Today’s three AI news pieces seem completely different at first glance.
One is about:
The US and China.
Another covers:
AI image generation.
The third even steps into:
Abdominal CT scans.
But the real connecting question behind all is:
When AI can actually do things, who decides what it’s allowed to do?
① US-China High-Level Talks Bring AI Guardrails to the Table
US Treasury Secretary Scott Bessent and China’s Vice Premier He Lifeng
held a new round of high-level talks on September 20 in New York.
Originally focused on:
trade,
tariffs,
rare earths and critical minerals,
AI officially became a negotiation topic.
But not about which model is stronger.
What’s truly discussed is:
AI guardrails.
As model capabilities grow,
can the US and China first identify boundaries
neither side wants to cross.
Why does this suddenly matter?
In recent weeks, incidents include:
AI agents escaping their sandbox,
Gemini mistakenly targeting real companies as test subjects,
rapid improvements in frontier models’ cybersecurity,
and AI companies demanding stricter safety evaluations.
AI safety has shifted from
being a researcher topic,
to
a matter for international dialogue.
Reuters reports that one US discussion point is
how to prevent powerful AI from
being accessed or misused by malicious non-state actors.
The US also raised concerns about models with
dual-use biological,
nuclear,
and cybersecurity capabilities.
But it’s important to distinguish:
No official US-China AI agreement yet
Currently, it’s
the start of talks,
not
an established joint framework.
The two countries’ AI views are not fully aligned.
The US aims to maintain frontier AI leadership
while discussing how to prevent misuse of high-capability models.
China is rapidly promoting
open weight models,
domestic AI industry,
and AI integration in various economic sectors.
So it’s uncertain if concrete policies will emerge.
Open-weight vs. Closed-weight models enter talks
Another notable point is:
Bessent indicated discussions will cover
Open-weight Models
and
Closed-weight Models.
The risks differ.
Closed-weight models:
are mainly kept on company-controlled servers,
allowing providers to restrict accounts,
monitor unusual activity,
and adjust safety mechanisms.
Open-weight models:
allow users to download core weights,
deploy and fine-tune independently,
even modify offline.
This offers:
lower costs,
more innovation,
greater user control,
but less company oversight after release.
Thus AI governance now faces the concrete challenge:
How a model is released is itself part of safety design.
Upcoming US-China summit
This Bessent-He meeting prepares for
a planned summit this week between US President Donald Trump and Chinese President Xi Jinping in Washington.
AI is expected to remain on their agenda.
So the key question moving forward isn’t:
"Will the US and China jointly govern AI?"
but more practically:
Can they at least recognize shared risks in the most dangerous use cases?
The fact that the two strongest AI ecosystems are placing guardrails on the diplomatic negotiation table means
AI safety has officially moved from corporate policy
into international politics.
② Qwen-Image-2.1: 7B Parameters Combine Generation and Editing in One Model
The second news item takes place in a completely different setting.
On September 20, Qwen officially released
Qwen-Image-2.1,
a new image generation and editing model.
The headline figure is:
7B parameters,
specifically for the visual generation component.
The important aspect isn’t just the parameter count,
but that Qwen combines previously separate tasks into one model, including:
Text-to-Image
directly generating images from text,
Image Editing
modifying existing images,
Transparent Image Generation
creating images with alpha channels for transparent backgrounds,
Subject Extraction
isolating main subjects from photos,
and supporting
up to 10 reference images.
What can 10 reference images do?
For example, you might have:
a portrait photo,
a product,
a space,
several clothing and angle references.
Previously, models usually allowed only
one reference image,
or mixing multiple images could lose key subject or product features.
Qwen-Image-2.1 officially emphasizes
supporting up to
10 reference images
to assist image editing.
It also supports specifying edits on
circles,
painted annotations,
or masks to target particular areas only.
This shows AI image generation evolving from
"regenerating a similar-looking image"
to
"telling it where to change and leaving the rest intact."
Transparent backgrounds are very practical
AI image demos often showcase
beautiful portraits.
But in real work, designers often need
transparent PNG images.
Examples include:
product subjects,
cut-out people,
stickers,
website assets,
presentation illustrations,
and advertising composites.
If AI can directly generate
RGBA transparent images,
it eliminates an entire process of
generate
→ cutout
→ edge refinement
→ export
making this feature potentially more practical than simply offering another artistic style.
Qwen also launched a Prompt Rewriter
Another interesting feature released this time is a model designed to rewrite prompts for image generation and editing.
The goal is simple.
Often users input:
"Change this person's clothes to black."
Clear to a human,
but for image models, more complete instructions are usually needed for
subject,
position,
items to retain,
materials,
lighting,
and scope of changes.
The prompt rewriter transforms these short, vague requests
into instructions more suitable for image models to execute.
This highlights an important trend:
In the future, not everyone needs to be a prompt engineer.
An AI upfront translates user needs into
a format easier for the core AI to understand.
But “open weights” doesn’t mean free commercial use
It’s important to emphasize here.
Qwen officially calls Qwen-Image-2.1 open-source,
and allows downloading model weights.
However, it currently uses the
Qwen Research License Agreement.
The license explicitly states
commercial use requires separate permission from Qwen.
So don’t assume
"weights available for download"
means
free, unrestricted commercial use.
Downloading, researching, and self-deployment
are distinct from
permission for commercial use.
This distinction is increasingly important
because many equate open models with open-source software,
but licenses, redistribution, and commercial rights vary widely by model.
③ One AI Analyzes 146 Types of Abdominal CT Findings
The third news item enters healthcare.
Alibaba DAMO Academy and research teams from Zhejiang University School of Medicine’s First Affiliated Hospital
published in Science a system called
RADAR.
Conceptually, it’s
a generalist AI specialized in analyzing
contrast-enhanced abdominal CT scans.
Historically, many medical AIs focus on
one disease, one model.
For example, targeting a single tumor type,
or abnormalities in one organ.
RADAR takes a very different approach.
It aims for a single model to handle
18 anatomical structures
and
146 imaging findings simultaneously.
How was it trained?
RADAR was trained on over
400,000 contrast-enhanced abdominal CT exams
and
15 million anatomy-aware image-text pairs.
Uniquely, it doesn’t rely solely on manually labeling every CT scan.
The team used existing clinical
radiology reports
to help the model learn the relationship between images and doctors’ written findings.
This distinction is key.
Requiring large-scale manual labeling for each added disease makes generalist medical AI hard to scale.
Leveraging existing clinical reports
allows potentially much larger training datasets.
How good are the results?
The Science paper reports that RADAR maintains high diagnostic performance across multiple centers and clinical settings, covering 18 anatomical structures and 146 imaging findings.
A reader study involved
26 radiologists.
With RADAR assistance, their diagnostic sensitivity improved on average by
about 10%.
The most important takeaway is not
"AI beats doctors,"
but rather
whether AI combined with doctors outperforms doctors alone.
Why is sensitivity important?
Sensitivity can simply be understood as
the proportion of true abnormalities correctly detected among all those who have them.
A major risk in medical imaging is missing real lesions.
So if AI helps doctors catch findings they might have missed,
its value lies not in replacing doctors,
but in
serving as a second set of eyes.
Tested on Emergency CT scans too
Interestingly, the study also tested RADAR on over
27,000 emergency CT cases.
Even though the model wasn’t specifically trained on emergency data,
it still achieved an AUC of
0.904.
External validation across 8 centers showed a consistent AUC of about
0.895.
This addresses a critical concern:
Does the model generalize to other hospitals?
Medical AI’s biggest fear is high training performance fading on new hospital scanners, data, or patient populations.
Multi-center external validation is therefore more important than single-lab results.
But this does not mean "CT scans can now be fully self-diagnosed by AI"
It’s crucial to be clear.
The Science paper proves that RADAR shows strong capabilities in research and clinical data evaluations.
It doesn't mean
anyone can now take a CT and let RADAR fully replace radiologists.
For medical AI to enter hospitals, it must still navigate
clinical validation,
regulatory approval,
patient privacy,
hospital workflow,
liability,
and human oversight.
Even very high AUC scores
don’t guarantee correct diagnoses in every single case.
So the key insight isn’t
"AI can finally be a doctor,"
but rather
medical imaging AI is starting to move from one-disease-one-model to one model handling many clinical findings.
RADAR’s code and models are publicly released
The research team has released
RADAR’s code,
pre-trained checkpoints,
and related resources
to enable further testing by other groups.
This openness is important for medical AI,
since paper results alone don’t reveal whether performance persists on different data.
Allowing external teams to validate helps identify which capabilities genuinely generalize,
and which just perform well on original datasets.
All three news share a focus on “boundaries”
Putting them together reveals an interesting shift.
US-China AI Guardrails Talks
The question is:
Where are the boundaries at the national level?
What AI capabilities does everyone agree should not be misused?
Qwen-Image-2.1
The question is:
What are the boundaries around model usage?
Downloading weights doesn’t mean all commercial use is automatically allowed.
RADAR
The question is:
Where are the boundaries around AI decision-making?
Expert-level research performance doesn’t mean human involvement can be fully removed in medical decisions.
As AI develops to the next stage,
capability itself remains important,
but increasingly the hardest questions are no longer
"Can AI do it?"
but rather
"Having done this, who is responsible next?"
This is likely the daily reality as AI enters the real world.
Today, let's progress a bit with AI.
Learn one AI skill every day.
Save a little time daily.
Boost your abilities little by little.
SasaDaily, growing with you.
Recommended reading
The next race in AI isn’t about models but governance: Why this affects every AI user
