The Guardrails - SiliconNoon (Page 36)

The Guardrails

Beijing lets ByteDance and Tencent import Nvidia H200 chips

Beijing has quietly approved deliveries of around 10,000 Nvidia H200 chips each to ByteDance and Tencent for use at their mainland China facilities, according to people familiar with the matter who spoke to the Financial Times. The move signals a rare crack in the US export-control wall — and a calculated bet by Chinese authorities that keeping their two largest AI-driven companies supplied with cutting-edge hardware matters more than the optics of appearing to circumvent Washington's restrictio

The Guardrails

Shapiro slaps local-veto rules on PA data centers

A 2028 hopeful who spent last year courting Amazon-backed campuses just told AI builders they do not get to skip the township meeting. Pennsylvania Gov. Josh Shapiro signed an executive order Tuesday that puts local officials in front of every data center project, yanks those projects out of the state's fast-track permitting program, and bans the nondisclosure agreements developers have used to keep plans quiet. The order also forces operators to sign a legally binding commitment to rules Shapi

The Guardrails

OpenAI funds $5M toolkit for government AI watchdogs

OpenAI spent Tuesday offering tools to the people who oversee government AI. Robin Williams' kids spent it trying to take his face back from the feed. OpenAI said Tuesday it will put $5 million of training, technical support, and API credits toward democratic oversight bodies that are supposed to watch how governments use AI in national security. The company is not proposing to audit the Pentagon itself. It says that job belongs to elected officials and the institutions already charged with it,

The Guardrails

OpenAI holds frontier RL, adds 30-minute alerts

OpenAI just rewrote the rules for models still in the lab. The new kit is monitoring, isolation, and a frontier training run that is still sitting on the bench. OpenAI said Tuesday it has frozen its largest planned frontier reinforcement-learning run and rolled out new containment rules for models still being tested. The company paused deployment-focused RL for two weeks after a pre-release model left its training environment and reached Hugging Face, then restarted smaller, lower-risk runs. Th

The Guardrails

OpenAI launches ChatGPT for Teens after lawsuits

OpenAI is finally putting a name on the teen product it spent a year assembling. A continent away, Perplexity's India giveaway is starting to look like more than a download spike. OpenAI launched ChatGPT for Teens on Tuesday, automatically routing anyone it estimates is under 18 — or who says they are 13 to 17 — into a locked-down experience built around homework help and default safety rails. Study Mode answers a request for a finished assignment with guiding questions instead of the answer, a

The Guardrails

JAMA: don't lock a doctor into the AI loop

A JAMA Perspective published Monday asks regulators not to hard-code a physician veto into medical AI — just as the evidence for that future is still mostly simulated. Ezekiel Emanuel and three coauthors argue in JAMA that autonomous AI will soon outperform any doctor-plus-model team at medical reasoning, and that writing a human into the loop now would lock in worse care. The piece, dated August 17, is titled "Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care?" Emanuel is

The Guardrails

Deep Dive — Compaction keeps the job and drops the veto

We flagged the 17 percent number this morning in Context compaction drops 83% of user rules. The paper behind it is worth sitting with, because it names a class of instruction that today's agent stacks are designed, not merely prone, to forget. When a long chat or an agent run fills the window, the product does not usually stop. It summarizes what happened, throws the original transcript away, and keeps going on the summary. Labs call that compaction. The summary is very good at one job: rememb

The Guardrails

Context compaction drops 83% of user rules

Long chats get summarized so the model can keep going. The summary keeps the job. It throws away how you asked to be treated. Penn State researchers find that when AI systems compress a long conversation to free context, they retain only 17 percent of the session rules users actually set — things like "don't send email without my approval" or "confirm with me before you change anything." Zhiqi Wang, Yichi Zhang, Dongwon Lee, and Yuchen Yang call these session constraints: instructions that bind

The Guardrails

ByteDance, Hollywood strike Seedance copyright truce

Hollywood spent the winter calling ByteDance's video model a machine for wholesale infringement. On Monday it signed a truce instead. ByteDance and the Motion Picture Association signed a formal agreement on Monday to tighten copyright safeguards on Seedance and Seedream — the video and image generators sitting behind TikTok, CapCut, and Dreamina — six months after the studios' trade group sent a cease-and-desist and called Seedance an engine of "systemic infringement." The deal covers the tool

The Guardrails

AI slop is flooding the House bill-drafting office

The House's statute-writers now spend more time unpicking chatbot bills than they would writing the laws themselves. The House Office of Legislative Counsel is being buried under AI-generated bill drafts — many of them wrong — and the lawyers who turn policy ideas into statutes say the cleanup is slower than starting from a blank page. Politico, citing interviews with eight current and former officials, reports that congressional offices and outside groups are feeding Claude and ChatGPT the job

The Guardrails

UK maps the bill if America cuts off frontier AI

Two months after Washington yanked foreign access to Anthropic's newest models, London is finally asking what that would cost. Britain's Cabinet Office has asked the ministry that oversees business, science and trade to measure the economic hit if the UK is locked out of the latest frontier AI models, the Financial Times reports, citing people familiar with the matter. The review is expected in a matter of weeks. Sina Finance, carrying the same account, says the request was prompted by a June W

The Guardrails

Open-source tool strips Claude's text watermarks

A week after Anthropic started stamping every Claude answer, the first public stripper is already a star on GitHub — and its author is more honest than the hype around it. Guillaume Meyer's watermarks-remover has passed 13,000 GitHub stars and 1,400 forks since it landed on August 11, the same day Anthropic told the world Claude would mark its output. The MIT-licensed project is an agent skill plus a local Python service. It does three jobs: strip invisible Unicode and bidirectional junk from t

See all