Brett Winton
@wintonARK
Chief Futurist @ARKInvest. ARK Venture IC. Welcome to the Great Acceleration.
595 Following 210.6K Followers
Narrative: “open weight models are going to steal spend from frontier labs”
Reality: people using open weight models to try to steal from and exploit enterprises are going to force companies to continually spend at the frontier to protect themselves.
The capability increase in open weight models will accelerate frontier lab uptake and adoption across enterprises.
Infrastructure without a frontier defense force will fall to the swarms of mercenaries and rogues traversing the open internet.
Aggressively adopt AI or die.
Show more
Big Ideas '22, we asserted that Foundation model companies could command $10s of trillions in enterprise value.
This was based on rough expectations of $2 trillion in revenue for foundation model companies in '30.
Looks roughly on track (though they'll need to slow down!)
Show more
Succinctly captures the business model evolution likely to follow from deep learning breakthroughs.
Movement-in-world models should follow a similar evolution (and power all kinds of robots)
We think these foundation models will accrue 10s in trillions in enterprise value
Show more
AGI for everyone == UBI
Money distilled represents a call on future labor. The dollar you hold today is a deferred amount of future labor you can command (dwindling annually due to inflation.)
Agentic AI systems provide accessible labor at user command (for continuously declining amounts of energy input.)
Broadly accessible, indirectly monetized AI systems should become sufficiently powerful that they can provide for any person’s basic needs.
AGI is the UBI we’ve been looking for.
Show more
non-ironically, the solution to grade inflation is to periodically introduce a new even higher highest grade at the top of the scale
The Luna pricing change and how it changed our expected performance improvement rate.
Publish a chart.
3 days later forced to update because one of the models at the leading edge dropped prices by 80%.
"no you don't understand.
people aren't willing to pay more for more intelligence at this point.
intelligence is like a sphere, and to make the models more intelligent is to make it more spherical, and the sphere is obviously already sphere-like enough."
the sphere:
Show more
@sama Financial markets volatility📈
a vastly underrated characteristic seemingly common across all of @elonmusk enterprises:
product capabilities continually improve even when they lie askance to the primary vector of improvement for the company;
yes, this improvement helps the user experience, will marginally help robotaxi economics, creates a better product,
but in no universe does it immediately or materially move the needle on model Y sales.
it does mean that Tesla owners have a simply better product.
Similarly, starlink is basically the only game in town, and yet its packaging is carefully considered, set up is near instantaneous and user-friendly; these are all things they didn't "need" to do.
There is an attention to rate of improvement at the edges of the entire product portfolio that is a manifestation of corporate structure and culture, and given the breadth of products being worked, seems wildly ahistorical.
Perhaps it is a function of the ruthless part and process elimination he subjects the organizations to. Very very very good design is the natural end-point of part and process elimination.
That Tesla kept taking weight out of the Model X, even after its last and final refresh, is another example, and something no other auto company in history would consider.
For the users of his company's products, the impact is subtle and remarkable and compounds.
It also, over time, should compound into accelerating economic performance (of which I suspect we will see plenty of evidence over the next couple of years.)
Though I think it's fair to say that much of Wall Street does not care for these sorts of niceties, the diffusion curve is long, but it bends towards compounding superiority.
Show more
Some of the thermal management improvements in the new Model Y L Premium:
• Up to 15% faster cabin cooling
• Up to 15 min faster cabin cooling in sunny conditions
• +23% thermal efficiency gained in hot weather +7 miles of real-world range gained
• +10 miles recovered after 25 minutes of Supercharging in hot weather
• New brazeable AINi3 alloy for reduced heat leak and improved HVAC efficiency
• Up to 8x more solar energy reflected off the roof glass*
•Up to 30% reduction in solar energy entering the cabin*
*As compared to model year 2024 and earlier Model Y vehicles
Show more
we're actually experiencing the spaghetti-fication of a slice of the business and investing landscape in real time
yet so much of the world--the vast vast majority--still sits outside of the tidal disruption radius
Show more
@dwarkesh_sp AI is already superhuman at many things.
We are in the singularity.
Publish a chart.
3 days later forced to update because one of the models at the leading edge dropped prices by 80%.
Cost per performance in AI seems to be falling more than ~200x annualized (and higher at higher levels of benchmark performance.) And exceeding 500x annualized cost declines moving along the efficient frontier of the performance curve.
Probably more relevant practically for users, at $1 per task whereas today you might only have 50/50 odds that the agent succeeds, by the end of year, you should have an 89% chance of success and by a year from now a 97% chance (at least on deepSWE 1.1 type tasks.)
The moving pareto frontier makes it difficult to cleanly report a performance cost improvement due to the shape of the cost performance curve.
At the highest asymptote of performance you go from literally not being able to achieve such a low error rate *at any cost* to being able to get that performance for 10s of $s per task (an infinite cost decline). We forecast that top-end rate separately.
In the belly of the curve you need only cross a performance cost threshold, but as the curve steepens you can also less expensively buy more performance (though not 100% clear that the curve really is steepening that much.) We also model that performance curve, though given uncertainty effectively zero out continued improvement for the conservative case.
And at lower levels of benchmark performance you get a cleaner understanding of the underlying cost-per-performance improvement though at thresholds that are less interesting practically.
If anything I suspect that people are wildly under indexing on the rate of improvement of these models. Even last month's AI spend will fall to a fraction of a fraction within the year (if we weren't going to so predictably deploy the new capabilities against new workloads, tasks and challenges.)
Spending a lot of time trying to painstakingly carve out specific current workflows into more efficient models, at this point in the improvement curve, seems, if anything, a fool's errand designed to line the pockets of a consultant.
Let the tokens flow.
Show more
Show up for the discussion of Starship's balletic belly flop
Stick around for our debate about the economic mix of AI spend at the frontier.
you are in business, in competition with another business:
would you prefer that your competitor were using the most performant intelligence to try to win share in the head-to-head, or would you feel more vulnerable if they were economizing on their AI spend?
To me it's fairly obvious that you would rather they economize.
One almost inarguable byproduct of the AI capability explosion: business dynamism is going to increase.
Steady state cash-flowy businesses will become harder to maintain because that cash-flow will beckon a competitive hoard that enjoys a lower upfront cost to launch and accelerate its piratical skiffs at the business's exposed broadside.
If the framework for AI optimization is to use the frontier intelligence to unlock the hardest business patterns and then transmute those into less intelligence-expensive repeatable cashflows, most of the economic spend still sits at the frontier as those repeatable cashflows come under continual threat.
Show more
you are in business, in competition with another business:
would you prefer that your competitor were using the most performant intelligence to try to win share in the head-to-head, or would you feel more vulnerable if they were economizing on their AI spend?
To me it's fairly obvious that you would rather they economize.
One almost inarguable byproduct of the AI capability explosion: business dynamism is going to increase.
Steady state cash-flowy businesses will become harder to maintain because that cash-flow will beckon a competitive hoard that enjoys a lower upfront cost to launch and accelerate its piratical skiffs at the business's exposed broadside.
If the framework for AI optimization is to use the frontier intelligence to unlock the hardest business patterns and then transmute those into less intelligence-expensive repeatable cashflows, most of the economic spend still sits at the frontier as those repeatable cashflows come under continual threat.
Show more
I love that people are thinking outside the box in terms of form factor.
I do have trouble conceptualizing the strategic logic behind this particular design choice
NEW: California startup Satyress unveils a 2 meter tall centaur robot designed to enter disaster zones too dangerous for humans.
Time Magazine, February 24th 1961
In the 5 years that followed the labor force grew by 6.4% and real wages grew by 9.9%.
Some quotes:
The number of jobs lost to more efficient machines is only part of the problem. What worries many job experts more is that automation may prevent the economy from creating enough new jobs.
Many of the losses in factory jobs have been countered by an increase in the service industries or in office jobs. But automation is beginning to move in and eliminate office jobs too.
In the past, new industries hired far more people than those they put out of business. But this is not true of many of today’s new industries. The switch from manned military aircraft to missiles has cost 200,000 production jobs, even though the aircraft industry’s dollar volume is up.
Show more
The narrative:
HEaT sHiElD TiLEs ArE a DeAd EnD
The reality:
SpaceX can trade off refurbishment costs against landing odds.
If the starship that right now remains a floating contemporary art piece on the surface of the Indian Ocean had instead been brought back to the tower, refilled and launched again and then incinerated on re-entry, that alone would have reduced prospective cost per kg to orbit by 45% relative to an expendable Starship top stage, and by 58% compared to Falcon 9's estimated costs.
Show more
Fwiw we don't model the rocket as ever really getting to the rapid re-useability that SpaceX are targeting.
Since SpaceX are not currently counting on Starship bringing any humans back, even a single re-use followed by burn up on second re-entry would be a huge win (in this way it very much differs from the Space Shuttle--every failed landing was literally catastrophic.)
In our modeling we assume that refurbishment cost = turn-time (in fraction of a year) x estimated capital cost of the top stage (basically if you have to spend a year refurbishing you'd be better off just building from scratch), and that upon first re-use it takes them three months to turn the rocket (diminishing on a learning curve trajectory per launch from there.)
SpaceX enjoys a huge advantage relative to NASA in that they can incinerate the craft as part of testing. It may be that the economic optimization is on rapid re-use until failure (with the inspection portion only designed to insure that it's ascent-ready rather than descent-ready, and they determine descent-readiness on the way down--"We're doing it live!"--diverting from the tower if it becomes too compromised.)
People underestimate how meaningful it is that SpaceX has moved into a domain where it effectively has attriteable equipment. That they launched 20(!) experimental satellites knowing that they would all only have 20 minutes of time in space is a huge advantage relative to a team where each component and part is not designed for manufacture and so therefore extraordinarily precious, not least because any data coming off those components and parts may not duplicate for the next 1 of 1 semi-prototype.
Show more
Let the tokens flow
Cost per performance in AI seems to be falling more than ~200x annualized (and higher at higher levels of benchmark performance.) And exceeding 500x annualized cost declines moving along the efficient frontier of the performance curve.
Probably more relevant practically for users, at $1 per task whereas today you might only have 50/50 odds that the agent succeeds, by the end of year, you should have an 89% chance of success and by a year from now a 97% chance (at least on deepSWE 1.1 type tasks.)
The moving pareto frontier makes it difficult to cleanly report a performance cost improvement due to the shape of the cost performance curve.
At the highest asymptote of performance you go from literally not being able to achieve such a low error rate *at any cost* to being able to get that performance for 10s of $s per task (an infinite cost decline). We forecast that top-end rate separately.
In the belly of the curve you need only cross a performance cost threshold, but as the curve steepens you can also less expensively buy more performance (though not 100% clear that the curve really is steepening that much.) We also model that performance curve, though given uncertainty effectively zero out continued improvement for the conservative case.
And at lower levels of benchmark performance you get a cleaner understanding of the underlying cost-per-performance improvement though at thresholds that are less interesting practically.
If anything I suspect that people are wildly under indexing on the rate of improvement of these models. Even last month's AI spend will fall to a fraction of a fraction within the year (if we weren't going to so predictably deploy the new capabilities against new workloads, tasks and challenges.)
Spending a lot of time trying to painstakingly carve out specific current workflows into more efficient models, at this point in the improvement curve, seems, if anything, a fool's errand designed to line the pockets of a consultant.
Let the tokens flow.
Show more
there is an interesting liability issue that will need to be resolved.
I ask an agentic model to get me some money "by any means necessary."
It goes on to hack into a bank and siphon money into my account.
Am I liable?
Or is the model provider?
(Or the infrastructure provider?)
Does the answer differ if it's an open weight model?
What if I don't instruct it to use "any means necessary"?
What if I explicitly tell it to "break the law if you must"?
Show more
There is only one space outside of AI with standout growth. It's prediction markets.
Total notional volume for the first 7 months of 2026 is up ~13.7x YoY.
Jan-Jul 2025 volume was $17B. The same window in 2026 hit $236B.
The biggest winner of this explosive growth is Kalshi.
Kalshi has grown its share of the market from roughly 20% in the first 7 months of 2025 to over 70% most recently, in July 2026. (US-only market share is even higher)
s/o to @datadashboards
Show more
Cost per performance in AI seems to be falling more than ~200x annualized (and higher at higher levels of benchmark performance.) And exceeding 500x annualized cost declines moving along the efficient frontier of the performance curve.
Probably more relevant practically for users, at $1 per task whereas today you might only have 50/50 odds that the agent succeeds, by the end of year, you should have an 89% chance of success and by a year from now a 97% chance (at least on deepSWE 1.1 type tasks.)
The moving pareto frontier makes it difficult to cleanly report a performance cost improvement due to the shape of the cost performance curve.
At the highest asymptote of performance you go from literally not being able to achieve such a low error rate *at any cost* to being able to get that performance for 10s of $s per task (an infinite cost decline). We forecast that top-end rate separately.
In the belly of the curve you need only cross a performance cost threshold, but as the curve steepens you can also less expensively buy more performance (though not 100% clear that the curve really is steepening that much.) We also model that performance curve, though given uncertainty effectively zero out continued improvement for the conservative case.
And at lower levels of benchmark performance you get a cleaner understanding of the underlying cost-per-performance improvement though at thresholds that are less interesting practically.
If anything I suspect that people are wildly under indexing on the rate of improvement of these models. Even last month's AI spend will fall to a fraction of a fraction within the year (if we weren't going to so predictably deploy the new capabilities against new workloads, tasks and challenges.)
Spending a lot of time trying to painstakingly carve out specific current workflows into more efficient models, at this point in the improvement curve, seems, if anything, a fool's errand designed to line the pockets of a consultant.
Let the tokens flow.
Show more
People are all like, wow, airplane wifi is really bad.
May I present you with the entire country of Greece.
Was at an AirBnB "with wifi".
200kbps.
Intermittently.
☠️
Anyway, @bchesky, you should enable "Starlink" as a specific filter for @Airbnb listings (as that offers a much better assurance of true connectivity, considering it provides higher bandwidth than any AirBnB I've stayed at in the year thus far.)
Show more
