Sergiu Nistor

@yulintwt Well that's pretty much every deep learning paper ever. Very few papers provide mechanistic interpretability motivations. Most just provide intuition-based motivations, or just present the results on some benchmarks and call it a day. AI research is engineering, not research.

@ScholarshipfPhd hi

Say hi and I’ll recommend a research topic that perfectly fits your profile.

@GenAI_is_real You're confusing ML engineer with AI engineer. AI engineers use LLMs in the context of agents. ML engineers fine-tune LLMs, train other models, handle data pipelines, build training pipelines etc.

ml engineer is a legacy title. by 2030, you’re either an architect who knows how to orchestrate thousands of agents, or you’re just a casual user. the real job is 'system builder'—ensuring that the training-inference loop is tight and the data distribution doesn't drift. vibe coding is the entry point, but high-concurrency systems like sglang are the foundations that actually make those vibes scale.

From my observation of friends around me, those who’ve worked at frontier AI labs experience exponential growth. It’s not just technical. It’s a deeper shift in how they view the world, trends, and themselves. Being immersed in an environment full of other exceptional people led to exponential growth. There’s a clear lesson here for startups: hire the very best, put them together, and you get compounding effects. And for each individual, find the environment that puts you on an exponential curve.

if you want to get into research this is actually one of the best times to do it. as a beginner you don’t need to wait for opportunities to come to you or for the “right” people to guide you...you can just start. pick a domain you’re curious about (RL, mech-interp, diffusion models, etc.) and begin running experiments + learn about ml systems, inference, infra while building projects. i remember as a beginner some of the experiments i worked on (and i am still experimenting!) involved analyzing what changes inside a model during post-training, track where the model changes internally, how those changes evolve across checkpoints and which layers causally implement the new behaviors, looking at internal geometric shifts as training progresses, understanding how reward hacking happens by creating synthetic datasets, intentionally pushing models to reward-hack in an environment and then analyzing what’s going on internally. in this economy the only real proof of work is what you’ve built, the messy experiments you run and the knowledge you gain while fixing and debugging things. everything else is just a proxy.

@mishig25 @xenovacom I'd totally love this. I've been looking for the hardcover version of it for quite a while.

Who wants the hardcover? h/t @xenovacom for the design

Just finished the Data Analytics Specialization from DeepLearning.AI. 📊 A more relaxed, practical pause from deep learning. Focused on applied stats, analytics workflows, and making insights clear and actionable for stakeholders.

First cert of the year 🚀 Completed DeepLearning.AI’s Mathematics for ML & Data Science Specialization. - Strengthened understanding of eigenvectors - Reinforced Bayesian fundamentals Now reading Bishop’s "Pattern Recognition & ML" New year, stronger fundamentals 📈

@tom_doerr Well we already got LocalStack (localstack.cloud) for running locally a lot of AWS services.

S3 server for local development and testing github.com/johannesboyne/…

@MartinShkreli Could you please provide the reference for the numbers?

55 billion minutes spent on AI sites last month. 64% chatgpt, 15% gemini. deepseek and grok the only two on a trajectory to come even close.

@eliebakouch I rlly need a hardcover version of this

Training LLMs end to end is hard. Very excited to share our new blog (book?) that cover the full pipeline: pre-training, post-training and infra. 200+ pages of what worked, what didn’t, and how to make it run reliably huggingface.co/spaces/Hugging…

Christmas is close, but the learning continues! I completed the Machine Learning Specialization from DeepLearning.AI and found the reinforcement learning and recommender system modules especially strong. Andrew Ng’s clarity explaining complex ideas stands out. Onward!

@Rondo_ina_Condo @DeepLearningAI I honestly would love to get your perspective on this when you're done with it!

@SergiuNistor6 @DeepLearningAI I literally just started like 48 hours ago. So I can't answer that cuz I don't know. Definitely get back to you on this 😅

Just wrapped up the Deep Learning Specialization by @DeepLearningAI 🎉 Big takeaway: the course really nails why models are built the way they are. It connects math function design directly to goals like memory and attention, so architectures actually make intuitive sense.

@Rondo_ina_Condo @DeepLearningAI Best of luck! What do you think about the mathematical interpretations for the GRU and LSTM memory gates, and for self-attention, and transformer positional encodings? Personally it really changed the way I think about these models. It makes much more sense now.

@SergiuNistor6 @DeepLearningAI Congratulations. Im working on this specialization also

@embuildsai @DeepLearningAI However, there are more specializations and courses on the platform, for all levels. It's a gold mine really.

@SergiuNistor6 @DeepLearningAI how was the process working on this? I’ve used this site before and felt like it takes a while.

@embuildsai @DeepLearningAI Unfortunately I am a bit of a special case since I already have experience in the field. I think it is indeed quite hard, but they do start with the basics. I think it's more of an intermediary/advanced specialization, as it explains concrete methods and models in an intuitive.

@dwarkesh_sp is the new @lexfridman