lamb retweetledi
It's okay to be obsessed with something everyone else thinks is boring. The world runs on people who care way too much about things most people don't think about at all.
English
lamb
1.1K posts
lamb
@lamb356_
Building things | just here to help https://t.co/gjhtVUURbB
United States Katılım Aralık 2025
1.2K Takip Edilen79 Takipçiler
lamb retweetledi
banging away at a hobby project with the help of AIs. the goal: make BLAKE3 more efficient than SHA256 for almost all use cases in wasm/js/web. why? Because if all software switches to BLAKE3 by default, users will be safer, and more empowered to share data trustlessly.
English
lamb retweetledi
lamb retweetledi
The “genius girl” who previously worked at DeepSeek and was recruited by Lei Jun for Xiaomi AI is now on Twitter as well. It feels like more Chinese AI talent is realizing they can come here, speak for themselves, and build influence directly. I’m all for the added interaction and transparency.
Fuli Luo@_LuoFuli
MiMo-V2-Pro & Omni & TTS is out. Our first full-stack model family built truly for the Agent era. I call this a quiet ambush — not because we planned it, but because the shift from Chat to Agent paradigm happened so fast, even we barely believed it. Somewhere in between was a process that was thrilling, painful, and fascinating all at once. The 1T base model started training months ago. The original goal was long-context reasoning efficiency. Hybrid Attention carries real innovation, without overreaching — and it turns out to be exactly the right foundation for the Agent era. 1M context window. MTP inference for ultra-low latency and cost. These architectural decisions weren't trendy. They were a structural advantage we built before we needed it. What changed everything was experiencing a complex agentic scaffold — what I'd call orchestrated Context — for the first time. I was shocked on day one. I tried to convince the team to use it. That didn't work. So I gave a hard mandate: anyone on MiMo Team with fewer than 100 conversations tomorrow can quit. It worked. Once the team's imagination was ignited by what agentic systems could do, that imagination converted directly into research velocity. People ask why we move so fast. I saw it firsthand building DeepSeek R1. My honest summary: — Backbone and Infra research has long cycles. You need strategic conviction a year before it pays off. — Posttrain agility is a different muscle: product intuition driving evaluation, iteration cycles compressed, paradigm shifts caught early. — And the constant: curiosity, sharp technical instinct, decisive execution, full commitment — and something that's easy to underestimate: a genuine love for the world you're building for. We will open-source — when the models are stable enough to deserve it. From Beijing, very late, not quite awake.
English
lamb retweetledi
Tether is enabling you to train and run billion-parameter AI models directly on everyday devices
They're literally reinventing the term "smart phone"
This means you will have a FULLY PRIVATE offline personal AI in your pocket
Pump it full of your private information and have it manage everything, including real time feedback on messages you're sending
I think this is an understated breakthrough
Paolo Ardoino 🤖@paoloardoino
Tether AI breakthrough Tether AI team just released new version of QVAC Fabric to include the World’s First Cross-Platform BitNet LoRA Framework to Enable Billion-Parameter AI Training and Inference on Consumer GPUs and Smartphones. Background Microsoft's BitNet uses one bit architecture to dramatically compress models. Traditional LLMs operate on full-precision computation, where weights are stored as complex, high-resolution numbers. The innovation of BitNet is that it shrinks these weights into a tiny ternary range of only -1, 0, and 1. significantly reducing memory usage and computation. LoRA, is a parameter-efficient fine-tuning technique that reduces the number of trainable parameters by up to ninety-nine percent. Together they slash memory and compute requirements. Yet BitNet has mostly been limited to CPU or CUDA NVIDIA backends, and lacked the support of LoRA fine-tuning. Enters QVAC Fabric: the unlock Today, with QVAC Fabric LLM, is the first time BitNet LoRA fine-tuning and inference work cross-platform across GPU vendors and operating systems using Vulkan and Metal backends. That means support for AMD, Intel, Apple Metal and also Mobile GPUs. And for the first time ever, BitNet inference runs efficiently on smartphones using mobile GPUs. On flagship devices, GPU inference is 2 to 11 times faster than CPU while using up to 90% less memory than the full precision models. The biggest unlock: QVAC Fabric LLM support for BitNet LoRA fine-tuning on heterogeneous GPUs. Our team was able to demonstrate this by fine tuning models up to 3.8 billion parameters on all flagships phones such as Pixel 9, S25 and iPhone 16 and up to 13 billion parameter models on the iPhone 16. Github repositories: github.com/tetherto/qvac-… : general QVAC Fabric codebase github.com/tetherto/qvac-… : specific QVAC Fabric's BitNet knowledge base, architecture docs and pre-built binaries What does it mean? What used to require dedicated GPUs now runs on consumer hardware. This breakthrough is the first real-world signal of a local private AI that can truly serve the people. And this is just the beginning. In the next months and years Tether will relentlessly continue to invest significant amounts of resources and capital to continue to research and develop open-source intelligence that can scale and evolve on local devices, providing maximum utility and privacy to its users. The era of Stable Intelligence has just begun. Free as in freedom.
English
lamb retweetledi
threadguy endorses me on twitch and crude oil rallies 2.5% since close this guy is more influential than jerome powell
English
lamb retweetledi
gpt-5.4 has ramped faster than any other model we've launched in the API: within a week of launch, 5T tokens per day, handling more volume than our entire API one year ago, and reaching an annualized run rate of $1B in net-new revenue.
it's a good model, try it out!
English
lamb retweetledi
32× efficiency improvement in just the last 3 months, that’s the crazy jump from GPT-5.2 to GPT-5.4! 37 cents/task is essentially almost at human-level efficiency (target was 24 cents/task). This was inconceivable a year ago when o3 cost $4500/task on ARC-AGI-1, 12,000x improved!
Jesse🔸⏹️@PoliticalKiwi
GPT-5.4 (High) has now cleared 90% on this benchmark at a cost of just $0.37/task So that's a 32x efficiency improvement in the last three months, or 12000x since December 2024
English
lamb retweetledi
Congrats to the @Kimi_Moonshot team!
This is awesome.
Great to see this level of research coming from open-source frontier model labs.
I liked the paper so much I built a Rust implementation of it ;)
Full AttnRes + Block AttnRes with two-phase inference, built using Burn (tensor library and Deep Learning Framework, in Rust, by @Tracel_AI).
Runs on CPU, CUDA, Metal, wgpu.
Includes an interactive TUI that trains a model live and visualizes depth attention evolving from uniform to selective in real time.
Repo link and more on what is implemented in the comments.
Kimi.ai@Kimi_Moonshot
Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers. 🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. 🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale. 🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead. 🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains. 🔗Full report: github.com/MoonshotAI/Att…
English
lamb retweetledi
lamb retweetledi
This is wild.
143 million people thought they were catching Pokémon. They were actually building one of the largest real-world visual datasets in AI history.
Niantic just disclosed that photos and AR scans collected through Pokémon Go have produced a dataset of over 30 billion real-world images. The company is now using that data to power visual navigation AI for delivery robots.
Players didn't just walk around with their phones. They scanned landmarks, storefronts, parks, and sidewalks from every angle, at every time of day, in lighting and weather conditions that staged photography would never capture. They documented the physical world at a scale no mapping company with a fleet of vehicles could have replicated on the same timeline or budget.
Niantic collected this systematically, data point by data point, across eight years, while users thought the only thing at stake was catching a rare Charizard.
The most valuable AI training datasets in the world aren't being assembled in data centers. They're being built by people who have no idea they're building them.
NewsForce@Newsforce
POKÉMON GO PLAYERS TRAINED 30 BILLION IMAGE AI MAP Niantic says photos and scans collected through Pokémon Go and its AR apps have produced a massive dataset of more than 30 billion real-world images. The company is now using that data to power visual navigation for delivery robots, letting them identify exact locations on city streets without relying on GPS. Source: NewsForce
English
lamb retweetledi
the cure for cancer leads directly into the cure for aging
2026 is the year its happening
Satya Nadella@satyanadella
We’ve trained a multimodal AI model to turn routine pathology slides into spatial proteomics, with the potential to reduce time and cost while expanding access to cancer care.
English
lamb retweetledi
ORCA Dexterity just open-sourced 3 anthropomorphic robotic hands, tendon-driven with self-dislocating joints for reliability.
3D-print the free STL files, add off-the-shelf motors, assemble in under 8 hours for ~$2,200, and you have full hardware for dexterous manipulation research in any lab.
One of their variants is the orcahand touch. It comes with up to 83 taxels per fingertip, 1mm resolution, and 0.1N force detectability.
ORCA Dexterity@orcahand
it's time to drop three new #opensource robotic hands! this time with tactile sensors! Tweak it, 3D print it, and use them in your robotics and physical AI research! Here are some wild examples ↓↓↓
English
lamb retweetledi
lamb retweetledi
@rasbt @teortaxesTex Nice! My autoresearch would love some markdown version of this - pool of ideas.
English
lamb retweetledi
The most comprehensive platform for realtime crypto trading is now open for everyone for FREE forever.
Enjoy 😱
MMT@MMT_Official_
Today we're opening up MMT - everyone now has access to professional grade tools for FREE. - Market Profile / TPO - Custom Session TPO - Hyperliquid MBO Profile - Aggregated Heatmaps - HD Heatmaps - Liquidation Heatmap - Hyperliquid Liquidation Heatmap - Hyperliquid Stop Loss Heatmap - Hyperliquid Take Profit Heatmap - Aggregated Footprints - Filtered Footprints - Dual Cluster Modes - Bucketed Trade Size Groups - Aggregated Indicators - Aggregated CVD - Aggregated Open Interest - Net Longs/Shorts Indicator - VWAP Suite - Volume Bubbles - Custom Scripting - Community Indicators - Aggregated Orderbooks - Aggregated DOM - Orderbook Imbalances - Orderbook Depth Overlay - 1s Time-Frames - Custom Time-Frames Available now, for everyone – only in MMT
English
lamb retweetledi
metadao's tam is literally trillions
v few people can see that right now
research directed at applied decision markets will become v important
we're still in the 1st inning
Daniel Barabander@dbarabander
There is only one form of capital formation I'm aware of that can scale with AI: @MetaDAOProject AI will cause an explosion in long tail companies. More long tail = more necessities for programatic bolted-in accountability. The friction of the DE C-corp cannot keep up.
English
lamb retweetledi
Soon the market will wake up and we will have a $zec like runner for privacy in the AI space
While running models locally is the ultimate form of privacy we need alternatives if we want to utilise SOTA’s without fully compromising on privacy
It will be a bigger narrative than $zec
English