ozgung

11 ms·

181.
by ozgung 1y ago
This is not about _Large_ Language models though. This explains math for word vectors and token embeddings. I see this is the source of confusion for many people. They think LLMs just do this to statistically predict the next word. That was
182.
by ozgung 1y ago
All smartphones are also smartmicrophones.
183.
by ozgung 1y ago
Location: Izmir, Turkey Remote: Yes Willing to relocate: Yes Technologies: Python, C/C++, Swift, PyTorch, Metal/OpenGL and shader development, OpenCV Résumé/CV: https://ozgungenc.com Email: ozgungenc[at]gmail[dot]
184.
by ozgung 1y ago
Same. It acts like its output tokens are for free. My input output ratio is like 1 to 10 at least. Not counting "Thought" and it's internal generation for agentic tasks.
185.
by ozgung 1y ago
This. ChatGPT also agrees with you: "74 GB weight read is per pass, not per token." I was checking the math in this blog post with GPT to understand it better and it seems legit for the given assumptions.
186.
by ozgung 1y ago
In this blog post: podcast: 22 times video: 13 times blog: 6 times rogan: 5 times youtube: 4 times clubhouse: 4 times paper: 14 times conference: 0 times journal: 0 times These physicists love podcasts.
187.
by ozgung 1y ago
So, on August 21, 2025, everyone in the world suddenly decided that AI was a bubble, no matter what day had said the day before. And they decided to start a PR campaign to announce this to the general public. CEOs were overselling, now they
188.
by ozgung 1y ago
By Amara's Law and Gartner Hype cycle every technological breakthrough looks like a bubble. Investors and technologist should already know that. I don't know why they're acting like altcoins in 2021.
189.
by ozgung 1y ago
I think the author is confusing the attention mechanism with attention in the general sense. He refers to a phenomenon called “attention drift”, but neither Google nor LLM searches return any relevant references to that term. In fact, ChatG
190.
by ozgung 1y ago
This made me realize OpenAI is actually in the Artificial Humans business right now, not just AI. I am not sure if this was what they wanted. They have to deal with real humans. Billions of conversations with billions of people. In the Soci
191.
by ozgung 1y ago
"Why did the sun climb a tree?" Claude Opus 4.1: - To get to a higher branch of astronomy - Because it wanted to reach new heights - To see the dawn of a new day from a better view ChatGPT 5 Thinking: After thinking for 26 seconds
192.
by ozgung 1y ago
Good point. Maybe not for the CEO yet, but a manager without people to manage is a useless thing. So I think corporates will invent new Bullshit Jobs[1] for humans, to keep them in their organization chart. [1] https://en.wikiped
193.
by ozgung 1y ago
Everyone calls them Interviews but they are not really interviews. They are exams. Oral exams, live coding exams, system architecture exams, take-home exams, behavioral examinations, code review exams, extended essay writing exams, case stu
194.
by ozgung 1y ago
One key element missing from all these model cards are the model size/number of parameters. Without that info we are in the dark. We can't predict the future of AI. How does the intelligence scale with the increasing #parameters?
195.
by ozgung 1y ago
I asked GPT5 directly about fake system prompts. > Yes — that’s not only possible, it’s a known defensive deception technique in LLM security, sometimes called prompt canarying or decoy system prompts. …and it goes into details and even
196.
by ozgung 1y ago
This may be like saying “don’t think of an elephant”. Every time they say JSON, llm thinks about JSON.
197.
by ozgung 1y ago
Also this coding deception rate bar tries to decieve us. https://imgur.com/a/QkriFco
198.
by ozgung 1y ago
GPT-5 should mean a brand new model/architecture trained from scratch.
199.
by ozgung 1y ago
> 1000x employees are present in all companies There are 168 hours in a week, which is the max available for 1000x robots. So by OP's definition, a typical 1x employee works only for 10,08 minutes, and TikToks the rest of the week.
200.
by ozgung 1y ago
Learning Neural Networks from Hinton was transformative for me. It must be transformative for the industry, they still use Coursera lecture notes for citing RMSprop. I completed the very first MOOC offerings in Fall 2011, Machine Learning b
201.
by ozgung 1y ago
These products are basically public beta. Some features are even experimental. They are released to the public because (1) companies have to gain early market share (2) they need actual user data. Sam Altman firing drama from last year was
202.
by ozgung 1y ago
That's also because of seeing them as technology under development. The overblown expectations are because of their potential in the future. The glass is half full today. It was almost empty just a few years ago. The water level is ris
203.
by ozgung 1y ago
These examples show that we have a serious social issue, and it's not limited to teachers. People misuse LLMs. We engineers understand that LLMs are products under development. They only work correctly under certain circumstances, and
204.
by ozgung 1y ago
What a time to be alive
205.
by ozgung 1y ago
Writing with LLM is like doing a mapping from human prompt -> document. The longer the prompt, the closer the output document to the human intent. If the prompt is short (= human intent is low), the output must be filled with fluff, gene
206.
by ozgung 1y ago
Location: Izmir, Turkey Remote: Yes Willing to relocate: Yes Technologies: Python, C/C++, Swift, PyTorch, Metal/OpenGL and shader development, OpenCV Résumé/CV: https://www.linkedin.com/in/ozgungenc Emai
207.
by ozgung 1y ago
It's not about the products or revenue probably. It's about its stocks. Nvidia market cap is 4.262 Trillion dollars and for META it's $1.9T (in the six place among the tech companies). After the IPO, OPAI is guaranteed to be
208.
by ozgung 1y ago
That's the real problem with the current process. It assumes all the candidates are just random people without any credentials. So candidates need to prove that they can code each and every time. Even a Phd degree doesn't count. Y
209.
by ozgung 1y ago
How do you hire a civil engineer? Do you ask them to solve a series of construction puzzles for few hours?
210.
by ozgung 1y ago
ChatGPT's original answer was surprisingly close. Both humans and LLMs fail to multiply those numbers in their heads. Some people can multiply them using a special algorithm. Either for mind or for pen and paper. Your example shows tha

More ›