applied cybernetics // maths music theory econ @harvard @swarthmore
Manhattan, NY
- Tweets 23,353
- Following 1,757
- Followers 5,925
- Likes 106,678
i believe we can generalize this by finding a canonical set of (language-based) control problems and measuring the rate of frontier compression
> can a frontier model find a cheap controller of durably sufficient performance
i attempted a smaller instance last week—seems yes
This is directionally consistent with our internal data @cognition. Many tasks within software engineering are now "intelligence saturated." Once any model can do the task well, cost and speed are all that matter.
1
5
711
much of the app layer looks like control problems—you can then size the frontier market by addressing which always need frontier thinking for their operations
most ai applications today
-ai customer service
-legal contracts
-classification
look like control problems
> you imperfectly observe the environment, take an action, and get a reward
do you always need frontier AI to operate these controllers?
1
230
i fundamentally believe this will not prove to be a scaled business
the largest and most important exhaust will find the data they sit on is of strategic value—higher returns to eschewing the opportunity or vertically integrating
eg would apple or meta do this?
obv not
The emergence of data-native companies:
Capital markets haven't yet woken up to the opportunity to financialize the data sitting unaccounted for on balance sheets.
Data exhaust will be a new source of capital and a new driver of business models, akin to advertising.
6
52
4,538
capture a private operating env—where observation and actuation are proximal; indeed a pure complement to intelligence explosion—and let accelerating intelligence in performance/cost economics accrue to you
(few such env seem to exist)
AI doesn't erode moats, it enhances them
A universally scarce resource is human attention (and feedback as a corollary). When every other barrier is broken down, the value of what's left is amplified
This is also more "defensible" than other types of data (RL envs, internet data, old company logs,... which are static). Instead, real-time interaction patterns are dynamic. They are not yet publicly available, and can't be deduced from what already is.
1
29
4,092
Soren Larson retweeted
Blows my mind that people are just having this realization now? Of course having a differentiated (and compounding) ability to directly access and actuate on valuable economic information is the most important thing—that’s capitalism! Adding valuable economic information to general purpose technologies has always been the game. Unfortunately, most “data” companies don’t think about it this way.
Models capabilities are data bound,
Therefore the further the;
F(model + compute + training recipe)
is from the data source, the more polluted and expensive it gets
Kind of like an inverse square law
1
3
63
3,217
insurance has another medium to long term application:
> DISAGGREGATION
today you buy a service from someone because you trust them and their SLA
but that trust is expensive
insured ai-assisted service provision could move app layer services to infra and likely reduce costs
WSJ: insurance gives teeth to frontier audits.
Right now, we don’t have a path to trusted audits of frontier labs. Frontier labs have committed to third-party safety evaluations, but, today, labs set the rules and pick their own evaluator. They can change the rules or auditor at a whim. That doesn’t work.
Markets give audits teeth. When labs buy insurance, insurers require audits. Insurers would pay the bill in case of an incident, so they will pick a rigorous auditor rather than a lenient one. Insurers would pay for the audits, so auditors don’t have to worry about making the frontier lab “look bad” if they find something serious.
Interestingly, more rigorous audits help the frontier labs. One reason why labs commit to safety audits is to earn trust from the public that they are being responsible. Current audits don’t earn trust. Audits with teeth would earn trust.
Enjoyed my conversation with @jeaneaglesham at WSJ: wsj.com/tech/ai/ai-runs-into…
18
1,170
really neat to see this
humbly tho, i think the more interesting graphic is the blended human/agent labor total
depending on cost of human labor, it may well be that a dumber model with worse completion rate is actually more economical to use live
here's chat's reformulation
1
22
1,198
self driving will likely be fragmented
to the extent FSD delivery unit economics pencil v drones you need some physical world integrations layer to enable every FSD OS to securely hook into commerce
supply side moats that don’t compete on consumer economics in dangerrrr
Austin is now the heart of autonomy. Got Chinese food delivered fully autonomously for the first time via @goautolane. Stays warm. Super reliable. Super fast. No need to tip. And the food was awesome. When will other cities catch up…
5
1
11
2,174
an infra-lead neutral model aggregator that monetizes in outcomes and ‘senses’ via user-reported outcomes like meta Conversions API will absolutely crush
satya’s
>every firm is a learning loop
is self serving and not a path to real cost efficiency irrespective of corpo orthodoxy
Literally today an investor asked me if we’re seeing companies lock-in with one model provider (I said no)
It’s increasingly clear that where business spend on AI can, should, and is going:
Model-agnostic
Frequent switching
Managed centrally via firm defaults or through cost optimized routing
2
2
24
2,022
simulation: outcomes aggregators deliver more accurate outcomes than autarky-pilled continual learning orgs
hypersoren.xyz/posts/router-…
2
217
ive long set forth that agents must Perform Labor in order to get paid a take rate
ben thompson’s
>Agents (who perform no differentiated labor) Are The Final Aggregator
seems patently overzealous
if the service u provide the platform can replicate what rly r u doing?
see blog
30% agent share with a 3% take rate by 2030 is extraordinarily aggressive.
Repeat this arithmetic with 5-10% agent share and the entire endeavor looks futile / indefensible.
Unless, of course, agents don't implement a take rate, which, for many reasons, can't scale to a stable equilibrium.
Rather, if agents implement ad partnerships like the one Amazon has already struck with OpenAI, allowing the largest retail platforms to integrate directly into off-platform agentic purchasing workflows while retaining control over user relationships and their advertising revenue, then everyone's incentives align, and DAU (not eCommerce share or take rate) is the only determinant of revenue.
1
1
8
735
state machines and trajectories are in the air in new york
nearly everyone in ai applications seems to be studying them—not only for their agent framework adjacencies but for their capacity to cheaply simulate otherwise expensive (or irreversible) actions through latent state
I anticipated control engineering—and a resurgence of cybernetics—to matter well beyond current applications because systems are becoming overwhelmingly information-rich. You can only see their state in snapshots, and these fields are for regulating what you can't fully observe.
9
1
89
3,667
are most scaled ai applications control problems and if they are will they need frontier intelligence to persistently operate them?
(Probably not)
most ai applications today
-ai customer service
-legal contracts
-classification
look like control problems
> you imperfectly observe the environment, take an action, and get a reward
do you always need frontier AI to operate these controllers?
1
3
385
coasean singularity folks awfully quiet what happened lol
Field research 100% going to start looking more like espionage. As our ability to process information with AI rapidly accelerates, our incentive to reveal it is dramatically reduced. The dark forest era of corporate america is upon us.
3
17
1,646
most ai applications today
-ai customer service
-legal contracts
-classification
look like control problems
> you imperfectly observe the environment, take an action, and get a reward
do you always need frontier AI to operate these controllers?
10
7
66
5,894
everyone is on the hunt for sensors that
- front run existing markets
- privately legibilize net new signal
- maximize surplus generation and capture
below from usv earlier this week
Unlike bio, physical sciences/engineering have a powerful simulation infrastructure so "simulator + lab in the loop" is the paradigm. very rich idea space.
2
3
76
3,945
taps sign
the data industry for ai is temporary and it’s worse positioned than the data for finance one
recall bill gurley
> if financial sophistication ranged from 1-10 and really smart person in new york is 8.5, the average silicon valley person is a 2
this reversal hasn’t happened
the terminal value of data intermediaries is low—the data for ai industry is worse than that for finance, where at least for finance it’s possible for data streams to become required beta
no such candidacy for data for ai companies
14
3
138
23,682
