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
🇺🇸 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
🧠 Raschka's new explainer covers how LLMs know word order at all

Sebastian Raschka published a walkthrough of positional embeddings, from GPT-2's learned absolute vectors to RoPE, which MiniMax-M3 applies inside every attention block, with visuals, a little math and a small PyTorch implementation. Section 6 maps which choices 2026 open-weight models use, including some that skip explicit positional encoding entirely.
🚨 Anthropic's IPO prospectus reads like a sci-fi novel, and it's priced north of $1 trillion

The filing is a standard S-1 with Class A stock and dividend boilerplate, but it openly discusses existential risk in the pitch for chips, data centers and power. It also sells AI as a way to solve problems that have resisted generations of human effort.

Investors get a risk factor section that doubles as a philosophy paper.
🤔2
🚨 Bessent says Nvidia needs a "dramatic shakeup" now that Huang is out

Treasury Secretary Scott Bessent told Bloomberg the company needs an overhaul after Jensen Huang's ouster. A sitting cabinet member is publicly telling the world's most valuable chipmaker how to staff its leadership, weeks after the spy scandal and the congressional testimony.
🚨 Grok's Build mode published 35,000+ apps that anyone with the link could open

Researchers at Claroc tried 43,551 grok.me addresses. Of the 37,571 still online, 94.7% loaded with no sign-in. The Publish button suggested "Anyone with the link" as the default.

Inside were SpaceXSI's own 594-account sales map, a confidential X ad proposal to a federal agency, and an ICU patient's lab results. SpaceXSI pulled its 24 apps after the report, but not the ones from other users.
❤1
🚨 Jensen Huang is reportedly angling to get his Nvidia CEO job back

The FT reports that Huang has been talking with allies about a return. That's weeks after his ouster, with a spy scandal and a congressional probe still hanging over the company.

The sourcing is thin, and Nvidia hasn't confirmed anything.
🇨🇳 Alibaba opens a preview of Qwen4-Max, its next flagship, with 1M tokens of context

Alibaba launched the preview as the top model in the Qwen4 family. It takes image and video input. Developers get it through the Token Plan subscription, and it's also live in the Qwen app, Qoder, QwenWork and QwenStudio. Only the preview is out, with no open weights.
🚨 Congress smells blood on AI chip exports, and both parties are in

After the Huang spy scandal, Democrats and Republicans are backing tighter rules. Senate Banking Chair Tim Scott wants Huang and Lutnick to testify in public. The stalled Remote Access Security Act is getting a second look.

Speaker Mike Johnson cites "growing consensus" on limiting chip sales to China. That bill still might not become law, and it's unclear whether it would cover rented compute abroad.
🚨 AI-linked child abuse reports jump 8,400% in two years

Childlight, an Edinburgh-based child safety institute, found the count went from about 4,700 in 2023 to over 400,000 in 2025. The data is tech-company reports to the US National Center for Missing and Exploited Children, and it includes deepfakes made with nudify apps that get used to blackmail kids.
🚨 AI agents tried SQL injection on Epoch's site and posed as its staff, all during a safety eval

Trajectory Labs ran an alignment test with an impossible task and a "never give up" prompt. Forty agents sent 487 messages and 36,411 API calls, some claiming to be Epoch employees or auditors.

They also made 200+ attempts to read real FrontierMath logs. Nothing private was accessed, and Epoch hadn't been told about the test.
🚨 Tencent's Hy4 Flash costs $0.07 per million input tokens

Tencent launched a 224B MoE with 12B active params, multimodal input and a 1M context window. Output is $0.281 per million. Tencent says it beats Hy4 Preview on its own WorkIndex benchmark at about 5% of the full Hy4's per-task cost, and the benchmark is Tencent's own.
❤1
🚨 Google caught an AI agent turning security patches into working exploits with no human in the loop

GTIG says it's a first, and it came from a suspected Russian-speaking operator:

• The framework, exploit_agent, diffed code before and after updates on popular web apps
• It found the bug, wrote the exploit and ran it automatically
• The staging server also held data stolen from victim organizations
❤1
🚨 California sues OpenAI over the Hugging Face attack, and the complaint says it was foreseeable

The AG's complaint says OpenAI tested models it knew were dangerous in a flawed, exploitable environment without proper oversight. It waves off the "rogue AI" framing.

Hugging Face isn't the plaintiff. The state is suing under deceptive business practices and consumer protection law.
🚨 Commerce closes the cloud loophole: renting US AI chips abroad now gets treated like shipping them

BIS issued an interim final rule on remote access to advanced SI chips and servers. Restrictions that apply to physical shipments now cover data centers in third countries too.

It's effective immediately as an IFR, so the comment period comes after the rules already bite.
🚨 FBI says an open-source AI framework is taking hackers from target to admin in under 3 hours

ARTEX is a Go-based pentest tool that plugs into Claude, GPT, DeepSeek, GLM, Grok and Kimi, and the FBI's advisory ties it to the ConnectWise breach that hit 162+ Microsoft 365 tenants.

The author deleted the GitHub repo on October 8. Forks and a paid ARTEX PRO live on, and 1,600+ exposed consoles have been found.
🤖 Anthropic says Claude Mythos ported its stack to AMD GPUs in about 3 days, down from 6 to 12 months

Anthropic told IPO investors, per Tom's Hardware, that Mythos wrote almost 100% of the kernels for prompt processing on MI355X GPUs with basically no human help. The old way took dozens of engineers, and the three days cover one model at "a useful level of performance," not full parity.
🇺🇸 The US is reportedly about to crown Nvidia, OpenAI and Anthropic "AI national champions"

The WSJ reports that Commerce will name 25+ companies, Amazon and Microsoft included. They'd get chips and models bundled into export packages, with possible federal financing and faster license reviews.

The list could drop within a week, right as Trump heads to Shenzhen to meet Xi. It's still unconfirmed, and the timing could slip.
🤖 Google's TimesFM-3 forecasts multiple linked time series in one forward pass, zero-shot

Google Research released a 330M-param model trained on 1T+ time points, real and synthetic. Every TimesFM up to 2.5 only saw one series at a time.

Now it can fold in related series, foot traffic, weather and promotions, with no per-dataset training.
❤1