Good morning! Here's what's happening in AI today:
NVIDIA launches Alpamayo 2 Super, an open reasoning model for autonomous vehicles
Runway brings FLUX 3, generating up to 20 seconds of video with audio
Liquid AI ships an agentic model that runs entirely on-device
How to set up the only productivity stack for creators
4 new AI tools worth trying today
AI MODEL
NVIDIA released Alpamayo 2 Super, an open reasoning model built for autonomous vehicles that reasons through complex driving situations rather than just seeing them, now commercially available for robotaxis, trucks, shuttles, and delivery vans.
The model adds 360-degree awareness, high-level driving decisions, and automated reasoning labels, built specifically for real-world driving rather than a general-purpose vision model adapted for cars.
It's released under an open commercial license, so teams can inspect it, fine-tune it, and deploy it directly instead of working through a closed API.
NVIDIA is positioning this as a backbone that extends beyond passenger vehicles, into the wider field of mobile robots that will eventually need the same kind of reasoning.
Most driving-assistance models are trained to react to what a camera sees, this one is built to reason through a situation before acting on it, which is a meaningfully different bar for safety-critical systems. Making it open and commercially usable means smaller robotics and AV companies get access to reasoning capability that would otherwise require building from scratch. Whether it holds up outside NVIDIA's own benchmarks, on messier real-world roads, is what actually determines if this becomes an industry backbone or stays a strong reference model.
AI RESEARCH
Cursor open-sourced Mixture-of-Kittens, an MoE training megakernel for NVL72 GPU clusters that fuses all Mixture-of-Experts communication and computation into a single, fully deterministic kernel, running up to 2.37x faster than the strongest public baselines.
The kernel folds several separate steps that normally run one after another into a single deterministic pass, which is what produces the throughput gain rather than a routine hardware upgrade.
Cursor's own benchmarks show it outperforming existing baselines across multiple model families, including Kimi, GLM, Qwen, and DeepSeek, not just a single cherry-picked comparison.
It's fully open-sourced, so any team training large Mixture-of-Experts models on this hardware class can adopt it directly instead of building equivalent infrastructure from scratch.
Training infrastructure rarely gets public attention, but the actual cost of training frontier-scale models is dominated by exactly this kind of communication overhead between experts. A 2.37x throughput gain, if it holds up outside Cursor's own benchmarks, translates directly into either faster training runs or meaningfully lower compute cost for anyone training MoE models at this scale. This one's squarely for teams doing large-scale model training, not a tool a general reader tries today, flagging that plainly since it's a deliberate exception to this issue's usual lineup.
AI MODEL
Liquid AI released LFM2.5-2.6B, an agentic model that runs entirely on-device, planning and calling tools through multi-step tasks on phones, laptops, PCs, and robots without any data leaving the device.
Because everything runs locally, the marginal cost of each run is essentially zero, no per-token API bill stacking up every time the model completes a task.
It scores comparably or better than models up to nearly four times its size on tool-use and instruction-following benchmarks, a real efficiency claim, not just a smaller model doing less.
The model ships under an open-weight license, so it can be customized on a single GPU for a specialized task rather than treated as one-size-fits-all.
An agent that plans and executes multi-step tasks without sending anything to a server changes the privacy math entirely, sensitive data never has to leave the device it started on. The efficiency claim matters just as much for cost, running something locally at zero marginal cost is a different budget conversation than metered API calls for every action. This is crucial to watch closely if you're building anything that needs to run on hardware you don't control the network for.
HOW TO AI
You save a lot. Articles you meant to read, transcripts, research for a project that stalled, half-finished notes. A week later you couldn't tell anyone what was actually in any of it.
Claude Code can read it all, sure. But dump 145 documents on it and it searches one at a time, slow, and blind to how any of it connects.
The fix: three tools, chained together. Graphify turns a folder into a connected map of ideas. Obsidian holds that map somewhere you can actually browse. Claude Code reads the map instead of the pile.
Run the whole pipeline in one sentence
Download the official Claude Code docs, map them with Graphify,
then turn that into an Obsidian vault.Claude Code fetches the source material on its own, then Graphify does the real work: reading every page, extracting the actual ideas inside them, not just filing pages as files. One run on the Claude Code docs turned 145 source documents into 591 separate ideas, wired together by 685 connections, sorted into 67 topic groups.
Open it and the map stops being abstract
Every dot is one idea, every line a real connection, dot size showing how central that idea is to everything else. Ask a question afterward and Claude Code doesn't just search for a keyword, it already knows how ideas relate, because it has the map instead of a pile of files.
Query it directly with graphify query "how does X relate to Y?", or drop the whole thing into your main vault as one folder. 145 files in, one connected map out, three sentences total. Read the full guide.
P.S. You can access all my prompts, workflows and AI trainings if you upgrade here
Ant Ling released the open weights for Ling-3.0-flash, with official BF16 and FP8-quantized versions available so teams can pick whichever fits their hardware and deployment needs.
Lumo launched version 2.0, letting people upload a file, ask a question, and get a chart back without leaving the conversation or sending their data to a company that trains on it.
Kiro open-sourced Kiro Crew, a persistent development workspace that runs on a person's own hardware and works wherever they are, grown out of an internal tool shaped by hundreds of contributors.
🖼️ Osaurus: Image and video generation are now live, powered by Venice, drawing on 215 models starting at $0.03 per image, with real files saved straight to your Mac.
🔬 Goodfire: Silico is now publicly available, letting anyone interpret and train their own models at frontier scale.
🐛 Lemma: Artifacts uses an agent's own prompts, traces, and private context to decide what actually deserves to surface as an issue, then learns from every one you validate or dismiss.
🎬 Motion: Point Claude at a video you like and it matches the style, kinetic type, sound design, and all.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial