a Microsoft director privately called AI scraping "the largest theft of labor in human history."
but the more interesting number is 93%.
new court filings say Microsoft’s own data showed Copilot cutting clicks to NYT sites by up to 93% vs traditional Bing search.
AI learned from the web, then learned to answer without sending us back to it.
that’s not just a copyright fight. it’s a feedback-loop problem.
my bet: AI search eventually needs source attribution + revenue sharing built into the product, or it slowly kills the content supply it depends on.
Excel has a world championship.
AI is already beating the people who win it.
PCWorld gave ChatGPT a “very hard” finals spreadsheet.
it finished the whole thing in 6m35s. the human winner hadn't finished after 40 minutes.
the interesting part isn't that AI learned Excel.
agents are starting to use software the way we do: read the file, reason through it, click, type, fix mistakes.
that means “being good at software” is moving up a layer.
less memorizing every formula.
more knowing what outcome you want + catching when the machine is wrong.
the next Excel power user might barely know Excel.
OpenAI spent millions attacking a $1M math problem.
the prize clearly wasn't the point.
~10,000 AI agents worked in parallel for 88 hours on a 90-year-old problem.
we keep imagining AI as one genius chatbot. the stranger future is intelligence you scale like cloud compute.
We spent years laughing at humanoid robots falling over.
China's military researchers are already planning for the version that doesn't.
Reuters found studies modeling humanoids working alongside troops in urban combat.
The technical part people miss: the hard problem isn't making a robot dangerous. it's getting one to see, balance, manipulate things and adapt in a messy world.
Factories are already training exactly that stack.
The robot race was never going to stay cute for long.
I think keyboards are about to lose a surprising amount of screen time.
Google just turned Gmail, Docs and Keep into things you can talk to.
not dictation.
You can ask "what gate is my flight?", ramble out a document, or dump a messy thought and let AI organize it.
The technical shift here is bigger than voice:
AI is moving from an app you open to a layer sitting inside the apps you already use.
Give it a few years and "using software" might look a lot more like telling it what you want than clicking through menus.
blog.google/products-and-pla…
Meta's comeback isn't the 61 on Artificial Analysis.
It's $0.55 per task.
Muse Spark 1.3 is basically in GPT-5.6 Sol territory on AA, while costing ~42% less per task.
Meta also says it uses ~20% fewer tool calls + ~25% fewer tokens than 1.2.
For agents this feels like the metric that actually matters:
not $/token.
not one-shot IQ.
$/finished task.
And if the open weights land anywhere close to this, decentralized inference gets a lot more interesting.
Google might be quietly proving something pretty important with Gemini 3.8 Flash.
the interesting part isn't another benchmark bump.
it's that "Flash" models are getting good enough at coding + agents that the old cheap-model vs smart-model split is starting to blur.
if a fast/cheap model handles 90% of the coding loop, the expensive frontier model becomes an escalation path instead of the default.
that's probably where AI products are heading anyway:
cheap agent does the work.
expensive model gets called when shit gets hard.
Anthropic casually dropped a pretty insane detail yesterday:
They audited their production RL environments and >10% had issues. reward hacking, broken tasks, bad configs, etc.
Kinda funny that we're spending billions making agents smarter while discovering the test harness itself can't always be trusted 😭
Agent infra is going to become a security discipline of its own.
ngl the reported NVIDIA + Hugging Face deal is way more interesting than just another $12.9B AI acquisition.
NVIDIA already owns the shovel. Now it wants one of the biggest places where everyone goes to find the gold.
Models themselves are getting cheaper + more interchangeable. The moat is slowly moving to everything around them: compute, distribution, tooling, community.
Also makes the decentralized AI thesis a lot less abstract imo.
If the whole AI stack keeps vertically integrating, neutral infrastructure starts becoming pretty valuable.
reuters.com/technology/nvidi…
Imagine joining a live story where #AI directs real people in real time — without a fixed script.
It watches the room, tracks every character, injects plot twists, sends private instructions, and keeps the story coherent as humans improvise.
5 strangers join a room. Nobody gets the full script. Nobody knows how the story ends.
Each person gets a character, private motivations and information nobody else sees.
Then they perform the story live.
Behind the scenes, an AI director continuously decides:
→ who should learn what → when to inject a new event → which character gets a hidden instruction → whether tension is too low → where the story should branch → how to recover when someone completely goes off-script.
Technically, this starts looking less like prompt engineering and more like #GraphEngineering.
You could model it as:
Actors → Character Agents → Scene State → Memory → Event Engine → Director → Narrative Branches
But here’s the weird part:
Humans are nodes in the graph too.
And humans are terrible deterministic systems.
They improvise. They misunderstand things. They leave. They contradict the AI. Sometimes they create a better plot than the system planned.
So the AI needs persistent world state, character memory, real-time multimodal understanding, low-latency orchestration, conflict generation, safety constraints and recovery logic…
while somehow not making the performance feel AI-generated.
That might be the hardest part.
If this works, AI doesn't replace actors or creators.
It becomes the invisible director + world engine behind them.
I’m seriously thinking about building something like this.
Would you join a live story where the AI is directing the world, but the humans decide what actually happens?
🤖 Made with AI
James retweeted
🔴 SOME CHINESE DEVELOPERS JUST HUMILIATED THE ENTIRE PAID AI VIDEO INDUSTRY WITH A FREE TOOL
they released LongCat-Avatar, an open-source AI that turns a photo and audio file into a realistic talking video with synchronized lip movements.
you can generate videos that run for minutes, completely free.
no camera or studio needed. upload the image, add the audio, and let the model do the rest.
it’s open source, FREE to use, and the repo is public.
I’ll leave the repo in the comments.
James retweeted
Motionsites MCP is insane.
your AI agents can now Access 700+ Animated Websites, Backgrounds, design prompts, sections, and apps.
ai tools can write code. they just don't know what good looks like. now they do.
👉 motionsites.ai/mcp
James retweeted
Bittensor $TAO Subnet 80 just launched: @OpenRoboto
Robotics is entering an interesting phase.
Instead of a few companies building behind closed doors, OpenRoboto is creating an open environment where models can improve together.
Open weights. Data. Benchmarks.
Small step today.
Potentially a huge shift for how intelligent machines are created.
James retweeted
Distributed training has never been easy. Building a distributed system that makes economic sense on Bittensor is harder still.
Today, we’re introducing Connito: a network for collaboratively building composable specialized AI.
Signal shows us that just by training a particular selection of experts as a partial model, we can improve the performance of the whole system.
Learn about our architecture and results: connito.ai/blog/distributing…
