Daniel

@kthai1618 This is very noble and good

These models and datasets are licensed for noncommercial use only. Open Source Pangram should NOT be used to enforce any AI use policies. But if you need a more accurate model with a state-of-the-art low FPR, please reach out! We also offer $5k API research grants.

@sxhealth101 don't work out because you lose electrolytes

Here is what one ejaculation contains: • Zinc – ~1–3 mg • Magnesium – ~10–15 mg • Calcium – ~20–30 mg • Potassium – ~50–100 mg • Protein – ~0.2–0.5 g • Fructose (natural sugar for energy) – small amounts • Amino acids (building blocks for the body) • Enzymes that support sperm function • Hormonal compounds in tiny traces It is not just fluid. It is a nutrient-rich biological material your body worked to produce. Repeated loss without recovery can add up over time. Retain your semen.

@awnihannun :(

I joined Anthropic as a member of the technical staff. Excited to work on frontier modeling at a place with unwavering values and a generational mission.

@alisaqqt boycott :)))))))

@XDanielDev I don't think they will open source it

Wan 2.7 is planned to launch within March — and it’s a major all-around upgrade over 2.6. Wan 2.7 will support: - first-frame & last-frame video generation - 9-grid image-to-video - subject + voice reference - instruction-based video editing - video recreation / replication A more powerful and comprehensive creative workflow is on the way.

@xbxnxdxcxtx @xai

@deepfates @snwy_me @mil0theminer wow

@XDanielDev @snwy_me @mil0theminer You argue like a bitch

apple's 50 year anniversary letter is fully AI generated

@snwy_me @mil0theminer

@XDanielDev @mil0theminer i ask again: are you able to comprehend concepts?

@snwy_me @mil0theminer the benchmarks they made from 2024 or what, I don't keep up with the snake oil industry

it is literally fucking math. pangram did not invent this they just executed it very well. like i said, there are many mathematical features of discrete sequences (i.e. text) that can be measured and used to find their similarity to others. AI texts share common features that human texts do not particularly contain. this is the basis for "yes, it is possible to create a classifier that is highly accurate". none of this is marketing or debatable, it is literal objective fact based in mathematics lol pangram is the highest-ranking classifier on benchmarks that are industry-standard.

@orphcorp @ersatz_0001 @snwy_me @mil0theminer in the most obscure and hidden way possible, of course

@XDanielDev @ersatz_0001 @snwy_me @mil0theminer if you had ever used Pangram you'd know it assigns a low/medium/high confidence score to all it's detections

@snwy_me @mil0theminer I really have no idea why (if you do know what you're talking about) you keep pushing this as if it's evidence of reliability? are you being paid?

if you read more than a few choice words you'd see that it's an embeddings classifer. AI generated texts have very distinct (albeit non obvious to human readers) features that cluster together within embeddings, where human texts do not cluster in the same way. this makes for a very reliable classifier. do you want to try rebutting that?

@snwy_me @mil0theminer what? are you trying to prove how much technical detail is included by copying the marketing?

>Our model is a slightly modified transformer-style architecture (Vaswani et al., 2017). The classifier is trained on a mixture of human examples and synthetic examples generated by LLMs to closely match the content of the human examples, using a method called mirror prompting that we detail in Section 4.2. learn to read maybe?

@snwy_me @mil0theminer omg is this the revolutionary Transformer I've been hearing all about? the one that they detail as "a slightly modified transformer-style architecture". there is nothing fake about pangram, but there is nothing special about it either, its fucking marketing

do you understand how academic paper work lmao “its outdated” maybe try being able to understand what any of it says before you dismiss it? do you have any like idea of how transformer language models work? how we can quantify metrics that aren’t immediately obvious by looking at the text? how a neural network learns features? what shannon entropy is? its not fake because you can’t grasp it lol

@mil0theminer @snwy_me this isnt even a paper with actual technical detail

@XDanielDev @snwy_me bro go look at her page, I think she knows how papers on AI work

@snwy_me @mil0theminer are you trying to sound smart or what? the paper is outdated and doesnt prove anything, its marketing

@XDanielDev @mil0theminer are you able to comprehend like concepts and stuff

@snwy_me @mil0theminer omg a paper from 2024. that solves everything, it must be perfect

@XDanielDev @mil0theminer try reading about it arxiv.org/pdf/2402.14873

@ersatz_0001 @snwy_me @mil0theminer yes, the absence of any confidence score makes it essentially say it is 100% confident

@XDanielDev @snwy_me @mil0theminer This is the share of the document that is detected as AI-generated, not confidence

@snwy_me @mil0theminer is this a joke or do you actually believe tools that return a 100% confidence

@mil0theminer pangram almost never false-positives; it almost always only false-negatives

@jarredsumner Want to get people on your side? Give the weights

Today’s version of this story is scarier. In 2013, even if you could get everyone’s data, grep and keyword search was the best you could do. The technology to do civilization-scale surveillance didn’t exist. And with LLMs, that changes.

In 2013, when government tried mass domestic surveillance, the tech industry responded with HTTPS everywhere. In 2026:

@john__allard This is a cult

anthropic is a crown jewel of American AI supremacy and the envy of our very capable international adversaries… it’s genuinely difficult to imagine a higher-leverage own-goal than designating them a supply chain risk another display of hesgeth’s profound incompetence