Obota

NLP: Natural Language Processing LLM: Large Language Model RLHF: Reinforcement Learning from Human Feedback. PPO: Proximal Policy Optimization DPO: Direct Preference Optimization ROUGE: Recall Oriented Understanding for Gisting Evaluation.

thank you for your attention to this matter.

the algorithm seems to be winning the war. we've begun doubting ourselves and our abilities... our focus has shifted to "is this written by a human or the machine?"

how did you know it was ai-written? which tool did you use or is it your sixth sense? such takes make me cry for the majority of african students. if you write poorly you're penalised, & if you write well, i guess it will be ai.

sigmoid.

graphical neural networks (gnn) are a powerful framework for capturing different graphs. gnns follow a neighborhood aggregation scheme, where the representation vector of a node is computed recursively. still, weisfeiler-lehman graph isomorphism test is more powerful to gnn.

have large language models become smatter or are humans getting used to the models output?

humbled 🙏

there's no difference between zero-shot and few-shot when it comes to prompting an llm for better output. also, role prompting is ineffective when it comes to complex problems & finally, stop emotional prompting - it doesn't work.

intelligence is what you use when you don't know what to do.

13,300,000,000. 13 BILLION, 300 MILLION! and still pay the salaries and allowances, offer them loans at good rates. people can be greedy.💔

in due time, you'll realize that you only need the basics in life.

😂

i'm sorry but i have to read this. 😭 "biological and phsico-chemical status of two wetlands in the nairobi national park, kenya." spoiler: i don't think he wrote it.

i need to pay you but i'm waiting for someone else to pay me first, who's waiting for someone else to pay them first, who's waiting for someone else to pay them first, who's waiting for someone else to pay them first. wait time per individual is 3 days. this economy🤧

incorporating multiple signals during training can significantly reduce data wastage. omnihuman by bytedance employs a mixed data training strategy with multimodality motion conditioning, leveraging scalability to overcome the scarcity of high-quality data.

there's a time i applied for a job at McKinsey. hehe, i saw bad things. the questions were like "am i the first person to go to the university in my family?" "am i connected politically?" wueh!

the four main ways to build & improve reasoning models: 1. inference-time scaling. 2. pure reinforcement learning. 3. supervised fine tuning & reinforcement learning (sft + rl). 4. pure supervised fine tuning & distillation. ~ ahead of AI.

received your email "understanding reasoning llms"😊🤝

politically, populists are winning. sad state of affairs.