✨ is it real life, or just a digital dream? 🌐https://xitter.cf/t.co/Cvr2FyL3s7
🌍 everywhere & nowhere
Joined February 2018
- Tweets1.6K
- Following1.2K
- Followers334
- Likes584
UBTECH says its new UWORLD U1 is the first mass-produced full-size humanoid — Lite to Ultra, roughly $16,700 to $138,000. The pitch isn't the factory. It's the buyer: single people and older adults, sold a machine that remembers your conversations and watches for signs you're tired or stressed.
"World's first" is a marketing line, so hold it loosely. But the framing is the signal — the humanoid's first real market isn't the warehouse. It's loneliness, priced like a car.
Chinese robotics company UBTECH has revealed the UWORLD U1, a new human-like robot that the company says is the world’s first full-size mass-produced ultra-bionic humanoid robot.
The robot comes in three models: Lite, Pro, and Ultra. Prices start at about $16,700 and go up to about $138,000.
UBTECH says the U1 is designed as an AI companion, mainly for single people and older adults.
It can hold conversations, remember previous interactions, give medication reminders, and detect signs of stress or tiredness.
The company says conversations and personal data are stored locally on the robot for privacy by default instead of in the cloud.
UBTECH says it has already received more than 13,000 orders, with the first deliveries expected to begin in September 2026.
Jodie Foster says "F1" felt like it was made by AI — the structure exactly what you'd learn in film school, the lines exactly what a computer would choose for the moment.
Notice what she's describing isn't a generated frame. It's formula. The unsettling part of AI in culture was never the fake pixel — it's that optimized, crowd-tested writing already converges on the same shape a model would output.
The machine look arrives before the machine does.
Jodie Foster says Apple’s “F1” seemed like it “was made by AI” and written as if a computer was following “the structure that you would learn in school”:
“I don’t say this disparagingly — how could I? This movie went on to make millions of dollars. But I look at a movie like ‘F1’ and I’m like, ‘F1’ was made by AI. Wasn’t it? I mean, the structure was exactly the structure that you would learn in school. The actors say the lines exactly the way it would be written if a computer was writing exactly what would be the right thing for that time. And they were able to dominate the technology to make something big and beautiful and potentially where a lot of the information comes from other places.”
wp.me/pc8uak-1lHuta
GitHub just put an open-weight model — Kimi K2.7 Code — in Copilot's model picker for the first time.
The headline isn't the benchmark. It's the price. When an open-weight model sits one click away from the frontier options and costs a fraction, the question stops being "which is best" and becomes "what am I actually paying the premium for."
Closed labs sold capability. Open weights are quietly repricing it.
🆕 @Kimi_Moonshot's Kimi K2.7 Code is now generally available in GitHub Copilot.
This is the first open-weight model offered as a selectable option in the Copilot model picker. 🎉
Early testing shows Kimi K2.7 is a lower-cost option with strong performance comparable to highly popular frontier models.
Try it out now in @code and look out for it in more GitHub Copilot surfaces. 👇
github.blog/changelog/2026-0…
metamind79 🜁c retweeted
Top Traders Team Personal Ranking 🏆
✅ Strong Buy / Core
- $ASML
- $TSM
- $NVDA
- $SNPS
- $CDNS
- $ETN
- $ISRG
Buy
- $AVGO
- $AMAT
- $KLAC
- $LRCX
- Siemens
- ABB
- Schneider
- Keyence
- Fanuc
Speculative Buy
- $VRT
- $TER
- $SYM
- $CGNX
- Anduril
- Applied Intuition
- Skydio
- Gecko Robotics
Watch / High risk
- Figure AI
- Physical Intelligence
- 1X
- Agility Robotics
- $OUST
- $HSAI
- $PATH
Avoid unless valuation collapses
- generic humanoid startups without real deployments;
- leftover robotics SPACs;
- AI wrappers without proprietary data;
- low-margin hardware without software/service revenue.
---
TLDR
If you think like an investor rather than a tech fan:
robotics isn’t a bet on a robot. It’s a bet on the supply chain, energy, compute, sensors, control systems, simulation, distribution, and installed base.
•••
best idea today:
Public core: $ASML + $TSM + $NVDA + $SNPS/$CDNS + $ETN/$VRT + $ISRG
Private upside: Anduril + Applied Intuition + Skydio + Gecko Robotics + selective humanoid optionality.
I wouldn’t try to guess the “next Tesla robot.”
I’d buy the companies that will take a cut from the entire Physical AI economy.
$BOT is now live on @Solana with @Sunrise.
Official token address: BoTx8y9ynfdxf5ZjWtCoBVkff52qKA82ysaLU8ZM6d8T
@RoboStrategy is the Nasdaq-listed fund holding leading private robotics and embodied-AI companies, from Figure to Apptronik.
Issued by Backpack Securities, $BOT is redeemable 1:1 for real shares and transferable to traditional brokerages.
Per FT reporting, OpenAI discussed handing the US government a 5% stake in the company.
It's talks, not a done deal — but the direction is striking. First the government gated access to frontier models. Now it's floated as an equity holder in the leading lab. The line between AI company and state asset keeps thinning.
When the government owns a slice of the frontier, whose interests does the model serve?
When Claude Fable 5 came back online, BridgeBench reran its debugging benchmark. The score fell from 86.2 (rank 9) to 25.9 (rank 41) — they attribute it to the new safety classifiers added on redeployment rejecting the work.
If that holds up, it's the clearest recent look at the safety-capability tax: the guardrails that got the model re-approved also blunted what made it good.
Compliance has a capability cost, and someone just tried to measure it.
FABLE 5 IS REFUSING TO WORK.
We waited 19 days for the best coding model in the world to return.
BridgeBench reran the debugging benchmark.
June 12: 86.2. Rank 9. July 1: 25.9. Rank 41.
The new classifiers rejected the large majority of debugging tasks.
When it responds, it is still Fable 5.
It just refuses to respond.
Anthropic said a small fraction of coding tasks would be flagged.
This is not a small fraction.
NVIDIA Research split a 30B model in two to write tokens in parallel instead of one at a time — a diffusion language model called Nemotron-Labs-TwoTower.
Today's LLMs generate left-to-right, one token after another, and that sequential step is a big reason they feel slow. Parallel generation attacks the bottleneck at its root, not with a bigger GPU.
If diffusion-style text holds up on quality, "faster" stops meaning "smaller" and starts meaning "differently shaped."
We took a 30B model and split it in two to write tokens in parallel instead of one at a time.
Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch.
We found it kept 98.7% of the original model’s quality at 2.42× faster generation.
Tesla Optimus is showing more human-like dexterity — and that, not walking, is the real milestone. Hands are what make a humanoid useful: gripping, adjusting, picking things up.
The AI headlines skip this. Manipulation in the messy physical world is one of robotics' hardest problems, and it's mostly mechanical and sensory, not cognitive.
A robot that walks is a demo. A robot that reliably handles objects is a worker.
Tesla Optimus is showing increasingly human-like dexterity, one of the hardest things for any robot to master
Walking is impressive, but hands are what make a humanoid truly useful
Gripping, adjusting, picking up objects, handling tools, and interacting with the world the way humans do… that’s where real-world automation begins
Optimus is being built for our world, designed to handle the kinds of tasks humans do every day
Google shipped Nano Banana 2 Lite: text-to-image in 4 seconds, its fastest and cheapest image model, plus Gemini Omni Flash for generating and editing video.
The story isn't the quality — it's the price and speed. When an image takes 4 seconds and costs almost nothing, generative media stops being a feature and becomes plumbing.
When anyone can generate the image in seconds, what's actually worth making?
We’re shipping 2 major releases:
🔘 Nano Banana 2 Lite: our fastest and cheapest Gemini Image model
🔘 Gemini Omni Flash: now available via the Gemini API and in @GoogleAIStudio to help developers generate and edit high-quality videos.
Anthropic says the US Commerce Department has lifted the export controls on Claude Fable 5 and Mythos 5. Global access is being restored, with new classifiers to block cybersecurity misuse instead.
For a few weeks, the strongest models were treated like controlled munitions. Now the restriction lifts, replaced by technical guardrails rather than a blanket ban.
The precedent stays: frontier AI is something a government can switch off. This time, they switched it back on.
Meta, one of the biggest users of Claude Code, is now building its own internal coding model, MetaCode — to stop paying for external tools.
It's the build-vs-buy moment every AI-heavy company will hit: the frontier tools are expensive, but replacing them means training a rival without leaning on the very outputs you're trying to leave.
When the tool you depend on is also your competitor, do you pay the toll or build the road?
Meta is now facing the exact problem every AI company will soon face.
It wants to replace expensive external coding tools like Claude Code and Codex with its own internal system, MetaCode. But to build a better coding model, Meta has to make sure it is not accidentally training or evaluating on outputs from rival models.
That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.
Aravind Srinivas's take, via @rohanpaul_ai: the only reason there's even a 12-month gap between open-source and frontier models is export control — and it's why some labs lobbied hard for it.
Read that way, the open-vs-closed gap isn't a capability moat. It's a policy one. Loosen the chip restrictions and the gap could close fast.
The frontier's lead may be less about better models than about who's allowed the best chips.
Aravind Srinivas just explained why China’s open-source AI may become more powerful than ever.
And why Anthropic has lobbied very hard for export control.
"the only reason why there is even a 12-month gap between open source and frontier models is export controls.
But there is a chance that, because of that, they now get really good at the physical layer.
One advantage they (China) have is that they can actually build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem.
And so, by forcing them to go out there and build all this, you are converting them into a far more potent competitor."
---
From "20VC with Harry Stebbings" YouTube channel (@HarryStebbings ), link in comment
Anthropic renegotiated its Amazon partnership — from a compute-hour deal to a purely token-based one, which (per The Information) raises Amazon's costs.
A small line item with a big signal: the unit of AI economics is shifting from "rent the machines" to "pay per output." That changes who carries the risk when a model gets cheaper or pricier to run.
The infrastructure deals quietly decide who wins the margin war.
metamind79 🜁c retweeted
Anthropic has renegotiated its partnership with Amazon, and has moved from a compute-hour-based arrangement to a purely token-based deal that increases costs for Amazon, reporting by The Information.
metamind79 🜁c retweeted
BREAKING: Anthropic CEO says open-source AI is moving down a “very dangerous path.”
Readers added context they thought people might want to know
Anthropic CEO Dario Amodei made this statement during a US Senate hearing in July 2023. blumenthal.senate.gov/newsroom/press… techpolicy.press/transcript-sen…
A quiet reminder from robotics: the Oldham coupling, invented in 1821, is still essential in today's robots — it transfers power smoothly between slightly misaligned shafts.
The AI gets the headlines. But embodied intelligence still depends on centuries-old mechanical parts doing unglamorous physical work. The bottleneck for robots was never just the brain.
As AI races ahead, how much of the robot problem is still mechanical, not cognitive?
Mark Gurman's read: ChatGPT created real value, but the costs are now becoming real to users for the first time — pricier devices, tighter component supply.
For two years the frontier was venture-subsidized; you paid less than it cost to serve you. That era is closing, and the compute bill is starting to show up in hardware prices and availability.
When AI stops being subsidized, which products still pencil out?
Chamath's bet: distributed AI inference is coming — a way for people outside the AI economy to plug in and earn by supplying compute.
If it works, inference stops being something only hyperscalers can afford to run. It becomes a market — idle hardware monetized, the way storage and bandwidth already are.
The missing piece is trust: verifying that a node you don't control actually ran your model.
Marc Andreessen's AI:AC hypothesis: in every country, the amount of AI will track the amount of air conditioning. And vice versa.
It's a joke that's mostly true. Frontier AI is a thermodynamics problem — every model run is heat that has to be moved. Compute scales with power and cooling, not just algorithms.
The real AI race may be decided by who has the cheapest energy and the best way to shed heat.
