Khalid Ahmed Ebrahim

AI Engineer building AI agents, mastering LLM fine-tuning & RAG. Skilled in LangChain, Autogen, Vector DBs, & ML solutions across industries. #AI #ML #DL
#LLMs

[Khalid Ahmed Ebrahim @K7alooda]

خد بالك من حاجتين:
الـ scheduler بيحدد المحطات، يعني مستويات الـ noise اللي هتقف عندها.
والـ sampler هو طريقة السواقة بينهم، ومنه اللي بيضيف noise جديد كل خطوة ومنه لأ.

ونصيحة: خلّي الـ CFG = 1 وخد الإعدادات من الـ model card.

#الذكاء_الاصطناعي
#ComfyUI #AIVideo

[Khalid Ahmed Ebrahim @K7alooda]

نصيحة لو بتشتغل على موديلات diffusion للصور أو الفيديو: دوّر الأول على الـ Lightning LoRA بتاعته ⚡

أغلب الموديلات زي Wan 2.2 و Qwen-Image بينزل لها واحدة (غالبًا من lightx2v)، وبتقلل الـ denoising steps لـ 4 بس بدل عشرات الخطوات 🧵

#GenerativeAI #DiffusionModels #AI

[Khalid Ahmed Ebrahim @K7alooda]

أول ما DDPM نزل كان بياخد حوالي 1000 خطوة، لأنه Markov chain وكل خطوة لازم تبقى صغيرة.
وبعدها DDIM كسر الفكرة دي ونزل لحوالي 50.
الـ flow matching خلّى الطريق أقصر، والـ distillation علّم الموديل ينط في 4 أو 2 أو حتى 1 🤯

#FlowMatching #StableDiffusion #LoRA

[Khalid Ahmed Ebrahim @K7alooda]

Know all three, and a 4th option shows up: combine text and audio features in one model.

With a narrow toolbox, you see the solution you know.
With a wide one, you see the problem first, then the trade-offs.

Which would you pick?

#SentimentAnalysis #AudioAI

[Khalid Ahmed Ebrahim @K7alooda]

One problem, three fields, three solutions.

Sentiment analysis is a classic NLP task: a sentence in, a label out (positive, negative, or neutral).

But what if the input is audio? The approach you reach for first depends on what you've studied. 🧵

#AI #MachineLearning

[Khalid Ahmed Ebrahim @K7alooda]

Speech processing:
Use a pretrained speech model (wav2vec 2.0, Whisper's encoder) to turn the audio into embeddings, then train a classifier on them.

It captures how something is said and much of what is said, but costs more compute.

#SpeechRecognition

[Khalid Ahmed Ebrahim @K7alooda]

Would GPT-5.6 have been just as powerful if Anthropic hadn't announced Mythos 5 & Fable 5?
Maybe frontier labs don't only choose when to ship. Maybe they also choose how much capability to expose.
What do you think?
#AI #OpenAI #Anthropic #Mythos #Fable #Sol #Terra #Luna