to be honest at this point this just seems to work haha

Try Brilliant's tutor for free over here: https://brilliant.org/DeepLearningwithYacine/ . You’ll also get 20% off an annual Premium subscription to unlock all courses. I'm receiving so much questions from students about whether they should be doing a PhD in a tech related topic (CS / AI / Robotics) and the answers is kind of different for everyone, but also kind of the same? so here is my answer packaged up with a nice example of the latest such question I received about whether or not you should do a PhD in tech in 2026. also btw I've dropped out of my PhD a while ago so I might not be the best to answer these sort of questions I found that the video from this professor really good : https://www.youtube.com/watch?v=m8fdL_eSYgs&t=406s Also if for beginners: 📌 learn to code from full-stack to AI with Scrimba https://scrimba.com/?via=yacineMahdid (extra 20% off pro with my link, great resource, I love the team) Table of Content: - 0:00 : should you do a phd??? - 1:02 : latest variation of the "PhD question" - 2:48 : the importance of the stated goal - 4:33 : there is benefits to a PhD - 6:09 : benefit 1: deep uninterrupted focus - 6:54 : benefit 2: playing with big toys - 8:02 : benefit 3: network with the top dogs - 9:18 : learning? - 11:28 : answering the damn question - 13:00 asking the right damn question blog about goal setting: https://www.yacinemahdid.com/p/how-to-stop-feeling-lost-in-tech blog about networking: https://x.com/yacinelearning/status/1990168103284801633?s=20 also do check Rajan blog posts they are very informative on his type of mindset: https://www.rajan.sh and do read Alex L. Zhang thought here they are golden: https://alexzhang13.github.io Enjoy my guys! 🌹 FTC disclaimer: This video was sponsored by Brilliant ---- Join the newsletter for weekly AI content: https://yacinemahdid.com LinkedIn: https://www.linkedin.com/in/yacinemahdid/ Twitter: https://x.com/yacinelearning ___ Have a great week! 👋

does a hyper hyper network exist you think

today we're finding out how you actually build a frontier open AI model with Chris Alexiuk, senior product research engineer at NVIDIA! nemotron 3 ultra is 550B params, trained in 4-bit, ~6x faster inference than comparable open models, and every artifact is released openly weights, datasets, recipes, even the failed runs! we'll walk through how a model like this actually gets built end to end: the hybrid mamba-transformer architecture choices, the synthetic data pipelines, the multi-environment RL and multi-teacher distillation (MOPD) behind its agentic skills, and the infrastructure that holds it all together at scale!

today we are taking a deeper look at one of the overlooked levers in LLM post-training: diversity, with Ryan Bahlous-Boldi, PhD student at MIT! ryan and his team noticed that standard RL post-training with GRPO quietly collapses your model onto a very tight set of potential answers space, which is exactly the wrong thing when you're running test-time search on top of it like alpha evolve. so they built Vector Policy Optimization (VPO), a drop-in replacement that trains the model to output a whole set of solutions covering the Pareto front of a vector reward instead of chasing one scalar! we'll get into the details of how it works, the pretty neat results on a few benchmark, why their stochastic scalarization setup makes it work, and the evolutionary search experiments where VPO cracked problems GRPO couldn't do at all! I’ve got a pretty good feeling about this research direction!

today we are taking a deeper look at one of the overlooked levers in LLM post-training: diversity, with Ryan Bahlous-Boldi, PhD student at MIT! ryan and his team noticed that standard RL post-training with GRPO quietly collapses your model onto a very tight set of potential answers space, which is exactly the wrong thing when you're running test-time search on top of it like alpha evolve. so they built Vector Policy Optimization (VPO), a drop-in replacement that trains the model to output a whole set of solutions covering the Pareto front of a vector reward instead of chasing one scalar! we'll get into the details of how it works, the pretty neat results on a few benchmark, why their stochastic scalarization setup makes it work, and the evolutionary search experiments where VPO cracked problems GRPO couldn't do at all! I’ve got a pretty good feeling about this research direction!

Benchmarking LLM in 2026 is now a high stake task that requires deep expertise on LLM architecture, Prompt Engineering, Harness, Sandbox, Scoring, Hardware, Serving, and more. It’s a much different exercise than in the past that have a big impact on the performance of the final model. Not only that but benchmarking also is starting to have deep cybersecurity concern as frontier models have gained the capabilities of chaining together exploit of various kind (in order to cheat on benchmark). Underelicitation, Safety concern, cost, all plays a role in figuring out if the new models are the right tool for the job and has downstream consequence across the economy. Well in this tutorial we are exploring LLM evaluations with a researcher who spends his days taking benchmarks apart: Florian Brand! he wrote benchmark deep-dives, co-authors the open-models coverage at Nathan Lambert's Interconnects, and now works on evals as a research engineer at Prime Intellect. we'll go through a gentle intro to why evals matter for beginners, dig into why two scores on the "same" benchmark often aren't comparable, and explore where evals are heading in the age of agents + a live walkthrough of the evaluation module in the Prime Intellect lab! Pretty packed session, but highly informative! Enjoy 🌹 important links: 👉 florian twitter: https://x.com/xeophon 👉 prime intellect twitter: https://x.com/PrimeIntellect 👉 learn more about evaluation at prime intellect here:https://www.primeintellect.ai/blog/hosted-evaluations Check out Arcee (American open research lab and sponsor of this interview): ♥️ https://www.arcee.ai/ Also trinity large thinking is available on open router: ♥️ https://openrouter.ai/arcee-ai/trinity-large-thinking#providers Also also for beginners: 📌 learn to code from full-stack to AI with Scrimba https://scrimba.com/?via=yacineMahdid (extra 20% off pro with my link, great resource, I love the team) Table of Content: 0:00:00: are AI Benchmark broken? 0:05:45: Florian Brand background 0:09:00: what motivates florian to work on evaluation? 0:13:33: what is the mirrorcode benchmark about? 0:18:20: cheating in agent benchmark is insaneeeee 0:24:08: LLM benchmarks in era of agents 0:26:30: what’s up with the pelican man 0:28:27: evals are about capabilities 0:31:46: components of running evals 0:35:30: the volume of things to audit is huge!!! 0:40:20: expert answers are wrong hahahaha 0:46:00: api providers aren’t the same 0:48:00: benchmark narrow capabilities (synthetically) 0:50:56: link between eval and environment 0:53:45: small validated benchmark or massive bench? 0:56:11: what is your flow to review a benchmark? 0:58:30: tracking work capabilities with evaluation 1:00:20: slide deck in industry is all vibecoded 1:03:30: harness impact in the evaluation 1:07:39: hardware/sandboxes impact evaluation too! 1:11:00: “is it going to get worse?” 1:12:40: all components influence the final score 1:13:50: training models on different harnesses? 1:17:20: is the model just the weights or it’s all of it? 1:19:30: how to craft benchmark that prevent to cheating and undereliciting models in 2026 1:23:19: ways agents cheat and steal 1:26:00: correct elicitation of capabilities is important 1:36:00: building evaluation on prime intellect 1:45:10: how do you design interactivity benchmarks? 1:48:40: do you think evals are well set to reflect real world performance? 1:52:50: what will the benchmarking landscape will look like in 1 year Papers & references to check out: 📌 why benchmarking is hella hard: https://epoch.ai/gradient-updates/why-benchmarking-is-hard 📌 benchmarking in 2026 situation: https://florianbrand.com/posts/benches-2026 enjoy folks hope this stuff is useful🌹 ---- Join the newsletter for weekly AI content: https://yacinemahdid.com Join the Discord for general discussion: https://discord.gg/QpkxRbQBpf ---- Follow Me Online Here: Twitter: https://twitter.com/yacinelearning LinkedIn: https://www.linkedin.com/in/yacinemahdid/ ___ Have a great week! 👋

it is really comforting that the one constant is forgetfulness

catastrophic forgetting!

class incremental is where it's at.

we need like 32 more mandy lu in the world

the amount of physicist turned ML scientist is off the chart.

usually when you are diving to patch that gap it's much easier than learning about the material the first time.

earlier this year I took two of my friends ludwig and lazarz on a big discussion about how to leverage ai for learning purposes in general. what I like about these two is that they are their own man and will say what they think without reserve. what I also like is that they have quite different learning background with ludwig being an almost pure hustling self-thaught learner and lazarz followed an academic path driven mainly by his own curiosity. it was a great conversion which I think might be helpful for the kids out there stressed out of their mind by what to do with their self growth in this ai age. enjoy! 🌹 ps: also do follow my guys on twitter they are net gain: - ludwig: https://x.com/ludwigABAP - lazarz: https://x.com/Laz4rz Also also for beginners: 📌 learn to code from full-stack to AI with Scrimba https://scrimba.com/?via=yacineMahdid (extra 20% off pro with my link, great resource, I love the team) **Table of Content:** 0:00 - man I love learning 2:30 - ludwig wild background 5:50 - “my guy where the hell this drive came from” 8:32 - lazarz inspirational background 12:00 - that kid that understand everything first shot discussion 21:18 - the benefit of formal university education 30:00 - “you don’t need [to know] much to make money” 36:10 - “ma man we need this shipped get it done” plus research is social 42:32 - “getting into an university for networking is a great idea” 44:50 - [ludwig] how to decide what is the next thing to learn? 49:20 - “in most area that aren’t math the map of knowledge is fucked” 58:48 - “the brain is one hell of a pattern matching machine” 1:00:08 - [lazarz] how to decide what is the next thing to learn? 1:08:30 - adding value to the world (and yourself) is the way to go get lucky 1:18:39 - how do you deal with the ai doom feeling 1:20:45 - “if you have no goals and you think it’s pointless well sorry to break it to you but it has always been pointless” 1:35:35 - why LLM are fundamentally different from human intelligence? 1:43:10 - non-neuron cells intelligence is super smart 1:48:45 - is neuroscience still useful analogy for creating ai system? 2:02:17 - five years from now what is the one skill you glad you have improved? ---- Join the newsletter for weekly AI content: https://yacinemahdid.com Join the Discord for general discussion: https://discord.gg/QpkxRbQBpf ---- Follow Me Online Here: Twitter: https://twitter.com/yacinelearning LinkedIn: https://www.linkedin.com/in/yacinemahdid/ ___ Have a great week! 👋

kinda wild how this would work you really have to get an intuition on the hyper params.

a loss graph with no spike 😌

try a bunch and guesstimate is a solid tip.

client: "🤬🤬🤬🤬" you: "what you mean??"

I've asked more dumb questions in my life than I can remember the answer for.

go do something that you are unqualified wow man what an advice.

you can just not know things

📌 learn to code from full-stack to AI with Scrimba https://scrimba.com/the-ai-engineer-path-c02v?via=yacineMahdid (extra 20% off pro with my link, great resource, I love the team) how do you actually learn AI in 2026, without wasting months on the wrong things? I sat down with Mandy Lu, a Stanford PhD student in applied math who's worked across basically every corner of the AI world: research in Fei-Fei Li's lab at Stanford, AI for healthcare and climate at Google Research, plus stints at Apple and NVIDIA. We talk about how she'd learn AI today if she were starting over, the fundamentals most people skip, and why finding someone to "mimic" beats grinding textbooks. If you're a beginner trying to break into AI, start here. 👇 📬 first steps: her Substack: where she breaks down the fundamentals and how to actually learn this stuff: → https://mandyslu.substack.com/profile/posts check here twitter too: → https://x.com/mandylu ⏱️ table of content 0:00 - mandy is the best 2:03 - my honest opinion on learning ai 3:17 - mandy background 6:25 - mandy current ai project 7:27 - how would mandy learn ai in 2026 16:06 - next first steps Also: 🔬 mandy's work: research & publications: → https://scholar.google.com/citations?user=wXRU2jUAAAAJ&hl=en&oi=ao google research | earth AI (climate, floods, agriculture): → https://research.google/research-areas/google-earth-ai/ google research | AI for diabetic retinopathy screening (600,000+ screenings worldwide): → https://blog.google/company-news/inside-google/around-the-globe/google-asia/arda-diabetic-retinopathy-india-thailand/ apple | the memories feature she worked on: → https://support.apple.com/en-ca/guide/iphone/iphd4f70e68f/ios NVIDIA | Jetson TX2 embedded GPU: → https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-tx2/ 📚 stanford courses mandy recommends: CS336: Language Models from Scratch (publicly available, full LLM engineering) CS229: Machine Learning / deep learning theory CS224N: NLP with deep learning CS231N: Computer Vision ---- Join the newsletter for weekly AI content: https://yacinemahdid.com Join the Discord for general discussion: https://discord.gg/QpkxRbQBpf ---- Follow Me Online Here: Twitter: https://twitter.com/yacinelearning LinkedIn: https://www.linkedin.com/in/yacinemahdid/ ___ Have a great week! 👋 #AI #MachineLearning #LearnAI #ArtificialIntelligence #DeepLearning

teaching really is an integral part of learning

just help people out guys

the amount of talent in the arcee team is beautiful to see man. like they are over there being 100% open about how their model is done and give it fully open weight for everyone to tweak. there is no where to hide and they answer every questions with such calm I love these guys.

today we are taking a deeper look at one of the overlooked levers in LLM post-training: diversity, with Ryan Bahlous-Boldi, PhD student at MIT! ryan and his team noticed that standard RL post-training with GRPO quietly collapses your model onto a very tight set of potential answers space, which is exactly the wrong thing when you're running test-time search on top of it like alpha evolve. so they built Vector Policy Optimization (VPO), a drop-in replacement that trains the model to output a whole set of solutions covering the Pareto front of a vector reward instead of chasing one scalar! we'll get into the details of how it works, the pretty neat results on a few benchmark, why their stochastic scalarization setup makes it work, and the evolutionary search experiments where VPO cracked problems GRPO couldn't do at all! I’ve got a pretty good feeling about this research direction!

to be honest I think they will keep climbing up.

the fine line that open weight LLM company have to walk on is very thin: too much open and there is no way you can fund the next iteration and it's effectively game over. too much close and like what's the point of having an open weight model anyway?

it's kind of crazy the amount of moving parts in a pre-training report, it takes so much inner scientific intuition to figure out where to bring the model next!

today we are taking a deeper look at of the secret behind LLM pretraining which is synthetic data pipelines with Joel Niklaus machine learning engineer at Hugging Face! joel an his team ran 90 controlled experiments and burned over a trillion tokens to figure out what actually makes good pretraining data and in huggingface fashion provided all of their findings/artifacts openly! We’ll check the different structured format they used, the multiple ablations they ran and the counter intuitive outcomes they got out of the result! Will be a fun one (recorded too no worries about it)

today we are taking a deeper look at of the secret behind LLM pretraining which is synthetic data pipelines with Joel Niklaus machine learning engineer at Hugging Face! joel an his team ran 90 controlled experiments and burned over a trillion tokens to figure out what actually makes good pretraining data and in huggingface fashion provided all of their findings/artifacts openly! We’ll check the different structured format they used, the multiple ablations they ran and the counter intuitive outcomes they got out of the result! Will be a fun one (recorded too no worries about it)

this manual experiment is for the wacky ideas that are purely taste driven and make 0 sense.

actually I love this idea, would make it very easy for ai agents to know how to build better experiments.

today we are exploring the strange geometry of neural networks where diverse task experts are densely packed around pretrained weights with Yulu Gan from MIT CSAIL! his new paper "Neural Thickets" shows that simple random perturbation of the weights (RandOpt) can rival gradient-based post-training methods like GRPO. we'll walk through the paper step by step with a gentle introduction for beginners, dig into why these thickets emerge at all from first principles, and explore what this means for post-training, distillation and continual learning! ps: check out the project blog over here (very visual): https://thickets.mit.edu/

kind of interesting that knowing where to dig deeper is still a very human-centric trait.

it’s funny because the two things mentioned are the hardest to get right in a research project 😮‍💨

it does make sense though if whenever things are digitalized you can leverage powerful ai system there will be an incentive for physical labs to digitalize the process in some shape or form…

today we are exploring LLM evaluations, and the surprisingly fragile numbers the whole field runs on, with a researcher who spends his days taking benchmarks apart: Florian Brand! he wrrote the benchmark deep-dives where he picks well-known benchmarks apart sample by sample to figure out what they actually measure, co-authors the open-models coverage at Nathan Lambert's Interconnects, and now works on evals as a research engineer at Prime Intellect. we'll go through a gentle intro to why evals matter for beginners, dig into why two scores on the "same" benchmark often aren't comparable, and explore where evals are heading in the age of agents + a live walkthrough of the evaluation module in the Prime Intellect lab! come hang out and ask your evals questions! 🌹 ps: you can check out his deep-dive on why benchmarking is so hard over here: https://epoch.ai/gradient-updates/why-benchmarking-is-hard he has laid out in a previous blog post 5 key characteristics of what CL should be fundamentally. we’ll review these along with a gentle introduction to continual learning for beginners, but also explore his current research on the topic through the prime intellect residency! come hang out and ask your continual learning questions! 🌹 ps: check out his blog on the subject over here: https://x.com/carnot_cyclist/status/2041479655035679163

kind of true that usually teacher even for human is to do the thing well. learning on your own is where you figure out the issues and how to recover.

in-context learning is such a magical thing when you think of it.

prioritization is such an important element as you become an adult.

it's like the simplest method ever man.