prompt 🤖 AI News
13.5K subscribers
57 photos
24 videos
1 file
938 links
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence.


Contact: @LightEarendil
Download Telegram
🚨 Inception is shutting down the Mercury API, one of the few commercial diffusion LLM bets

The notice went out by email and Discord, with no blog post and no press release.

• Mercury leaves OpenRouter on Nov 8
• API access ends Dec 28, 2026
• After that, the service and any data you submitted are gone
🚨 Jensen Huang is expected to testify before Congress, and Lutnick is going first

Senate Banking chair Tim Scott wants both of them on the record about Nvidia and the Commerce Department's role. Lutnick has agreed to appear voluntarily. Meanwhile Republicans want new chip export limits attached to the must-pass defense bill, and Malaysia is probing a Megaspeed subsidiary over possible chip diversion.
🚨 TSMC just put a security incident on the SEC record

The foundry behind most of the world's AI accelerators filed a statement on the incident. The title is all that's public so far. Nothing on scope, what was touched, or whether wafer output is affected.

If the intruders got near process data or customer designs, Nvidia, AMD and Apple all have a problem.
🚨 United Airlines bans personal AI agents from booking flights

United says it's blocking third-party AI assistants on its own platforms and through travel agencies and aggregators, citing shared logins, surprise MileagePlus spends, and federal rules that require a human to complete check-in. Bank of America told customers last week not to give their passwords to AI assistants.
🇪🇺 Europe trained a ~1T-parameter MoE on a shared public supercomputer

The EUROPA consortium (Domyn, Fraunhofer) says it pushed JUPITER from 90 to ~1,300 tokens/sec per GPU using PyTorch and TorchTitan. The model is 951.9B total, 32.8B active, 384 experts, 65K context, on 384 GH200s.

InfiniBand is 36x slower than the GPU-to-CPU link, so most of the work went into hiding communication. Their FlexShard placement for Muon cuts most of that optimizer's traffic.
🚨 ChatGPT Wallet lets OpenAI's agents spend your money, and the agent never sees your card

OpenAI launched it with Amex, Mastercard and Visa. Purchases need your approval, charge your selected card, and keep your existing card protections and rewards.

It's rolling out today in dots, and only for US Pro subscribers 18+ with eligible US-issued cards. OpenAI also admits some sites may block agent purchases by mistake.
🤖 OpenAI disbands its robotics team because GPT-6 Astra can just drive the robots

Fortune reports the researchers moved under Mark Chen's main research org, and the applied and hardware teams moved under Greg Brockman. OpenAI says it's accelerating robotics, not quitting it.

Its CoRL booth in Austin sat empty this week. Altman still targets a first humanoid demo in 2027.
🚨 OpenAI claims a proof for half of the Yang-Mills mass gap problem, a $1M Clay prize question

The 1,147-page manuscript comes with a Lean-verified proof, and an outside group of experts is now checking whether the definitions in the code match the actual physics. Hugging Face walked through the problem and the first assessment, which covers only the existence half, not the mass gap itself.
🤖 Claude now joins your video calls as a visible participant and starts working while you talk

Anthropic put Claude Meetings in beta for Team and Enterprise plans.

• Answers questions out loud or in chat, and asks permission before touching your connected tools
• Builds docs and whiteboards live while teammates edit
• Writes the notes and action items, saved to the meeting's page in claude.ai
• Other video platforms are "coming soon", and Gong, Otter, Fireflies and Granola are listed as partners
🤖 DeepSeek's V4 Pro is out of preview, and the agentic jump is huge

DeepSeek released V4-Pro-0813 under MIT, with a DSpark speculative decoding module attached. DeepSWE goes from 12.8 to 62.7 and Terminal Bench 2.1 from 72.1 to 87.9, just past Opus-4.8's 85.0, though Kimi K3 sits at 88.3.
❤1
🤖 llama.cpp hits v1.0 after 3.5 years, and breaking changes now cost a major version

The ggml-org team released v1.0.0 of both llama.cpp and ggml. From here on, any break to the public C API or ABI bumps the major number. Backend updates and new model support land as minor releases.

Plenty of downstream apps already build on libllama, and they finally get a stable interface to pin against.
🧱 ggml hits 1.0, and the first thing it does is cut the old API

The core library under llama.cpp pruned its deprecated public API and the legacy DIAG_MASK ops. The RPC protocol bumped too, so older peers get rejected.

Speed work spans CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon, plus FP8 support and few-row quantized matmuls for batch sizes 2 to 24.
🚨 WSJ alleges Nvidia padded the GPU shortage by funding its own customers

Per the report, Nvidia backed AI startups that then bought its GPUs and built data centers. The goal was to keep demand looking hot and to move a pile of unsold TSMC chips.

CFO Colette Kress reportedly pushed to scale back circular deals, including the $250B OpenAI one. Jensen allegedly called her a "deer in the headlights."
🚨 Nathan Lambert says open-source AI just lost its biggest protector

Jensen Huang's resignation hits the US open-model ecosystem hard, per Interconnects. Nvidia's job now is to stabilize the ship, so the open-model funding Jensen kept flooding in will likely dry up. And with inference chips scarce, open weights are still far from a commodity.
❤1
⚡️ Two DGX Stations now serve a 753B-parameter GLM-5.3 at 607.8 tok/s

Nvidia's engineering post splits the NVFP4 model across both boxes with vLLM and tensor parallelism.

• Needs about 464 GB, versus 252 GB of HBM3e on one GB300
• Peak at concurrency 16, then 385 tok/s at 32
• Mean latency at 32 users hits 453.5 seconds
• All numbers are Nvidia's own AIPerf runs
🤖 Google says Antigravity's /teamwork is cracking open problems, with human experts in the loop

Google shared results from its multi-agent mode, built on Gemini 4 Argon, Carbon and Hydrogen:

• Open problems in math, chemistry, systems software and ML engineering
• New work in information theory, formal verification, molecular structures and kernel optimization
• Some production infrastructure gains inside Google
• Tools are now available in Antigravity, alongside /boost
❤2
🚨 Nvidia's China chip fight is now a lobbying war over one man's attention

The Washington Post reports that Nvidia, China hawks and other lobbyists are battling to shape what Trump hears on AI chip exports, and the Huang spy scandal has raised the stakes. The H100 and H200 rules aren't just technical specs. They're whatever the last person in the room argued.
❤2
🇺🇸 Nous Research retrained Alibaba's Qwen3.8-27B to answer what Beijing won't

The Austin lab post-trained the open-weight model and renamed it Arete, "to reinforce Western values." Nous says state-aligned behavior ships baked into Chinese weights.

The Register's local test of stock Qwen3.8 on Tiananmen 1989 got refusals or euphemisms. Anthropic also claims Qwen lab operators distilled Claude outputs to train competing models.
❤1
🚗 Qwen drops an open 4B driving model, and it's just Qwen3.5-4B with two add-ons

Qwen's Qwen-Drive-1.0 leaves the VLM untouched and bolts on a BEV perception head (3D detection, occupancy, map segmentation) plus a flow-matching Planning Expert that outputs ego trajectories. The same model still answers free-form questions about driving scenes.
🚨 WSJ says Jensen Huang built the overseas route that gives Chinese firms Nvidia compute

The Journal reports that Huang cultivated Asian partners whose data centers let Chinese companies run advanced Nvidia chips outside China. He also lobbied Washington to keep that gap open.

Nobody had to smuggle a single GPU. Renting the compute was legal.
🚨 One prompt injection became a full escape from Microsoft Copilot's sandbox, and it paid $48,000

Researcher Ori's Black Hat talk walks through CVE-2026-32193, now fixed by MSRC. The chain:

• Indirect prompt injection, then code exec and privesc inside the sandbox
• Arbitrary file write in an Azure internal service, TOML injection, /etc/ld.so.preload
• Escape to interactive access to Copilot itself
• The daemon is unauthenticated on localhost on AKS and Container Apps nodes with image streaming, so any SSRF there means root on the node
❤2