Even with personas, avg correlation between twins & real humans was just ~0.2. Good enough to beat ML benchmarks (XGBoost), still too low for high-stakes personalization.
1/ #DigitalTwins hit ~75% “accuracy” predicting individual responses—so did simpler models using only demographics. What matters more? Correlation. And that’s where richer data helped—especially in social & cognitive domains.
Can AI “digital twins” truly simulate human behavior?
A new mega-study tested 2,000+ twins built from rich personal data—500+ questions per person—across 19 experiments.
Results challenge assumptions about accuracy, bias & social science use cases
DL and/or enjoy a 🧵
arxiv.org/pdf/2509.19088
