*https://www.youtube.com/watch?v=DGWtSzqCpog
**https://300.ya.ru/v_ijRmV7GO
timestamps
00:00:00 Введение в проблему
- Файл с уценкой от менеджера по продукту в Anthropic уничтожил рыночную капитализацию на 285 миллиардов долларов за 48 часов.
- Anthropic выпустила набор плагинов для Claude Cowork, один из которых занимается юридической проверкой контрактов.
00:00:39 Реакция рынка
- Плагин с открытым исходным кодом содержит 200 строк структурированных подсказок Markdown.
- Акции Thomson Reuters упали на 16%, LexisNexis — на 14%, Legal Zoom — на 20%.
- Продажи распространились на частные инвестиции, акции Ares Management, KKR и TPG упали примерно на 10%.
00:01:18 Влияние ИИ на юридический анализ
- Искусственный интеллект снижает стоимость юридического и финансового анализа, что создаёт проблему с маржой для фирм, взимающих дополнительную плату.
- Модель лицензирования SaaS на одно место, на которой строится экономика корпоративного программного обеспечения, даёт трещину.
00:02:00 Структурная проблема отрасли
- Anthropic выпустила набор стартовых плагинов, которые любая компания может легко настроить.
- Текстовый файл может описать основной рабочий процесс индустрии юридической информации, что указывает на структурную проблему модели ценообразования.
00:03:09 Критика модели SaaS
- Модель SaaS, где каждый человек, прикасающийся к инструменту, должен заплатить лицензионный сбор, ломается при внедрении ИИ.
- Среднее соотношение цены и прибыли в индустрии программного обеспечения сокращалось в течение нескольких месяцев.
00:04:02 Аргументы против замены ПО ИИ
- Дженсен Хуан утверждает, что ИИ не заменяет ПО, а работает на нём.
- Чем больше агентов ИИ, тем больше программной инфраструктуры требуется.
00:05:34 Аналогия с печатными СМИ
- Интернет разрушил модель доступа к контенту печатных СМИ, но контент сохранился.
- Программное обеспечение сохраняет свои данные, но модель доступа к ним меняется.
00:06:59 Противоречия на рынке
- Инвесторы разрабатывают два тезиса: расходы на инфраструктуру ИИ неприемлемы или внедрение ИИ устареет устаревшие бизнес-модели ПО.
- Оба тезиса не могут быть правдой одновременно.
00:08:13 Пример KPMG и Grant Thornton
- KPMG требует от Grant Thornton UK снижения платы за аудит за счёт ИИ.
- Grant Thornton сначала сопротивлялся, но в итоге согласился на 14% скидку.
00:09:02 Искусственный интеллект как инструмент переговоров
- Искусственный интеллект используется в деловых переговорах для снижения цен.
- KPMG не автоматизировала аудит, а применила ИИ как рычаг в переговорах.
- Изменение цен на фондовом рынке может произойти быстро.
00:09:39 Влияние ИИ на экономику профессиональных услуг
- ИИ меняет экономику, делая прежние цены неоправданными.
- Пересмотр цен на аудит может привести к изменениям в судебных издержках и других профессиональных услугах.
- Будущее за бережливыми командами, а не за масштабным внедрением ИИ.
00:11:02 Роль данных и подотчётности
- Информационные системы остаются важными, несмотря на ИИ.
- Подотчётность и взаимоотношения с поставщиками остаются ценными.
- Сложность рабочих процессов с ИИ делает подотчётность ещё более важной.
00:12:38 Трансформация SaaS-компаний
- SaaS-компании должны перейти от универсального интерфейса к агентской архитектуре.
- Необходимо перестроить продукт, ценообразование и выход на рынок.
- Переход требует быстрой архитектурной трансформации.
00:14:02 Инженерные разработки и кризис ресурсов
- Корпоративные SaaS-компании тратят много ресурсов на инженерные разработки.
- Разработчики поддерживают универсальные платформы, что ограничивает создание новых продуктов.
- Агентурная разработка меняет экономику создания ПО.
00:15:21 Снижение стоимости разработки ПО
- Стоимость разработки ПО приближается к нулю благодаря ИИ.
- Экономика покупки и сборки меняется в пользу создания ПО на заказ.
- ИИ может создать пользовательскую CRM-систему за один день.
00:16:29 Проблема артикуляции потребностей
- ИИ должен уметь преобразовывать расплывчатые потребности в работоспособное ПО.
- Проблема артикуляции остаётся узким местом для SaaS-компаний и создателей альтернатив.
- Вице-президент по продажам часто не может чётко сформулировать свои потребности.
00:17:25 Агентский поиск и контекст
- Агенты могут изучать контекст и задавать уточняющие вопросы.
- Они наблюдают за шаблонами использования и постепенно уточняют своё понимание потребностей человека.
- Для SaaS-компаний окно возможностей ещё не закрыто.
00:18:23 Преимущество SaaS-компаний
- SaaS-компании могут выиграть время благодаря обработке данных и подотчётности.
- Важно переосмыслить рабочие процессы, а не просто внедрять ИИ поверх существующих подходов.
- Пример: использование ChatGPT для проверки электронных писем вместо написания вручную.
00:19:13 Развитие ИИ-технологий
- Темп развития ИИ-технологий непостижим: Codex может создавать целые приложения.
- OpenAI запустил Frontier — платформу для корпоративных агентов.
- Context эволюционирует, и агенты становятся лучше в поиске контекста.
00:20:09 Гиперускорение ИИ
- ИИ-модели дают устаревшие советы из-за гиперускорения технологий.
- Необходимо фундаментально переосмыслить рабочие процессы для успеха.
- Время для перехода сокращается с каждым новым обновлением.
00:21:13 Переосмысление рабочих процессов
- Опыт использования ИИ-систем меняет представление о возможностях.
- SaaS-компании и работники умственного труда должны переосмыслить свою архитектуру и рабочие процессы.
- Переоценка цен на рабочие процессы продолжается.
00:22:13 Индивидуальные решения
- Важно относиться к ИИ как к фундаментальному сдвигу, а не как к второстепенному фактору.
- Упражнения помогают переосмыслить роль в контексте ИИ.
- Переходный период сокращается, и переоценка цен не прекращается.
00:23:06 Будущее ИИ
- Искусственный интеллект быстро развивается, и текущие новости могут устареть.
- Всем придётся приложить усилия для адаптации к новым технологиям.
- Автор верит, что зрители смогут адаптироваться к изменениям.
In this video
Chapters
Transcript
The 200-Line File That Moved Markets
0:00
A 200line prompt just killed $285
0:03
billion in market value. That’s right. A
0:06
markdown file, not a product, not a
0:08
platform, a markdown file from some
0:10
product manager at Anthropic erased $285
0:13
billion in market cap on the stock
0:16
market in just 48 hours. On January
0:18
30th, Anthropic released a set of
0:20
plugins for Claude Co-work, its desktop
0:22
AI tool. One of them handles legal
0:24
contract review. It can triage NDAs. It
0:27
can flag non-standard clauses against a
0:29
negotiation playbook and generate a ton
0:31
of compliance summaries. The kind of
0:33
work that until last week required a
0:35
parallegal, maybe a Westlaw
0:37
subscription, something that had to do
0:38
with billable hours, right? The plugin
0:40
is open source. Anyone can read it. And
0:42
when people did, they found roughly 200
0:45
lines of structured markdown prompts.
0:47
First year law school content dressed up
0:49
with some clever workflow logic. It’s
0:51
basically a fancy prompt. It shipped
0:53
with a little disclaimer. All outputs
0:55
should be reviewed by licensed
0:56
attorneys. I’ll bet by Monday morning.
0:59
Thompson Reuters had posted its biggest
1:01
single day stock decline on record. It’s
1:03
down 16%. RELX, the parent company of
1:06
Lexus Nexus, fell 14%. Legal Zoom
1:09
cratered at 20%. You get the idea. The
1:12
selling spread to private equity from
1:13
there. Aries Management, KKR, and TPG
1:16
all dropped about 10%. If AI compresses
1:20
the cost of legal and financial
1:22
analysis, then every firm charging
1:24
premium fees for that analysis has a big
1:27
big margin problem because they can’t
1:29
charge that much. Not next year and
1:31
maybe not now. But here’s what almost
1:33
nobody is saying clearly enough. The
1:35
markdown file itself is not the cause.
1:39
It just revealed what has been going on
1:41
for a while. the per seat SAS licensing
1:44
model, the financial bedrock that the
What Anthropic Actually Shipped
1:46
entire enterprise software economy has
1:49
been built on for 20 years, it was
1:51
already cracking. The market just hadn’t
1:53
priced it in yet because frankly, Wall
1:55
Street doesn’t understand AI that well.
1:57
So, this crash wasn’t really about
1:59
Claude. And we should be precise about
2:01
what actually happened because the
2:03
narrative has already crystallized into
2:06
anthropic crash to the software market.
2:08
And that framing, while it’s fun for
2:09
headlines, misses the real structural
2:12
story. What Anthropic actually shipped
2:14
was a set of open source starter
2:16
plugins, basically templates that any
2:18
company can customize for their own
2:19
workflows very easily. The legal plugin
2:21
was one of 11. It was very competent,
2:24
but it’s not by itself revolutionary.
2:26
Any decent prompt engineer could have
2:28
assembled something comparable in an
2:30
afternoon. So why did it move $285
2:33
billion? because the plugin made visible
2:36
what the market has been quietly
2:38
worrying about for months. If a text
2:41
file can approximate the core workflow
2:43
of a $60 billion revenue legal
2:46
information industry, then that whole
2:49
pricing model that the industry is built
2:51
on has a big structural problem. Not a
2:54
competitive problem, not a better
2:55
product, a structural problem. The kind
2:59
that doesn’t get solved by shipping
3:01
faster or hiring better salespeople.
3:03
Thompson Reuters charges per seat. Lexus
3:05
Nexus charges per seat. Westlaw charges
3:08
per seat. The entire enterprise software
3:11
economy from Salesforce to Service Now
3:14
to Adobe runs on a model that says every
3:16
human who touches this tool must pay a
3:19
license fee. That’s how these companies
Why Per-Seat SaaS Was Already Cracking
3:21
make their money. That’s how they
3:23
forecast their revenue. That’s how Wall
3:24
Street values them. That model works
3:27
when humans are the bottleneck. It
3:29
breaks when AI agents can do the work
3:32
without logging in. And the signals were
3:34
already everywhere if you knew where to
3:36
look. The software industry’s average
3:38
forward price to earnings ratio has been
3:40
compressing for months from X 8 months
3:43
ago to about 2x right when the sell-off
3:45
hit. That is the largest 4-month
3:48
valuation compression since the 2002.com
3:51
bust. Earning season has already been
3:54
ugly. Software companies are missing
3:55
revenue estimates at rates not seen
3:58
since the postcoid correction and
3:59
broader tech continues to beat. The AI
4:01
companies are fine, right? The per seat
4:03
model was under pressure before
4:05
anthropic shipped this little prompt
4:07
file. So the cloud plugin, it didn’t
4:10
start the fire. It just showed everyone
4:11
the building was already burning. Now I
4:13
got to be honest, plenty of smart people
4:15
think that this sell-off is a big
4:17
overreaction. And they might be right
4:19
about the selloff, but they would be
4:21
wrong about what it means. Jensen Hang
4:23
speaking at the Cisco AI Summit a few
4:25
days before the crash offered the
4:27
strongest version of the
4:28
counterargument. This notion that the
4:30
software industry is in decline and
4:31
being replaced by AI is the most
4:33
illogical thing in the world, he said.
4:35
And do you know why? Hong’s argument is
4:37
very simple. AI doesn’t replace
4:39
software. AI runs on software. The more
4:42
AI agents you deploy, the more software
4:45
infrastructure you need. More databases,
4:47
more APIs, more middleware, etc. So
4:49
every AI agent that replaces a
4:51
parallegal still needs West Law’s data.
4:54
It still needs a CRM. It still needs
4:56
document management. If anything, AI
4:58
should increase the total amount of
5:00
software the economy uses. Jensen’s not
5:02
wrong. He’s also not making the argument
5:04
he thinks he’s making. Nobody’s serious
5:06
is arguing that the world needs less
5:08
software. The argument is that the world
Jensen Huang’s Argument — and Why It Misses the Point
5:10
no longer needs to pay for software the
5:12
way it currently pays for software. So
5:14
Jensen is defending the product and he’s
5:16
right to do so. The market is attacking
5:19
the pricing model. Those are very
5:20
different things and confusing them is
5:22
how incumbents lose transitions they
5:24
should have survived. Print media made
5:27
this same mistake. Newspapers had
5:29
content people wanted. They had local
5:32
information, investigative journalism,
5:34
weather. The internet didn’t make that
5:35
content worthless. What the internet did
5:38
was destroy the access model. the idea
5:40
that you had to buy a whole newspaper to
5:42
get the one section you cared about and
5:45
that advertisers would pay premium rates
5:47
to reach readers with no alternative.
5:49
The content actually survived. The
5:51
business model didn’t. And that’s why so
5:54
many newspapers are in trouble. Print
5:55
media’s content did eventually get
5:57
commoditized. Anybody can publish now.
6:00
software’s content like proprietary
6:02
databases like structured workflows,
6:04
decades of accumulated enterprise data
6:07
that hasn’t been commoditized and it
6:09
actually probably won’t be. Thompson
6:12
Reuters case law database isn’t
6:14
something a startup vibe codes in a
6:16
weekend. Salesforce’s customer
6:18
relationship data is irreplaceable for
6:20
many of their clients. Adobe’s creative
6:22
tool ecosystem has a pretty deep moat.
6:25
So the data is safe, but the per seat
6:28
access model for that data is not. And
6:32
the companies whose entire financial
6:34
identity is built around per seat
6:36
licensing, they’re about to face the
6:38
hardest strategic question in enterprise
6:40
software. How do you repric your most
6:43
valuable assets without destroying your
6:45
revenue in the transition? Bank of
6:47
America’s Vivic Arya published the most
6:49
revealing analysis of the crash. He
6:51
called the sell-off internally
6:53
inconsistent. And he’s right in a way
6:55
that tells you something important about
6:56
where the market’s head is at right now
6:58
on software and on AI. Investors were
The Print Media Parallel
7:00
simultaneously running two thesis.
7:03
Thesis one, AI infrastructure spending
7:05
is unsustainable and the capex boom will
7:07
collapse. Thesis 2, AI adoption will be
7:10
so powerful that it renders established
7:12
software business models obsolete. Both
7:15
cannot be true. If AI is powerful enough
7:17
to crash $285 billion in software market
7:21
cap, the infrastructure required to run
7:23
that AI is underbuilt, not overbuilt,
7:26
the SAS apocalypse is paradoxically the
7:28
strongest possible demand signal for
7:30
continued AI infrastructure investment.
7:33
And yet both trades were profitable in
7:35
different hands at different moments.
7:37
The Deep Seek sell-off punished Nvidia
7:39
last year. The SAS correction punished
7:41
Salesforce at almost the same time this
7:43
year. Wall Street does not resolve
7:45
logical contradictions. It rotates
7:47
between them. One week the market prices
7:49
in an AI winter, the next it prices in
7:52
an AI revolution so total that legacy
7:55
software can’t survive it. The
7:57
contradiction persists because no single
7:59
firm needs to hold both positions. The
8:01
market as a whole holds them and the
8:03
market as a whole has no obligation to
8:05
be coherent. The incoherence is the real
8:08
story, not the crash, the incoherence.
8:10
But this is not really a story about
8:12
stocks. It’s bigger. While everyone was
8:14
watching Thompson Reuters stock price, a
8:17
quieter story broke that almost no one
8:19
paid attention to. And that tells you
8:20
about where we’re all headed more than
8:22
any given stock chart. KPMG, one of the
8:25
big four accounting firms, pressured
8:27
Grant Thornton UK, which is its own
8:29
auditor. Yes, the big four have to have
8:31
auditors, to cut their audit fees. The
8:33
demand was to pass on cost savings from
8:36
AI. Grant Thornton initially resisted,
8:39
arguing that quote, «High highquality
What the Market Got Wrong: Two Contradictory Theses
8:40
audits rely heavily on expert human
8:42
judgment and that fees reflect the cost
8:45
of people.» PMG’s response, «Per the
8:47
Financial Times, lower your prices or
8:49
we’ll find a new auditor.» And Grant
8:51
Thornton blinked. PMG’s international
8:53
audit fees dropped from $416,000 in 2024
8:57
to just $357,000
8:59
in 2025. They got a 14% discount. And
9:02
that story matters to me more than
9:04
Thompson Reuters stock price. And I want
9:06
to tell you why. The SAS apocalypse was
9:08
just a market event. Traders were
9:10
repricing stocks based on a change view
9:12
of the future. They do that all the
9:13
time. The KPMG negotiation is an
9:16
operating event. A real company using AI
9:19
as a lever in a real business
9:21
negotiation to extract a real price
9:23
reduction from a real counterparty. The
9:26
stock market repricing that could
9:27
reverse tomorrow. The KPMG precedent
9:30
won’t. Think about what KPMG actually
9:32
did. They didn’t automate their audit.
9:34
They didn’t replace Grand Thornton with
9:36
AI. They used the existence of AI. The
9:39
fact that everyone now know these now
9:41
knows these tasks can be done more
9:43
cheaply as a negotiating weapon. The
9:46
threat isn’t we’ll replace you with AI.
9:48
The threat is we both know AI changes
9:51
the economics. So your old prices,
9:53
they’re not justified anymore. That’s
9:55
the playbook and it works in every
9:57
knowledge work fee negotiation. Now if
9:59
audit fees get renegotiated on the basis
10:02
of AI cost savings, legal fees can be
10:04
next, then consulting fees, then
10:06
implementation fees, then design fees,
10:08
then every form of pro-services billing
10:11
that currently scales only with the
10:13
number of humans touching the work. You
10:15
cannot use that scaling assumption. Lean
10:17
teams are the future. The cascade
10:19
doesn’t require anyone to actually
10:20
deploy AI at scale. It just requires
10:23
buyers to point at that SAS apocalypse
10:25
and say, «We know the world changed. So,
10:28
let’s talk about your assumption that
The KPMG Story Nobody Covered
10:30
the work is done per human and let’s
10:33
talk about your rates.» The big four are
10:35
a sign of things to come. When they talk
10:37
about not automating their own work, but
10:40
just negotiating down the cost of
10:41
services, that is a big operating
10:44
mechanism that is going to shake the
10:47
industry. It’s not really the markdown
10:49
files. its fee negotiation leverage
10:51
spreading like wildfires through the
10:53
professional services economy like a
10:55
crack through an iceberg. All of those
10:57
assumptions that humans have to do the
10:59
work are shattering. The software did
11:01
not die. The data systems underneath
11:04
enterprise software, Thompson Reuters
11:06
case law databases, Salesforce’s
11:08
customer graphs, SAP’s resource planning
11:11
lo SAP’s resource planning logic,
11:13
Adobe’s creative workflow ecosystem.
11:16
Those all represent decades of
11:18
accumulated, structured, proprietary
11:21
information that no markdown file comes
11:23
close to replacing. Those data systems
11:26
will continue to exist. They must. The
11:28
economy runs on them. And there’s a
11:30
second edge that the market panic has
11:32
really overlooked, the single ringable
11:35
neck. Enterprises don’t just buy
11:37
Salesforce because it’s the best
11:38
possible CRM. You can make the case for
11:41
a lot of other software that’s better.
11:42
They buy Salesforce because when
11:44
something goes wrong at 2 am on the
11:46
night before the board meeting, there’s
11:48
a phone number to call and a contract
11:50
that says somebody is accountable. That
11:53
accountability layer, the vendor
11:54
relationship, the SLA, the legal
11:56
liability, the proservices team that
11:59
shows up when the system breaks, that is
12:01
enormously valuable to big
12:03
organizations. And no amount of agentic
12:05
AI eliminates the need for it. If
12:07
anything, the complexity of AIdriven
12:09
workflows makes that accountability even
12:11
more important, not less. So, the data
12:14
edge is real. The accountability edge is
12:17
real. What died is the pricing model
12:19
that sits over the top. The idea that
12:21
you can charge every human who touches
12:22
the software a nice convenient fat per
12:25
se license fee and that your revenue
12:27
scales linearly with that headcount. If
12:29
one AI agent can do the research that
12:31
previously required 10 parallegals with
12:33
10 separate Westlaw loginins, Thompson
12:36
Reuters doesn’t lose the value of their
12:37
data, they lose nine seats of revenue.
12:40
The data becomes actually more important
12:42
in an AIdriven world. It’s the fuel the
12:44
agents run on. But the per seat access
Fee Negotiation as the Real Contagion
12:46
model, that’s just broken. Here’s what
12:48
the investor thesis actually comes down
12:50
to. The markdown file represents an
12:52
existential threat if and only if these
12:55
SAS companies run business as usual. If
12:57
they just bolt AI features on top of
12:59
their existing UI, if they just add a
13:00
chatbot, then the market’s right.
13:02
They’re dead. The market is right to
13:04
repric them. The survival path is
13:06
actually fundamentally different. And
13:08
it’s the one Thompson Reuters,
13:09
ironically, is attempting with co-consel
13:12
to pivot from a one-sizefits-all
13:14
interface that humans navigate to an
13:16
agentic first architecture that AI
13:18
agents navigate and charge for the value
13:20
of the data and the accountability
13:22
rather than the number of humans logging
13:23
in. That’s not a feature update. That’s
13:26
a rebuild of the product, the pricing,
13:28
and the go to market simultaneously
13:30
while your stock price is cratering.
13:32
Whether the incumbents pull it off is a
13:34
$285 billion question. Literally, they
13:37
have the data edge, they have the
13:38
ringable neck edge, and those are real.
13:40
But pivoting from UI first to agentic
13:42
first is the kind of architectural
13:44
transformation that does tend to kill
13:47
companies that attempt it too slowly.
13:48
And the clock is running at a speed that
13:51
nobody in enterprise software has ever
13:54
experienced. There’s a second angle to
13:56
this that most SAS apocalypse analysis
13:58
completely misses and it might matter
14:00
even more than the pricing question.
14:02
Think about what enterprise software
14:04
companies really spend their money on.
14:05
Engineering. Thousands of developers
14:07
maintaining, updating, debugging, and
14:09
extending one-sizefits-all platforms
14:11
designed to serve every possible
14:13
customer configuration. Docuine employs
14:16
thousands of developers. That’s the real
14:18
cost of enterprise SAS. Not the servers,
What Actually Survives: Data and Accountability Edges
14:20
not the sales team, but the army of
14:22
engineers keeping a general purpose
14:24
system alive for millions of users who
14:26
each use it just a little bit
14:28
differently. Now, think about the
14:29
opportunity cost. Every developer
14:32
maintaining a legacy SAS UI is a
14:35
developer that is not building custom
14:37
agentic workflows. Every sprint spent
14:39
adding features to a one-sizefits-all
14:41
product is a sprint not spent rethinking
14:43
the product for an agent first world.
14:45
The companies that crashed this week,
14:47
they’re not just facing a pricing model
14:49
crisis, they’re facing a resource
14:51
allocation crisis. Their most valuable
14:53
people are maintaining the old thing
14:56
when they need to be building the new
14:57
thing desperately. And the transition
14:59
requires doing both of those
15:01
simultaneously within the same budget.
15:03
This is where Agentic software
15:05
engineering changes the math in a way
15:06
that most people haven’t fully
15:08
internalized. The cost of building
15:10
software is falling to zero. Not slowly
15:13
and not theoretically. It’s happening
15:15
right now. Cursor shipped a system that
15:17
generates a thousand code commits per
15:19
hour with no human involvement. Strong
15:21
DM published a production framework that
15:23
states code must not be written by
15:25
humans and code must not be reviewed by
15:27
humans. That is not laughable. In 2026,
15:29
that is what is happening. A researcher
15:31
at OpenAI spent $10,000 on codeex tokens
15:35
and automated his entire research
15:37
workflow. These aren’t demos. These are
15:40
operational systems running in
15:42
production. When building software cost
15:44
starts to approach zero, the economics
15:45
of buy versus build flip for the first
15:48
time in a long time. The entire
15:50
enterprise SAS value proposition was
15:52
predicated on the idea that it’s cheaper
15:54
to buy a general purpose tool than to
15:57
build a custom one. That was true when
15:59
software engineering was expensive and
16:00
slow. When an AI agent can build a
16:02
custom CRM in an afternoon, calculus can
16:05
reverse for some folks. Why pay
16:07
Salesforce per seat fees for a tool
16:09
designed to serve every company on earth
The Articulation Problem in Agentic Software
16:11
when you could have a tool designed to
16:12
serve your company? That is the promise
16:14
of vibe coding. That is the promise of
16:16
vibe engineering. Now you might wonder
16:19
is that how it actually works? The
16:21
honest answer is it depends. And what it
16:23
depends on is the hardest problem in the
16:25
entire stack. It’s harder than
16:26
intelligence. It’s harder than coding.
16:28
And it’s harder than pricing models. It
16:30
depends on whether an AI agent can take
16:32
the vague, implicit, half-articulated
16:35
thing a human actually wants and turn it
16:37
not just into workable software, but
16:39
very quickly into workable software with
16:42
minimal sustainment costs. I’ve
16:44
mentioned in a previous video that I am
16:46
skeptical of this long term, especially
16:48
for enterprises. Remember how we talked
16:50
about companies hiring for a single
16:52
ringable knack and paying for enterprise
16:54
data access? Those remain edges. And
16:58
anyone who wants to engineer their way
17:00
forward into a cheaper CRM and not
17:02
Salesforce must confront them. But they
17:05
also must confront the articulation
17:07
problem. And that is a real bottleneck.
17:09
Not just for SAS companies, but for
17:11
anyone who wants to build your own
17:13
alternative. When a VP of sales says, I
17:15
need a better way to track the pipeline.
17:17
That sentence contains less than 5% of
17:21
the information required to build a
17:22
useful tool. Frankly, less than 1%. The
17:26
other 95 or 99% is buried in how the
17:29
team actually works. What the unspoken
17:31
conventions are, which exceptions matter
17:34
and which don’t, how this quarter’s
17:36
priorities differed from last, what
17:38
better means in context. Now, a skilled
17:40
product manager will spend weeks
17:42
extracting that information through
17:44
interviews, observation, iteration.
17:46
Whether an agent can do the same thing,
17:48
not just write the code, but understand
17:50
the need deeply enough to write the
17:52
right code is one of the biggest
17:53
questions in software right now. I am
17:55
skeptical that we’re there yet, except
17:57
in a few cases where you have
17:59
extraordinary context availability
Bolt-On vs. Rebuild: The Individual Version
18:01
across the enterprise. But Agentic
18:03
Search is making progress on exactly
18:05
that problem. Agents can explore
18:07
context. They can ask clarifying
18:09
questions and they do now. And they can
18:11
observe usage patterns and iteratively
18:13
refine their understanding of what a
18:14
human actually needs. So, it’s starting
18:16
to come, but the question is timing. For
18:19
SAS incumbents, this means the window
18:21
has not yet closed. Their data edge and
18:24
their accountability edge really do buy
18:26
them time, but only if they use that
18:28
time to pivot to Agentic first, rather
18:31
than bolting AI onto the existing UI and
18:34
saying a prayer. Here’s the thing that
18:35
connects the SAS apocalypse to your
18:37
actual life. The same dynamic that is
18:40
threatening enterprise SAS companies,
18:42
the difference between bolting AI on top
18:44
of your existing approach and actually
18:46
rethinking how you work from the ground
18:48
up applies to every individual knowledge
18:51
worker that is watching this video. If
18:53
you’re using chat GPT to proofread
18:55
emails you could have written anyway,
18:58
you are bolting AI on the top. If you’re
19:00
using Claude to summarize documents you
19:02
could have read anyway, you’re bolting
19:03
AI on the top. If you add Copilot to
19:06
your IDE, but your development workflow
19:08
is just the same as it was two years ago
19:10
or even five months ago, you’re bolting
19:12
AI on the top. And just like the SAS
19:14
companies that are bolting AI features
19:16
onto their existing products and hoping
19:18
the market does not notice, you are
19:20
decorating a structural problem in your
19:22
own career rather than solving it. The
19:25
pace right now is almost
19:26
incomprehensible. 20 minutes after Opus
19:29
4.6 dropped, Codeex dropped. And Codeex
19:32
can ship entire desktop apps if properly
19:35
prompted end to end from scratch. OpenAI
19:37
isn’t done with Codeex though. They also
19:39
launched Frontier in the same week as
19:41
they dropped Codex 5.3. Frontier is an
19:44
enterprise agent platform. So that means
19:46
that it you can use Frontier to deploy
19:49
enterprise agents securely across your
19:52
entire data ecosystem. Remember when I
19:54
said that context was evolving and
19:56
agents were getting better at searching
19:58
for context and learning from context
19:59
clues how to build good software?
The Hyper-Acceleration Nobody Can Keep Up With
20:01
Frontier is part of why Claude Co has
20:04
gone from an interesting demo to a $285
20:07
billion market event. If you ask an AI
20:10
model right now to help you figure out
20:12
how to use AI, you will get advice
20:15
that’s 6 months out of date. Even the AI
20:17
cannot keep up with itself. This is what
20:19
hyper acceleration feels like. And that
20:21
word does sound like marketing. It
20:23
sounds like hype. I will have people in
20:24
the comments who say I’m overhyping, but
20:26
you got to live through it. And then it
20:28
sounds like another Tuesday. The gap
20:29
between I use AI tools and I’ve
20:31
rethought how I work around extremely
20:34
rapidly evolving AI capabilities is all
20:37
of our individual versions of what
20:38
happened in the SAS market. The first
20:40
approach feels really productive.
20:43
Bolting on AI lets you feel like you’re
20:45
keeping up. The second approach
20:47
fundamentally rethinking how you work
20:50
from the ground up. That’s what changes
20:52
outcomes and the window to make that
20:54
transition keeps compressing every time
20:56
there’s a new update, which frankly is
20:58
every few days. If you haven’t tried
21:00
Opus 4.6 and experienced what a good
21:03
million token context window feels like,
21:06
you’re already out of date. If you
21:07
haven’t used Cloud Co-work or Codeex or
21:10
played around with OpenAI Frontier,
21:12
please try them. Not because any one
21:14
tool is the answer, but because the
21:16
experience of using these systems
21:18
changes your mental model of what’s
21:20
possible. And your mental model of what
21:22
is possible is the thing that determines
21:25
whether you are bolting on AI in your
21:27
own career and praying or whether you’re
21:29
rebuilding for an AI future that is
21:31
coming like a title wave. The SAS
21:33
companies that survive the SAS
21:34
apocalypse will be the ones that rethink
21:36
their architecture before the market
21:38
makes them. The knowledge workers who
21:40
thrive through the transition will be
21:42
the ones who rethink their workflows
21:45
before the boss forces them to. It’s the
21:47
same dynamic. It’s the same urgency.
21:49
It’s just at a different scale. The per
21:51
seat SAS pricing model is broken. The
21:53
data and accountability underneath it
21:55
are not. And the same logic applies to
21:57
you. Your skills, your domain expertise,
What You Need to Do Right Now
22:00
the thing that makes you passionate
22:02
about work, that didn’t break. But the
22:05
assumption that you can just take that
22:07
to work and not use AI or only use AI a
22:10
little bit or use AI in a chatbot, that
22:12
is broken. And you’re going to need to
22:14
look at how you fundamentally rethink
22:16
your workflows to get there. And that is
22:19
exactly what I’m putting together in the
22:21
exercises that go with this video on my
22:23
Substack. got a bunch of exercises that
22:25
help you think about how you can take
22:27
your unique role and essentially do the
22:30
repricing, do the rebuilding that the
22:32
SAS companies are talking about, but at
22:34
individual scale for your individual
22:37
workflows, how you think about AI, not
22:38
as a bolt-on, but as a fundamental
22:40
shift. A 200line markdown file did not
22:43
decide who wins and loses, but it did
22:46
compress a transition that everybody
22:48
expected to take 5 years into a 48 hour
22:50
repricing event. And the repricing
22:52
hasn’t stopped. It’s just getting
22:54
started. The clock is ticking. It’s not
22:56
stopping. And I want you to make good
22:59
decisions with your career. And we’ll
23:02
have to see if the SAS companies make
23:03
good decisions with their futures as
23:06
companies because by the time you watch
23:07
this, whatever the stock market price
23:09
says, the AI that you hear about in this
23:12
video will already be overtaken by some
23:16
other news. That is how fast we’re
23:17
moving. AI isn’t stopping, and we’re all
23:20
going to have to dig in to get through
23:21
this together. I know you can do it.