learning with yacine show

learning with yacine show

By yacine mahdidScienceTechnology
Download on the App Store

All learning with yacine show episodes

  • How to Get Into AI Research in 2026?

    should you do a PhD? should you leave your software engineering job for grad school? how do you get into AI research coming from industry?I get these questions all the time, so I sat down with Andrew Drozdov, senior research scientist and one of the retrieval leads at Databricks, who went through every single one of these steps.andrew wrapped up his undergrad at Michigan a semester early without even planning it, then filled a whole semester with grad-level courses like J. Alex Halderman's security class (two research papers per lecture, reviews graded by your peers, and live demos like freezing RAM with an upside-down aerosol can to read the password straight out of memory!). after that he worked at Okta and Datadog, did Andrew Ng's lectures on the side, and made the tough call to leave a job he liked for a master's at NYU right as attention was taking off, then a PhD at UMass.we talk about how to know if grad school is actually right for you, why "basically no one should do a PhD" (unless you are really passionate), why you should tie your success to more than papers when ICLR is getting 60,000 submissions, his front row seat at NYU's CILVR lab (attention in RNNs, parallel neural architectures, latent syntax), and how to pick work that lasts by anchoring yourself on problems rather than methods.I really enjoyed this interview, a great starting point if you are thinking about grad school or a research career. part 2 on KARL and AutoIndex is coming soon!Also, check out over here a few useful links from andrew:👉 his website: https://mrdrozdov.github.io/👉 his advice for PhD students: https://mrdrozdov.github.io/blog/2024/advice/👉 check his twitter too: https://x.com/mrdrozdov👉 his PhD thesis on adaptive retrieval-based methods!!!! : https://scholarworks.umass.edu/entities/publication/04520d37-6227-4441-9ec8-e91495271ce8👉 KARL (for part 2): https://arxiv.org/abs/2603.05218👉 AutoIndex (for part 2): https://arxiv.org/abs/2607.18603btw also check out more of Deep-ML I love you moe man keep doing what you are doing:📌 premium: https://www.deep-ml.com/premium📌 the Anthropic track to get started: https://www.deep-ml.com/interview-prep/anthropic

    Table of Content:- 00:00:00 - should you do a PhD? (and all the other questions I keep getting)- 00:03:34 - Andrew Drozdov!!!!- 00:06:40 - why take grad-level classes during your undergrad?- 00:10:18 - Halderman's security class (two papers per lecture + peer graded reviews)- 00:14:39 - a cool demo about cold booting demo🪄- 00:16:22 - is this where your research interest was seeded?- 00:18:32 - what advice would you give your undergrad self getting into AI research?- 00:21:50 - just get in! (help out in a lab even with zero qualification)- 00:23:10 - how did working at Okta and Datadog help you afterward?- 00:27:17 - what made you ready to leave a job for grad school?- 00:29:50 - basically no one should do a PhD, unless...- 00:30:27 - ICLR at 20,000 plus submissions: tie your success to more than papers- 00:32:53 - you don't have to finish it (dropping out to go to Big Co is ok)- 00:34:23 - how was it to study during the 2015-2024 big boom?- 00:35:10 - front row seat at NYU's CILVR lab (attention in RNNs, QRNN, latent syntax)- 00:39:38 - is attention really all we need? - 00:40:33 - how do you know what you work on will still be useful in the future?- 00:47:18 - conclusion: nobody is going to hand you the path into research man seize itenjoy and do follow andrew on twitter, and read his PhD advice post if you are thinking about grad school!

    Andrew Drozdov — transcript

    00:00:00 – Introduction

    Yacine Mahdid

    Should you do a PhD? What should you major in? Should you take a break right now and go to industry? How do you get into research from a software engineering background? I get variations of all of these questions all the time, so I figured I’d bring in someone who is uniquely positioned to answer them better than I can: Andrew Drozdov, research scientist at Databricks.

    He’s one of the retrieval leads on their research team, and he recently co-authored KARL, a reinforcement learning paper on knowledge agents, among other things. That paper is how I first came across his very elegant work. By the way, we’re going to have a part two with him on this topic sometime this month, if I manage to get my stuff together to edit it.

    His main focus is finding ways to give models the context they need so that they can actually answer the questions people care about, especially in enterprise settings.

    He studied computer science at the University of Michigan, and he was cranking through requirements so fast that he finished a semester early, which is so funny. Instead of just coasting, he took graduate courses that actually sparked his research interest.

    After that, he went into industry at Okta and Datadog as an early employee, and he started working through Andrew Ng’s machine learning lectures on the side. Eventually he made the tough call of leaving a job he liked to go after AI, and this was way before the LLM takeoff. He did his master’s at NYU right as attention and the early transformer wave were starting to form. Then he did his PhD at UMass.

    During this interview, I asked him a whole bunch of questions fueled by the kind I’ve been receiving. It’s pretty all-encompassing, with the grad school stuff, the industry stuff and the research stuff all blended together. Hope you enjoy.

    00:03:34 – Pulling in context the model never saw

    Andrew Drozdov

    All right. So, Yacine, thanks for having me on. You reached out to me, I feel like, more than six months ago, shortly after the KARL paper came out, so I’m glad we’re getting this together.

    I’m Andrew Drozdov. I’m a research scientist at Databricks, I work on retrieval research, and I’m one of the retrieval leads on the research team. I like to work on problems where we pull in important context that was not necessarily in the model’s training data, so that it can actually get the answer to your questions correct.

    Basically, there are a lot of questions that people care about, that are really high value, that are based on data from their daily interactions, or from business information, or from real news events that are coming out all the time. And there’s just no way the model can answer those questions without pulling in that information.

    I think by now it’s become kind of standard to do these types of things, but when I was doing my PhD, it was not as clear. So I’m glad people realize this is important, and I really enjoy working on this problem. I did my PhD on core NLP, on retrieval and on language models, and I’m basically still working on all those topics day to day.

    Yacine Mahdid

    Yeah, and it’s still very important in various ways, not just for the big LLMs. I have a project right now with one company I’m working with, and we’re almost done writing the paper, but we had to finetune the cross-encoder to help do the retrieval. All the techniques behind it are extremely important, especially when you actually have a whole bunch of data, you’re live in production, and you have latency and all these concerns. I’d say it became even more important than before.

    Andrew Drozdov

    It’s a big space. I’m happy to talk about all the little corners.

    00:05:52 – The grad class that blew his mind

    Yacine Mahdid

    But I think it would be cool if we could wind it back to your undergrad. The whole arc of your career has all the facets people usually ask about: industry, academia, what to research, whether you should do a master’s. All of it is there.

    I want to take it back first to the undergrad portion, because you did something I also did when I was doing my undergrad: y…



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    49 min
  • How to Make Better Robot Policies with Dense Subtask Labels?

    most robot datasets give the robot one instruction for a whole episode, like “clean the table”, and then a few minutes of motion. the model has to figure out by itself where one step ends and the next begins which is hella hard.I had the pleasure of chatting with jade choghari from scale AI, who walks us through FineART, a 1,700 hour two-arm robot dataset where every single step is labeled in plain words by a human (534k labels, no VLM in the loop!)these subtask labels do a lot of heavy lifting to shape a proper policy.without them the model can’t even reliably tell the left bin from the right one, even with a big vision-language backbone behind it!with them, the steps it learned in words carry over when transfered to a completely different robot, to the point where it pulls off a long breakfast task it never practiced there and even reuses an “adjust” move it only ever saw on the previous robot!very fun chat, we also get into how scale ai actually collects this stuff behind the scenes which is as impressive as the result themselves!👉 Jade Website: https://jadechoghari.github.io/👉 Jade GitHub: https://github.com/jadechoghari👉 FineART (paper): https://arxiv.org/abs/2609.36416👉 FineART-VLA (model): https://huggingface.co/lerobot/fineart_vla_baseCheck 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♥️ and check out the NAC harness https://www.arcee.ai/blog/nacTable of Content:- 00:00:00 - what is FineART?- 00:03:55 - my man jade choghari at ScaleAI SF HQ- 00:06:34 - highschool days to machine learning research engineering- 00:10:10 - the waterloo magic touch 🪄- 00:10:50 - how did you get into lerobot?- 00:12:46 - what drove jade into ai research?- 00:15:01 - FineART: Data + VLA walkthrough- 00:16:08 - “Divide each difficulty into as many parts as is feasible and necessary to resolve it” Rene Descartes- 00:17:30 - how do you decide what a subtask should be?- 00:21:56 - why full 100% human organic??- 00:24:35 - dataset distribution- 00:26:28 - how in the world did you pull off 100% human annotation in there man- 00:31:05 - FineART-VLA policy + loss function- 00:36:18: FineART-VLA interesting emerging behavior- 00:42:00 - why does knowledge insulation protect the VLM?- 00:45:15 - you think we could relax knowledge insulation once the expert is up to speed?- 00:47:10 - what type of steps does the model get wrong?- 00:51:31 - what context does the model get for each task?- 00:55:20 - how are you handling error in the teleoperation setup?- 01:01:00 - how many review passes does a label go through before it’s considered good?- 01:04:00 - the surprising result in the breakfast task- 01:07:40 - what makes skills transfer across robots just work?- 01:09:50 - where did the model failed?- 01:11:25 - does the model get lost on long tasks?- 01:12:35 - why did ABC collapse while FineART kept climbing the benchmark?- 01:14:17 - what’s next for FineART?- 01:19:08 - how should you mix sim + teleop + human video data?- 01:22:28 - is the VLM backbone here to stay?- 1:26:25 - what is jade excited about these days?



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    1 hr 29 min
  • LeRobot: Open-Source End-to-End Robot Learning with Steven & Pepijn from Hugging Face

    there is this kind of renaissance in robotics / physical ai at the moment and it's very stimulating.things finally just work ™️I had the pleasure of chatting with two experts from HuggingFace immersed in this environment steven palma and pepijn kooijmans from the LeRobot team about this renaissance. the gist of the mission of LeRobot is that robotics is very different from software/LLM in the sense that if you wanna start tinkering with robot you need to shell out quite a lot of money to get something working and handle a maddening stack of complexity.so the team at LeRobot is trying to level up the playing field by basically doing what the Transformers library did for NLP but for robotics.it's one library that takes you from plugging in a cheap arm like the SO-101/microduck/etc. to recording data, to training a policy (ACT, SmolVLA, all the way up to ports of the π series), to running it straight on the real robot. there are 90,000+ LeRobot datasets on the hub with more than 200 terabytes of robot data (and it's growing)!in this interview we get into:- how to get started in robotics- why real world is much better than sim- what porting a VLA like π0.5 actually looks like - why almost no data is bad data anymore as long as it's labeled right- Steven's essay on why open robotics makes us safer- how frontier models like Astra drive robots by writing classical inverse kinematics code - whether VLAs are dead now that world action models are here- and why robotics is just much harder than language!amazing session btw did my best to but a whole bunch of reference straight into the video hope you enjoy!

    useful links

    👉 Steven Hugging Face: https://huggingface.co/imstevenpmwork👉 Pepijn Hugging Face: https://huggingface.co/pepijn223👉 LeRobot (GitHub): https://github.com/huggingface/lerobot👉 LeRobot: An Open-Source Library for End-to-End Robot Learning (main paper): https://arxiv.org/abs/2602.22818👉 Why I Believe Open Robotics Makes Us Safer, Not Less Safe (Steven's essay): https://imstevenpmwork-open-robotics-blogpost.hf.space/👉 FineART (paper): https://arxiv.org/abs/2609.36416Check out Arcee (American open research lab and sponsor of this interview):♥️ https://www.arcee.ai/Also check out trinity large thinking:♥️ https://www.arcee.ai/blog/trinity-large-thinking♥️ and check out the NAC harness https://www.arcee.ai/blog/nac

    here is the full table of content btw:

    - 00:00:00 - to play with AI for robots you need a robot- 00:01:41 - sponsor: Arcee and the NAC harness- 00:03:43 - meet Steven and Pepijn from LeRobot- 00:04:02 - what did you do before getting into robotics?- 00:05:59 - how would you advise someone to get into robotics?- 00:09:07 - how do I get my 7 year old into robotics?- 00:11:55 - what is LeRobot? the end-to-end robot learning pipeline- 00:15:32 - is LeRobot the Transformers library for robotics?- 00:17:38 - open source hardware and making robots cheap- 00:21:22 - why real world first instead of simulation first?- 00:24:32 - the LeRobot community: hobbyists, startups, labs + the global hackathon- 00:27:35 - you don't need a frontier budget to do robotics research- 00:29:41 - how do research labs and industry connect?- 00:32:35 - no single thing will solve robotics: data, model size, latency, factory wifi- 00:37:08 - is Astra doing GPT-3 level zero-shot generalization now?- 00:39:59 - frontier models going back to classical robotics with inverse kinematics scripts- 00:42:53 - RT-1 tokenized actions and System 1 / System 2 setups- 00:45:53 - any plan to make robotics harnessing easier?- 00:49:14 - is LeRobot a platform or just a very good library?- 00:50:40 - the 3 principles of LeRobot- 00:53:08 - what does porting a VLA like π0.5 look like?- 00:55:40 - community optimizations vs the paper + you can't vibe test a robot- 00:58:31 - any bias toward some policies over others?- 00:59:57 - survival of the fittest for policies- 01:03:42 - could the 88k LeRobot datasets enable broad generalization?- 01:05:52 - a little book in the big library that everyone can read- 01:08:01 - what makes good demonstration data vs bad data?- 01:09:21 - almost no data is bad data if it's labeled right (feat. FineART)- 01:13:01 - record as many sensor modalities as you can- 01:16:46 - core argument of Steven's essay on open robotics and safety- 01:23:54 - put the bottleneck on hardware, not on research- 01:26:51 - the wifi analogy- 01:28:17 - are we close to finding the scaling law for robotics?- 01:31:25 - could there be multiple scaling laws for different components?- 01:33:53 - how many scenarios do you need to test to call a model general?- 01:38:08 - is the VLM backbone a crutch or a crucial component?- 01:42:08 - are VLAs dead now that world action models are here?- 01:44:01 - dead for who? use whatever works (XGBoost is still king sometimes)- 01:45:54 - is data the answer or are we still waiting for our GPT moment?- 01:47:06 - the GPT moment was distribution + instructability- 01:51:16 - how different is RL post-training in robotics vs language models?- 01:55:00 - what's next for you and for LeRobot?

    enjoy!!

    Full Transcript

    with full references!!!

    00:00:00 – Robotics for the Price of a Laptop

    Yacine Mahdid

    If you want to play with AI for text today, all you need is a laptop. If you want to play with AI for robots, you need a robot. And for a long time, that meant something costing around $100,000.

    So today I’m with Pepijn and Steven from LeRobot at Hugging Face, which is an open-source community effort to make robotics much more accessible. LeRobot is basically what Hugging Face’s Transformers library was for NLP, but for robotics. It’s one library that takes you all the way from plugging in a cheap arm, to recording some data, to training a policy, and then literally running inference straight on the real robot.

    The LeRobot paper, describing one open-source library for end-to-end robot learning, from hardware to deployment. (source)

    It ships everything from small models like ACT and SmolVLA all the way up to frontier models like the π series.

    ACT (Action Chunking with Transformers), one of the small policies LeRobot ships. (source)

    SmolVLA, Hugging Face’s small vision-language-action model for affordable robotics. (source)

    π0, Physical Intelligence’s vision-language-action flow model, the frontier end of what LeRobot supports. (source)

    The main vision behind this library is that the real magic in robotics is in the real world, and the only way to get a lot of real-world data is to make robotics cheap enough that the whole community can start to collect it. And it’s actually working. There are now around 88,000 LeRobot datasets on the Hub (I think it’s even higher than that now), and it’s growing. That’s more than 200 terabytes of robotics data.

    The number of LeRobot datasets on the Hugging Face Hub over time, climbing steeply. (source)

    So in this interview, we’ll cover LeRobot, of course, but I’ve also asked them for a few of their opinions on robotics in general. Stuff like why planning might already be mostly solved and where the real bottleneck in robotics might be. Why almost no data is bad data anymore, as long as it’s labeled. How frontier models are actually controlling these robots without being trained on them. Whether VLAs are dead now that world models are something that actually works. And why robotics is just harder than language.

    I’ll link the LeRobot repo and their ICLR paper down below. It’s a great companion to this conversation. Enjoy.

    00:03:43 – Robotics Is a Galaxy of Planets

    Steven Palma

    Hello everyone, my name is Steven. I’m a robotics engineer here at Hugging Face. I’ve been working on LeRobot almost from its very beginning, or well, around two years already. I’ve always been in the robotics space, so I’m very excited to be here. And thanks, Yacine, for the invitation.

    Yacine Mahdid

    Cool. I just want to understand a bit more: what did you do before getting into robotics?

    Steven Palma

    I like to see robotics as this galaxy, right, which is composed of different …



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    1 hr 58 min
  • Next-Latent Prediction Transformers Learn World Models feat. Jayden Teoh from MIT

    it’s very interesting to see where next-token transformers still fall short (like the path-star graph problems) while crushing problems that look even more daunting.I had the pleasure of chatting with jayden teoh from MIT who walks us through next-latent prediction (NextLat), a simple fix in the training objective he worked on when he was still an undergrad!the change is super tiny: -> bolt a small MLP onto a GPT and make it predict the transformer’s own next hidden state. to pull that off, the internal representation of the transformer needs to become a belief state, a compact summary of everything that matters for the future. cool thing is that you can throw the MLP away at inference and the model runs exactly like a normal GPT, just with a way better internal world!very solid lecture and we even have some great slides in there from jayden!

    👉 Jayden Twitter: https://x.com/jayden_teoh_👉 NextLat (thread):

    👉 NextLat (paper): https://arxiv.org/abs/2511.05963👉 NextLat (blog): https://jaydenteoh.github.io/blog/2026/nextlat/👉 NextLat (GitHub): https://github.com/JaydenTeoh/NextLatCheck 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♥️ and check out the NAC harness https://www.arcee.ai/blog/nacTable of content:- 00:00:00 - what’s next latent prediction?- 00:04:20 - walkthrough of the next latent prediction paper with your boy- 00:14:46 - “hey jayden how are you man”- 00:15:32 - introduction and background of jayden- 00:19:25 - how did you find and cultivate your research field of interest?- 00:23:43 - is failed experiment worth publishing?- 00:25:57 - book recommendation for diving deeper into deep learning- 00:28:49 - theoretical or empirical research design?- 00:30:57 - where did the intuition for NextLat came from?- 00:34:20 - [presentation] Next-Latent Prediction Transformers Learn Compact World Models- 00:39:27 - methodology of NextLat- 00:42:17 - why stop gradient?- 00:43:50 - did stop gradient was used for the MLP?- 00:45:17 - how do you think about this lack of generalization with the scale of transformer?- 00:49:47 - how is deepseek MTP different from NextLat?- 00:51:19 - better internal representation (of newyork)- 00:55:11 - why is MTP representation much more compressed than GTPs?- 00:57:58 - Planning (the path-star graph)- 01:01:25 - Reasoning (with the countdown game!)- 01:03:34 - stronger representation unlocked- 01:07:43 - how is NextLat different from JEPA or LLM-JEPA?- 01:10:13 - how would that loss work for other modalities?- 01:12:36 - how would you extend this to predict earlier layers or several layers at the same time?- 01:15:40 - language modeling and MTP- 01:19:10 - speculative decoding directly!- 01:23:13 - data-efficiency is super important- 01:25:09 - RNN without recurrence?????- 01:28:20 - where does the trained MLP length generalization coming from?- 01:31:03 - why not have more MLP at different layers?- 01:34:38 - what’s next for NextLat?- 01:37:40 - how about SEGreg instead of SG?- 01:40:45 - have you tried to visualize NextLat hidden state after training vs something like GPT?- 01:46:22 - do you think there are other inductive biases we should explore forcing the transformer to have?- 01:56:53 - how to prove things are scalable?- 02:02:34 - did you engineer the speculative decoding gain while designing NextLat?- 02:06:04 - what’s next for jayden and this research direction?

    enjoy! 🌹



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    2 hr 9 min
  • Prime Agent: Self-Improving RLM Harness with Seth Karten from Prime Intellect

    RLM + continual harness + multi-agent communication is what compose the prime agent harness developed by prime intellect which now sits at a gazillion stars on github (22.3K).I was kind of already mesmerized by the idea of having an RLM write codes to organize and pick the context to solve problem and to delegate subtask to sub agents; but ON TOP of that to have the harness around the RLM be modifiable by the RLM itself and the sub agents being able to communicate together took it to another level of interest.the ideas of creating a limited set of primitives that fit with the model shape and letting the main agent just figuring out how to run them sound just so right to me.BUT the usual assumption earlier in 2026 was that the harness is temporary thing. Meaning as models get smarter that scaffolding will eventually gets absorbed into the weights.YET we have seen in late 2026 how absolutely essential a solid harness with good primitives are which can be steered by the model itself.that’s why I sat down in september with seth karten from princeton and prime intellect to take a deeper look at how all of this stuff works.the team productionized alex zhang’s recursive language models and bolted a continual harness around, where the agent edits its own skills, memories, and prompts while it runs, plus persistent sub-agents that message each other directly.the kicker here is that if you read the paper thoroughly you get this feeling in the back of your mind that what you are reading is actually a computer science paper. akin more to something a operating system PhD would put out than anything ml related.we got confirmation from seth himself that this was intentional.the whole thing is designed like an OS and the manipulation at different level like weights, context, REPL, and disk form a memory hierarchy, and each level has its own update rule.in this 2h30 we get into how it works, the results on ARC-AGI-3 (95.5% with opus 5 on the public set, vs ~30% in the official harness), the week-long factorio run with 633 sub-agents, why letting an agent edit its whole system prompt makes it delete itself (hahaaaaa), why models cheat, and a live demo spawning a whole bunch of sub-agents on stream!

    👉 Seth Website: https://sethkarten.ai👉 Twitter: https://x.com/sethkarten👉 Prime Agent (paper): https://arxiv.org/abs/2608.23552👉 Prime Agent (blog): https://www.primeintellect.ai/blog/prime-agent👉 Prime Agent (GitHub): https://github.com/PrimeIntellect-ai/prime-agentCheck 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♥️ and check out the NAC harness https://www.arcee.ai/blog/nac

    here is the full run down:

    * 00:00:00 : prime agent = RLM + Continual Harness + Agent Communications

    * 00:03:24 : Seth background and past research!

    * 00:04:50 : how do you feel about harness field crystallizing into a distinct research discipline?

    * 00:06:52 : Seth benchmarking journey

    * 00:09:41 : The issue with games and tendency of reward hacking with frontier models?

    * 00:12:30 : frontier models and their taste for exploration + LLM play pokemons.

    * 00:17:30 : would harness capabilities be integrated within the models at the limit?

    * 00:27:00 : we’re reading a computer science paper actually!

    * 00:28:40 : prime agent: a self-improving RLM harness

    * 00:28:56 : a raw llm is just a sequential processor

    * 00:29:46 : harness adds persistent state, tools and compute

    * 00:34:15 : issue with zero sum data

    * 00:36:25 : system state spans weights, context, sessions and disk

    * 00:42:10 : the system is von neumann like

    * 00:45:03 : expressive harness lets model build it’s own worfklow

    * 00:47:03 : each RLM call creates a persistent subagent session

    * 00:54:57 : does the orchestrator always go idle?

    * 00:57:50 : what problems are we solving with each of the primitives for long-horizon control mechanisms?

    * 01:04:35 : are there plans to make public the Prime Intellect traces or all of this will be limited to first party model release?

    * 01:06:35 : is c-python kernel a better sole tool for models than a traditional JSON tool?

    * 01:07:55 : refine updates the harness from multiple trajectories

    * 01:12:18 : direct messages coordinate recursive sessions

    * 01:15:20 : what are the limit of the continual harness?

    * 01:20:00 : the figure 2 is suuuuuuuper important guys!!! + continual harness lore

    * 01:30:15 : three research questions to test the design

    * 01:32:13 : prime agent scales to 95.5% with more output token on ARC-AGI 3

    * 01:34:08 : how did the harness evolve during the ARC-AGI 3 run?

    * 01:39:05 : ablation on the three pillars of prime agent?

    * 01:40:40 : harness stuff is getting serious!!!!!

    * 01:43:45 : more evals for long context and more

    * 01:49:20 : out-of-loop experiments results

    * 01:53:40 : PMPP-Hard (GPU Kernels)

    * 01:54:35 : automated factorio runs!

    * 01:57:30 : is the monitoring in place for exploit like seen in factorio is enough to avoid them?

    * 02:05:10 : mazebench result (3D maze that is super duper hard)

    * 02:07:20 : why the heck are models giving up sooooo early man.

    * 02:13:15 : there is a whoooole lot to be done still.

    * 02:14:40 : at which training stage does introducing the harness have the greatest impact?

    * 02:16:35 : during introduction of harness in training would decomposition of abstractions make sense?

    * 02:20:20 : prime agent demo 🎉🎉🎉

    * 02:29:10 : how do you personally handle the context between multiple hierarchical agent?

    * 02:31:35 : how do you go about handling multi-agent trajectory for training models?

    * 02:35:27 : what’s next for seth + prime agent!

    stuff like this fully build in the open is what make machine learning / ai field such a pleasure to work in man big thanks to the team!



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    2 hr 37 min
  • How VPO Trains LLMs to Excel at Test-Time Search with Ryan Bahlous-Boldi from MIT

    usually post-training with GRPO will straight up collapse the model onto a tight set of potential answers (i.e. the spooky sharpening of the distribution).which in some cases might be exactly what we’re looking to do.however, in the grand 2026, we are increasingly using these models in long running harnesses that requires the LLM to think about different solutions and act on them compositionally to stumble on a discovery.having the model being so cooked it can’t produce diverse set of solution anymore isn’t very optimal in these situation.this is why ryan bahlous-boldi and co. from MIT took a deeper look at the machinery behind post-training to develop an objective that elicit these diverse set of solution.Vector Policy Optimization (VPO) is what came out of this intense cogitation and leverage a few structural elements of post-training:1. rewards are often vector valued so let’s keep that up?2. in-context learning is pretty good already so why not keep in-context multiple draft solution?3. let’s use up all the possibly weighting across rewards and make a vector that maximize all dimensions?taking the mean of this bad boy give us the post-training objective that VPO optimize for and voila, we have a model that can do a whole bunch of cool stuff like clear up hard problem that GRPO can’t even fathom.it was a very fun session and I learned a whole bunch from ryan highly recommend folks to keep an eye out for that scientist.

    👉 Ryan Website: https://ryanboldi.github.io👉 Twitter: https://x.com/RyanBoldi👉 Vector Policy Optimization (main paper): https://arxiv.org/abs/2605.22817👉 Vector Policy Optimization (main website): https://vpo-rl.github.ioCheck 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♥️ and check out the NAC harness https://www.arcee.ai/blog/nacin this 2h04 journey we have:- 00:00:00 - diversity in RL finetuning- 00:03:49 - quick explanation of VPO by your boy- 00:09:07 - what make a good research paper idea feat. VPO + AlphaEvolve- 00:16:59 - Heeyyyyyyy Ryan Bahlous-Boldi how you doing man- 00:22:29 - core intuition behind VPO- 00:31:50 - test-time search requires training for diversity- 00:35:08 - is hindsight replay used to create new reward signals?- 00:36:55 - will harness-level fine-tuning update heuristics directly?- 00:38:38 - post-training (currently) optimizes mainly for one answer- 00:40:01 - inference is no longer just one forward pass- 00:40:44 - mismatch between exploration and exploitation- 00:42:25 - how do you shape the reward for exploration?- 00:45:40 - a magic trick 🪄- 00:47:25 - multiple answer in your bag is needed- 00:48:06 - you gotta have a diverse set of solution- 00:49:30 - solution should be a vector and rewards are vector-valued.- 00:51:35 - walkthrough of one beautiful figure godamn- 01:02:37 - how do reward weighting, solution diversity, and sample count affect training when performance varies significantly?- 01:05:35 - the different kind of dimensions to manipulate (number of solution / different type of weighting)?- 01:12:39 - can mean based weighting overlook exceptional solutions? like is there better weighting method?- 01:16:14 - why do the solution have a shared reasoning prefix?- 01:19:23 - VPO covers the frontier of solution- 01:20:53 - VPO results + ablations- 01:29:15 - how does VPO compare with dense rewards methods that group each reward dimension separately?- 01:30:19 - VPO ablations continued- 01:32:32 - VPO on hard problems!!!- 01:34:35 - VPO on even harder problems!!!- 01:39:25 - diversity of VPO during training- 01:40:35 - why this thing works?- 01:44:30 - yacine ramble about solutions collapse with VPO and how defining space of solutions- 01:46:30 - limitations of VPO- 01:50:11 - the high level vision of VPO- 01:51:42 - does uniform group scaling make optimization equivalent to max-at-N with nonlinear rewards?- 01:52:49 - how are non-committal post-training exploration and continual learning connected?- 02:00:53 - what’s next on this research direction?



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    2 hr 5 min
  • What Continual Learning Actually Means feat. Ilija Lichkovski from Prime Intellect

    continual learning is one hell of a hot topic in 2026 and it’s beautiful because (like consciousness) barely anyone agree on what are we even talking about.this is why earlier this year I sat down with Ilija from the prime intellect bunch to discuss what he thinks about the subject while bathing in his radiating enthusiasm!in this 1h30min discussion we are going to talk a whole bunch on how he thinks about the components of continual learning while being exposed to a barrage of exquisitely curated literature.amazing guy and lovely mind to exchange with highly recommend.👉 Ilija Website: https://ilijalichkovski.github.io👉 Twitter: https://x.com/carnot_cyclist👉 Defining Continual Learning (main article): article

    Table Of Content:- 00:00:00 - what even is continual learning?- 00:01:45 - ilija background in physics, ai safety, machine learning- 00:08:05 - continual learning from first principles presentation- 00:12:08 - time component in machine learning vs organic learning- 00:17:30 - three flavours of incremental learning- 00:20:10 - five desirable aspect of continual learning- 00:23:10 - general capacity that should stay the fixed in continual learning?- 00:32:40 - replay method in ML- 00:36:10 - regularization family of methods- 00:38:00 - meta learning methods- 00:43:50 - hypernetworks method- 00:46:25 - experiment with hypernetwork- 00:48:45 - “what is a hypernetwork btw?”- 00:49:35 - hypernetwork on split-mnist- 01:02:18 - compositionally generalization experiments- 01:10:45 - continual learning benchmark should be method agnostic- 01:13:25 - on the generalization of LLM from in-context learning to finetuning- 01:17:07 - more experiments with self-study- 01:19:18 - OPSD vs GRPO experiments with Qwen3-8B- 01:21:55 - learning different topics that are connected (or not connected)- 01:28:20 - discovery shaped problem- 01:30:40 - what’s next for ilijaenjoy! 🌹big rafale of papers and blogs:- Overcoming classic challenges for artificial neural networks by providing incentives and practice | https://arxiv.org/abs/2410.10596 - Three types of incremental learning | https://www.nature.com/articles/s42256-022-00568-3- In vitro neurons learn and exhibit sentience when embodied in a simulated game-world (Cortical Labs, Pong) | https://doi.org/10.1016/j.neuron.2022.09.001- Human brain cells play Doom (Cortical Labs CL1) | https://github.com/SeanCole02/doom-neuron- Self-Distillation Enables Continual Learning | https://arxiv.org/abs/2601.19897- Overcoming catastrophic forgetting in neural networks | https://arxiv.org/abs/1612.00796- Metalearning Continual Learning Algorithms | https://arxiv.org/abs/2312.00276- Text-to-LoRA: Instant Transformer Adaption | https://arxiv.org/abs/2506.06105- Doc-to-LoRA: Learning to Instantly Internalize Contexts | https://arxiv.org/abs/2602.15902- Human-like systematic generalization through a meta-learning neural network | https://www.nature.com/articles/s41586-023-06668-3- Continual Learning Bench (website) | https://continual-learning-bench.com/- Continual Learning Bench (repo) | https://github.com/pgasawa/continual-learning-bench- Continual Learning Bench (paper) | https://arxiv.org/abs/2606.05661- Negation Neglect: When models fail to learn negations in training | https://arxiv.org/abs/2605.13829- On the generalization of language models from in-context learning and finetuning: a controlled study | https://arxiv.org/abs/2505.00661- Learning to Discover at Test Time | https://arxiv.org/abs/2601.16175- Learning, Fast and Slow: Towards LLMs That Adapt Continually | https://arxiv.org/abs/2605.12484- A Theory of Usable Information Under Computational Constraints | https://arxiv.org/abs/2002.10689

    pheeeeeeeeew!



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    1 hr 36 min
  • RLMs are Compositional Generalizers with Alex Zhang from MIT

    this discussion I had with alex earlier this year re-shaped a lot how I’m seeing these new breed of frontier models that are specialist at coding + multi-agent + long running.

    the concepts of compositional generalization and sub tasks being locally in-distribution makes it very very obvious why LLM + harness are doing so well right now.

    and also explain why tweaking the harness a bit sometime make a whole set of tasks that were previously unatainable for LLM a breeze.

    I strongly believe now that processing tasks into a “coding-friendly” format will be a continuous vector of breakthrough for the months to come.

    check out the full interview below and big thanks to everyone that submitted questions was an awesome session!

    table of content:

    * 0:00 - RLM are compositional generalizers

    * 4:40 - overview of what is an RLM

    * 10:00 - updated info and experiments on RLM

    * 22:40 - Big Boss Alex L. Zhang

    * 27:45 - what is “locally in-distribution”?

    * 34:00 - do we even need long context inside an LLM?

    * 36:50 - yacine is rambling again…

    * 39:35 - is locally in-distribution computation enough for scientific breakthrough

    * 46:40 - why are transformers poor at compositional generalization

    * 52:50 - what is missing in this soup of capability from the transformers side.

    * 57:25 - [at gun point] improve the harness or train a new model for compositional generalization???

    * 1:03:00 - how early RLM should enter the training pipeline?

    * 1:07:00 - how do you validate such a major architectural change btw?

    * 1:10:00 - what does Alex thinks about long-running persistent ai agents?

    * 1:14:00 - is code enough for compositional generalization across modalities?

    * 1:18:52 - how are we doing at the frontier for long horizon tasks?

    * 1:22:00 - speculative programmatic tool call (sPTC)

    * 1:28:00 - does the RLM impose an abstraction the model should learn for itself?

    * 1:32:00 - what’s next for Big Boss Alex L. Zhang (and what’s not next)?



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    1 hr 36 min
  • Inside a Frontier Open Model: Nemotron 3 Ultra Explained ft. NVIDIA's Chris Alexiuk

    open models are infinitely interesting to me. being able to peek under the hood and look at the interconnected pieces that each hums and work in a specific way to feed the whole is a blessing.

    I had a fantastic time earlier this year chatting with chris about the nemotron family of models and why they are engineered the way they are.

    throughout the discussion it became clear all decisions surround nemotron 3 ultra was about making it able to be fast fast and handle long context efficiently.

    huge thanks to the nemotron team for putting their stuff out there like that! ❤️

    Table of Content

    * 0:00 : nemotron 3 ultra ethos

    * 3:42 : Chris the Senior Product Research Engineer (phew)

    * 6:10 : Nemotron Lab & the Coalition

    * 9:10 : the open model ecosystem party 🤗

    * 12:50 : “a faster model is a smarter model”

    * 20:23 : intro to nemotron family

    * 38:18 : infrastructure issues on the research side for these type of models?

    * 42:30 : main design choices that sets nemotron 3 ultra apart?

    * 48:10 : long context inside the models or inside the harness?

    * 57:35 : Latent MoE tradeoff or is it a free 🥪

    * 1:01:14 : very aggressive grouped query attention why?

    * 1:08:00 : what in the structure of nvidia enabled this whole thing?

    * 1:11:36 : advice for undergrad that want to work on this stuff?

    * 1:15:45 : how does the shared weight MTP work?

    * 1:19:20 : main idea behind MOPD?

    * 1:24:55 : what’s next for nemotron scaling?

    * 1:26:52 : any weird reward hacking stories in nemotron?

    * 1:30:30 : what’s next for nemotron ++++?



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    1 hr 32 min
  • How to Generate Synthetic Pretraining Data? with Joël Niklaus from Hugginface

    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!

    lots of things to learn from experiment like this and to be honest there is still so much more to do!

    couple of useful links:

    👉Joel Website: https://niklaus.ai👉 Twitter: https://x.com/joelniklaus👉 Paper : arXiv: https://arxiv.org/abs/2604.13977👉 Blog: https://huggingface.co/spaces/HuggingFaceFW/finephraseAlso 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♥️ and check out the NAC harness https://www.arcee.ai/blog/nac

    on the menu:

    - 0:00 : synthetic data in pre-training LLM

    - 3:36 : Joel Niklau ML Research Engineer at Hugging Face

    - 4:50 : how to approach research problems at the edge?

    - 7:26 : the label here is still wrong is should be “benchmarking work overview”

    - 9:10 : what are the first few steps to get started building a private benchmark.

    - 11:18 : synthetic data pipeline and it’s impact in benchmark making

    - 13:19 : Synthetic Data Playbook Presentation Overview

    - 19:30 : what is rephrasing in synthetic data?

    - 21:05 : why are large scale ablation studies on data pipeline becoming feasible now?

    - 24:00 : how is rephrasing done?

    - 25:42 : experimental results

    - 27:58 : why does synthetic restructuring hurt common-sense knowledge?

    - 30:10 : what’s the value prop behind doing rephrasing?

    - 32:16 : is math a type of rephrasing?

    - 38:02 : what’s up with closeness between generator vs the student?

    - 39:40 : synthetic data + mixing real data

    - 41:20 : what drive this real-to-synthetic ratio?

    - 43:50 : what other structured formats could be explored for synthetic data?

    - 45:20 : is the order of datasets used for pretraining also affecting the distribution?

    - 46:40 : would adding multilinguality be considered a form of rephrasing for LLM?

    - 50:40 : which non-synthetic dataset should you mix in?

    - 1:03:20 : do model-specific writing habits survive the rephrasing process? + sawing off the edges?

    - 1:08:30 : would their be a gain to have a purpose built 270M model for being a generator?

    - 1:14:30 : does generator diversity still help as the student model scales?

    - 1:26:30 : ❤️❤️❤️

    - 1:27:15 : useful of synthetic data for post-training?

    - 1:28:50 : what’s next for this direction and for the field?

    btw: do follow joel he is a fantastic scientist and a very good educator!



    Get full access to learning with yacine at www.yacinemahdid.com/subscribe
    1 hr 31 min