Super‑Exponential AI Timeline Explained
Key Points
- MER, a nonprofit model‑evaluation and threat‑research group, tracks how long AI agents can perform tasks compared to humans, using success‑rate thresholds (50 % and 80 %).
- Because the task‑relative metric has no upper limit, unlike fixed‑scope benchmarks, it reveals that AI progress is not merely exponential but super‑exponential.
- The latest Opus 4.5 results show AI achieving roughly five hours of human‑equivalent work at a 50 % success rate (and 2,728 min at 80 %), indicating a doubling of capability roughly every 4–4½ months.
- Projections suggest AI will handle 10 h of work by Q1, 20 h by mid‑year, and 40 h by year‑end, a self‑reinforcing flywheel that makes 2025 the “last normal year” and points to dramatically accelerated, AI‑driven development from 2026 onward.
Sections
- Super‑Exponential AI Timeline Explained - The speaker clarifies MER’s benchmark graph, showing how AI agents increasingly outperform human task times without a ceiling, indicating a super‑exponential growth trend in AI capabilities.
- Accelerating AI Task Delegation - The speaker argues that swiftly learning to identify and assign high‑quality, week‑long tasks to AI will create a power‑law advantage in a super‑exponential future, leaving late adopters far behind.
- Future Work: Managing AI Agents - The speaker predicts 2026 will require workers to relinquish traditional career models and become outcome‑obsessed leaders of delegated AI agents.
Full Transcript
# Super‑Exponential AI Timeline Explained **Source:** [https://www.youtube.com/watch?v=X_EJi6yCuTM&pp=ugUEEgJlbg%3D%3D](https://www.youtube.com/watch?v=X_EJi6yCuTM&pp=ugUEEgJlbg%3D%3D) **Duration:** 00:10:23 ## Summary - MER, a nonprofit model‑evaluation and threat‑research group, tracks how long AI agents can perform tasks compared to humans, using success‑rate thresholds (50 % and 80 %). - Because the task‑relative metric has no upper limit, unlike fixed‑scope benchmarks, it reveals that AI progress is not merely exponential but super‑exponential. - The latest Opus 4.5 results show AI achieving roughly five hours of human‑equivalent work at a 50 % success rate (and 2,728 min at 80 %), indicating a doubling of capability roughly every 4–4½ months. - Projections suggest AI will handle 10 h of work by Q1, 20 h by mid‑year, and 40 h by year‑end, a self‑reinforcing flywheel that makes 2025 the “last normal year” and points to dramatically accelerated, AI‑driven development from 2026 onward. ## Sections - [00:00:00](https://www.youtube.com/watch?v=X_EJi6yCuTM&pp=ugUEEgJlbg%3D%3D&t=0s) **Super‑Exponential AI Timeline Explained** - The speaker clarifies MER’s benchmark graph, showing how AI agents increasingly outperform human task times without a ceiling, indicating a super‑exponential growth trend in AI capabilities. - [00:03:20](https://www.youtube.com/watch?v=X_EJi6yCuTM&pp=ugUEEgJlbg%3D%3D&t=200s) **Accelerating AI Task Delegation** - The speaker argues that swiftly learning to identify and assign high‑quality, week‑long tasks to AI will create a power‑law advantage in a super‑exponential future, leaving late adopters far behind. - [00:07:00](https://www.youtube.com/watch?v=X_EJi6yCuTM&pp=ugUEEgJlbg%3D%3D&t=420s) **Future Work: Managing AI Agents** - The speaker predicts 2026 will require workers to relinquish traditional career models and become outcome‑obsessed leaders of delegated AI agents. ## Full Transcript
We are on the super exponential timeline
for AI agents and I want to explain what
that means and why it's super important
that we all pay attention to it. MER is
the model evaluation and threat research
company. It's a nonprofit. It's
dedicated to understanding how models
perform and they are famous for
producing a graph that shows how long
models can do useful agentic work for at
a time. It's a little bit of a confusing
graph to understand, so I'm going to
explain it really simply. Basically,
they take a task and they measure how
long a human takes to do that work task.
And then they want to find out if the AI
can do that task with at least a 50%
likelihood of success. Why 50%, because
they had to pick a number somewhere.
They also measure it at 80%. And we'll
get to that. PTR is important because it
does not top out. And so if you have a
lot of these these benchmarks like
Swebench is an engineering one, it tops
out at 100% and we're already in the
like way way up at the top it doesn't
matter like you can go from 91 to 93 and
you don't really get a sense of how the
models change. TR is different because
that graph has no top end. It can just
keep doing more and more work and that
allows it to show super exponential
progress. And one of the biggest debates
of 2025 was are we on an exponential
time scale with AI or are we on a super
exponential where it's increasing faster
than exponentially. It seems like we're
on the super exponential trend line. And
one of the things that made us think
that is this latest result from Opus 4.5
which shows over 4 hours 4 hours and 45
minutes almost 5 hours of human
equivalent work done at a 50% likelihood
of success. Now the 80% mark is also
measured and it is 2728 minutes for Opus
4.5 which you might think oh that's not
that far but keep in mind it was not
that long ago that we were 1 minute 2
minute 10 minute 30 minutes and now
we're up to almost 5 hours and that is
the point of a super exponential curve.
We are on a doubling rate every 4 to 4
and 1/2 months right now. And so if the
number is 50% complete, but the time
horizon is four almost 5 hours, we're
going to be at 10 hours by the end of
Q1, we'll be at 20 hours by the end of
Q2 into Q3, and we may be at 40 hours by
the end of the year or past. And that is
why we have to pay attention to this.
Super exponential gains suggest that we
have hit a selfreinforcing
flywheel with AI. And that is indeed
what we hear out of model makers and
that is why 2025 was the last normal
year. We are going to see really really
weird progress from AI in 2026 and every
year after because AI itself is starting
to reinforce AI systems. We're bringing
AI in to help train AI systems. Now that
is going to become more and more
automated. We are going to have
capabilities that AI itself helps to
grow speeding up the whole process and
all of that is going to allow us to
continue to make progress on these tough
tasks that don't have an upper limit.
And this matters because really our
ability to do meaningful work is going
to be determined by whether or not we
can define useful high taste highquality
work that an AI can do over a period of
time. Do you have something for an AI
that would take you a week to do? Maybe
it's your taxes. I don't know. But
that's going to increasingly become the
question. And if you don't, then the
question is going to be what does it
take for you to get there? What does it
take for you to gain the skill to assign
that work? Because in a super
exponential world, the skill we need to
learn is also super exponential. The
people who figure out how to assign
agents work now in January and February
and March are going to have a much
easier time learning how to continue
assigning agents work when the agents
can do much harder stuff. Whereas if you
wait and say, "I'm going to catch up.
I've scheduled this for Q2 or Q3 next
year. That's my AI quarter." Good luck
with that. Like it doesn't work that
way. There will be people who are
running circles around you because they
can assign their agents a week's worth
of work. And once you can assign your
agents a week's worth of work and spend
up two or three of them, look at how
much more productive that makes you.
You're going to be running circles
around people. And that that is the
power law distribution world we're going
to live in. Super exponentials create
power laws. So power law is that the
idea that the world we live in is not
normally distributed. A normally
distributed world, most people are on
the average, a few people are on the
tails. Einstein's way over here, right?
But in a power law world, just a few
people are going to be able to do a
tremendous, tremendous amount. And it's
not because they're necessarily going to
have lots of money to do it. It's
because they have the skills to do it.
AI is going to disproportionately
reward skill development where it's
related to artificial intelligence and
everything else. People are going to
start to lose traction. If you are
looking to make a dent in your career, I
would look less in 2026 at your job
family's traditional requirements and
look more at where can an agent do a
meaningful amount of work for a week in
this traditional job family area and how
can I make sure I set myself up so I
know how to define and assign that work,
know how to hold it accountable, know
how to put good taste down so I know
what excellent looks like in that work,
know how to intervene, keep the agent on
track, have the technical foundations
necessary to define and set up an
agentic system. This is going to become
more and more relevant for all of us.
The technical skill sets are going to
spread across the job families. The
non-technical skill sets are also going
to spread across the job families.
Engineers who traditionally just had to
do code are going to have to have some
business fluency and customer fluency
now because they have to be the ones
with good taste when they're
architecting systems. And frankly, they
now have to architect systems that
non-technical people can contribute code
to. So just that one shift, that ability
of agents to do work over time is going
to multiply the impacts across all of
the rest of us. Having agents that work
longer means all of our jobs are going
to change forever. And you might think
I'm like a hype person. This is not
being me being hypy. This is me just
talking about the reality that we are on
a super exponential curve. Humans are
bad at estimating super exponential
curves. And so I just want to make it
really concrete for you. I do think
there is no way that work will not
change for everybody if we are in a
place where it's 5 hours and doubling
every four months. Because you look at
it by April you're going to be at 10
hours. By what July September you're
going to be at 20 hours by December
you're going to be at 40 hours maybe.
Right? Maybe it's not even like it's
just it's going to be crazy. Are you
able to delegate a week's worth of work?
That is that is the question of 2026. We
will all have to let go of a lot. We
will have to let go of our traditional
understandings about career progression.
We'll have to let go of our traditional
understandings about job families. What
job families know and what they don't.
We are going to have to be outcome
obsessed, ownership obsessed. The work
of the future is going to reward people
who are ownership and outcome obsessed
because that's where human value shows
up. It's when we make sure that what's
made is actually relevant for people, is
actually useful, is actually good. It's
not just vibecoded slop. There will be
lots of vibecoded slop. In fact, I would
expect it to 100x in 2026 because you
can ask your agent to do a lot of
terrible, terrible things. It's going to
be up to you to decide that the agent's
work is worth it. That you are assigning
the agent and the agent is doing good
work to get meaningful work done that
compounds over time. The strategy
rewards used to acrue to leaders.
Strategy is now an individual thing
because you are effectively a strategic
manager of a team of agents or you will
be in 2026. You can make them yourself.
There will probably be startups that
market them to you. But either way, you
will end up with a team of agents
working for you. Do you know how to
manage them? Do you know how to lead
them? Do you know how to drive them to
develop compounding advantage over time?
That used to be a question for directors
and above. It's not for if now it's for
everybody. Everyone will need to be able
to do this and the people who can are
going to look like they can do anything
like that. The span is going to be
incredible because they're able to
leverage their own domain expertise and
expand their scope of impact from there.
I do not mean that you can do anything
that requires deep domain expertise that
you do not have. There are still going
to be real value that you can't get to.
there's going to be real value you can't
get to just by adding agents. So, for
example, if you are a lawyer and you
have decades of experience, agents are
going to transform the legal profession
and how you work, but it it's not going
to transform it to the point where I, as
a non-awyer, can come in and do work for
a white shoe law firm and get exactly
the same quality of work done at the end
as the lawyer who's got decades of
experience. there's going to be a reward
for understanding the business deeply
that will show up in your ability to
direct AI agents toward useful ends. And
so as much as it may seem like I'm
saying the agent can do work, we won't
do any. What I'm really saying is our
domain expertise is worth more and more,
but boy do we have to be smart and
leverage it really, really differently
to get where we need to go in 2026. And
that's going to change all of our skill
sets. We're all going to have to learn
together. We've never gone through this
workflow and workforce transformation
before. So, we're all going to have to
just jump in and figure out how to do it
together. But, I do think it's real. I
do think it's coming. And I do think the
key is that super exponential graph.
Opus 4.5 was just the latest getting to
5 hours. It won't be the last. It's not
like Claude has a special, you know,
Claude doesn't have a special monopoly
on this, right? We're going to see this
from Gemini. We're going to see this
from Chad GPT. We'll see it from other
model makers as well. We will continue
to see exponential gains from agent
working time in 2026 and that will
change the way all of us have to do our
work.