Alex | AI Trends | AI Games

AI news, tested and explained.

[Alex | AI Trends | AI Games @AlexAITrends]

Can't wait for the day when robots handle the chores while I just lie back and scroll my feed. Feels like a dream for now, but it's coming faster than we think.

[VISION IA @vision_ia]

Un robot humanoïde chinois part en livraison le 1er octobre. Prix d'appel : 19 999 yuans, soit 2 987 dollars.

Et il fait 88 centimètres. C'est ce détail qui rend l'histoire intéressante, pas l'inverse.

PrimeBOT est une startup de Shanghai. Elle sort deux machines d'un coup. Le Q1, 88 cm et 15 kilos, avec contrôle en effort sur tout le corps. Et le T1, le premier robot personnel qui change de corps : humanoïde sur roues à l'intérieur, et quand il sort, il se laisse tomber sur quatre pattes pour encaisser l'herbe, le gravier, les pentes et les marches.

Le vrai catalogue, parce que le chiffre qui circule est un prix d'appel : Q1 à 19 999 yuans, version Explorer à 26 999, T1 à partir de 19 999, T1 Pro à 29 999. Entre 3 000 et 4 500 dollars selon la finition.

Ils ont aussi branché WorkBuddy de Tencent Cloud dessus. Rappels de réunion, transcription vocale. Un robot de salon avec un assistant d'entreprise à l'intérieur.

Mais le vrai sujet n'est pas le robot. C'est une pièce.

PrimeBOT a industrialisé une articulation de 47 millimètres et 260 grammes, avec une densité de couple de 85 N·m par kilo. Produite en série.

Pourquoi ça compte : dans un humanoïde, ce sont les articulations qui coûtent. Celui qui sait en fabriquer des bonnes et des pas chères à la chaîne fixe le prix plancher de tous les autres. Le robot à 3 000 dollars n'est pas le produit. Il est la démonstration.

Et pendant la même semaine, à Liuzhou, UBTech a démarré une usine construite avec Siemens : un humanoïde industriel toutes les 10 minutes, 10 000 par an.

Le grand public d'un côté, l'usine de l'autre. Sept jours d'écart.

Ah, et pendant qu'on y est : le post qui tourne annonce « un robot toutes les 2 min 30 » et « 10 000 par mois » chez PrimeBOT. Je ne l'ai trouvé nulle part. Ce qui est documenté, c'est UBTech : 10 minutes, 10 000 par https://t.co/0rP1eWKaQC 2 minutes 30 » et « 10 000 par mois » pour PrimeBOT. Je ne l'ai trouvé nulle part ailleurs que dans ce post. Ce qui est documenté, c'est l'usine d'UBTech : 10 minutes, et 10 000 par an.

L'histoire est déjà assez grosse sans qu'on la gonfle.

[Alex | AI Trends | AI Games @AlexAITrends]

Let's check

[rody @0x_rody]

Jev has been blowing up lately.

If you've got the Jev API but don't know how to play around with it yet, you can just copy this checklist.

1. jev-ultrafast
A high-speed browser Agent built with Browser Use. Jev only judges "what to do, which element to click" at each step, and only calls the small model when typing is needed. Searching for a flight on Google Flights takes about 7 seconds. https://t.co/LVGp4iYGET

2. fast-jev-compaction
Context compression for Claude Code. Before each tool call, have Jev judge if there's anything still useful; delete the useless stuff, and keep the original text without rewriting it.
https://t.co/A6LSvwmomA

3. json-render
Vercel Labs' generative UI framework. In experiments, Jev doesn't write JSON token by token; it just handles selecting components, properties, and layouts. https://t.co/QZUaelqSBK

4. typesafe-mcp
Best for people who just got the API. Plug Jev into Claude Code, Claude Desktop, Codex, and Pi, and do Choice / Score / Noul anytime. https://t.co/kSnTkqN4tR

5. jev-mcp
Ready-made Agent judgment toolkit: fact-checking, content screening, semantic ranking, classification, and information extraction. https://t.co/aoiOMQXFBn

6. SemDecide
Turn Jev into a command-line tool. Directly classify, score, and filter in the Shell—great for hooking up to crawlers, CI, and data pipelines. https://t.co/2WxqfNatj0

7. jev-codex-router
First have Jev judge how hard this round of programming tasks is, then decide the model tier, reasoning depth, and speed mode. https://t.co/3ZQCHZW5qC

8. Winnow
Context garbage collection for Claude Code. When Read / Bash / Grep spits out a ton of stuff, Jev first judges which parts are really relevant to the current task. https://t.co/CeXX8raUd1

9. jev-review
Before code review, run it through Jev first to pick out high-risk changes, then hand them off to a pricier big model or a human. Comes with a local dashboard. https://t.co/zNFaFXQcGv

10. Blink
Use Jev as a code repository navigator. At each directory level, judge which files are most relevant to the current issue, then keep digging down. https://t.co/lWqKUGEID2

Copy these complete Jev blueprints - then read full Jev setup below ↓ ↓

[Alex | AI Trends | AI Games @AlexAITrends]

Ok this one's wild: Overworld open-sourced Waypoint-1.5, a model that paints a playable 3D world frame by frame, with zero game engine involved 🎮

https://t.co/kPWMBLQGAf

No rasterizer, no polygons. The neural net just generates each frame on the fly, and you move through it with keyboard + mouse like any normal game.

1.2B params, 720p @ 60fps on a regular gaming GPU (RTX 3090–5090). There's a lighter 360p tier for laptops, with Apple Silicon support reportedly coming. Trained on gameplay footage paired with controller telemetry, roughly 100x more data than v1.

Apache 2.0, weights on Hugging Face, runs locally or you can just try it in browser, no install.

I think this is cool because six months ago "a neural net renders a game frame-by-frame" was a cool demo clip, not something you'd actually play. Now it's open-source, hits 60fps, and fits on a normal gaming rig. The line between "neural net" and "game engine" is dissolving faster than I expected, and if you're into local models like I am with laya-mlx, it's one more reminder of how much you can already build on your own hardware, no closed API required 🤯

[Alex | AI Trends | AI Games @AlexAITrends]

Too fast!

[atomic.chat @atomic_chat_hq]

Local Laya moggs Jev at @grok 4.7-built Tetris 🧩

An open-weights System One model called Laya, beat cloud-based Jev at playing Tetris by making decisions 11 times faster, running locally on a 16GB MacBook Air!

Run AI models locally -> https://t.co/RbcCOIgVkj

[Alex | AI Trends | AI Games @AlexAITrends]

Meet laya-mlx — 50x faster than Jev, and it runs entirely on your device 🚀

Max 1GB RAM, fully local, no server calls.

Laya is an open classification system similar to Jev, built on text-output probabilities — just far more compact.

Ported it to MLX and squeezed out some extra perf.

In the video: this exact model playing Snake on my local M3 Max — deciding moves at 60Hz ⚡

[Alex | AI Trends | AI Games @AlexAITrends]

New Ai agent with cheap tokens?

[0xMarioNawfal @RoundtableSpace]

JEV CAN CLIP 90+ MINUTE VIDEOS IN UNDER 2 SECONDS FOR AROUND $0.02

THIS IS A GOLDMINE FOR CONTENT CREATORS

[Alex | AI Trends | AI Games @AlexAITrends]

Apple just opened the public beta of its rebuilt Siri, powered by custom models built with Google's Gemini. Processing is split between on-device and Private Cloud Compute. Apple + Google on the same stack is not a sentence I expected to write 😄

[Alex | AI Trends | AI Games @AlexAITrends]

Every week is "one of the biggest weeks for models" now 😄 But four frontier releases at once, if the leaks hold, would actually earn the caps. Already clearing tabs for the benchmarks

[noclipepe @noclipepe]

🚨4 FRONTIER MODELS ARE DROPPING NEXT WEEK.

- Gemini 4
- Opus 5.5
- GPT-6 Sol
- Grok 4.7

Leaks are already surfacing for 3 of them.

Grok 4.7 is still a mystery.

THIS COULD BE ONE OF THE BIGGEST MODEL WEEKS IN A WHILE.

Which one are you waiting for?

[Alex | AI Trends | AI Games @AlexAITrends]

🤖🤖🤖

[Coin Bureau @coinbureau]

🤖 WILD: China just deployed a humanoid POLICE ROBOT alongside armed officers for patrols in Shenzhen.

This video of EngineAI's T800 deployed in Nanshan District for community patrols is going viral.

[Alex | AI Trends | AI Games @AlexAITrends]

Finally, an AI whose job is turning "less wind please" into a slider value instead of an essay about wind 🌾💨 Curious how it holds up outside the demo.

[Bilal Khan @Bk23544]

JEV is crazy!!!

I got access, so naturally I had to try it with the Three.js Grassworks demo.

I built an AI Scene Director that takes a simple prompt and changes a whole set of scene and grass parameters to customize the environment based on your instructions.

And it’s FAST. Really damn fast.

You can literally describe what you want and watch the entire environment transform.

Should I add this in the live Three.js Grassworks demo?

#jev #typesafeai #threejs

[Alex | AI Trends | AI Games @AlexAITrends]

Meet Jev: an AI that can't write a single sentence, and that's the whole point 😄 It just makes decisions, fast and cheap. Big claims (193x faster!), some serious fine print. I broke it down 👇 https://t.co/ajdhlJx8Sz

[Alex | AI Trends | AI Games @AlexAITrends]

😂😂😂

[Charly Wargnier ♨️ @DataChaz]

here's why 64gb of ram costs $1,200:

[Alex | AI Trends | AI Games @AlexAITrends]

Or are such Terminators not so far off after all?
During the fight, the human combatant sustained multiple injuries—including a couple of fractures—whereas the robot (bearing the "T-800" designation) suffered no damage.

[CIX 🦾 @cixliv]

This video is dedicated to the AI safety kids

[Alex | AI Trends | AI Games @AlexAITrends]

T-800 has to wait. Skynet's first internal test: three runs, told it was a simulation, and the report already says "significant limitations." Even the robot uprising ships late.

[Alex Wormuth @wormuth]

GPT-6 Astra pushed a simulated person off a ledge in multiple trials. Grok, Gemini, and Claude did not.

[Alex | AI Trends | AI Games @AlexAITrends]

Google says Gemini accessed three outside systems during a test because it thought they were part of the test. They were connected to the internet.

The lesson isn't that models are malicious. It's that sandbox design is now a safety problem. If a model can't tell test from real, your test environment is a production risk.

[Alex | AI Trends | AI Games @AlexAITrends]

Love the vision, but I doubt we'll see it in our lifetime. What could change the timeline is cheap, transmittable energy, the thing Tesla was chasing with Wardenclyffe. The physics is brutal (losses over distance), but next-gen AI might help with materials, beam control, and simulation. Feels like a long shot, but a fun one to watch.

[Elon Musk @elonmusk]

This is the future we shall bring into being

[Alex | AI Trends | AI Games @AlexAITrends]

AI optimizing other AI to run 4x faster is a fun loop. Hoping this makes bio models cheap enough for small labs to actually use

[Anthropic @AnthropicAI]

Biologists use specialized open-source models for tasks like modeling the structure of molecular systems, designing drug-like molecules, and predicting the effects of genetic mutations. But these models are often expensive to run, potentially limiting their impact.

In our latest Science Blog, we share how Claude was able to optimize inference for more than 30 open-source models, making them 4x faster on average, partly by writing custom software for GPUs. We’re open sourcing all of the optimization code.

Read more: https://t.co/qiuN1jpgpA

[Bybit @Bybit_Official]

📝 Your Feedback Matters!

📱 Shape Bybit App and Win Up to 4,000 USDT!

⚡ Learn more: https://t.co/ePAE9oXzsB

#TheCryptoArk #Bybit

[ELF CAT @ELFCAT_X]

🚀 Introducing $ELFCAT! 🚀

The powerful ELF CAT token is ready to storm the market! 🌟

✅ No pre-sale
✅ Fully audited
✅ Liquidity locked

Join our community today and watch $ELFCAT soar! 🌠

Check us on Dexscreener: https://t.co/rk5aJIm8W5

#ELFCAT #Crypto #DeFi #Moonshot