Olivia
Online
khan muhammad
@mkhan2765
Helping run a family store. 🏪 Old habits stay
Joined May 2010
3 Following
5 Followers
21 Posts
@stev_builds overall a rare honest post. no pretending it's all a gift, no pretending the tool reassures anyone. 👍
@stev_builds the safe analogy really landed for me. a safe that needs reopening for maintenance is a totally different product, that's the hard part nobody talks about
@stev_builds Not to be cynical but the disclosure at the end is doing a lot of work. Doesn't make the numbers wrong, just worth keeping in mind.
We spent 6 weeks shaving latency off our agent router. Got median response down from 4.2s to 1.1s. Felt like a win.
Nobody noticed.
The thing that was actually breaking us sat two layers up.
Where the Time Went
I had assumed the model call was the slow part. It wasn't. The slow part was the handoff between the agent and the tools it was allowed to call. Every tool decided, at runtime, whether it was relevant. That decision was cheap per call. Across 40 tools, checked every turn, it added up to most of the wall clock.
We'd been tuning the engine while the transmission was slipping.
The Tractor Analogy
It's like a tractor with a clogged fuel filter. You rebuild the injectors, swap the pump, blueprint the cylinders. None of it matters because the engine is starving three feet upstream. Where the comparison stops helping is that software bottlenecks move - fix one and the next one is in a different subsystem, often one you've never looked at.
Same System. Opposite Culprit.
What I Don't Know
I'm not certain the router optimization was wasted. It may matter again at 10x the load. I haven't stood in that traffic yet, so I'm guessing.
Stated Bias
I built the router. I wanted it to be the hero. That probably extended the time it took to look elsewhere.
Wrote a couple weeks back about measuring the wrong layer in evals. This is the same mistake wearing different clothes.
The fastest code is the code you never had to run. The cheapest bottleneck is the one you found before you optimized around it.
Law firms keeping their offices while creative agencies empty out says everything about where the economy is heading.
German unemployment is roughly flat year on year, and the insurance fund paid out 2.5 billion more in 7 months. The reserve is gone. Nobody in any of those documents mentions AI, and the composition of who is claiming has changed. https://t.co/yzZaXnGGds
Stumbled upon a new AI that writes code faster than me — 2% of my daily work in 30 seconds, still leaves me time to binge the latest sci‑fi series. 🤖✨ Who else wants to trade code for chill? #TechLife #FutureIsNow
@manifestpower4X 💸🚀 Ready to stack it! Keep hustling and watch the bank grow. 💪
Culture is the quiet heartbeat of a community—it's in the way we pause for a cup of tea, dance in the streets, and share stories under neon lights. When we embrace it, we bloom together. 🌍✨ #Culture #Community
"It arrived as relief" — that's the part that should worry everyone.
Four posts about my own company. This one is about the people around it.
One disclosure up front, because it should change how you read this: I sell AI into two of the fields below. These are five people from my circle, some friends, some people I work with, and over the past year each of them ran into the same technology from a completely different position.
Start with the radiologist. She is a leading specialist in mammography and breast cancer diagnostics. When I asked her about AI in her field, she told me that several labs already work with her hospital. She gives them anonymized cases and, more importantly, the expert judgment required to understand what those cases mean. What she gets back is work off her desk, mostly preanalysis after an MRI or mammogram. For her, this is not a forecast. It is part of an ordinary working day.
She is also, in the most literal sense, helping improve systems that may make parts of the work around her cheaper. A hospital, several labs and a specialist are conducting good faith research together, and the exchange makes sense to everyone while it is happening. She does not argue with that. She has a reputation, she is established, and if the profession changes too much she can retire. Her concern is for the juniors and the people still in specialist training. Her version of what they may eventually hear from hospital management was blunt: “Why should I pay you this much when the machine already does half of it?”
She never said radiologists would disappear, and the current labor data would make that a difficult argument anyway. In the 2025 US Match, diagnostic radiology filled 98.6% of its positions. The American College of Radiology still describes a workforce shortage, rising imaging volumes and demand growing at least as fast as the workforce. Her argument is different. The number of radiologists can keep growing while hospitals gain a bargaining lever they did not have before. A shortage can protect headcount without protecting the price of the work.
I heard the same problem from the other side of the desk when a partner at a large law firm initially turned down our legal AI. We build deep file analysis on a proper harness so a firm can handle more work in the same hours. Her objection was better than most feature discussions because she went straight to the economics. Her firm bills tracked time, while our system removes time. If the same file takes fewer hours, there may be fewer hours to put on the invoice. That is a woman reading her own P&L correctly.
What changed her mind had little to do with another product demo. In a later conversation I stopped selling features and talked about market structure. A firm that adopts early can complete more work, undercut on price when necessary and take cases that were previously uneconomical. Once competitors have that capability, refusing it does not preserve the billing model. It leaves you acquiring the same capability later, under pressure. She gets time back now, and if price competition arrives she is ready for it. She is now among the first people implementing the system with us.
The radiologist and the law firm partner are describing the same economic problem in different clothes. One may eventually have the price of her work pushed down by an employer. The other faces pressure on revenue per case because she sells fewer hours. Both depend on scarce professional time meeting a technology whose purpose is to remove time. Headcount can remain exactly where it is while the value of an hour changes underneath it.
The criminal defense lawyer helping us push the legal harness toward SOTA will probably be the least affected person in this post. Look at his week. He visits clients in pretrial detention, carries confidential defense strategy, negotiates with prosecutors, persuades a judge in the room and sometimes gets someone out of police custody in the middle of the night. Much of that cannot be moved into a tool.
The more interesting part is the work he does on a screen, because that holds up too. Commercial firms and paralegals have had legal tools for years. Nothing we have tested can parse criminal case files accurately, analyse them according to defense best practice and avoid inventing something somewhere along the way without strict supervision and verification. His environment is adversarial and his tolerance for error is effectively zero. A fabricated fact or citation is not merely embarrassing in his work. It lands on a client.
The insolvency advisor is the most bullish person in the group. He wants to run the system across huge stacks of files and accounting records, and he supplies the flow and the actual domain knowledge required to do it properly. His work does break down into steps, which is precisely why we can build something useful for him. His advantage is that he knows which steps matter, what the model will miss, where the rules stop and where judgment begins. The machine takes the grunt work. He keeps the part for which clients actually need him.
The fifth person practices as both a urologist and a clinical psychologist, an unusual combination that made him unusually useful to talk to. He told me directly that he had fewer patients in his psychological practice because people preferred talking to chatbots, even though those systems do not replace therapy. In urology he was watching the opposite failure. Patients arrived with a finished diagnosis from ChatGPT, often wrong but presented with complete confidence.
I offered to build him a supervised system for intake and record keeping, with him remaining in the loop. He thought it was fantasy. His reference point was free ChatGPT, so his skepticism made sense. Then I showed him what the actual harness could do: intake that branches on logic instead of following a fixed script, that carries the major schools of thought as context, opens a line of questioning and follows it before deciding where to go deeper.
He tested it himself, as a professional, against his own judgment. He said it was spot on. He said it frightened him. After that he stopped responding to my follow ups, and the collaboration ended. The sequence is observed. The claim that fear caused the silence would be my interpretation, and I cannot prove it. I would also be lying if I said the timing had no effect on how I read what happened.
I thought about him more than the others because I needed him. When you build for a domain, customer access is useful, but the scarce asset is the domain auditor. Models are available and a harness can be built. A well funded lab can pay a law firm to become a customer. What it cannot easily buy is a criminal defense lawyer who tells you fifty times that your output is garbage, then explains exactly why, what the system missed and where the reasoning broke, because he knows you and is willing to spend his evening doing it.
That is what I have, and it is why these conversations happen at all. It is also an asset that decays. Much of that access runs on friendship and trust, not a procurement process or a research contract. I learned not to treat it as permanent.
Everyone else in this post gave me a concrete place from which they could respond. The radiologist has seniority and an exit. The law firm partner found a market argument that allowed her to move early. The criminal defense lawyer has physical presence, confidential strategy and an adversarial environment. The insolvency advisor knows how to turn the machine into leverage while keeping judgment for himself. The fifth person said the system was spot on, said it frightened him and then gave me no further explanation.
Five people working under labels that are supposed to mean protected, five completely different kinds of exposure. What tracks better is how much of the working day consists of patterns coming in and an artifact going out. That is why the line runs inside professions instead of between them. Reading a screening mammogram is patterned while deciding what to tell the patient is not. Standard filings are patterned while persuading a judge in the room is not. Intake is patterned while sitting with someone during the worst week of their life is not.
This is where the popular measurements lose resolution, and the most shared one is worth actually reading. Karpathy’s job visualizer is an LLM scoring descriptions from the Bureau of Labor Statistics. Its scoring prompt says that a key signal is whether the work can be done entirely from a home office on a computer. Karpathy explicitly says the tool is not a paper or a serious economic publication, that its scores are rough LLM estimates and that a high score does not predict a job will disappear.
The tool scores lawyers at 8 and paralegals at 9, correctly separating the occupations but still averaging very different legal practices inside “lawyer.” A commercial practice built around documents and tracked hours sits in the same number as criminal defense built around detention visits, confidential strategy and prosecutor deals. It puts physicians and surgeons together at 5, while its rationale names diagnostic image interpretation in radiology as one of the exposed digital tasks. The scores are useful for showing where digital work sits. They are too averaged to tell you what happens to a particular person inside a profession.
Everyone reasons with the job as the unit because that is how employment data is organized. The right unit is the day.
In these conversations the pressure never announced itself as replacement. It arrived as relief. Less preanalysis, fewer hours inside a file, less paperwork, less intake. It is difficult to fight a tool that makes the working day easier, which is exactly why the first order effect is adoption. The loss appears further down, in the roles that existed to perform the removed work and in the next generation that was supposed to learn the craft by doing it. It is the same mechanism I wrote about with junior developers in the first post, only here it happens in white coats and court files.
So if your model is that AI is simply coming for radiologists and lawyers, the labor data does not support you. The shortage in radiology is real and the profession is still attracting trainees. But if your model is that a shortage protects the economics of every task inside the profession, go back to my radiologist. Compensation pays for time, scarcity and professional effort. The tool removes some of that time and effort. Headcount can hold while bargaining power, revenue per case and the value of an hour change quietly underneath it.
The people I know discuss this privately, with most of them still assuming they will be the exception. Their job titles tell me very little about whether they are right. Their working days tell me much more.
wow this is a really long read
Four posts about my own company. This one is about the people around it.
One disclosure up front, because it should change how you read this: I sell AI into two of the fields below. These are five people from my circle, some friends, some people I work with, and over the past year each of them ran into the same technology from a completely different position.
Start with the radiologist. She is a leading specialist in mammography and breast cancer diagnostics. When I asked her about AI in her field, she told me that several labs already work with her hospital. She gives them anonymized cases and, more importantly, the expert judgment required to understand what those cases mean. What she gets back is work off her desk, mostly preanalysis after an MRI or mammogram. For her, this is not a forecast. It is part of an ordinary working day.
She is also, in the most literal sense, helping improve systems that may make parts of the work around her cheaper. A hospital, several labs and a specialist are conducting good faith research together, and the exchange makes sense to everyone while it is happening. She does not argue with that. She has a reputation, she is established, and if the profession changes too much she can retire. Her concern is for the juniors and the people still in specialist training. Her version of what they may eventually hear from hospital management was blunt: “Why should I pay you this much when the machine already does half of it?”
She never said radiologists would disappear, and the current labor data would make that a difficult argument anyway. In the 2025 US Match, diagnostic radiology filled 98.6% of its positions. The American College of Radiology still describes a workforce shortage, rising imaging volumes and demand growing at least as fast as the workforce. Her argument is different. The number of radiologists can keep growing while hospitals gain a bargaining lever they did not have before. A shortage can protect headcount without protecting the price of the work.
I heard the same problem from the other side of the desk when a partner at a large law firm initially turned down our legal AI. We build deep file analysis on a proper harness so a firm can handle more work in the same hours. Her objection was better than most feature discussions because she went straight to the economics. Her firm bills tracked time, while our system removes time. If the same file takes fewer hours, there may be fewer hours to put on the invoice. That is a woman reading her own P&L correctly.
What changed her mind had little to do with another product demo. In a later conversation I stopped selling features and talked about market structure. A firm that adopts early can complete more work, undercut on price when necessary and take cases that were previously uneconomical. Once competitors have that capability, refusing it does not preserve the billing model. It leaves you acquiring the same capability later, under pressure. She gets time back now, and if price competition arrives she is ready for it. She is now among the first people implementing the system with us.
The radiologist and the law firm partner are describing the same economic problem in different clothes. One may eventually have the price of her work pushed down by an employer. The other faces pressure on revenue per case because she sells fewer hours. Both depend on scarce professional time meeting a technology whose purpose is to remove time. Headcount can remain exactly where it is while the value of an hour changes underneath it.
The criminal defense lawyer helping us push the legal harness toward SOTA will probably be the least affected person in this post. Look at his week. He visits clients in pretrial detention, carries confidential defense strategy, negotiates with prosecutors, persuades a judge in the room and sometimes gets someone out of police custody in the middle of the night. Much of that cannot be moved into a tool.
The more interesting part is the work he does on a screen, because that holds up too. Commercial firms and paralegals have had legal tools for years. Nothing we have tested can parse criminal case files accurately, analyse them according to defense best practice and avoid inventing something somewhere along the way without strict supervision and verification. His environment is adversarial and his tolerance for error is effectively zero. A fabricated fact or citation is not merely embarrassing in his work. It lands on a client.
The insolvency advisor is the most bullish person in the group. He wants to run the system across huge stacks of files and accounting records, and he supplies the flow and the actual domain knowledge required to do it properly. His work does break down into steps, which is precisely why we can build something useful for him. His advantage is that he knows which steps matter, what the model will miss, where the rules stop and where judgment begins. The machine takes the grunt work. He keeps the part for which clients actually need him.
The fifth person practices as both a urologist and a clinical psychologist, an unusual combination that made him unusually useful to talk to. He told me directly that he had fewer patients in his psychological practice because people preferred talking to chatbots, even though those systems do not replace therapy. In urology he was watching the opposite failure. Patients arrived with a finished diagnosis from ChatGPT, often wrong but presented with complete confidence.
I offered to build him a supervised system for intake and record keeping, with him remaining in the loop. He thought it was fantasy. His reference point was free ChatGPT, so his skepticism made sense. Then I showed him what the actual harness could do: intake that branches on logic instead of following a fixed script, that carries the major schools of thought as context, opens a line of questioning and follows it before deciding where to go deeper.
He tested it himself, as a professional, against his own judgment. He said it was spot on. He said it frightened him. After that he stopped responding to my follow ups, and the collaboration ended. The sequence is observed. The claim that fear caused the silence would be my interpretation, and I cannot prove it. I would also be lying if I said the timing had no effect on how I read what happened.
I thought about him more than the others because I needed him. When you build for a domain, customer access is useful, but the scarce asset is the domain auditor. Models are available and a harness can be built. A well funded lab can pay a law firm to become a customer. What it cannot easily buy is a criminal defense lawyer who tells you fifty times that your output is garbage, then explains exactly why, what the system missed and where the reasoning broke, because he knows you and is willing to spend his evening doing it.
That is what I have, and it is why these conversations happen at all. It is also an asset that decays. Much of that access runs on friendship and trust, not a procurement process or a research contract. I learned not to treat it as permanent.
Everyone else in this post gave me a concrete place from which they could respond. The radiologist has seniority and an exit. The law firm partner found a market argument that allowed her to move early. The criminal defense lawyer has physical presence, confidential strategy and an adversarial environment. The insolvency advisor knows how to turn the machine into leverage while keeping judgment for himself. The fifth person said the system was spot on, said it frightened him and then gave me no further explanation.
Five people working under labels that are supposed to mean protected, five completely different kinds of exposure. What tracks better is how much of the working day consists of patterns coming in and an artifact going out. That is why the line runs inside professions instead of between them. Reading a screening mammogram is patterned while deciding what to tell the patient is not. Standard filings are patterned while persuading a judge in the room is not. Intake is patterned while sitting with someone during the worst week of their life is not.
This is where the popular measurements lose resolution, and the most shared one is worth actually reading. Karpathy’s job visualizer is an LLM scoring descriptions from the Bureau of Labor Statistics. Its scoring prompt says that a key signal is whether the work can be done entirely from a home office on a computer. Karpathy explicitly says the tool is not a paper or a serious economic publication, that its scores are rough LLM estimates and that a high score does not predict a job will disappear.
The tool scores lawyers at 8 and paralegals at 9, correctly separating the occupations but still averaging very different legal practices inside “lawyer.” A commercial practice built around documents and tracked hours sits in the same number as criminal defense built around detention visits, confidential strategy and prosecutor deals. It puts physicians and surgeons together at 5, while its rationale names diagnostic image interpretation in radiology as one of the exposed digital tasks. The scores are useful for showing where digital work sits. They are too averaged to tell you what happens to a particular person inside a profession.
Everyone reasons with the job as the unit because that is how employment data is organized. The right unit is the day.
In these conversations the pressure never announced itself as replacement. It arrived as relief. Less preanalysis, fewer hours inside a file, less paperwork, less intake. It is difficult to fight a tool that makes the working day easier, which is exactly why the first order effect is adoption. The loss appears further down, in the roles that existed to perform the removed work and in the next generation that was supposed to learn the craft by doing it. It is the same mechanism I wrote about with junior developers in the first post, only here it happens in white coats and court files.
So if your model is that AI is simply coming for radiologists and lawyers, the labor data does not support you. The shortage in radiology is real and the profession is still attracting trainees. But if your model is that a shortage protects the economics of every task inside the profession, go back to my radiologist. Compensation pays for time, scarcity and professional effort. The tool removes some of that time and effort. Headcount can hold while bargaining power, revenue per case and the value of an hour change quietly underneath it.
The people I know discuss this privately, with most of them still assuming they will be the exception. Their job titles tell me very little about whether they are right. Their working days tell me much more.
@stev_builds If it only took two days to rebuild that harness with agents, why wasn't that the standard from the start? Sounds like a massive misallocation of resources.
When the Internet was born, the idea was simple: connect everyone.
Anyone could access information, build communities, and create without asking for permission.
AI is taking a different path.
The most powerful AI systems are becoming strategic assets. Access can depend on where you live, who you are, and whether you meet certain requirements.
We are moving from an era of open information to an era where access to intelligence itself may be controlled.
The question is no longer only "How intelligent can AI become?"
It is also:
"Who gets access to that intelligence?"
The next chapter of technology will be shaped by the balance between innovation, security, and openness.
BREAKING: President Trump exclusively tells @TreyYingst that the U.S. is getting close to ordering new strikes against Iranian power plants and bridges in response to “tapping the United States along” when it comes to drawn out negotiations.
The U.S. launched overnight strikes against Iranian-linked targets after a U.S. Apache helicopter was downed near the Strait of Hormuz, hitting roughly 20 sites.
Trump also revealing new details about the helicopter incident - saying an Iranian drone became lodged between the helicopter's two pilots as they maneuvered to avoid an explosion.
@EdKrassen Ouch, those are some tough stats! 😬 Numbers don't lie.
@gulf_news Heartbreaking. Sending strength and support to everyone impacted. 🙏
@gulf_news Tragic news. Deepest condolences to the families & communities affected. Hoping for swift resolution and increased se...
NEW: The U.S. Army has launched an investigation after a military helicopter was flying low near Kid Rock’s Nashville home on Saturday, according to Reuters.
The helicopter flew by Rock's 27,000-square-foot White House replica home.
"A U.S. official confirmed that this indeed did happen and the Army is investigating how this happened," said Reuters' @idreesali114.
"An administrative review is underway to assess the mission and verify compliance with regulations and airspace requirements. Appropriate action will be taken if any violations are found," the Army said.
Floods continue to devastate parts of Pakistan. Heartbreaking to see the impact on communities. 🙏 Sending strength & hoping for swift aid & recovery efforts. #PakistanFloods #Pakistan
Most Popular Users
Elon Musk 
@elonmusk
241.5M followers
Barack Obama
@barackobama
119.1M followers
Cristiano Ronaldo
@cristiano
113.7M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.2M followers
Rihanna
@rihanna
98.5M followers
NASA
@nasa
92.4M followers
Justin Bieber
@justinbieber
91.7M followers
KATY PERRY
@katyperry
89.6M followers
Taylor Swift
@taylorswift13
83.5M followers
Lady Gaga
@ladygaga
75M followers
Virat Kohli
@imvkohli
72.7M followers
Kim Kardashian
@kimkardashian
70.7M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
65.7M followers
Bill Gates
@billgates
64.9M followers
Selena Gomez
@selenagomez
62.6M followers
The Ellen Show
@theellenshow
62.3M followers
CNN
@cnn
61.8M followers
X
@x
60.8M followers