I just got my first ad for an AI Skill as a service for Claw www.instagram.com/reel/DVHXndM...

Increasingly training AI is like raising a child The pretraining is pre-teen years, instruction tuning teenage years, RL/TTC like college years

A study shows pretraining LLMs with positive AI discourse can reduce misalignment rates from 45% to 9%, highlighting the data's influence on AI behavior. This alignment pretraining approach offers a method for enhancing AI safety without sacrificing performance. https://arxiv.org/abs/2601.10160

I want to build its bones and blood not its character

Outer Wilds, if you haven't played it yet, is one of the best games of all time. It's a fun space exploration game with a timeloop twist, and 90% of all game progress happens in your head, as you learn things about the universe and your situation. Absolutely play it if you haven't yet.

the difference between Anthropic vs other labs is Anthropic is planning for how their products will **fit into** society, whereas everyone else is just making products

Model that goes through your 50+ open tabs, infers why you have been unable to close each one, and converts each into a to-do item or a reminder connected to a long-term goal. In theory you can manually review all the goal entries and reminders; actually it will do that and remind as needed.

the reason for the absurd levels of secrecy from the big labs is because they have no moat

If you haven't yet changed how you teach and evaluate thanks to AI, you probably need to. This is not something that educators asked to happen, but it is worth noting what is out there - in this case an OpenClaw tool designed to cheat.

Does anyone do hyper parameter tuning as a model is actively training (learning rate or loss function weighting or drop out params etc) and is there an AI model that monitors training runs and tweaks it as it goes?

Just re-predicting that we’ll see an optical compute breakthrough this year

I love Alysa Liu's story. It's such a powerful example of intrinsic motivation and real freedom to choose. She quit at 16 when she felt her choices were being made for her. She came back to it entirely on her own. You can see her joy. It reminds me so much of what I wrote about Monsters, Inc and UBI

When imagining backlash against AI due to job risks, I think people think of Dune’s Butlerian Jihad or Luddites. But what those fights actually looked like during the previous Industrial Revolutions were about regulation, redistribution, nationalization, unions & safety nets. Could expect similar

Many benchmarks use LLMs as a judge of correctness, typically a smaller, cheaper model. This paper shows weaker judges are not able to evaluate smarter models. A benchmark is really a triplet of dataset, model, judge & judges are increasingly the bottleneck being saturated. arxiv.org/pdf/2601.19532

Power On: Tim Cook, like he did before the Apple Watch and Vision Pro, is dropping hints about Apple’s new next big thing: AI wearables centered around Visual Intelligence. www.bloomberg.com/news/newslet...

I’m seeing she’s diagnosed with ADHD and now everything about her makes sense to me

Claude Code is the most enthusiastic I’ve felt about a technology since the iPhone. I expect the level of disruption and transformation it creates will be of similar impact. The adoption curve across industries is the main open question I have.

Dear Apple TV and Netflix Please find a new color grading for your shows Thanks

It’s funny seeing rightwing Gen Alpha think the late 90s era was some anti-woke comedy free for all Pushing against “political correctness” was big business back then— woke was very much a cultural force then (complimentary)

Alysa Liu is an allegory for AI Her philosophy is performance for arts’ sake, and she happens to be great at her craft As humans, we can still do things we love for the sake of personal pride and joy, irrespective of how it might be judged on some rote economical value scale

A lot of people, including me, were wondering how an IRB could ever approve this study. The answer is that no IRB did. The person who “signed off” on approval from the only ethics board that reviewed it had resigned three years earlier. His signature was used without his knowledge.

Angie Craig: "They removed everyone from the ICE facility 30 minutes before we arrived — this is not the intent of the law. DHS is violating the law" Ilhan Omar: "What happened today is a blatant attempt to obstruct members of Congress from doing their oversight duties"

He's two years old and my nephew still can't predict the next token with anything near the accuracy of a frontier model. Could have trained a full model in six months

Did a demo of apple’s gaussian splatting research and built a custom splat renderer with interactivity, at a local meetup last week (in Raleigh) youtu.be/3MAkONN4i78 visionOS viewpoint m.youtube.com/watch?v=s1_C...

Really neat article soon benchmarks will be useless and the only useful metric is which models are solving unsolved problems phys.org/news/2026-02...

Code is now worthless The only thing that matters is meeting some actual need Food shelter entertainment safety exploration

This is pretty wild. A 3d rendering engine that uses svg as its output viewport www.reddit.com/r/singularit...

LLM improvements are causing a lot of existential crises. My advantage is I’ve been through so many that I’m immune. Highly recommend as a strategy globally, albeit not locally

Research shows fine-tuning vision-language models on harmful data causes unexpected misalignment across tasks, showcasing AI safety fragility. Many of the harmful behaviors remain, underlining the need for improved continual learning frameworks to ensure alignment. https://arxiv.org/abs/2602.16931

Well done Claude Opus 4.6! METR estimate that Claude Opus 4.6 has a 50%-time-horizon of around 14.5 hours (95% CI of 6 hrs to 98 hrs) on software tasks. metr.org/time-horizons/

I’ve shifted my thinking about AI the past 2 weeks away from where is AI going and how, to what real problems can I help solve with Ai? Just about anything you can think to do on a computer AI can help you do it, now it’s more a matter of finding the limit of the things you can even imagine trying

Fewer than 39,000 Cybertrucks were sold in 2024, and just over 20,000 found homes in 2025.

"We, the people, must redeem The land, the mines, the plants, the rivers. The mountains and the endless plain— All, all the stretch of these great green states" — Langston Hughes

This agreement allows DHS agencies like CBP & ICE to essentially skip the competitive bidding process. Palantir's software is considered dangerous primarily because its technology facilitates mass surveillance & enables targeting which erodes civil liberties. Gift link => archive.ph/1yX5v

Llama.cpp joins Hugging Face github.com/ggml-org/lla...

At least twice Amazon Web Services went down after they let their agentic Kori coding bot run loose on the back end.

Has anyone used Zvec before? I like that it supports dense, sparse, and multi-vectors. Any feedback?

it's claude code but for killing people

ICYMI: As memory shortage persists, vendor price quotes are not long remembered

What do we get first… Every city has a self driving car service Or A humanoid robot that can get in any car and drive it?

I expect we’ll see an optical ai breakthrough this year too btw

If you’re wondering what the red/white colors are like that on the parachute, it’s a numerical encoding that says “dare mighty things”

Percy's latest finding, captured on the day of her fifth landiversary with her SHERLOC WATSON camera 🔍 #Mars Feb. 18, 2026 - Sol 1777 Credit: NASA/JPL-Caltech 🧪🔭

Pretty weird blobby things on mars

Yep. Since the word "women" was included in their DEI target list, this brain-dead process resulted in the instant cancellation of _ongoing research_ into a potential breast cancer prevention candidate. Not treatment. Prevention. They're just sick with misogyny, along with every other bigotry.

Wrote about how DOGE bros blew up already approved grants by feeding project names into ChatGPT and asking it "is this DEI, tell me in less than 120 characters."

this is what sixteen thousand tokens a second looks like by the way

16000 tokens per second on a decent model. This type of speed is the future. Opens up an entirely new class of user experiences

Meta says it is "explicitly separating" Quest VR from its Worlds platform, making Worlds "almost exclusively mobile", letting it better compete with Roblox (Jay Peters/The Verge) Main Link | Techmeme Permalink

Introducing Gemini 3.1 Pro across the developer ecosystem. It's a smarter model for your most complex coding problems. Start building with Gemini 3.1 Pro today via the Gemini API in Google AI Studio, Vertex AI, Gemini CLI, Android Studio and our agentic development platform, Google Antigravity.