most "convert everything into a skill" advice stops at storage. nobody's covering what happens once the pile gets big. the notes section of my guide already runs into a small version of this problem, years of scattered vault entries that just sit there until you go looking for something specific. AgeMem tackles the harder version of that same problem, teaching the agent to actively manage what it keeps, updates, or throws out over time, not just search through a pile when asked. concretely: it gives the agent tool-based control over storing, retrieving, updating, summarizing, or discarding memory, trained across 3 stages of reinforcement learning because memory decisions produce sparse, awkward reward signals normal training can't handle. why it matters: across 5 long-horizon benchmarks, it beat setups that treat long-term and short-term memory as separate bolted-on pieces, on both context efficiency and memory quality. it's an ACL'26 highlight paper too. paper: full article below 👇

a small team spent 7 years building digital humans for real companies before open sourcing any of it. Duix.Avatar started as a way to clone someone's face and voice from real video, cutting the cost of a usable digital human from six figures down to about $1,000. 10,000+ businesses used the paid version before any of the core tech became public. the open source release runs fully offline, no cloud calls, clone your appearance and voice locally, then drive a talking avatar with just text or audio, in 8 languages.

a small team spent 7 years building digital humans for real companies before open sourcing any of it. Duix.Avatar started as a way to clone someone's face and voice from real video, cutting the cost of a usable digital human from six figures down to about $1,000. 10,000+ businesses used the paid version before any of the core tech became public. the open source release runs fully offline, no cloud calls, clone your appearance and voice locally, then drive a talking avatar with just text or audio, in 8 languages.

setting up a research-grade weather model used to mean hours of install docs. now you just ask an agent. Earth2Studio is NVIDIA's framework wrapping pretrained weather models, GraphCast, Aurora, FourCastNet3, AIFS, behind one shared interface. the loop: ask Claude Code to run the install skill > it configures the environment and recommends a model > point it at a data source, GFS, ERA5, HRRR > run the forecast, swap the model next time, same pipeline. each wrapped model and dataset carries its own license from its own provider, worth checking before you redistribute anything you generate. the skills, if you want to try it: npx skills add NVIDIA/skills --skill earth2studio-install

setting up a research-grade weather model used to mean hours of install docs. now you just ask an agent. Earth2Studio is NVIDIA's framework wrapping pretrained weather models, GraphCast, Aurora, FourCastNet3, AIFS, behind one shared interface. the loop: ask Claude Code to run the install skill > it configures the environment and recommends a model > point it at a data source, GFS, ERA5, HRRR > run the forecast, swap the model next time, same pipeline. each wrapped model and dataset carries its own license from its own provider, worth checking before you redistribute anything you generate. the skills, if you want to try it: npx skills add NVIDIA/skills --skill earth2studio-install

anthropic's moat just got a lot smaller. researchers across NUS, Stanford, Oxford, and Princeton built on anthropic's own agent-skill format, then showed the format alone isn't what does the work. the number that stands out: an agent with every skill dumped straight into its context scored 65.6% on a long-horizon retail task. the same agent, the same skills, but with a working memory deciding which one to invoke and when, hit 83.6%. same skill library, 18 points apart. different setup than what i wrote about, but the same underlying idea: having the skill isn't the unlock, knowing when to reach for it is. wrote the full guide on doing this with everything you already know: books, docs sites, videos, the open web, your own notes. full article below 👇

Claude and ChatGPT can finally make clean decks for you. You don't touch a single slide. It builds the whole thing. Here's the exact workflow: Step 1: Open ChatGPT or Claude
Step 2: Go to Settings → Connectors → search "Gamma"
Step 3: Click Connect (no API key needed)
Step 4: Prompt it like normal: "Build me a campaign launch deck I can share with my team"
Step 5: Switch on the Gamma connector and hit send
Step 6: Watch the full deck generate inline, no new tab
Step 7: Click the Gamma link to edit further You never leave the chat. Ask for changes and the deck updates in the same thread. I pointed it at my own newsletter and asked Claude for a Q1 2027 growth and content strategy for Simplifying AI. Then, "turn it into a Gamma presentation." It thought through the whole strategy, then handed me a clean deck. One chat and I didn’t have to do a single slide.

anthropic's moat just got a lot smaller. researchers across NUS, Stanford, Oxford, and Princeton built on anthropic's own agent-skill format, then showed the format alone isn't what does the work. the number that stands out: an agent with every skill dumped straight into its context scored 65.6% on a long-horizon retail task. the same agent, the same skills, but with a working memory deciding which one to invoke and when, hit 83.6%. same skill library, 18 points apart. different setup than what i wrote about, but the same underlying idea: having the skill isn't the unlock, knowing when to reach for it is. wrote the full guide on doing this with everything you already know: books, docs sites, videos, the open web, your own notes. full article below 👇

anthropic's moat just got a lot smaller... researchers across NUS, Stanford, Oxford, and Princeton built on anthropic's own agent-skill format, then showed the format alone isn't what does the work. the number that stands out: an agent with every skill dumped straight into its context scored 65.6% on a long-horizon retail task. the same agent, the same skills, but with a working memory deciding which one to invoke and when, hit 83.6%. same skill library, 18 points apart. different setup than what i wrote about, but the same underlying idea: having the skill isn't the unlock, knowing when to reach for it is. wrote the full guide on doing this with everything you already know: books, docs sites, videos, the open web, your own notes. full article below 👇

what happens when you build a voice model that was never actually built around speech? most TTS systems map text to phonemes, then phonemes to sound. Bark skips that step entirely, text goes straight to audio tokens, GPT-style, the same way a language model predicts the next word. output runs about 13 seconds at a time by default. install with: pip install git+

what happens when you build a voice model that was never actually built around speech? most TTS systems map text to phonemes, then phonemes to sound. Bark skips that step entirely, text goes straight to audio tokens, GPT-style, the same way a language model predicts the next word. output runs about 13 seconds at a time by default. install with: pip install git+

nvidia's moat just got a citation problem exposed. researchers built a way to trace a bad citation in a deep research report back to the exact agent that broke it, not just flag that something's wrong. applied to nvidia's #1-ranked# open source system, 84.7% of the final report's errors trace back to the orchestrator, even though it's nowhere near the worst individual agent on its own. worth noting: two one-line prompt fixes raised citation recall by 5 points, but recall still sits around 64-70% after that, so this isn't solved.

nvidia's moat just got a citation problem exposed. researchers built a way to trace a bad citation in a deep research report back to the exact agent that broke it, not just flag that something's wrong. applied to nvidia's #1-ranked# open source system, 84.7% of the final report's errors trace back to the orchestrator, even though it's nowhere near the worst individual agent on its own. worth noting: two one-line prompt fixes raised citation recall by 5 points, but recall still sits around 64-70% after that, so this isn't solved.

this repo beat Runway Gen-3 and Luma 1.6 in blind human evaluation. it also needs 60GB of VRAM for one video. - text alignment: 61.8%, best of the 6 - motion quality: 66.5%, best of the 6 - overall preference score: 41.3%, ranked #1# a 13B parameter model, largest open-source video model at release, using a decoder-only LLM as the text encoder instead of CLIP+T5. this is the original HunyuanVideo though. Tencent's own HunyuanVideo-1.5 replaced it since, same lineage, far lighter to run.

this repo beat Runway Gen-3 and Luma 1.6 in blind human evaluation. it also needs 60GB of VRAM for one video. - text alignment: 61.8%, best of the 6 - motion quality: 66.5%, best of the 6 - overall preference score: 41.3%, ranked #1# a 13B parameter model, largest open-source video model at release, using a decoder-only LLM as the text encoder instead of CLIP+T5. this is the original HunyuanVideo though. Tencent's own HunyuanVideo-1.5 replaced it since, same lineage, far lighter to run.

most physical ai research can't be reproduced because everyone's test rig looks different. OpenArm stack, save this: - hardware: full CAD, STL, STEP, Fusion 360 assemblies, CERN-OHL-S-2.0 licensed - control: CAN library for motor comms plus ROS2 integration - teleop: bilateral control with real force feedback through the leader arm - simulation: Isaac Lab and MuJoCo environments to train before touching hardware - data: a dataset format, recording API, and Dora dataflow nodes for collection and inference what makes it different: OpenArm Cell standardizes the background, lighting, and camera placement, so a lab in Tokyo and a lab in Berlin get comparable results.