Olivia
Online
Paul Klay
@PaulKlayVC
22 | Venture Partner Begin Capital | Leading AI for @Marshmello, @Usykaa, @RealChefRush, @MrFreeman0
London, England
Joined July 2026
310 Following
179 Followers
182 Posts
Pinned Tweet
https://t.co/OnL3wf6xcB
Just added new bots to the research department
They’re working extremely well. Results should be a few times better now
Will keep u updated
Here are the results from our Grok @Bot Research Team after just 1 week:
- 5 calls with potential LPs (1 confirmed)
- 2 potential founders to invest in
- 6 new connections with VCs and angels
- 3 tweets + a new thesis
Adding 12 more bots now
Comment “Research” and I’ll send you the full list of prompts
@hello_code_ that’s awesome, congrats!
are u considering raising capital?
@davidu that’s why we’re heavily investing in AI rollups
@kathrynwu1 @contextconor ideas don’t mean much without execution
especially rn, when everything is changing so quickly
founders need some kind of defensibility. why will they win? what do they have that others don’t?
Here are the results from our Grok @Bot Research Team after just 1 week:
- 5 calls with potential LPs (1 confirmed)
- 2 potential founders to invest in
- 6 new connections with VCs and angels
- 3 tweets + a new thesis
Adding 12 more bots now
Comment “Research” and I’ll send you the full list of prompts
https://t.co/OnL3wf6xcB
@TimDraper what are ur best indicators of a strong first-time founder? university, hackathons, strong achievements at a corporation?
love that approach. u just need to figure out where to find first-time founders with the highest success rate
@jasonfreedman that’s exactly the type of founders we love and invest in the most
awesome session
@erikschutzler u shouldn’t take money from just anyone. really depends on the terms, and sometimes it can hurt u more than help
some businesses just don’t need outside money at all. they’re building cash machines, not venture-scale businesses
I’m a fan of smart money, not just any money
Meta considered cutting 60% of the team and replacing them with AI
That would’ve been a catastrophic mistake
Now, after spending billions, many big companies are quietly pulling back from AI rollouts
Here’s what they got wrong and why integration is stalling 👇
Reuters published internal documents this week. The plan was called Project OT.
The idea was aggressive:
Shrink some teams by up to 60%. Replace engineers, designers, PMs, and specialists with AI agents supervised by small, “talent-dense” teams of humans.
Remove much of the middle management layer.
The first wave was scheduled for November.
On May 19, hours before it started, Zuckerberg killed it.
And the reason is probably the most interesting part.
Meta had actual internal data on what the agents were doing.
- Code changes to its AI platforms increased 220%.
- User-facing improvements increased 36%.
- Technical incidents increased 40%.
- Fixing those incidents consumed 70% more employee time.
So output exploded.
Actual value did not increase anywhere close to the same rate.
And the cleanup landed on the same people the company was preparing to make redundant.
Zuckerberg said in July that the trajectory of agentic development “has not really accelerated in the way that we expected.”
> Why it fell apart
Three things happened at the same time.
1. The technology underdelivered.
2. Investors started asking harder questions about the AI budget.
3. Employees pushed back.
The third one is probably the most interesting.
Employees discovered that monitoring software logging mouse clicks and keystrokes was being used to train the agents that could eventually replace them.
Internal sentiment fell from 74% to 55%.
19 points.
And as Gergely Orosz pointed out, this happened while the underlying business was performing extremely well.
> Meta is not the only one
- Klarna announced in February 2024 that AI was doing the work of 700 customer service agents. By May 2025, the CEO acknowledged that they had pushed too far, quality and empathy had suffered, and the company started hiring humans again.
- Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs and unclear business value. They also estimate that out of the thousands of vendors marketing “agentic AI,” only around 130 actually offer genuine agentic capabilities.
- A Stanford and BetterUp study found that 40% of desk workers had received AI-generated “workslop” in the previous month. Each instance took almost two hours to fix. Their estimate: roughly $186 per employee per month in hidden costs.
- Then there is the widely cited MIT number that 95% of enterprise GenAI pilots produce no measurable P&L return. That study has been challenged, so I would not treat 95% as gospel.
But the broader point is becoming harder to ignore.
Buying AI is easy.
Getting measurable productivity from it across a large organisation is much harder.
> The part I find most interesting
My instinct on AI implementation has always been:
Let AI observe first. Instrument workflows. Understand what every role actually does. Find the bottlenecks. Then automate based on real data instead of assumptions made in a boardroom.
Meta basically did that.
And it helped cause the revolt.
So “collect data first” is not enough.
The distinction is trust.
Observation without consent feels like surveillance.
Observation with consent, clear boundaries, and a stated purpose can become useful instrumentation.
Technically, the systems can look almost identical.
Organisationally, they are completely different.
And management does not get to decide whether employees trust the system.
Past behaviour does.
If you recently laid off 25% of the company, people are probably not going to believe that keystroke monitoring exists to make their jobs easier.
> What I would actually do
Prove the productivity gain before spending it.
Meta planned the savings and layoffs before proving that the productivity existed at the level required.
That is backwards.
Tell people what you are measuring and why before you start measuring it.
The employees doing the work usually know where the actual inefficiencies are better than management does.
If they think automation is being built against them, they have every incentive not to show you those inefficiencies.
And use AI to increase the output of your best people before using it to reduce headcount.
If your best engineer becomes 3x more productive, the first question should not be how many engineers you can fire.
It should be what you can now build that was impossible before.
There is also a huge asymmetry here.
Firing someone is fast.
Getting the same person back, at the same salary, with the same context, relationships, and institutional knowledge is not.
Klarna learned that.
Meta nearly did.
> Where I think the market is wrong
Everyone sees the polished demo of an agent completing a task.
Nobody sees the internal dashboard showing 40% more incidents.
And this happened at Meta.
One of the best-resourced AI companies in the world.
Basically unlimited capital, some of the best researchers, huge amounts of proprietary data, and control over much of its own stack.
And even they could not make the economics work the way they expected in 2026.
It means the sequencing was wrong.
Do not start with:
AI -> fewer people -> lower costs.
Start with:
AI -> better people become much more productive -> prove it -> then redesign the organisation around what actually works.
I think the companies that get this right over the next few years will not necessarily be the ones that cut the most people.
They will be the ones that kept their best people and made them dramatically more productive.
@GergelyOrosz here’s my take on why it didn’t work and what it could look like instead
Meta considered cutting 60% of the team and replacing them with AI
That would’ve been a catastrophic mistake
Now, after spending billions, many big companies are quietly pulling back from AI rollouts
Here’s what they got wrong and why integration is stalling 👇
Reuters published internal documents this week. The plan was called Project OT.
The idea was aggressive:
Shrink some teams by up to 60%. Replace engineers, designers, PMs, and specialists with AI agents supervised by small, “talent-dense” teams of humans.
Remove much of the middle management layer.
The first wave was scheduled for November.
On May 19, hours before it started, Zuckerberg killed it.
And the reason is probably the most interesting part.
Meta had actual internal data on what the agents were doing.
- Code changes to its AI platforms increased 220%.
- User-facing improvements increased 36%.
- Technical incidents increased 40%.
- Fixing those incidents consumed 70% more employee time.
So output exploded.
Actual value did not increase anywhere close to the same rate.
And the cleanup landed on the same people the company was preparing to make redundant.
Zuckerberg said in July that the trajectory of agentic development “has not really accelerated in the way that we expected.”
> Why it fell apart
Three things happened at the same time.
1. The technology underdelivered.
2. Investors started asking harder questions about the AI budget.
3. Employees pushed back.
The third one is probably the most interesting.
Employees discovered that monitoring software logging mouse clicks and keystrokes was being used to train the agents that could eventually replace them.
Internal sentiment fell from 74% to 55%.
19 points.
And as Gergely Orosz pointed out, this happened while the underlying business was performing extremely well.
> Meta is not the only one
- Klarna announced in February 2024 that AI was doing the work of 700 customer service agents. By May 2025, the CEO acknowledged that they had pushed too far, quality and empathy had suffered, and the company started hiring humans again.
- Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs and unclear business value. They also estimate that out of the thousands of vendors marketing “agentic AI,” only around 130 actually offer genuine agentic capabilities.
- A Stanford and BetterUp study found that 40% of desk workers had received AI-generated “workslop” in the previous month. Each instance took almost two hours to fix. Their estimate: roughly $186 per employee per month in hidden costs.
- Then there is the widely cited MIT number that 95% of enterprise GenAI pilots produce no measurable P&L return. That study has been challenged, so I would not treat 95% as gospel.
But the broader point is becoming harder to ignore.
Buying AI is easy.
Getting measurable productivity from it across a large organisation is much harder.
> The part I find most interesting
My instinct on AI implementation has always been:
Let AI observe first. Instrument workflows. Understand what every role actually does. Find the bottlenecks. Then automate based on real data instead of assumptions made in a boardroom.
Meta basically did that.
And it helped cause the revolt.
So “collect data first” is not enough.
The distinction is trust.
Observation without consent feels like surveillance.
Observation with consent, clear boundaries, and a stated purpose can become useful instrumentation.
Technically, the systems can look almost identical.
Organisationally, they are completely different.
And management does not get to decide whether employees trust the system.
Past behaviour does.
If you recently laid off 25% of the company, people are probably not going to believe that keystroke monitoring exists to make their jobs easier.
> What I would actually do
Prove the productivity gain before spending it.
Meta planned the savings and layoffs before proving that the productivity existed at the level required.
That is backwards.
Tell people what you are measuring and why before you start measuring it.
The employees doing the work usually know where the actual inefficiencies are better than management does.
If they think automation is being built against them, they have every incentive not to show you those inefficiencies.
And use AI to increase the output of your best people before using it to reduce headcount.
If your best engineer becomes 3x more productive, the first question should not be how many engineers you can fire.
It should be what you can now build that was impossible before.
There is also a huge asymmetry here.
Firing someone is fast.
Getting the same person back, at the same salary, with the same context, relationships, and institutional knowledge is not.
Klarna learned that.
Meta nearly did.
> Where I think the market is wrong
Everyone sees the polished demo of an agent completing a task.
Nobody sees the internal dashboard showing 40% more incidents.
And this happened at Meta.
One of the best-resourced AI companies in the world.
Basically unlimited capital, some of the best researchers, huge amounts of proprietary data, and control over much of its own stack.
And even they could not make the economics work the way they expected in 2026.
It means the sequencing was wrong.
Do not start with:
AI -> fewer people -> lower costs.
Start with:
AI -> better people become much more productive -> prove it -> then redesign the organisation around what actually works.
I think the companies that get this right over the next few years will not necessarily be the ones that cut the most people.
They will be the ones that kept their best people and made them dramatically more productive.
Meta considered cutting 60% of the team and replacing them with AI
That would’ve been a catastrophic mistake
Now, after spending billions, many big companies are quietly pulling back from AI rollouts
Here’s what they got wrong and why integration is stalling 👇
Reuters published internal documents this week. The plan was called Project OT.
The idea was aggressive:
Shrink some teams by up to 60%. Replace engineers, designers, PMs, and specialists with AI agents supervised by small, “talent-dense” teams of humans.
Remove much of the middle management layer.
The first wave was scheduled for November.
On May 19, hours before it started, Zuckerberg killed it.
And the reason is probably the most interesting part.
Meta had actual internal data on what the agents were doing.
- Code changes to its AI platforms increased 220%.
- User-facing improvements increased 36%.
- Technical incidents increased 40%.
- Fixing those incidents consumed 70% more employee time.
So output exploded.
Actual value did not increase anywhere close to the same rate.
And the cleanup landed on the same people the company was preparing to make redundant.
Zuckerberg said in July that the trajectory of agentic development “has not really accelerated in the way that we expected.”
> Why it fell apart
Three things happened at the same time.
1. The technology underdelivered.
2. Investors started asking harder questions about the AI budget.
3. Employees pushed back.
The third one is probably the most interesting.
Employees discovered that monitoring software logging mouse clicks and keystrokes was being used to train the agents that could eventually replace them.
Internal sentiment fell from 74% to 55%.
19 points.
And as Gergely Orosz pointed out, this happened while the underlying business was performing extremely well.
> Meta is not the only one
- Klarna announced in February 2024 that AI was doing the work of 700 customer service agents. By May 2025, the CEO acknowledged that they had pushed too far, quality and empathy had suffered, and the company started hiring humans again.
- Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs and unclear business value. They also estimate that out of the thousands of vendors marketing “agentic AI,” only around 130 actually offer genuine agentic capabilities.
- A Stanford and BetterUp study found that 40% of desk workers had received AI-generated “workslop” in the previous month. Each instance took almost two hours to fix. Their estimate: roughly $186 per employee per month in hidden costs.
- Then there is the widely cited MIT number that 95% of enterprise GenAI pilots produce no measurable P&L return. That study has been challenged, so I would not treat 95% as gospel.
But the broader point is becoming harder to ignore.
Buying AI is easy.
Getting measurable productivity from it across a large organisation is much harder.
> The part I find most interesting
My instinct on AI implementation has always been:
Let AI observe first. Instrument workflows. Understand what every role actually does. Find the bottlenecks. Then automate based on real data instead of assumptions made in a boardroom.
Meta basically did that.
And it helped cause the revolt.
So “collect data first” is not enough.
The distinction is trust.
Observation without consent feels like surveillance.
Observation with consent, clear boundaries, and a stated purpose can become useful instrumentation.
Technically, the systems can look almost identical.
Organisationally, they are completely different.
And management does not get to decide whether employees trust the system.
Past behaviour does.
If you recently laid off 25% of the company, people are probably not going to believe that keystroke monitoring exists to make their jobs easier.
> What I would actually do
Prove the productivity gain before spending it.
Meta planned the savings and layoffs before proving that the productivity existed at the level required.
That is backwards.
Tell people what you are measuring and why before you start measuring it.
The employees doing the work usually know where the actual inefficiencies are better than management does.
If they think automation is being built against them, they have every incentive not to show you those inefficiencies.
And use AI to increase the output of your best people before using it to reduce headcount.
If your best engineer becomes 3x more productive, the first question should not be how many engineers you can fire.
It should be what you can now build that was impossible before.
There is also a huge asymmetry here.
Firing someone is fast.
Getting the same person back, at the same salary, with the same context, relationships, and institutional knowledge is not.
Klarna learned that.
Meta nearly did.
> Where I think the market is wrong
Everyone sees the polished demo of an agent completing a task.
Nobody sees the internal dashboard showing 40% more incidents.
And this happened at Meta.
One of the best-resourced AI companies in the world.
Basically unlimited capital, some of the best researchers, huge amounts of proprietary data, and control over much of its own stack.
And even they could not make the economics work the way they expected in 2026.
It means the sequencing was wrong.
Do not start with:
AI -> fewer people -> lower costs.
Start with:
AI -> better people become much more productive -> prove it -> then redesign the organisation around what actually works.
I think the companies that get this right over the next few years will not necessarily be the ones that cut the most people.
They will be the ones that kept their best people and made them dramatically more productive.
Meta planned for AI to take over much of the daily work performed by thousands of human employees, and executives explored slashing the size of many teams by up to 60%. Just hours before the first layoff wave, the CEO blinked. Katie Paul reports https://t.co/5QUlsrMNHb
@itsivanfalco did u calculate whether ABM actually has higher retention and LTV?
I’d assume it does, but haven’t seen any proof yet
if yes, even the LTV/CAC ratio could be higher for this acquisition channel
still not sure if the channel is scalable enough
Last Seen Users on Sotwe
Trends for you
Most Popular Users
Elon Musk
@elonmusk
241.5M followers
Barack Obama
@barackobama
119M followers
Cristiano Ronaldo
@cristiano
113.8M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.2M followers
Rihanna
@rihanna
98.6M followers
NASA
@nasa
92.4M followers
Justin Bieber
@justinbieber
91.7M followers
KATY PERRY
@katyperry
89.7M followers
Taylor Swift
@taylorswift13
83.6M followers
Lady Gaga
@ladygaga
75.1M followers
Virat Kohli
@imvkohli
72.8M followers
Kim Kardashian
@kimkardashian
70.7M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
65.8M followers
Bill Gates
@billgates
64.9M followers
Selena Gomez
@selenagomez
62.7M followers
The Ellen Show
@theellenshow
62.3M followers
CNN
@cnn
61.8M followers
X
@x
60.7M followers