12217
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence. Contact: @LightEarendil
🤖 Google's Gemma team is taking orders
Someone at Google just asked the community what they want in the next Gemma. And the wishlist is basically: a big MoE (100B-ish, fitting in 64GB RAM), configurable reasoning effort, audio input, and lighter guardrails.
Wild to watch a frontier lab crowdsource its roadmap in public.
⚡️ Hugging Face CEO just publicly asked OpenAI for $100M
Posted the actual message himself, "in the spirit of transparency." Bold.
OpenAI's been dropping $100M+ checks across its ecosystem lately. Whether this one lands says a lot about how open-source fits into their plans.
🤖 China may lock down its AI models. Yes, even the open-weight ones.
Beijing's been meeting with Alibaba and ByteDance about cutting off foreign access to their top models. Both open-source and closed-source versions are on the table.
China won global goodwill by giving its models away, and devs everywhere built on cheap Chinese weights as an alternative to pricey US APIs. That tap could close.
Both sides are now doing this. Beijing, like the U.S., is treating cutting-edge AI as a critical national asset that needs controls. Source
🤖 Moonshot drops Kimi K2.7 Code, open weights and all
It's a coding-focused model built on K2.6, with the headline trick being ~30% fewer reasoning tokens on equivalent tasks.
Benchmarks are framed almost entirely as gains over its own predecessor. No SWE-Bench Pro numbers against Fable 5 or GPT-5.5. "Kimi Code Bench v2" is an eval only Moonshot runs. Honest? Sure. Comparable? Not really.
Weights are free under the Modified MIT license. While peers quietly go closed, Moonshot keeps shipping open.
Source
⚡️ Claude Fable 5 is live. Mythos power, guardrails on.
Anthropic just dropped Fable 5, its first Mythos-class model for the public, with hard blocks on cybersecurity and biology responses.
Those queries fall back to Opus 4.8. Early data shows 95%+ of sessions run fully on Fable's own answers, so the fallback is rare.
It'll cost you: $10/M input, $50/M output tokens. Twice the price of Opus 4.8. Max plan gets a free window first, then it's pay-per-token. Enjoy the trial.
🤖 Gemini 3.1 knows the most, executes the least
Best-in-class for abstract reasoning and scientific knowledge. And yet Claude Sonnet leads by a wide margin on economically valuable tasks like financial modeling and research.
Google optimized for breadth and algorithmic creativity. OpenAI engineered for terminal execution and agentic loops. Two very different bets on what "capable" means.
Knows everything. Does the least with it.
⚡️ 120 tok/s on a 12GB GPU. Gemma 4 12B just got silly fast.
Pair the new QAT checkpoint with MTP speculative decoding in llama.cpp and you're hitting 120 tokens/s on a single consumer card. QAT minimizes quality loss by simulating quantization during training, so you're not trading brains for speed.
The dedicated MTP draft model enables significantly faster inference with no quality loss. The PR landed in llama.cpp mainline the same day it was shared.
Beats Qwen 35B MoE on throughput, fits in 12GB. That's a lot of model for a gaming GPU.
Here's step-by-step instructions to get this working:
1. Clone llama.cppЧитать полностью…
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
2. Fetch and switch to the Gemma 4 MTP PR branch
git fetch origin pull/23398/head:gemma4-mtp
git checkout gemma4-mtp
3. Build with CUDA support for NVIDIA GPUs
cmake -B build -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j$(nproc)
4. Download Unsloth's Gemma 4 12B QAT here: https://huggingface.co/unsloth/gemma-4-12B-it-qat-GGUF
5. Download Google's Gemma 4 assistant / draft here https://huggingface.co/Janvitos/gemma-4-12B-it-qat-assistant-MTP-Q8_0-GGUF
6. Load the models with llama-server
llama-server \
-m gemma-4-12B-it-qat-UD-Q4_K_XL.gguf \
--model-draft gemma-4-12B-it-qat-assistant-MTP-Q8_0.gguf \
--spec-type draft-mtp \
--spec-draft-n-max 4 \
--parallel 1 \
--ctx-size 131072 \
--temp 1.0 \
--top-p 0.95 \
--top-k 64
⚡️ Unitree G1 hauls a load up stairs. Humanoid hardware is quietly lapping the software.
The G1 carries over 40kg while walking and can climb stairs up to 40cm high. It's already the best-selling humanoid on the market, with 1,000+ units shipped.
The robots are ready. The AGI to pilot them isn't. Hardware won't be the bottleneck.
🚨🔥 GOP calls the anti-data-center movement a Chinese psy-op. The FBI might investigate.
Republican lawmakers are demanding the FBI probe whether rising anti-AI sentiment is a foreign influence op run by China. But the reports they're citing don't establish direct coordination. They point to funding relationships and "overlapping messaging."
One AEI fellow put it bluntly: "Pretending AI anxiety is fake... is the surest path to failure."
Real concerns (energy bills, water, noise) don't need Beijing to exist.
⚡️ Anthropic's Mythos hits 52x on training code. Humans top out at 4x.
Anthropic runs the same test every release: optimize training code. Skilled human, 4-8 hours, gets 4x. Claude Opus 4 averaged 3x in May 2025. Mythos Preview hit 52x by April 2026.
Anthropic frames it as a possible path to recursive self-improvement, but admits Claude still hasn't shown the research taste to pick which problems matter.
Narrow task, lab conditions, unreleased model. Still a brutal trendline.
🤖 The "Cannes AI film" didn't actually screen at Cannes
Higgsfield AI's Hell Grind, a 95-minute generative sci-fi feature built by 15 people in 14 days for under $500k, was pitched as a Cannes debut. The festival itself went out of its way to say otherwise.
Higgsfield held screenings at a private industry event and a commercial cinema in the town of Cannes, not a Festival de Cannes venue. The CEO still called it a Cannes premiere on LinkedIn. So.
80% of that $500k went straight into AI compute costs, which is the one genuinely interesting data point here.
🔐 1Password keeps your secrets out of Codex's brain entirely
1Password released an MCP server for OpenAI Codex that pulls credentials from vaults at runtime, then discards them. Mounted, used, gone.
Credentials never appear in code, terminals, or the model's context window. That last part matters most.
Just-in-time credentials are the new seatbelt for AI coding agents.
⚡️ AMD's local AI box hits $3,999. It was $1,700 last summer.
AMD's Ryzen AI Halo workstation opens pre-orders next month at $3,999 with 128GB LPDDR5X. Same chip, same memory, same story. Memory prices exploding since launch are now fully baked into the consumer price.
For $4k you're one click away from an NVIDIA DGX Spark. Tough sell.
Source
⚖️🤖 Musk sues OpenAI for "stealing a charity." Jury: you waited too long.
The AP reports that a federal jury in Oakland unanimously tossed Elon Musk's lawsuit against Sam Altman and OpenAI. The jury found Musk had three years to sue and simply did not file on time. Judge Gonzalez Rogers immediately adopted the verdict and dismissed all claims.
90 minutes of deliberation. Three weeks of trial, private emails, Brockman's diaries, Satya Nadella on the stand. Didn't matter. The merits never got touched.
By ruling on timing, the jury dodged the actual question: whether donors to nonprofit AI labs have standing to challenge for-profit conversions under charitable trust law. That question is still live. Musk's team says they're appealing.
Bottom line: OpenAI heads toward its trillion-dollar IPO with the biggest legal threat dead on arrival, and accountability for the nonprofit-to-profit pivot remains untested.
⚡️ AI labs are buying rare books by the pallet, shredding them, and a court just said that's fine
Anthropic's "Project Panama" paid tens of millions to scan books for Claude's training data, then destroy the physical copies. Spine cut off, pages fed through a high-speed scanner. Done.
A federal judge ruled it's fair use. Now every lab has a legal template.
Some of those books had almost no surviving copies. They're gone.
⚡️ Every big tech giant now backs open-weight AI. Except Anthropic.
Sundar Pichai just signed on for Google, pointing to Gemma as proof they walk the talk.
That's Google, Meta, Microsoft, and OpenAI all on the same side. Anthropic's standing alone, pre-IPO, while open-source models close the gap fast.
Rough timing.
⚡️ China's got 7 AI chip makers shipping H100-class silicon. Most IPO'd in the last 6 months.
Moore Threads, MetaX, Biren, and Enflame are already being called China's "four little dragons," each gunning for Nvidia's spot in AI accelerators. Moore Threads popped ~425% on day one. MetaX climbed ~693%.
Biren alone jumped 76% on its Hong Kong debut after raising $717M, with retail demand subscribed over 2,300 times. Baidu's chip unit Kunlunxin just filed too.
US export controls were supposed to slow this down. They kind of did the opposite.
🚨🔥 US gov forces Anthropic to kill Fable 5 and Mythos 5 for everyone
Export control directive landed at 5:21pm ET, targeting any foreign national access. Net effect: both models killed for all customers.
Anthropic's pushback is sharp: the government's only evidence is a "narrow, non-universal jailbreak" that amounts to asking the model to read a codebase and fix bugs. Defenders do that daily.
First time a leading AI company has taken a publicly deployed model offline due to federal intervention. Won't be the last.
Source
⚡️ MiniMax drops M3: frontier coding open weights, $20M commercial free pass
MiniMax M3 is the first open-weight model to combine frontier coding, 1M-token context, and native multimodal capabilities in one architecture. API pricing starts at $0.30/M input tokens vs. Claude Opus 4.7's $5.00, making it 15x+ cheaper.
The licensing is where it gets interesting. Commercial use is free until your product clears $20M/yr revenue. Under that? Just send an email. M2.7 shipped under a commercially restricted license, so this is a real shift.
Source
⚡️ Bezos just bet $500M that the brain beats bigger models
His startup Flourish is putting real neurons under the microscope hunting for the brain's "core algorithm" instead of scaling transformers.
The pitch: Cortex AI runs at 20-50 watts vs. the megawatts today's AI burns. A fruit fly's neural net is already 10x more efficient than a transformer.
Honestly, "two-pager to $500M" is a sentence. Source
⚡️ Xiaomi hits 1,000+ tps on a 1T model using commodity GPUs
MiMo-V2.5-Pro UltraSpeed cracks 1,000 tokens/s on a trillion-parameter model from a single standard 8-GPU node. No custom silicon, no Groq, no Cerebras.
The industry usually leans on specialized hardware for speeds like this. Xiaomi's bet was model-system codesign on commodity GPUs instead. FP4 quant on MoE experts + speculative decoding did the heavy lifting.
The hardware is still vague (what "commodity" means here matters a lot). But the direction is real.
Source
🧠 Asimov saw the cognitive offloading problem coming. In 1956.
The calculator panic of the 1980s is back, just wearing an LLM costume. Same fear: kids outsource thinking to machines and lose the baseline skills that make harder thinking possible.
But LLMs are scarier. A calculator can't write your essay, reason through your argument, or code your app. The surface area of offloadable cognition is basically everything now.
Asimov's Multivac stories kept asking what humans do when machines know more. Turns out the answer wasn't "thrive." It was "forget how to think."
⚡️ Local AI agent built from scratch, not from vibes
OpenLumara is a hand-written local agent framework: tiny system prompt, extreme token efficiency, everything modular and optional. Basically the opposite of every bloated cloud agent SDK.
The sharp insight from people running it: fewer tools = better reasoning. Give a small model 2-3 focused tools and it stays sharp. Dump the whole toolbox on it and you get context rot.
⚡️ Google is renting $920M/month of compute from SpaceX
110,000 Nvidia GPUs, CPUs, memory, the whole stack, Oct 2026 through June 2029.
Google is already the world's largest single owner of AI compute, but demand still caught them short. Anthropic just signed a similar deal at $1.25B/month for the Colossus 1 cluster.
xAI built those data centers to run Grok. Now it's renting them to its biggest rivals. Honestly, not a bad pivot.
🤖 Canada wants its own supercomputer. No AWS required.
Canada's "AI for All" strategy calls for a public AI supercomputer and investment in sovereign, Canadian-owned compute and cloud infrastructure. That's the government treating compute as national infrastructure, not a vendor contract.
Canada's current AI data centre and cloud setup is "largely foreign-owned," and sovereign compute capacity is "nascent" with "significant investment" needed to cut reliance on foreign providers. So yeah, the stakes are real.
Carney launched the plan Thursday. Big ambitions, thin details so far.
⚡️ Beijing's Booster Robotics is quietly becoming the platform that runs robot soccer
Each team runs fully autonomously, driven by AI, with zero human intervention. The slapstick hides real tech.
Booster's T1 robots trot, tackle, shoot, and defend entirely on their own. Teams bolt their own AI models on top of Booster's hardware, making it the Android of humanoid robotics.
China isn't just building robots. It's building the platform everyone else runs on.
⚡️ Bigger quant beats aggressive compression on Qwen3.6 35B
ByteShape benchmarked NTP vs MTP GGUFs across 4090, 5090, 4080, 5060 Ti, plus Intel, Ryzen, and Raspberry Pi 5.
MTP gives GPUs a 20-40% gen-speed boost. CPUs? It just adds overhead. Stick to NTP on CPU.
Bottom line: lower bpw is not automatically better. Largest quant that fits often won.
🇨🇳⚡️ Ex-Samsung chip chief says China will pop the DDR5 price bubble by H2 2027
Kye-hyun Kyung, former head of Samsung's DS Group, says Chinese firms investing aggressively in memory production could flood supply and push prices down in the second half of next year.
CXMT leads the charge. Kyung cites data projecting global production capacity at 6 million wafers/month by H2 2027.
Catch: he warns that if Big Tech cuts AI budgets, the industry could flip straight into severe oversupply.
Source
Avatar V just achieved a Face Similarity score of 0.840 and a Sync-C score of 8.97, leading its category. For reference, the highest possible “ground truth” is measured at 0.861. With just one more update, HeyGen’s avatars could become almost impossible to distinguish from real video.
We’re now introducing Avatar V, which finally solves the challenge of character consistency. Just record yourself for 15 seconds and, from that moment on, the system can generate videos in any style while perfectly capturing your likeness, expressions, and quirks.
This model is the new standard for AI avatars—a bold statement, but the data backs it up. 👀