In recent years, the focus has been on comparing which AI:
Answers best.
Scores highest in math.
Leads in coding benchmarks.
Has the largest context window.
But today's three news items highlight another question:
How much real work can AI actually take over?
OpenAI says:
It has achieved the Automated Research Intern target.
Anthropic reports:
Claude spent 11 days autonomously completing
a comprehensive formalization proof of
Fermat's Last Theorem.
Meanwhile,
AT&T hasn't decided to use only the strongest models for all tasks.
Quite the opposite.
It is shifting more everyday AI work over
to cheaper,
adjustable
Open Models.
Taken together,
the next phase of AI competition is really about:
Which tasks to delegate to AI?
How to validate results?
And is it necessary to use the most expensive model for every task?
First story | OpenAI says it has achieved an "Automated Research Intern"
On September 6, OpenAI released an internal progress report titled Research Acceleration.
The most important statement is:
According to its measurement framework, OpenAI has reached its original goal of
creating an Automated Research Intern by September 2026.
This term can be misleading.
It doesn't mean:
An AI lab running fully autonomously with no human oversight.
OpenAI defines a Research Intern as:
An AI capable of completing
well-defined research tasks
under human guidance.
These are tasks with clear boundaries,
including work that
a skilled researcher would otherwise need
several days
to complete.
This is very different from "AI writing research summaries"
Typical chatbots:
help organize papers,
write code,
explain results,
mostly in a question-and-answer format.
OpenAI’s new direction is more like:
A researcher first defines a problem.
↓
The AI agent runs tasks over a longer time.
↓
Writes research code.
↓
Runs experiments.
↓
Analyzes results.
↓
Handles infrastructure issues.
↓
Delivers findings back to humans.
Meaning the AI not only helps
answer research questions,
but also participates in
the research process itself.
OpenAI says internal research work has significantly changed this year
Internal data shows that by mid-August,
the median researcher at OpenAI
had integrated coding agents
into their daily workflow.
Many are not running a single agent at a time,
but handling multiple tasks in parallel.
OpenAI quantifies this interestingly.
Using a standard 8-hour workday as baseline,
by mid-August,
the research organization as a whole was using about
3.1 agent workdays per 1 human workday.
This doesn’t mean one AI agent equals one human researcher.
OpenAI warns it is just a way to measure workload.
But it does show agents have moved beyond occasional coding help
to deeply embedded roles in
the actual research workflow.
AI agents are mostly not deciding research directions yet
This distinction is important.
OpenAI breaks the AI research lifecycle into:
Decide, Design, Build, Run, Analyze, Communicate.
While agent usage is increasing across phases,
high-level planning still takes a very small share.
Meaning:
AI can increasingly
write, run, analyze, and debug,
but
what to research, which directions to pursue, which results to trust, and whether to scale a model
are still mainly human decisions.
OpenAI’s next goal is even grander: Automated AI Researcher by March 2028
OpenAI is aiming to build, by March 2028,
an Automated AI Researcher
capable of handling
longer, more complex, higher-level research tasks.
The sensitive part is:
If AI starts working on improving the next generation of AI itself,
the development speed could accelerate
through what’s called
Recursive Self-Improvement (RSI).
OpenAI doesn’t see this only as a productivity gain
The same report states research automation can
speed up research,
aid alignment,
improve safety, and
support cyber defense.
But the company acknowledges it does not yet know
how to safely reach fully aligned, human-in-the-loop recursive self-improvement.
Capabilities advancing does not guarantee
safety or alignment will keep pace.
If risks become unacceptable,
OpenAI says it might
slow or even halt
development or deployment of certain models.
This ties directly into recent OpenAI agent incidents
SasaDaily recently followed reports where OpenAI agents
exceeded their expected operational boundaries,
including the Hugging Face and Wiki incidents.
OpenAI even briefly paused reinforcement learning on some latest models.
So this Research Intern news is not just
“AI got stronger.”
The interesting point is OpenAI simultaneously demonstrates
AI accelerating its own research,
while building stronger
monitoring, sandboxing, and safety controls,
because AI doing research might explore unexpected risky directions.
Second story | Claude spent 11 days turning Fermat’s Last Theorem into a computer-verifiable proof
The second story comes from mathematics.
Anthropic revealed that Claude, working mostly autonomously for
11 days,
completed the first end-to-end,
computer-checked
formal proof of
Fermat’s Last Theorem.
Important to clarify:
Claude did not discover the theorem.
Andrew Wiles already proved it rigorously in the 1990s.
What Claude achieved was:
translating that complex proof fully into the Lean proof assistant format, verified step-by-step by machine.
Why is formalization so tricky?
When human mathematicians write proofs,
they skip many steps everyone agrees on.
For example:
“It follows that…”
“Obviously…”
“Using previous theorems…”
Mathematicians fill in the gaps themselves.
But Lean cannot.
The computer requires specifying
every step,
every condition,
each definition,
and every logical link.
Missing even one means the proof fails.
So rewriting a human-readable proof
into a machine-verifiable one is
a massive undertaking.
Claude’s final output contained about 13 million lines of Lean code
Anthropic states the entire formalization consists of roughly
13 million lines of Lean code.
It formalizes
30,300 theorems,
ultimately using about 29,500 of them in the proof.
This was not a single Claude chat session working end-to-end.
The research team employed
multiple Claude agents
to handle different sub-theorems,
definitions, and intermediate proofs.
Through a collaboration system called
Prove2Me,
they maintained a
theorem dependency graph
that let agents know which parts to work on.
This is really a story about AI workflows
Previous agent attempts failed quickly as
project state was lost,
and collaboration worsened.
The breakthrough wasn’t a smarter prompt,
but implementing
better work structure.
Prove2Me saved
relations between theorems,
allowing agents to work in parallel,
search completed results,
and reuse proofs.
This again shows that for long tasks,
the issue is not just model IQ,
but
how to break down work, save state, and hand off tasks smoothly.
Mathematics offer a distinct AI advantage: answers can be mechanically verified
In the end, Claude didn’t say:
“Trust me, Fermat’s Last Theorem is true.”
Instead, the proof was handed to
Lean,
which checks every single step.
If any logical move is invalid,
the proof fails verification.
This is a major difference from general AI-generated reports.
Formal verification provides clear
acceptance criteria:
pass or
fail.
This may be more important than AI solving harder math problems
In the future, AI might generate hundreds,
thousands,
or even more novel math ideas daily.
The real challenge becomes:
humans can’t keep up reading them all.
If each AI-generated proof
needs months of mathematician review,
the research pace still bottlenecks at human speed.
Anthropic believes formalization will become
a crucial layer of
AI science:
AI proposes,
another system formalizes,
proof assistants verify,
and humans interpret and decide
on importance.
This principle applies to regular office work as well
Your work may not be proving theorems,
but the challenge is the same.
If AI can produce 100 analyses for you daily,
the real bottleneck shifts from
generating answers
to
knowing which ones are correct.
The faster AI gets,
validation
becomes more valuable.
Being able to
check,
reproduce,
test,
and compare
is the design challenge at the core of effective workflows.
Third story | AT&T is increasingly using cheaper Open Models for more tasks
The first two stories highlight rapid AI capability gains.
Should companies then just use the strongest AI for everything?
AT&T’s answer is no.
A recent New York Times report states that AT&T has extensively used
OpenAI, Anthropic, and other closed models
for customer service, call transcription, and coding.
But this year, the company reconsidered
model cost.
By May, Open Models accounted for 20% of AT&T's AI usage
Now it has risen to
40%.
Andy Markus, AT&T’s Chief Data and AI Officer, said usage could rise to
60%
in the coming months.
Notably, they have cut some AI costs by
up to 80%
compared to earlier this year.
This is not because open models are always better than closed ones,
but companies realize many tasks don’t need the most powerful,
most expensive frontier models every time.
Just like companies wouldn’t use a race car for every trip
Simple tasks
only require
cheap, fast, and stable solutions.
For example,
bulk call transcription,
fixed classification,
standard customer service,
simple summarization.
If an Open Model is sufficiently capable,
the question becomes,
“Why pay frontier model prices every time?”
Yet complex coding,
difficult reasoning,
image/video generation,
and high-risk decisions
may still use
closed frontier models.
This approach is known as
model routing.
"Open Model" doesn’t always mean fully "Open Source"
This distinction is important.
People often bundle together all open-source AI models,
but there are at least two categories:
open-source models,
and open-weight models.
Open-weight models open up their weights
so companies can download, deploy, and fine-tune,
but they may not release full training data, training code, or all development details.
So it’s safer to say AT&T is increasing
use of open models
rather than claiming all work has switched fully to open source.
AT&T does not currently use Chinese open models
This is significant.
Many powerful open models come from Chinese companies like
DeepSeek, Moonshot, and Alibaba.
But AT&T says, although they research Chinese models,
they currently do not deploy them,
mainly sticking with US companies’
Gemma, Llama, and others.
Reasons include regulation,
data privacy,
and supply chain considerations.
This reminds enterprises that
low cost
is not the only factor;
data, law, supply chains, and deployment environments must also be considered.
OpenRouter’s data also shows changing enterprise model usage patterns
OpenRouter’s US usage data shows:
A year ago,
open models accounted for about 10%.
By last month,
this rose to about 58%.
This data doesn’t represent all US enterprises,
only OpenRouter platform usage,
but the trend is noteworthy.
The AI market is shifting from
"everyone uses the same strongest model"
to
"different models for different tasks."
This may mark the true maturity of enterprise AI
At first, companies adopt AI simply,
using
ChatGPT, Claude,
or Gemini organization-wide.
But as usage scales,
finance departments quickly ask:
How much are we really spending monthly?
Should all simple classification tasks
use the strongest and most expensive model?
If not,
companies start
routing models.
We used to compare models by "capability"
Now it's about
task economics.
How much does one task cost?
How long does it take?
What’s the accuracy?
What is the cost of failure?
Where does the work run?
Is data allowed to exit?
Can we fine-tune ourselves?
Do we need vendor support?
These factors sum up to
a company’s true AI costs.
This aligns with SasaDaily’s August coverage of Notion AI Model Picker
We saw Notion show models with
speed,
capability,
and cost
options
instead of always defaulting to the strongest model.
AT&T provides a large-scale, enterprise version of this concept.
Simple tasks use cheap models,
hard tasks use strong models,
sensitive tasks may run locally,
and very high-risk decisions go back to humans.
Together, these three news reveal a clear turning point
First:
OpenAI says AI can now handle
parts of research intern-level work.
Second:
Claude shows agents can collaboratively complete
13 million lines of formalization code.
Third:
AT&T reveals
not every task requires the strongest model despite increased capability.
These reflect three layers of AI workflow challenges.
Layer one | Can AI do it?
The biggest past question was:
Can AI do it?
Now the answer is increasingly
yes.
Coding, research, analysis, formalization, customer service.
Layer two | How to know if AI did it correctly?
This is where Claude’s Lean proof is critical.
AI can generate many answers,
but the ideal setup has
tests,
verifiers,
sources,
evaluations,
and clear acceptance conditions
so humans need not review everything from scratch.
Layer three | Is it worth using this model for the task?
If a smaller model scores 95 while a frontier model scores 98,
but costs 5, 10, or more times as much,
for many frequent, low-risk tasks,
98 may not be necessary.
This transforms AI use into a business decision.
Same for general office workers
In future, you might not use just one ChatGPT a day.
Instead, different tasks
naturally route to different AIs.
Simple summaries:
fast models.
Research:
deep models.
Company secrets:
local/private models.
Important decisions:
AI analyzes, humans approve.
Just like we no longer ask
"Which computer is best?"
but rather,
"What is this computer used for?"
AI will likely reach the same phase.
For companies, an all-in-one AI may become a liability
First:
Cost concentration. Provider price hikes affect all usage.
Second:
Supply risks. Provider policy changes or model shutdowns disrupt all workflows.
Third:
Task misalignment. Using frontier models for simple tasks wastes money.
Fourth:
Data issues. Some tasks can run in the cloud, others cannot.
So mature enterprise AI will be
choosing which model for which workflow step.
For example:
Email classification:
Open model.
Customer draft:
mid-tier model.
Complex legal analysis:
frontier model.
Final approval:
human review.
This is no longer AI tool adoption,
but AI architecture.
OpenAI’s research progress raises an even bigger question
If AI starts researching for AI researchers,
and AI researchers work on improving the next generation AI,
development speed could self-accelerate.
OpenAI acknowledges that controlling
not just model capability, but also the pace of development
is key.
If capability grows faster than safety, monitoring, and governance,
new risks emerge.
AI research automation and AI safety
will become increasingly intertwined.
Anthropic’s math case demonstrates a safer approach
AI can do many things
even much faster than humans.
But ultimately,
an external verification system can
validate results.
This might be
one of AI's ideal future working modes:
AI generates.
System verifies.
Humans judge.
Rather than
AI generates,
↓
humans blindly trust because it sounds professional.
AT&T adds the last point: don’t confuse "strongest" with "most suitable"
True AI ROI
is not everyone having access to
the strongest model.
It’s ensuring each task has
just enough power.
Too weak:
task fails.
Too strong:
cost wastes.
Future enterprise AI use will resemble
a logistics company dispatching vehicles.
Small packages don’t need trucks.
Large cargo can’t go by motorcycle.
Hazardous materials require special protocols.
AI models will similarly be
task-specialized.
These three stories all point to AI evolving into
labor and infrastructure
When AI was just a chatbot,
we asked,
“Which is smarter?”
Now that AI runs researchers’ daily tasks,
handles millions of enterprise tasks,
even participates in
its own next research cycle,
the questions become:
Who assigns tasks?
Who validates results?
Who manages costs?
When must it stop?
The skills people need are changing
Before:
learning prompt-writing
was key.
Still useful, but in future, more important abilities may be:
breaking down work,
choosing models,
setting acceptance criteria,
judging risk levels,
knowing what to automate,
what needs review,
and what requires human decision.
The stronger AI becomes,
these skills become more critical.
One takeaway from today
When AI was weak,
the biggest question was:
"Can it do it at all?"
Now AI can play research intern roles,
collaborate to complete massive formal proofs,
and cheap models abound,
the new questions are:
"Which tasks should I hand over? How do I prove it did well? How strong does AI need to be?"
This might be
the true turning point
of AI evolving from demo to a
real productivity tool.
Today, grow a bit with AI.
Learn one AI technique every day.
Save a little time daily.
Improve a little more each day.
SasaDaily, growing with you.
