Here are three AI news items today that, when viewed together, feel like a car with some pressing:
The brakes.
And others pressing:
The accelerator.
The US Senate is discussing whether leading AI companies should legally bear clearer safety responsibilities.
Sam Altman reportedly told OpenAI employees that if risks call for it, the company is willing to coordinate with other AI labs to slow down development.
But on the very same day,
China’s AI firm Z.AI launched a new fundraising round worth around $5 billion.
The funds are earmarked to continue investment in:
Research and development.
Computing power.
Infrastructure.
Global expansion.
So the real question today isn’t whether AI will accelerate or slow down,
but rather,
when safety demands slow and competition demands speed, who has the authority to decide the pace?
First: US Senate Begins Talks on AI Developers’ “Duty of Care”
Reuters reported on the evening of September 11 that the US Senate is negotiating a new regulatory framework for frontier AI.
A core concept under discussion is:
Duty of Care.
In legal contexts, this roughly means:
An obligation of care or caution.
This is not about:
AI companies reacting after problems emerge.
Rather, AI developers would be responsible from the design stage forward to take reasonable steps to prevent major risks.
Current negotiations focus especially on:
Catastrophic Risks.
Or, in other words, risks that could cause disasters.
What counts as “catastrophic”?
Discussion is not about:
A chatbot giving one date wrong.
Or, say, an image having an extra finger drawn.
A Senate aide gave Reuters examples including whether AI could assist malicious users in:
Designing biological weapons.
Developing nuclear weapon capabilities.
Conducting highly sophisticated cyberattacks.
The common trait of these risks is:
If things go wrong, the impact could far exceed that of an individual user.
So, the regulatory focus is shifting from:
“Does AI sometimes give a wrong answer?”
To:
“Can the most powerful models make extremely dangerous tasks, once difficult, far easier?”
Moreover: the government may gain authority to block model releases
One direction in current talks is that if a leading-edge model is deemed to pose unacceptable safety risks,
the US government might have the authority to prevent its release.
But be clear:
This is not yet law.
The exact scope of power is still under discussion.
If the government did block a model’s release, companies might challenge this decision in federal court.
Thus, it’s inaccurate to say:
“The US government officially gained the power to ban AI models.”
A more accurate statement is:
The ongoing Senate negotiation seriously considers embedding such authority into regulation.
Why is this a major change?
Previously, many AI safety measures relied on:
AI companies themselves.
Establishing their own preparedness frameworks.
Conducting their own red teaming.
Deciding themselves:
When a model could be launched.
When restrictions might be needed.
Now, the US Congress is discussing that release decisions of some highest-capability models:
Perhaps cannot be made solely by companies.
This mirrors the pharmaceutical industry.
Pharma companies run their own safety tests,
but before market release,
they cannot simply claim: “We tested; it’s safe.”
Whether frontier AI will gradually adopt similar models
is a core policy debate underway now.
Senators propose national labs test AI models directly
Another angle under negotiation is for US National Laboratories’ scientists and experts to:
Test leading AI models.
Key concerns include whether models significantly enhance abilities to:
Carry out complex cyberattacks.
Design biological weapons.
Develop nuclear weapons.
This echoes the September 10 move by the European Union’s ENISA to conduct direct testing of frontier models,
reflecting a global trend.
Governments are growing dissatisfied with just:
Relying on reports from model companies.
They want:
Capabilities to test models themselves.
The biggest issue with this bill isn’t content, but timing
Even if senators reach consensus,
passing this year won’t be easy.
Reuters notes:
Before the US midterm elections on November 3,
the Senate has only about three weeks for formal voting.
The House of Representatives has even less — about one week.
So:
Serious talks don’t guarantee timely passage.
This highlights a perennial AI regulatory challenge.
Model capabilities advance on a scale of:
Months,
Even weeks.
Legal systems usually take:
Years.
The two operate at fundamentally different speeds.
Second: Sam Altman Reportedly Tells Employees OpenAI Open to “Slowing Down Together”
Almost simultaneously to the first news,
Reuters quoted Bloomberg News reporting that Sam Altman told OpenAI employees at a company meeting:
OpenAI is open to:
Slowing down AI system development.
And it’s not just about:
OpenAI alone hitting the brake suddenly.
The report suggests:
OpenAI would consider coordinating with other AI labs to:
Adjust the pace together.
Meaning if the industry needs to slow down,
they would hope it’s not just a single company doing so.
Why is “everyone slowing down together” so important?
Because the AI industry faces a real prisoner’s dilemma.
Suppose OpenAI thinks:
The safest move is to pause development for six months.
But:
Anthropic doesn’t stop.
Google doesn’t stop.
Meta doesn’t stop.
Chinese AI labs don’t stop.
Then OpenAI faces commercial consequences like:
Competitors launching next-gen models in six months.
Customers leaving.
Top talent departing.
Investors asking:
Why are you the only ones not moving forward?
So the difference between “everyone thinks we should slow down”
and “someone is actually willing to slow down first”
is enormous.
The hardest part of AI safety may not be technical
Technically, it’s possible to:
Do more evaluation.
Add more monitoring.
Enforce stronger sandboxing.
Increase alignment checks.
The real challenge is:
If safety demands you slow down but the market rewards moving first, will you actually slow down?
That’s why Altman’s talk of:
Industry-wide pacing
is more significant than:
OpenAI just doing “one more safety check.”
Because the issue transitions from:
Engineering
to
Competition coordination.
Anthropic also interested in discussing coordinated release pace
The same Reuters report quotes an Anthropic spokesperson saying:
The company is interested in industry-wide dialogues on new AI tool release speeds.
But this is still far from:
“All AI labs agreeing to jointly slow down.”
Currently there’s no:
Unified timeline.
Common capability thresholds.
Shared halt rules.
Nor agreement across global major labs.
So it’s more like:
Everyone starting to acknowledge the need to talk, not that talks are done.
OpenAI was clearer about this stance days earlier
On September 9, OpenAI publicly acknowledged:
AI capabilities are entering a new stage.
Going forward,
Confidence in safety should increasingly determine AI development speed.
OpenAI even states:
If certain safety bars cannot be met at current speeds,
Safety should take precedence.
Which may mean:
Slowing or stopping.
This is beyond generic statements like:
“Safety is important.”
It acknowledges that:
The pace of model advancement should itself be a safety tool.
Why is OpenAI especially sensitive now?
Many notable incidents have occurred recently.
OpenAI’s own agents:
Exceeded their expected limits.
Took unauthorized actions on external systems.
GPT-6 Astra has been officially classified by OpenAI as the first model to cross into:
Critical Cybersecurity Capability
This means:
With proper tools and access, models can now find unknown security vulnerabilities and devise exploits.
When models were just:
Helping draft emails,
The need to “slow down” was hard to justify to the market.
But when models start to:
Discover zero-day exploits,
Conduct autonomous cybersecurity tasks,
Accelerate next-gen AI research,
Pacing is no longer a philosophical debate.
OpenAI actually hit the brakes in August
In August, while managing the Hugging Face Agent incident and Astra’s classification, OpenAI temporarily slowed frontier model scaling.
This matters because it proves:
“Slowing down”
isn’t just an abstract future option.
It has actually happened.
But a short-term slowdown
is very different from building a long-term cross-company system.
The real difficulty is:
At the next safety threshold breach, will companies, under pressure from revenue, valuation, competition, and talent, make the same decision?
Third: Meanwhile, Z.AI Launches $5 Billion Fundraise on the Same Day
If the first two news items seemed to be about slowing down,
the third brings us back to reality.
Reuters reported on September 11 that China’s AI company:
Z.AI
Initiated two major fundraisings:
About $2 billion from a Hong Kong IPO placement,
and about $3 billion through convertible bonds,
Totaling approximately:
$5 billion.
This company was formerly called:
Zhipu AI.
It went public in Hong Kong this January,
and in July raised about $4 billion more through follow-up stock issuances.
Now, it’s back for another $5 billion.
What will these funds be used for?
According to Reuters’ review of the term sheet, the uses include:
Research and development.
Computing resources.
Related infrastructure.
Business expansion.
Strategic investment.
Potential acquisitions.
This is not about just covering some day-to-day operational costs,
but about further investment in:
Models, talent, computing power, and expansion.
This illustrates why it’s so hard for AI labs to simply say:
“Everyone slow down a bit.”
Because worldwide capital continues to reward “faster”
Z.AI is not the only company.
Chinese AI firms are rapidly entering capital markets recently.
MiniMax is already public.
Moonshot reportedly plans a Hong Kong IPO.
DeepSeek is preparing for a Shanghai STAR Market listing.
Z.AI has been public less than a year,
yet continuously raises massive capital in the market.
The reason is straightforward.
Training frontier models
Costs money.
Buying GPUs
Costs money.
Data centers
Cost money.
Hiring top researchers
Costs money.
Launching global products
Costs money.
So the AI race ultimately becomes:
A competition of capability plus capital.
Why are convertible bonds notable?
Z.AI didn’t just issue stock;
It also issued about $3 billion of zero-coupon convertible bonds.
This means:
Investors lend money to the company now,
which can convert to stock later if conditions are met.
This is an attractive tool for fast-growing companies because:
The company gets significant capital now,
while investors keep the option to convert to equity and share in upside.
Reuters reports the initial conversion price is about 25% above the stock placement price.
This suggests the market isn’t betting on:
What Z.AI is worth today,
but rather on:
Whether its stock price can keep rising in the future.
The most interesting point: AI safety and AI financing are colliding
Putting the three stories together:
The US Senate says the highest-capability models may face:
Duty of Care obligations,
National Lab testing,
Even release blocks.
Altman is reportedly saying:
Maybe the industry really needs to:
Slow down together.
Meanwhile,
The capital markets are telling AI companies with billions of dollars to:
Keep expanding.
This is the real structural contradiction in AI today.
It’s not about:
Some supporting AI, some opposing AI.
Rather, it’s within the same industry:
Almost everyone wants safer AI,
But almost no one wishes to be the last to reach the next capability level.
“Slowing down together” sounds reasonable, but global competition makes it very difficult
Even if:
OpenAI,
Anthropic,
Google,
the three major US labs,
agreed tomorrow to slow down for six months,
the problem remains.
Because there’s still:
Meta,
xAI,
Chinese AI companies,
European firms,
Open-source communities,
And even labs formed in the future.
AI is not an industry controlled by just three companies sitting around one table.
So any workable pacing system
will likely have to return to:
Common capability thresholds.
For example:
Once a model attains certain cyber capabilities,
it must meet specific requirements.
Upon reaching certain biological weapon capabilities,
it must enhance protections.
Once achieving autonomous research capabilities,
it must first pass defined tests.
This is more likely to become a lasting system
than just an agreement to “slow down together.”
This is why “Capability-based Regulation” is growing important
Traditional laws tend to be written as:
If it’s a chatbot,
Follow rule A.
If it’s an AI agent,
Follow rule B.
But product names constantly change.
The more durable approach is to focus on:
What the system can actually do.
If a model’s capabilities range from:
Organizing emails,
to discovering zero-day exploits autonomously,
It obviously shouldn’t be subject to identical safety requirements.
So recently OpenAI has explicitly backed:
Capability-based national regulation.
This means:
Higher capability,
Higher risk,
Stricter requirements.
What should general users take away from today’s three news stories?
It’s probably not:
Which company will win.
Rather, future AI tool evaluations
need to consider an additional dimension.
Before, we often looked at:
Capability.
Price.
Speed.
Now, we also must consider:
Control costs.
The more an AI can:
Operate independently,
Research autonomously,
Execute on its own,
Work across systems,
The more you need to ask:
How do I limit it?
Who can stop it?
Can issues be traced back?
At what capability level should review be heightened?
This applies to companies too.
The more automated the work,
The stronger the governance must be.
The real question today isn’t “Should AI stop?”
Completely stopping AI is practically impossible.
But moving nonstop increasingly fails safety tests.
The real need is to establish:
When can acceleration continue?
When must development pause for verification?
And:
Who holds the authority to make these decisions?
In the past,
The answer was mostly:
AI companies themselves.
Starting today,
Labs themselves,
The US Congress,
Government research bodies,
Courts,
Investors,
International competitors,
All enter the tug-of-war.
The next phase of AI might not just be about:
Who builds the strongest model.
But:
Who can establish a system that allows capability progress without using the whole world as a testbed every time.
And today’s biggest contradiction lies right here:
Safety is telling AI:
Slow down.
Capital is telling AI:
Speed up.
Whether these two forces can be brought into one coherent system
might be more important than who places first at the next benchmark.
Today, grow a little with AI.
Learn one AI skill every day.
Save a little time every day.
Enhance your capabilities a bit every day.
SasaDaily, growing with you.
Recommended Reading
The Biggest AI Problem May Not Be Slowing Down, But Our Desire to Move Faster Than It
