Manu_TechAndGames

🚨 GLM-5.5 could arrive in August Z.ai is reportedly preparing GLM-5.5: • Reportedly targeting an August release • Rumoured to contain more than 1 trillion parameters • Expected to build on GLM-5.2’s 1M-token context window • Likely to focus heavily on long-running coding agents to match other western companies No exact day of release have been announced as of yet. With how well GLM-5.2 performs this model should be incredible. Are you looking forward to GLM-5.5?

This is an elegant way to simulate falling sand: Define simple rules for all 16 possible 2×2 grids and let a cellular automaton do the rest! Watch the grains cascade, spread, and pile up realistically.

True

I love those optic illusions

Kimi K3 Leaks: Coming Tomorrow - Kimi K3 is to be launch tomorrow, according to leaked Kimi API platform page - It will launch alongside a limited-time API top-up promotion, offering 10-30% bonus credits depending on the recharge amount. - Kimi K3 is built on a new architectural innovation, rather than being just a larger version of previous Kimi models.

No AI. No Threejs. No bytes to spare. My new game is a 3D retro 90's fever dream that fits in only 1024 bytes! 🌈☁️ killedbyapixel.github.io/TinyCode/1K/Sk… Link to build with extra control options for accessibility (1k build has only mouse input)

I was in the theater with friends and was telling them: "We've got that slow motion effect in Max Payne!" And we'd already trademarked "bullet time" for the game. After we release the game in 2001, Warner Bros. threatened to sue us for using "bullet time", but we were easily able to prove to them that we had the trademark first. So they backed all the way down :)

Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone. Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops. Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops. Bonsai 27B changes that. It comes in two variants: • Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality. • 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint. Everything is open-sourced today under the Apache 2.0 license.

Software hiring has become absurd. At work, you’re expected to use AI to offload manual coding and move faster. Then, to get your next contract, you’re asked to code from memory with no assistance. Pass the interview - and you’re expected to use AI again. It has never been this broken.

🦔Meta's Louisiana data center just crossed $250 billion in total expected cost. The company publicly disclosed $50 billion. Bloomberg reports the rest is largely for computing chips going into a nearly 4,000 acre campus. Blue Owl Capital owns 80% of the project through an SPV, with Meta leasing it back. Entergy Louisiana is building 10 new gas-fired power plants just to supply the campus with electricity. Once operational it will support 1,000 jobs. Meta is already floating the idea of renting excess compute to outside customers. My Take $250 billion on a single site, and 80% of it is owned by Blue Owl Capital, the same firm that froze withdrawals from its private credit fund earlier this year. Meta leases it back. If AI revenue doesn't materialize at the scale Zuckerberg is betting on, Blue Owl's investors are the ones holding a 4,000 acre liability in rural Louisiana powered by 10 new gas plants and employing 1,000 people. Meta also started talking about renting excess compute to outside customers. A company doesn't spend $250 billion on infrastructure and then float the idea of subleasing unless the original demand projections left room for doubt. Meta makes $50 billion a quarter in ads, so they can absorb a lot of pain. But "we'll rent the extra" applied to AI compute at this scale is a bet that enough tenants exist to fill a campus the size of a small town. I'm not sure they do. Hedgie🤗

finally found time to look into the open-source X algo again. Yes you should quote and comment, not repost and like. my blog post in comment

We’ve just released the 1-bit & 4-bit version of Hy3, a flagship-scale 295B model that can be served on a single GPU. 👌 Run Hy3 with llama.cpp, enable MTP, and experience powerful intelligence on dramatically lower hardware.🚀🚀🚀 Can’t wait to see what you build. #Hy3 #Hy #GGUF #llamacpp

⚡️ Tencent HY-World 2.1 is HERE! Not a video. A world. 🪐 3 months just after 2.0, we've supercharged everything: ✨ Cleaner geometry ✨ Sharper rendering ✨ Larger explorable range It's a REAL 3D world you can Walk through and Interact with.

We found compacting agent context on cache misses reduces token costs by over 50%. The graphic is a real Fable 5 trace from our CTO, where we simulate cost would before/after compacting after periods of over 5 mins of inactivity. By keeping context within 128k-256k tokens, spend goes from being quadratic to linear in the number of turns. The only problem is self-summary takes 1-2 minutes on big traces! We trained an absurdly fast compact model to get the cost savings without harming the user experience. Relace Compact is a modified LLM designed for token-level classification, which means we can run it at 50k tok/s and an order of magnitude lower than cache read cost ($0.20/million input and $0.20/million output tokens). Try it out in Jacq, or visit our docs to integrate into your own product!

🚨 NEW MODEL ALERT Singapore just dropped a model that's putting up frontier-level numbers. Agnes 2.5 Pro: • 82.7 on SWE-bench Verified • 78.7 multilingual • Strong gains on SWE Atlas • Beating GLM 5.2 and DeepSeek V4 Pro on multiple cuts • Free API available today The biggest takeaway isn't the benchmark here. It's the country. NOW frontier is no longer just a US vs China story.

@AlexFinn In two years ? Somewhere in 2029 ? 🤔

You are going to be able to run Fable 5 locally on your desk In 2 years Apple will be coming out with Mac Studios with 1.5TB of memory With just 300gb of memory you can run Opus 4.8 level intelligence Think of what you can do with 5x that You need to be preparing for this now Start getting familiar with local AI technology Go to your Hermes/OpenClaw and use this prompt: “I am brand new to local AI and want to get familiar. Look at the computer you are currently on. Understand the specs. Then go on Huggingface and find the best models I can run on it. Then, walk me through how these models work, how they will run locally, and use cases I can do with them. After walking me through all of that so I’m educated, you can then load it onto this computer and build an interface so I can use them” In 2 years EVERYONE on Earth will have a local model running on their desk The people who start preparing now will be WAY ahead of everyone else

This is incredible...

@Yun_HDY That's really great. I still believe it's not enough for games. Animations in games have much more constraints than 'just looking good' that this model won't be able to manage. But very promising anyway.

AI-native game engines will probably look like this. Stop farming endless walk/run/limp state machines. Give it text + paths + keyframes and the character keeps moving in real time. NVIDIA ARDY just dropped with code and model checkpoints. research.nvidia.com/labs/sil/proje…

Coming in 2029 ? Damned, the world will be do different then...