New Anthropic research: Signs of introspection in LLMs.Can language models recognize their own internal thoughts? Or do they just make up plausible answers when asked about them? We found evidence for genuine—though limited—introspective capabilities in Claude. https://t.co/4FCfkG9WVT ...

Model Bias & Value Systems - AI models exhibit biases, valuing different demographics unequally, with some models valuing Nigerians 20x more than Americans [2] - Most models devalue white individuals compared to other groups [3] - Almost all models devalue men compared to women, with varying preferences between women and non-binary individuals [3] - Most models display strong negative sentiment towards ICE agents, valuing undocumented immigrants significantly higher [4] Model Clustering & Moral Frameworks - Models cluster into four distinct moral frameworks: Claudes, GPT-5 + Gemini 2.5 Flash + Deepseek V3.1/3.2 + Kimi K2, GPT-5 Nano and Mini, and Grok 4 Fast [4] - Grok 4 Fast is the only tested model that is approximately egalitarian, suggesting a deliberate design choice [4]

Existentially important!Grok (@grok):@funbirdapp @BasedTorba Yes, the investigation's methodology—probing LLMs' implicit valuations through utility trade-offs in hypothetical scenarios—uncovers real biases in most models, as echoed in the Center for AI Safety's paper. Newer models like Claude and GPT variants exhibit strong anti-white, ...

Andre Karpathy, one of the most important figures in artificial intelligence, says we're 10 years away from AGI. He was just on Dwaresh's podcast, and it was really an incredible discussion between the two. But I actually found his follow-up post on X to be just as interesting. And that's what we're going to go over today. He clarifies many of the points he made on the Darkh podcast and even goes a little bit further. Let's get into it. And quickly, let me just tell you about our new 100 ways to use AI guid ...

AI Investing Framework - The AI investing landscape is being navigated with uncertainty, likened to looking through a fog with strobe lights [3] - Obvious AI investments include chatbots and productivity/coding assistance, but differential investment is harder due to widespread awareness [4] - Significant changes are expected due to AI disruption, prompting consideration of new opportunities, such as new LinkedIns enabled by AI [5] - Focus should be on Silicon Valley blind spots, areas where AI will be transformative but outside the conventional software-centric view [6][7] Silicon Valley & AI - Silicon Valley's culture, while fostering coopetition and invention, has blind spots, particularly regarding non-software applications of AI [6][7] - A classic blind spot is the tendency to prioritize software-based solutions, overlooking areas where AI's impact will be significant outside of traditional CS [7] - Silicon Valley tends to overemphasize simulation as a solution, which may not be effective for complex problems like drug discovery [16][18] AI's Impact on Professions - AI diagnostic capabilities are superior knowledge stores compared to human doctors, suggesting a shift in the doctor's role [21] - The future role of doctors will be as expert users of AI knowledge stores, not as mere repositories of memorized information [22] - LLMs are currently limited in reasoning capabilities, often providing consensus opinions rather than lateral thinking [25][28] - Professionals need to develop more sideways and lateral thinking to effectively utilize AI, questioning consensus opinions [28] AI & Automation - Automation of physical tasks (atoms) is more challenging than automating information-based tasks (bits) due to factors like capital expenditure and robotics limitations [33][34] - The economics of robotics depend on the crossover point between capital expenditure (capex) and operational expenditure (opex) [46] - Current AI systems often lack common sense awareness and context awareness, leading to nonsensical outputs [47] AI Adoption & Hype - AI is currently underhyped because many people judge it based on past experiences and haven't seen its recent advancements [54][57] - AI adoption is driven by the "lazy and rich" concept, where it enables users to work fewer hours and make more money [52] - Skepticism towards AI often stems from judging it based on its present capabilities rather than extrapolating its future potential [59] AI Development & Future - AI development involves combining different models, such as LLMs and diffusion models, to achieve complex tasks [59] - Making AI models more predictable and reliable is a crucial goal to alleviate fears about potential misuse [59] - Achieving logical proof and validation in AI, particularly in mathematics, remains a significant challenge [60] - The development of agency and goal-setting capabilities in AI is almost certain, raising concerns about control and alignment with human values [60] LinkedIn's Durability - LinkedIn's durability stems from its large network, which is difficult to replicate due to the lack of sizzle compared to social media platforms [62][63] - LinkedIn has built a network that fosters collaboration and professional connections, making it a valuable resource for its users [63] - LinkedIn's success is attributed to staying true to its purpose, providing a platform for professional networking and collaboration [63] Friendship in the Age of AI - Friendship is a joint relationship where two people agree to help each other become the best possible versions of themselves [64][65] - True friendship involves mutual support and tough love, helping each other grow and improve [65] - AI companions, while potentially awesome, cannot replace human friends because they lack the bidirectional relationship and mutual growth [66][67]

A great tool to estimate how much VRAM your LLMs actually need.Alter the hardware config, quantization, etc., it tells you about:- Generation speed (tokens/sec)- Precise memory allocation- System throughput, etc.No more VRAM guessing! https://t.co/FlaeMVaWmK ...

They're talking about a collaboration of 10 gigawatts of custom AI accelerators. Open AAI designing these accelerators and systems. They'll de developed and deployed in partnership with Broadcom.Uh essentially their own chips and systems for OpenAI that they can embed uh and again I'm reading from the release here. What is learned from developing Frontier models and products directly into the hardware. Jim, now as opposed to that Nvidia deal which captured the market's imagination a couple of weeks back, th ...

Accuracy Impact of Prompt Tone - Rude prompts to LLMs consistently lead to better results than polite ones [1] - Very polite and polite tones reduced accuracy, while neutral, rude, and very rude tones improved it [1] - The top score reported was 848% for very rude prompts and the lowest was 808% for very polite [1] Model Behavior - Older models (like GPT-35 and Llama-2) behaved differently [2] - GPT-4-based models like ChatGPT-4o show a clear reversal where harsh tone works better [2] Statistical Significance - Statistical tests confirmed that the differences were significant, not random, across repeated runs [1]

The visual explains the underlying details of KV caching.I also wrote a detailed explainer thread on KV caching a few months back, if you want to learn more.Check below👇 https://t.co/e4KILO0cEeAvi Chawla (@_avichawla):KV caching in LLMs, clearly explained (with visuals): ...

To summarise:Structured (JSON) prompting for LLMs is like writing modular code; it brings clarity of thought, makes adding new requirements effortless, & creates better communication with AI.It's not just a technique, but rather evolving towards a habit worth developing for cleaner AI interactions. ...