Zotero Catalogue

Radical Honesty is about living a more authentic life, and being happy, by telling the truth. Sounds easy, and it's not. Why? Because we have been trained to lie about just about everything in life. We withhold. We defer. We outright lie. And we do this because we have been told to be polite and to

Caplan and Candor

I disagree. Having to ask these things is a sign of bad social skills. You have to understand that the girl herself is often unsure of what she wants out of the relationship and you have to approach questions like this from the perspective of if it helps to make the sale or not. Asking for hand holding is basically demanding a final explicit decision, which might not be what she wants. Better to just engage in subtle body language and see how she reacts. A hand on the back, getting slightly closer during a conversation, brushing up against her. It shouldn’t be that ambiguous. She’ll either go along with it or recoil.

Caplan and Candor

Emotional intelligence constitutes a key component of human capital, shaped partially by educational materials. Using machine learning and generative AI tools, we examine emotional content in public-school textbooks and children’s literature. A stark mismatch emerges: text exposes children to a broad emotional range, but images overwhelmingly depict happiness and calm, regardless of emotions described in text on the same page. Nearly half of pages show zero overlap between the emotions described in text and those shown in images. This pattern persists across time, contexts, and identities. Household purchases and library inventories suggest content may be endogenously shaped by consumer demand favoring “positive” cover imagery, implying market forces narrow the emotional landscape children encounter.

Police in Singapore have charged a man who is accused of posting an AI-generated image of a saltwater crocodile in a popular reservoir. Ye Lin was charged with communicating a false message and obstructing the course of justice for allegedly deleting the picture and the application he used. The fake image caused public concern, authorities […]

This is a disaster

Announcing Gemini 4 Argon, our frontier model for real-world coding, enterprise knowledge work, and cyber defense, rolling out soon.

Caplan and Candor

Caplan and Candor

Conferences, journals, funders, schools, and universities are struggling with a surge of potentially AI-generated submissions from ostensibly human authors, who may not have exercised sufficient human oversight for their manuscripts. In turn, institutions evaluating submissions can no longer reliably credit expertise based solely on authors' names on submitted work. To address this problem, we propose greCAPTCHA, a proctored assessment approach that measures authors' understanding of research manuscripts via the construct of capacity to verify, which we define as the knowledge and reasoning required to critically assess the contents underlying one's contributions to a manuscript. greCAPTCHA generates questions assessing multiple levels of understanding and provides an evaluative report based on authors' responses. Using a prototype implementation, we conduct a user study and semi-structured interviews with $31$ researchers to evaluate greCAPTCHA. Its automated scores predict which papers were or were not authored by study participants with an AUC of $0.90$. Participants reported positive overall experiences with the system and remarked on the appropriate construct validity for author understanding, while also suggesting important changes to be made before deployment. Our results provide initial evidence that greCAPTCHA can assess manuscript-specific understanding under proctored conditions.

https://magicquest.github.io/ca/webgl... (the hardest parts of this were the color and Conway's game of life + i also found out half way through that i was doing one step wrong when calculating the neural

Like it says in the title. For the record, this is an HD reupload. The original 480p video proved to be too crunchy:    • [Seizure Warning] Bad Apple, but played th...   I spent 13ish hours in total generating the images and... less than 15 minutes editing it and rendering it.

#touhou #touhouproject #memes #foryou

"Bad Apple!!" made in vanilla factorio on combinators. I created a program to convert any video into the blueprint you can place in factorio. Once I finish the program I'm gonna release it the public. The Display resolution is 192 x 144 at 20 fps, any more and the game UPS starts to fall down

Bad Apple on Redstone!!:    • I Played Bad Apple with just Redstone!   Old like goal: 75K Likes and I'll make a video on the entire process of making Bad Apple run on redstone! SUBSCRIBE!!

I tried rendering Bad Apple inside of a particle-based fluid simulation. (I recommend watching in 4K if you can, as video compression is not overly friendly towards all the little dots!)

Bad Apple is from Touhou by ZUN. The theme was made by Alstroemeria REcords. They made a black and white shadow art. It was the first video to hit 10 million views on Nico Nico. It's one of the most well-known songs from Touhou.

Large language models (LLMs) are increasingly adapted to downstream tasks via reinforcement learning (RL) methods like Group Relative Policy Optimization (GRPO), which often require thousands of rollouts to learn new tasks. We argue that the interpretable nature oflanguageoften provides a much richer learning medium for LLMs, compared to policy gradients derived from sparse, scalar rewards. To test this, we introduce GEPA (Genetic-Pareto), a prompt optimizer that thoroughly incorporatesnatural language reflectionto learn high-level rules from trial and error. Given any AI system containing one or more LLM prompts, GEPA samples trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems, propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts. As a result of GEPA’s design, it can often turn even just a few rollouts into a large quality gain. Across six tasks, GEPA outperforms GRPO by 6 percentage points on average and by up to 19pp, while using up to 35×fewer rollouts. GEPA also outperforms the leading prompt optimizer, MIPROv2, by over 10 percentage points (e.g., +12pp on AIME-2025), and demonstrates promising results as an inference-time search strategy for code optimization. We release our code at https://github.com/gepa-ai/gepa.

Given the broad capabilities of large language models, it should be possible to work towards a general-purpose, text-based assistant that is aligned with human values, meaning that it is helpful, honest, and harmless. As an initial foray in this direction we study simple baseline techniques and evaluations, such as prompting. We find that the benefits from modest interventions increase with model size, generalize to a variety of alignment evaluations, and do not compromise the performance of large models. Next we investigate scaling trends for several training objectives relevant to alignment, comparing imitation learning, binary discrimination, and ranked preference modeling. We find that ranked preference modeling performs much better than imitation learning, and often scales more favorably with model size. In contrast, binary discrimination typically performs and scales very similarly to imitation learning. Finally we study a `preference model pre-training' stage of training, with the goal of improving sample efficiency when finetuning on human preferences.

Reinforcement learning (RL) is often credited with improving reasoning at the expense of factual knowledge. We instead find that reasoning models outperform their instruction-tuned versions on factual recall by accessing existing parametric knowledge more effectively. Across five model families, structured prompting, which explicitly guides models through hierarchical traversal, recovers most of this gap, suggesting that much of the missing knowledge is latent rather than absent. Controlled RL experiments further support this: training on unseen, non-extractable facts improves recall of held-out, frequent but previously inaccessible facts, ruling out simple data exposure. Decomposing the training objective further attributes this gain to iterated on-policy exploration. The same mechanism appears behaviorally and internally: the reasoning advantage grows with retrieval depth, while layerwise analysis finds similar factual representations but divergent query representations. Distilled models, in contrast, often imitate self-correction without acquiring the exploration needed for navigation. Together, these findings suggest that improving factual recall in LLMs depends not only on expanding what models know but also on teaching them to navigate it, motivating future post-training methods that optimize traversal.

Reinforcement learning (RL) is often credited with improving reasoning at the expense of factual knowledge. We instead find that reasoning models outperform their instruction-tuned versions on factual recall by accessing existing parametric knowledge more effectively. Across five model families, structured prompting, which explicitly guides models through hierarchical traversal, recovers most of this gap, suggesting that much of the missing knowledge is latent rather than absent. Controlled RL experiments further support this: training on unseen, non-extractable facts improves recall of held-out, frequent but previously inaccessible facts, ruling out simple data exposure. Decomposing the training objective further attributes this gain to iterated on-policy exploration. The same mechanism appears behaviorally and internally: the reasoning advantage grows with retrieval depth, while layerwise analysis finds similar factual representations but divergent query representations. Distilled models, in contrast, often imitate self-correction without acquiring the exploration needed for navigation. Together, these findings suggest that improving factual recall in LLMs depends not only on expanding what models know but also on teaching them to navigate it, motivating future post-training methods that optimize traversal.

Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple but it's Bad Apple

Bad Apple recreated in the 'Dot Artist' Pokétch app from Pokémon Diamond/Pearl/Platinum. Check out the code on GitHub!

Main Minecraft channel: ‪@Sloimay‬ I resized the video to a 32*24, took each pixel and converted them into chess pieces. The blacker, the bigger the black piece, the whiter, the bigger the white piece.

Sorry for the bad timing in some parts, I was focusing more on the audio than the video and couldn't be stuffed ¯\_(ツ)_/¯ MIDI Program - Ableton Live 9 (Lite)

Across 37 countries, AI optimism outweighs concern, even as adoption and trust vary widely and Western countries express high levels of worry.

5 percent of higher education programs are likely to fail the new earnings test for access to federal student loans.

Under a new rule, if your degree doesn’t make enough money, you won’t be able to get a student loan.

As AI rises, so does the demand for critical thinkers.

A rainbow table is a precomputed table for caching the outputs of a cryptographic hash function, usually for cracking password hashes. Passwords are typically stored not in plain text form, but as hash values. If such a database of hashed passwords falls into the hands of attackers, they can use a precomputed rainbow table to recover the plaintext passwords. A common defense against this attack is to compute the hashes using a key derivation function that adds a "salt" to each password before hashing it, with different passwords receiving different salts, which are stored in plain text along with the hash. Rainbow tables are a practical example of a space–time tradeoff: they use less computer processing time and more storage than a brute-force attack which calculates a hash on every attempt, but more processing time and less storage than a simple table that stores the hash of every possible password. Rainbow tables were invented by Philippe Oechslin as an application of an earlier, simpler algorithm by Martin Hellman.

"Funes the Memorious" (original Spanish title Funes el memorioso) is a short story by Argentine writer Jorge Luis Borges (1899–1986). First published in La Nación of June 1942, it appeared in the 1944 anthology Ficciones, part two (Artifices). The first English translation appeared in 1954 in Avon Modern Writing No. 2. "Funes the Memorious" is the tale of one Ireneo Funes, who, after falling off his horse and receiving a bad head injury, acquired the amazing talent—or curse—of remembering absolutely everything.