Driss Guessous

Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. https://lnkd.in/gY4NvS5h Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. Cost and latency are important in real-world use cases. Inkling's continuous thinking effort lets you pick your point on the cost/performance curve — reaching the same score with a fraction of the tokens. Inkling natively understands and reasons across text, audio, and images. It’s strong on audio in particular, ranking among the strongest open-weights models on VoiceBench, MMAU, and AudioMC. We’re grateful to our partners for their day-0 support across the open-source ecosystem: NVIDIA, Together AI, Fireworks AI, Databricks, Unsloth AI, Modal, Baseten, LightSeek Foundation, Inferact on VLLM, RadixArk on SGLang. Inkling is the first in a family. We’ve included some details of Inkling-Small, a lighter-weight model trained on a similar recipe, with full weights to follow. We hope you enjoy Inkling, and as always we’re keen to see what you build.

Excited to share that after 2 years on the PyTorch team @ Meta, I’ve started a new role at Thinking Machines Lab. What attracted me most: the mission to focus on the human side and customization of AI. Most AI today gets trained once and frozen, not adapting to the people actually using it. TML is betting on the opposite: AI that’s built with people, not instead of them, hypercustomized. Amazing what this small team has been up to so far. Check out thinkingmachines.ai — more to come very soon! Huge thanks to the PyTorch team for two great years. Onward 🚀

June 10 edit: ATTENTION: Laid off employees Here is how to message a recruiter and be MUCH more likely to get a response: DO: 1. Find a specific role(s) on the careers site first. 2. Message the contact in the sheet either (1) after applying or (2) with a link to this role. DO NOT: Cold email asking them to find a fit. Amanda @ Waymo has a good post on this: https://shorturl.at/mgCPT ---- June 8 edit: Hey all - I'm continuing to work through the messages on here, and you should hear back from me in the next 48 hours! Appreciate the patience :) ---- May 26 edit: The list of hiring companies is out now to everyone who was on the ex-Meta list that I was able to verify. If you had a .edu e-mail address, often the university blocked me from sharing a GSheet with you. So if you haven't received the hiring company list, please sign-up again on the ex-Meta list with a ***non .edu email***. ---- May 21 edit: Currently ~350 companies and 2.5k employees have signed up. Access will be shared by EOD Tuesday, once I have filtered out spam sign-ups. ---- May 20 is (1) my birthday 🥳 (2) Meta laying off 10% of its employees 🤦‍♂️. Every time Meta does a big lay off, I've created two registries to help ex-Meta people find new jobs. Last big lay-off, we had 5,000+ former employees and 600+ companies sign-up. We're doing this again for the 5/20 lay-off. So - if you are a company who wants to hire ex-Meta people, or you are a person who was laid-off, please use the forms below: 1. Companies - if you are interested in hiring former Meta employees, please register here: https://lnkd.in/gncnh2SX 2. Ex-Meta employees - if you want to receive the list of hiring companies, please register here: https://lnkd.in/gAmEc3Zz ---- FAQ: Do I make money off this, sell this, etc? No. What are the details of how this works? 1. Companies sign-up. 2. Ex-Meta people sign-up. 3. Ex-Meta people get shared access to the list of hiring companies and can ping the hiring managers. Is it anonymous? - For companies that sign-up, NO. Your info is shared with the laid off people. - For laid-off ex-Meta people that sign-up, YES. Nobody sees your info ever. I use it to filter out spam sign-ups. I'm ex-Meta, from an earlier lay-off, can I sign-up? Yes! If you are not part of Meta, can you still sign-up? Yes! Please use this form: https://lnkd.in/gTfqe8w6 However for this list, the contact info is NOT private - as it will be shared with the companies looking to hire people. As a laid-off employee when will I receive access to the company hiring list? By EOD Tuesday. Then I update the permissions every ~48 hours. If I'm laid off too, likely quicker since I'll get a minute to automate this better. -- Finally - how can you help? Share the post. Visibility of the lists is always the most helpful thing.

Week 3 nearly in the books at Reflection. A few quick thoughts — and some roles we are hiring for. On the company: The mission — It's big, bold, and resonates with everyone. There's a clarity to it that has people genuinely aligned and moving in the same direction. That feels great. The vibes — Incredibly collaborative and open. There's a ton of work to do, and people are eager to pitch in across teams. Infra, product, research, policy, legal — it all just works seamlessly, with effectively no boundaries. The talent — Has eclipsed my wildest expectations. We're hiring some of the best people in the space from OpenAI, Anthropic, Google, and Meta, and that's attracting more top talent. It feels like things are snowballing in a way that's building a really strong team and culture. I can't talk yet about what we're working on, but I'll be doing talks and interviews soon — stay tuned. What we're hiring for: - Safety — a safety lead and a safety PM. Core to what we do, and to doing it openly. - Policy — so much happening at the global stage; hiring the best in regulatory, safety and preparedness, and more. - Legal — great work to be done on open source, commercial deals, and beyond. - Comms and community — events, hacks, technical blogs, publications. Looking for someone to help drive it. - SWEs/MLEs — we're building a ton in open source, and hiring engineers who are builders first: people who engage communities around their projects and evangelize in the ecosystem. Truly unique work, and unapologetically open source. - GPU infrastructure & compute — We're developing foundational models at very large scale, and hiring engineers to build GPU and compute infra across multiple clouds. - Post-training — Actively recruiting candidates to lead post-training. Given our roots, we're heavily invested in RL — so if large-scale RL, environments, and shaping foundational model behavior excites you, this could be a great fit. If any of this resonates, please submit at: https://lnkd.in/gjkZgE8f Cheers! Bonus: https://lnkd.in/gXgEyQp7

Senior engineer is where most SWE careers settle, and that's fine. But if you're going to push for staff, you should know: at Pinterest, the jump from Senior to Staff is worth $328K. We looked at the Levels.fyi data to find how well different companies actually reward the promotion. Filtered for new offers only across the top four US tech metros (Bay Area, Seattle, LA, and NYC) to get a clean comparison. Senior is considered the terminal level at most tech companies, meaning you can stay there indefinitely with satisfactory performance. But it's not just that engineers choose to stay. Staff roles are also more scarce than Senior roles with some companies (depending on size) not even having a Staff level. Companies simply don't need as many staff engineers because the role demands a fundamentally different kind of work: owning technical direction across teams, driving org-wide decisions, and operating with a scope that goes well beyond shipping features. Across Levels.fyi, senior engineer submissions outnumber staff submissions roughly 3 to 1. That jump in responsibility is why the comp premium can be so significant, and why it varies so much by company. Pinterest came in at +82%, with staff engineers making $727K compared to $399K at senior. LinkedIn showed the largest dollar jump at +$345K. On the other end, Nvidia's staff premium was just 22%, one of the smallest gaps we found despite being one of the most in-demand companies in tech. The takeaway isn't about which company is "better." It's that the value of this promotion is entirely company-dependent. If the Senior to Staff jump is part of your career plan, where you do it matters as much as whether you do it. Explore the full data on Levels.fyi: https://lnkd.in/g_GJcFG8 Plus, if you’re in HR and want to create neat reports like this, check out our Data Explorer tool: https://lnkd.in/gvgc56FE

The mixed emotions of our last day at Meta.. w/Mark Saroufim (camera work by Victoria Dudin) Cheers!

Here's a side quest I've been working on for a little while: mujoco-torch (https://lnkd.in/eN82F9hP) I started this during my time at Meta in my spare time as a way to demonstrate torch.compile usage for physics simulation in #PyTorch. Since then, I worked during my weekends to lead it to a more mature state. It's now available on PyPI. `pip install mujoco-torch` Running physics in PyTorch (or JAX) instead of a traditional C simulator isn't just about speed -- it means your simulator lives in the same framework as your learning pipeline. Gradients flow through the physics, batching is a single torch.vmap call, and torch.compile fuses the entire simulation step into optimized GPU kernels with zero Python overhead. No serialization boundaries, no cross-framework data copies -- just tensors all the way down. ⚙️ What it does: mujoco-torch is a full PyTorch port of MuJoCo MJX -- forward dynamics, constraint solving (CG and Newton), frictional contacts, 30+ sensor types, and 12 collision functions covering all major geom types. It's numerically equivalent to MJX, verified at float64 precision. You can load (hopefully) any MuJoCo XML, move it to GPU, and simulate with a few lines of code. The end goal is E2E compilation of your simulation and training loop! (I've also added a small rendering util that alleviates all the headless EGL rendering nightmares you might have had with with MuJoCo!) 🔦 Demo: The repo also ships with an RL environment zoo built on TorchRL: HalfCheetah, Ant, Hopper, Walker2d, Swimmer, and Cartpole, all trainable with SAC or PPO out of the box. These environments served as end-to-end validation that the physics actually produces learnable behaviors. 🚀 Perf: with torch.compile + torch.vmap at large batch sizes, we see millions of simulation steps per second on a single GPU -- up to 3.6M steps/s on Half-Cheetah and 2.5M on Humanoid at batch size 32k. That's roughly 40x the throughput of MuJoCo C running sequentially on CPU. The library is designed for the regime where you're running thousands of parallel environments, which is exactly where modern RL training operates. Hope you'll enjoy it! #robotics #AI #ReinforcementLearning

An excellent introduction to microscaling formats (mxfp8, mxfp4, nvfp4) used in machine learning workloads, how they affect performance (both in in terms of inference throughput and Mean Absolute Error, which describes the loss in accuracy due to quantization and scaling.) https://lnkd.in/gUgZPEeQ by Vasiliy Kuznetsov and Driss Guessous via PyTorch

CUDA developers heading to GTC next week, here is your guide to many of the CUDA-focused sessions! See you there. https://lnkd.in/gPqsWpJM