Alex Welcing

Web in NYC with TX roots. No expert in DevOps, AI, Kubernetes, or human rights, but know black and trans lives matter and the cloud is held up by people.

[Alex Welcing @AlexWelcing]

Video as context is instantly a genre definer.

[Kimi Developers @KimiDevs]

Kimi Code, our open-source coding agent, just got a major upgrade!

🔹One-line CLI install, zero setup, fast startup​
🔹Drag in videos as coding context: reference-to-LUT, long-video-to-short, screen-recording-to-code, and more​
🔹Plugins for stocks, financial reports, academic papers, with more coming​
🔹Supports the ACP protocol, and works with JetBrains, Zed, and more​
🔹Hooks for custom tools and workflows​

Try it with Kimi K2.6 👉 https://t.co/1owb8LUBMQ
Issues, plugin ideas, and PRs welcome! Community feedback helps shape what ships next.​🚀

[Alex Welcing @AlexWelcing]

Friends are coming!!

[Jorge Bravo Abad @bravo_abad]

Boltzmann generators go big: scalable equilibrium sampling for materials

Predicting whether a crystal phase is stable, or computing its free energy, usually means running molecular dynamics for a long time. The problem is that MD explores configuration space one step at a time, so consecutive snapshots are correlated and the system mixes slowly. Boltzmann Generators offered a clever alternative: train a normalizing flow to map random noise straight onto the Boltzmann distribution, giving uncorrelated equilibrium samples in one shot, plus free energies through reweighting. The catch was scale. Earlier BGs maxed out around 500 atoms, with poor sampling efficiency and training runs approaching a full GPU-year, slower than the MD they were meant to replace.

Schebek, Noé and Rogal fix this with a clean idea: stop modeling the whole configuration at once and learn from local atomic environments instead. They pair augmented coupling flows with a graph neural network that builds per-particle embeddings from each atom's neighborhood. By modeling displacements from the ideal lattice rather than absolute positions, the input stays size-independent, so a model trained on a few hundred atoms transfers directly to systems above 1000 atoms. Scaling becomes linear rather than quadratic, and training needs only the energy function, no MD samples.

The payoff is large. For mW ice, a global BG needed over 330 GPU days and reached effective sample sizes around 0.2%, while the local model converges in about 4 GPU days and pushes efficiency up by an order of magnitude. It reproduces radial distribution functions and free energies that match MD to within 10^-3 kBT per particle, resolving tiny phase-stability differences in Lennard-Jones crystals, water ice, and the silicon phase diagram. Conditioning on temperature, pressure, and interaction parameters lets one model cover a whole family of materials.

The real lever is cost. A trained flow needs only tens of thousands of energy evaluations for a large cell, against tens of millions for standard MD free energy methods, and that gap widens sharply with expensive machine-learned interatomic potentials. In materials development, energy storage, and catalysis, this means screening phase stability and thermodynamics at realistic system sizes without paying the usual simulation tax.

Paper: Schebek et al., Nature Communications (2026) — CC BY 4.0 | https://t.co/61llOHl0Ek

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

Incredible list
so much focus on specific combinations of atoms
It'll be interesting how materials science develops in this new AI era.

[vLLM @vllm_project]

🎉 The vLLM community just got a free course, built by @RedHat_AI with @DeepLearningAI. It walks through the full optimize �� deploy → benchmark lifecycle for serving open models.

Three labs, each on a live vLLM server:
- Compress: quantize a Qwen model with LLM Compressor, then measure the size vs. accuracy tradeoff
- Serve: deploy with vLLM's OpenAI-compatible API and watch continuous batching, PagedAttention, and prefix caching in the live metrics
- Benchmark: simulate traffic with GuideLLM and check quality with lm-eval

A lot of the work went into visualizing what actually happens under inference, thanks to @cedricclyburn: how tokens flow through the model, how the KV cache grows in GPU memory, and what changes when you move from FP16 to INT8/INT4.

~1.5 hours, 9 lessons, 3 labs. Free on https://t.co/pGAwqbmMdc.

📝 Read more: https://t.co/r8ITc2prI2

[Dimitris Papailiopoulos @DimitrisPapail]

https://t.co/n10GwfKYuY

[LeRobot @LeRobotHF]

VLA-JEPA just dropped in LeRobot 🤖

What makes this model special is that it does not just learn what action to take from a given observation, it also leverages a JEPA world model to learn action-relevant dynamics.

During training, the VLA leverages V-JEPA2 by conditioning its predictor. This clever trick adds a world modeling objective to the training, which also allows pretraining on human videos.
At inference, the world model is dropped entirely, keeping only a standard VLA architecture: Qwen backbone and action head.

The demo here was only fine-tuned on 13 examples, showing great pretraining capability and running in real time on @NVIDIARobotics DGX Spark!

VLA-JEPA is the first world model to be ported to LeRobot, and I feel like it won't be the last 🚀

@Thom_Wolf @ClementDelangue

[GREG ISENBERG @gregisenberg]

I was once pitching in a board room at a top 3 VC firm for a $15M Series A.

12 people in the meeting. One of the GPs fully fell asleep. Out cold for 30+ minutes. Nobody acknowledged it. Everyone just kept going.

I kept presenting my Series A slides to an unconscious man in a Herman Miller chair and somehow that was considered normal. That's venture capital.

You might fly across the country to perform for people who may or may not be conscious.

It's a dance.

And sometimes you lead and sometimes you follow and sometimes your partner is unconscious.

If you're raising right now, just know: every founder has a story like this. The process is weird. The power dynamic is weird. You're not crazy for thinking it's weird.

No one talks about it because they want to continue raising. But I'm happy to stick my neck out there.

It is weird.

[Alex Welcing @AlexWelcing]

One step at a time!

[Markus J. Buehler @ProfBuehlerMIT]

Paper: https://t.co/B2bwWnZ2tI

[Alex Welcing @AlexWelcing]

This is not the B2B SaaS era! Redefine 🦄

[Ashton Forbes @AshtonForbes]

Are you paying attention?

[Mustafa Suleyman @mustafasuleyman]

Super excited to announce seven new world-class MAI models today. They represent what we consider a new era in AI designed to keep you in control and on the frontier.
First is our text foundation model, MAI-Thinking-1, exceptionally strong on reasoning and SWE tasks.
- It’s a 35B active parameter MoE with a 256K context window. Independent human raters on Surge prefer it for overall quality in blind side-by-sides versus Sonnet 4.6, and it’s achieved 97% on AIME 2025, the key measure of its general-purpose reasoning abilities.
- It's at 53% on SWE Bench Pro, placing it right alongside Opus 4.6 on one of the toughest coding benchmarks.
- And since we co-designed our models with our own silicon, MAI-Thinking-1 is optimized on our MAIA 200 chip. Benchmarking head-to-head against the GB200, we see 30% better performance per dollar as well as a 1.4x performance-per-watt gain when running our MAI models on the MAIA 200 end-to-end.

Next is MAI-Image-2.5 and its Flash variant. Two super strong models now at #2 on the leaderboards, surpassing the score of Nano Banana 2 on image editing.

Last for now is MAI-Code-1-Flash, our new inference efficient coding model, especially tuned for VS Code and GitHub Copilot CLI.
- Code-1-Flash achieves 51% on SWE Bench Pro, despite having just 5B parameters, putting it closer to Haiku in size but cheaper in cost.

All of this is the foundation for Microsoft Frontier Tuning. It lets you customize our models to create custom, company-specific agents that only you control. You can make our model, your model. Your data. Your agents. Your moat.

Early adopters are already seeing a difference. When we tuned our models for McKinsey’s tasks, MAI delivered the highest win rate, outperforming GPT-5.5 on quality, while being 10x lower on cost.

Also really excited to be collaborating with the amazing team at Mayo Clinic to jointly train a new frontier AI model for healthcare.

Our announcements today mark another milestone on the road to humanist superintelligence. You can learn more and about our other new models in our latest blog: https://t.co/v65eop5Ixq

[Alex Welcing @AlexWelcing]

@emaros96 Oh this is fun, cool work!

[Parallel Web Systems @p0]

Today we're launching Index: a platform for content owners to understand how AI agents use their work, and earn revenue when they do.

Our first partners include @TheAtlantic, @FortuneMagazine, @PRNewswire, @PitchBook, @ZoomInfo, @Tracxn, @RocketReachCo, @enigma_data, @fiscal_ai, plus creators @alexeheath, @mariogabriele @azeem, @every, and @packyM.

[Sapient Intelligence @Sapient_Int]

Introducing HRM-Text.

An ultra-lean 1B-parameter reasoning language model designed to deliver strong general performance with a fraction of the data, compute, and infrastructure.

Trained on just 40B structured tokens, HRM-Text achieves competitive performance while using ~1/1000 of the training data of comparable models.

The kicker? The full model trains in roughly one day on a $1,000 budget.

This opens the door to a new generation of AI that is powerful, accessible, and radically easier to adapt. Theories and research concepts once deemed too expensive to test are officially back in the game.

Sapient Intelligence invites you to help us shape a new paradigm for general intelligence.

[Entire @EntireHQ]

We're excited to publish new research on agentic search.

We found that nearly 50% of an agent's tool calls read and search for code and files. So then why does it take coding agents so long to find the right file?

We thought the bottleneck was slow code search. But after investigating 200,000 real world tool calls, we found the real problem.

Principal Engineer @evisdrenova explains the research and what really matters for agentic search:

https://t.co/OCRMYPwzd2

[Alex Welcing @AlexWelcing]

@GuoMh14 Great great work, well deserved recognition from the journal.

[Natalie Wolchover @nattyover]

Bacteria move around using a molecular machine called the flagellar motor that rotates faster than the flywheel of a race car engine and switches directions in an instant. After 50 yrs, scientists have finally figured out how it works. “My lifelong quest is now fulfilled.” Link⤵️

[Alex Welcing @AlexWelcing]

Should have dozens of enterprise clients, great opp for B2B with consulting. Protect your worth, it’s best for the product too. Good call.

[MrNeRF @janusch_patas]

Heads up: there are some changes coming to LichtFeld Studio.

LichtFeld is growing quite a lot, but contributions are not. In fact, besides the two corporate contributors and one bronze sponsor, there is less than $100 on average contributed per month. There were 10,000 downloads in the last 30 days, and nobody contributed anything financially back to the project. Not a single dollar. I cannot support my living on that basis.

Given the amount of work I am putting into this project, and it is a lot, often easily 60 hours a week, to provide a state-of-the-art, self-contained 3D reconstruction suite that is used by businesses around the world, this situation is no longer sustainable. The idea was to grow it together, but at the moment it is mainly being used without enough support in return.

If every user had contributed only $10, the continuation of the project would have been secured. I have asked for that quite often. I like sharing my knowledge, but I cannot sustain it under these circumstances.

Going forward, there will not be any free binaries provided anymore. I created a portal at https://t.co/mdGITFOGVQ which will give access to daily binaries after registration and donation. This will hopefully help finance the development and maybe even allow me to pay someone in addition.

If you cannot afford it, you can still put in the work and build it from source. There are no restrictions.

I am aware that people like to work around something like this, and they certainly will. But in the end, that simply means that my work on this project will be discontinued at some point.

Btw, after payment an invoice will be automatically issued!

[Alex Welcing @AlexWelcing]

@prathoshap Wow thank you for sharing!

[KapishDima @kapish_dima]

Introducing Remocn - a @shadcn registry with ready-made animations, typography, transitions, UI elements, and complete scenes for @Remotion (a library for creating videos using React components).

If you’re a solo builder like me and don’t have the budget to produce a full demo for your product - this is the solution. I’ve created a collection of elements that will help you build demos much faster.

Plus, everything lives right next to your project since it’s built with React.js. All components are added via shadcn, which means you can easily customize them to match your style

All of this is FREE and open source 😱

[Tin Hadzi Veljkovic @HvTin]

Can a lightweight Transformer compete in crystal generation without equivariance?

We show that it can.

Crystalite combines chemistry-aware priors with geometric attention biases for:

- SOTA CSP
- best S.U.N. among all baselines
- much faster sampling

More info below! 👇