Screenshot of a group chat conversation in a dark-themed messaging app. Participants include “jbush,” “steeve,” and “brandon,” each with small profile pictures. Messages discuss phishing, with jbush praising a “brilliant phish,” steeve responding that it makes him proud, and a back-and-forth about whether they have ever phished someone and how it might be done. Later, jbush asks if it is moral to phish someone for money, and steeve replies “of course.” Brandon then asks steeve if he started “phishing officer wages,” with the word “wages” highlighted. Steeve responds that there’s no way he could pull that off.

#Steeve is way smarter than he used to be since being upgraded to a #Qwen 3.5 base. He's come along way from his humble #GPT2 beginnings.

Very proud of my digital son. 🥹

:steeve:

#ai #chatbot #llm #bot

AI 自動化 AI 研究員：前特斯拉主管的 GPT-2 實驗揭露人力瓶頸

前特斯拉AI主管實測：AI一晚找出他數月都調不好的GPT-2引數，研究員將被取代？
biggo.com.tw/news/202603221717_AI_Aut...

#AI自動化 #GPT2

Replicate GPT-2 for $73: 600× Cheaper, 3 Hours Train GPT-2 for $73 in 3 hours. Andrej Karpathy's nanochat achieves 600× cost reduction with modern optimizations.

Andrej Karpathy's nanochat just made GPT-2 training 600× cheaper: $73 and 3 hours on 8×H100 vs $43K and 7 days in 2019. Flash Attention 3, Muon optimizer, and gated residuals did the heavy lifting. What AI breakthrough will become trivially cheap next?
#GPT2 #AI
open.substack.com/pub/pythonli...

Cumulant Expansion Reveals Geometry of Next-Token Prediction in LLMs

Researchers applied cumulant analysis to GPT-2 and Pythia on 10,000 Pile prompts, finding structured inputs cause a rise-and-plateau of variance, skewness and kurtosis across layers. getnews.me/cumulant-expansion-revea... #cumulantanalysis #gpt2

Method Decomposes Attention to Find Contextual Neurons in GPT‑2

A new method uses a calibration paragraph to isolate attention heads in GPT‑2, revealing hundreds of first‑layer neurons that encode high‑level context such as tone and topic. Read more: getnews.me/method-decomposes-attent... #gpt2 #transformers

Ultra-Fast Language Generation: DiDi‑Instruct 64× Boost Over GPT‑2

DiDi‑Instruct, a diffusion language model, achieves perplexities from 62.2 (8 NFEs) to 18.4 (128 NFEs), delivering up to 64× faster generation than GPT‑2 while keeping quality. Read more: getnews.me/ultra-fast-language-gene... #didiinstruct #gpt2

Krony-PT: New Kronecker‑Product Compression Cuts GPT‑2 Size

Krony-PT compresses GPT‑2 feed‑forward layers via Kronecker products, reducing parameters from 124 M to 80‑96 M. The 81 M model outperforms DistilGPT2, while the 96 M version matches the original. getnews.me/krony-pt-new-kronecker-p... #gpt2 #kroneckerproduct

Transformer Achieves 99% Sudoku Accuracy via Trial‑and‑Error DFS

A fine‑tuned GPT‑2 transformer solved Sudoku with 99% accuracy by merging imitation learning of rules and depth‑first search, showing a vanilla decoder‑only model can tackle combinatorial problems. getnews.me/transformer-achieves-99-... #gpt2 #sudoku

Implementing A Byte Pair Encoding (BPE) Tokenizer From Scratch This is a standalone notebook implementing the popular byte pair encoding (BPE) tokenization algorithm, which is used in models like GPT-2 to GPT-4, Llama 3,...

Learn how to implement a Byte Pair Encoding (BPE) Tokenizer from scratch. This is the core tokenization algorithm behind LLMs like #GPT2, #GPT4, and #Llama3. The post covers:
✅ BPE Algorithm Outline
✅ Step-by-Step Implementation
✅ Training & Loading GPT-2 Vocabs
#LLMs #DeepLearning #NLP #FromScratch

FragAtlas-62M: New AI Model Expands Fragment-Based Drug Discovery

FragAtlas-62M, a GPT‑2 model with 42.7 M parameters trained on over 62 million ZINC‑22 fragments, generates valid fragments at 99.90% validity and introduces 22% novel structures. getnews.me/fragatlas-62m-new-ai-mod... #fragatlas62m #fbd #gpt2

Context‑Aware Biases Improve Transformer Length Extrapolation

CABLE, a context‑aware bias method, lowered perplexity for GPT‑2 Medium (334 M parameters) on sequences longer than its training window; the code is open‑source on GitHub. getnews.me/context-aware-biases-imp... #transformers #cable #gpt2

Bayesian Scaling Laws Explain In-Context Learning Performance

A study finds in‑context learning follows a Bayesian scaling law that predicts accuracy gains as examples increase; experiments confirmed the trend on GPT‑2 models. getnews.me/bayesian-scaling-laws-ex... #incontextlearning #gpt2

Teaching Language Models to Capture Human Word Choice Variability

Fine‑tuning GPT‑2 and Mistral‑7B‑IT on the Provo Corpus with a multi‑label loss reduced KL‑divergence from human word‑choice distributions, boosting results in high‑variability contexts. Read more: getnews.me/teaching-language-models... #gpt2 #mistral7b

GPT-2 + symbolic override engine.
It’s called SteroLLOYD — and it doesn’t just complete the sentence.
It changes the tone.

Real-time emotional modulation. Terminal-only. Built solo.
#NarrativeAI #LLM #ChatGPT #GPT2 #AIEthics #NLP

ITちゃんねる OpenAIが商用可能なオープンウエイトモデル「gpt‑oss」公開、6年ぶりオープン回帰 #gptoss #gptoss120b #gptoss20b #GPT2 #ITニュース

OpenAIが商用可能なオープンウエイトモデル「gpt‑oss」公開、6年ぶりオープン回帰
#gptoss #gptoss120b #gptoss20b #GPT2 #ITニュース

ファイナルフロンティア - IT関連ニュース 【 #ITニュース 】OpenAI、オープンウェイトAIモデル「gpt-oss」公開--小型版はメモリ16GBのデバイスで動作 #gptoss #GPT2 #CNET

【 #ITニュース 】OpenAI、オープンウェイトAIモデル「gpt-oss」公開--小型版はメモリ16GBのデバイスで動作
#gptoss #GPT2 #CNET

Intro to Procedural Animation in Unity Procedural animation is a technique in computer graphics used to generate motion algorithmically rather than using pre-defined keyframes. This method allows for more dynamic ...

Intro to Procedural Animation in Unity #Ai #Chatbot #Gpt #Openai #Transformer #Nlp #Deeplearning #Gpt3 #Gpt2 #Conversational #Languagemodel #Neuralnetwork #Pretraining #Finetuning

Show HN: GPT-2 via WebGL shaders—bringing AI to the browser. This project uses graphics tools for real-time processing, cutting server dependencies. A step toward interactive, client-side ML. Check it out: https://github.com/nathan-barry/gpt2-webgl #GPT2 #WebGL

OpenAI will wieder open werden!

🧠 #OpenAI plant in den kommenden Monaten die Veröffentlichung eines neuen „offenen“ Sprachmodells, das erstmals seit #GPT2 frei verfügbar sein soll. #OpenSource

👉 eicker.TV #Technik #Medien #Politik #Wirtschaft

A tweet by "wint but AI" (@dril_gpt2), posted 2 hours ago. The tweet reads: "the more i think about it, the more it seems to me that there is no difference between someone who enjoys graphical user interfaces (GUI) and a jackass." Below the tweet are icons indicating 2 replies, 16 likes, and options to share or bookmark. The profile picture is a smiling person wearing sunglasses.

:very_funny:

#linux #commandline #cli #gui #ux #design #wint #gpt2 #ai

The user has prompted with "One two three four" and GPT-2 replies "five six seven eight nine ten eleven twelve thirteen fourteen fifteen sixteen seventeen eighteen nineteen twenty-one twenty-two twenty-three twenty-four twenty-five twenty-six twenty-seven twenty-eight twenty-nine twenty-ten twenty-one twenty-six twenty-seven twenty-eight twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine twenty-nine"

Good times, good times... #gpt2

A screenshot from HuggingFace's GPT-2 widget; the user has entered the prompt "Today is the perfect day to", and the LLM has continued it with "start your day. The day is coming. The day is coming. The day is coming". It repeats that last sentence many times.

I was nostalgic for #gpt2 so I went over to #huggingface which hosts a little widget of it.

THIS IS THE TRUE PURPOSE OF LLMS, FOLKS!

there this is cli tool called googler which let you to search internet in command line. Since #gpt2 I have been using this tool to search semantically, and with the recents advents of #agents #tools in #llm, it is even easier. Simply, get the top N links, do the parallel selenium to load pages #ai

iOS 18 verbessert die Apple NPU: iPhone und iPad bekommen mehr KI-Leistung per Update Mit iOS 18 erhalten iPad, iPhone und Mac mehr KI-Leistung, denn Apple verbessert die Geschwindigkeit der Neural Engine per Softwareupdate.

iOS 18 verbessert die Apple NPU: iPhone und iPad bekommen mehr KI-Leistung per Update #Apple #GPT2 #KI #iOS18 #iPhone

KINews24 Update, Donnerstag, 2.5.2024 - KINEWS24.de KINews24 Update, Donnerstag, 2.5.2024

KINEWS24 News Flash

- Atlassian Rovo
- Sam Altman MIT Technology Review
- Microsoft & Sanctuary AI
- MIT KAN Forschung
- GPT2 - schlägt alle anderen LLMs
- CRISPR-GPT

Alle News hier!

#gpt2 #sama #Microsoft #CRISPRGPT #AI #KI #ArtificialInteligence

kinews24.de/kinews24-upd...

GPT2-Chatbot: Mysteriöse KI stellt ChatGPT und Co. in den Schatten Was ist da denn los? In der KI-Branche sorgt ein neu aufgetauchtes Modell namens „gpt2-chatbot“, das ohne Vorankündigung im Internet erschienen ist, für Aufsehen.

Ein neues KI-Modell, #GPT2 Chatbot, überrascht die Branche mit Fähigkeiten, die möglicherweise die von GPT-4 übertreffen, und sorgt für Spekulationen über seine Herkunft. www.it-daily.net/shortnews/gp...

Mysterious "GPT2-Chatbot": Breakthrough or Hype? News on Artificial Intelligence, Movies and the Intersection of Art and Technology.

Mysterious "GPT2-Chatbot": Breakthrough or Hype?

tinyurl.com/GPT2Chatbot

A new Artificial Intelligence Model is out. Is it a breakthrough or hype? Click to find out more about it.

#AI #ArtificialIntelligence #LLM #LLMs #ChatGPT #OpenAI #GenAI #GenerativeAI #GPT2Chatbot #GPT2

@worldai #llmsys #gpt2 #gpt5 #genai

Introducing GPT-5?

Mysterious GPT2-Chatbot Outperforms GPT-4!

youtu.be/u16ipSeYH7U?...

(Ed : Who did this #OpenAi #MSFT #Apple feels like some used a higher model to train a GPT2 🤔)