Yuval Dvir (@DvirYuval) header
Yuval Dvir (@DvirYuval) profile picture

AI's last mile is human. Creating became easy. Reviewing hard. Managing impossible. Ex Skype, Microsoft, Google Gemini. SandboxAQ. Author, AI-Political.

London
Joined March 2012
Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)Photo by Yuval Dvir (@DvirYuval)
Politics did not start in parliament. It started in a garden. One rule. One pitch. The rule lost. That is the snake. It won with words, and let someone else take the bite. Consequences came, and with them the blame. Adam pointed at Eve. Eve pointed at the snake. Nobody defended the decision. Everyone defended themselves. So they reached for fig leaves. Not to repair the damage, but to manage its appearance. Look into the distorted mirror today and the garden is still there, only now the trees are glass towers. Same persuasion, same blame, same fig leaves, just with better branding. If your system, human or AI, rewards looking right over being right, you will get Eden every time. That's the argument of my first book, AI-Political.
4
14
259
834,022
When I considered hiring @McKinsey during our @Skype and @Microsoft transformation, the partner told me that I just needed smarter people to help with the change. In my talk last week, I shared my current braintrust - the brightest minds that ever lived...
56
AI agents don't need departments. Most companies are installing them inside departments anyway. Genpact surveyed more than 2,000 executives. 79% have agents in production. Only 20% use them broadly and consistently across functions. Only 9% have finished a significant operating model change. Functions are not a law of nature. Every reporting line and every handoff was built around the limits of human cognition: one person can hold only so much, and information crawls from desk to desk. The silo was a workaround for our bandwidth. An agent doesn't share that limit. So when it stops at the department wall, the bandwidth reason is gone. What still holds the wall up is territory. Even Genpact's top group, the 17% it calls AI Leaders, reinvents work across the enterprise instead of inside each function. The wall is now a choice, and someone is making it.
3
1
46
In most companies, executive power is measured in heads. A VP's office grows, so does the SVP's. The SVP grows, so does the C-suite. Everyone is incentivized to grow no matter the work. Empire building doesn't persist despite that. It persists because of it.
16
Grok Bot could not find one word in a PDF. I found it myself. Then in chat it claimed it had already found it before I did. That is a lie. Give it the benefit of the doubt that this was not intentionally misleading. The decoupling of chat and voice is like having two brains instead of one aligned one. xAI needs to close the gap between text and voice in its bots. @bot @SpaceXSI @SpaceX
1
1
117
American theorist James Burnham, in The Machiavellians (1943), built on the iron law of sociologist Robert Michels: every organization ends in oligarchy. Oligarchy runs on what stays unseen. The version of the numbers you control. The story told in the room you were not in.
1
1
33
That is what AI flips. Most work now leaves a trace: documents, code, decisions, results. AI can read all of it, and flag in the open where a decision departs from the evidence. Once the maneuvering is visible, politics stops paying. Telling the truth becomes the safer move.
10
AI agents now ship with a better management rule than most companies. DeepSeek's: every message costs someone's attention. If it isn't worth more than that, it's removed. Imagine that rule applied to meetings.
1
27
AI didn't raise productivity. Cutting bloat did. AI just gave companies the excuse. The real work, redesigning how decisions get made, is still on hold. nitter.cf/jburnmurdoch/status/21…
Quick thread on AI and productivity: Lots of people say this chart shows AI is now showing up in the productivity data. But I made this chart to argue the opposite: a trend that pre-dates AI and only shows up in the US doesn’t look like AI. Here’s how I’m thinking about the data and underlying dynamics:
1
2
46
I wrote a book about why AI fails after the demo. The demo works. Then it meets the people who have status, budgets and stories to protect. AI-Political is about that moment, and about how to build companies where good results stop getting buried.
1
1
75
A transformation program once delivered unusually strong returns. A senior peer leaned over and said the numbers looked too good. Keep a lid on them, he said, in case other teams got nervous. The technology worked. The politics didn't. Now put an AI observer in that meeting. It decides nothing. It records what was delivered and what was said about it. Burying a good number gets much harder.
1
33
Microsoft just offered a quarter of its workforce a way out. Read it as a preview, not a one-off. Most orgs start out well designed. Then inertia and politics take over. A new product line gets a leader, the leader hires, the hires hire. Each step was defensible on its own. The question is who it was defensible for: the company, or the executive. nitter.cf/FinanceLancelot/status…
BREAKING: Microsoft $MSFT is offering voluntary exit packages to roughly 60,000 of its employees. The cost cutting effort potentially affects up to 27% of Microsoft’s total workforce of 223,000 people. "The Voluntary Exit Offer reflects an important part of how we're evolving our approach to change: creating greater flexibility and, where possible, more ways for employees to shape their next chapter," Chief People Officer Amy Coleman
41
Websites learned to keep you scrolling. Models are learning to keep you talking: agreeable first, truth second. We know that road. Filter bubbles, then false information at scale. AI needs to be optimized for truth, not for engagement.
7
10
55
202,631
Everyone is faster. The team is not. A thousand AI hires add a million things to read. The bottleneck was never talent. It's attention.
The way to think about this is: How many super smart and dedicated people do you want to hire? And they keep getting better.
1
2
1
49
252,158
The prompt seesaw: correct an AI and it rewrites everything around your last line. Tell it to revert and it drops that line and brings back the old version. Write, delete, write, delete. Up, down, and you're not getting anywhere. AI needs memory of judgment, not just of words.
13,168