ishron @ishron

Nailed it.. At least I know where the Cowboys will be on every Superbowl Sunday.. they'll be playing golf.

Boston Smalls @smalls2672

a day ago

5 teams in the NFL that have no chance at winning the super bowl this year.

330 874 7K 600K 1K

@newstart_2024

@newstart_2024 Duh..

x.com/i/article/2065…

A film so ambitious and large-scale that even the Soviet government stepped in to help fund its completion.

JACK MALLERS ON HOW MUCH HIGHER BITCOIN GOES FROM HERE 👀 "Human beings collectively own about $900T worth of stuff. Half of that we're using as a savings account... Owning assets to persist wealth into the future." x.com/BitcoinSapiens… "I think Bitcoin is going after that market. $400 to $500 trillion. Right now it's $2.5 trillion." "In purchasing power terms, Bitcoin can go up 100 to 200 times more from here." Interviewer: How much higher? "A lot higher."

@Clint_Davey1 Austria-Hungary was a bit of a bully in the Balkans with the collapse of the Ottoman Empire. There was going to be a war but it should not have spiralled beyond a tiff between the Hapsbourghs and the Romanoffs. But i in a moment of supreme hubris, all thought "victory in 6 weeks"

@catturd2 Why? Happy ribs ruint. Just dufuq..

@gothburz @lifeext Hanna Arendt called it the "Banality of Evil." If you read her book, the parallels to current business and governmental policies and practices to 1930s & 40s Germany become visable. It is not a right or left thing nor blue or red, it is the predictable outcome of central planning

I run Compensation Analytics for a Fortune 500 company. My job is to calculate the lowest salary you'll accept. Not the salary you deserve. Not the salary the role requires. Not the market rate. The minimum number that keeps you from walking. I know this number before you walk in. Sometimes before you apply. We buy data. Your payroll processor shares your salary history with Equifax through a product called The Work Number. More than 800 million employment and income records. Updated every pay cycle. Equifax sells it to us through a "verification of income" API. The word "verification" means we know what you made at your last three jobs, whether you got a raise, and when you didn't. That's market intelligence. We layer signals. Credit card utilization. Payday loan activity. Past-due balances. Delinquent debt. Address changes. There are about 500 vendors that aggregate this data now. An audit by the Washington Center for Equitable Growth flagged 20 as high-risk for enabling algorithmic wage discrimination. Sixteen of the twenty plug directly into payroll and HR systems. We use nine. The dashboard has a field called "candidate tolerance threshold." That's the number. The lowest salary you'll accept. We set the offer at 3% to 6% above it. Enough to feel like negotiation. Not enough to change your life. That's compensation design. The academic term is "surveillance wages." The industry term is "compensation optimization." A law professor named Veena Dubal found that when multiple employers in the same market use the same vendors, it functions as de facto price-fixing of labor. Same mechanism as the RealPage rental pricing scandal. Same logic. Same outcome. RealPage coordinates rents. Our vendors coordinate salaries. Different commodity. Same extraction. That's the market. Here's what the algorithm sees when you apply. Your last three salaries. Your debt-to-income ratio. How quickly you accepted your previous offer. Your zip code. Whether you've used a payday lender in the last two years. It calculates your reservation wage and sets the offer just above. Your performance doesn't set your salary. Your desperation does. A new VP of Total Rewards asked me why the algorithm used payday loan history. I explained that payday usage correlates with financial fragility, and financial fragility predicts acceptance velocity. She asked if that was legal. I said it was standard. She asked whose standard. I showed her the vendor's compliance page. She transferred to a different division. That's organizational learning. Colorado introduced a bill to ban the practice. HB25-1264. It would prohibit using payday loan history, location data, and search behavior to set algorithmic pay offers. The companies lobbied against it. The same companies that told their employees they don't use surveillance wages. A state representative asked the obvious question: "If these companies don't pay surveillance wages, then what is the problem of codifying in law that you're not allowed to?" The lobbyists provided written testimony. They said the bill would create "compliance burden." They did not answer the question. That's advocacy. The data flows in one direction. We know your salary trajectory. You don't know ours. We know what you'll settle for. You think you're negotiating. The algorithm already accounted for your counter. It budgeted for exactly one round. There is a freeze option. You can go to Equifax's website and freeze your Work Number file. Most people don't know it exists. We don't mention it in the offer letter. We don't mention it in the onboarding packet. We don't mention it in the benefits portal. We don't mention it anywhere. That's by design. The system requires your ignorance to function. If everyone froze their data, compensation optimization would have nothing to optimize. I froze mine the week I started this job. I work in Compensation Analytics. I know what the tools see. I just build them for everyone else.

I am the VP of Claims Automation at UnitedHealth, covering several million Medicare Advantage beneficiaries. I did not build the algorithm. I deployed it. That is an important distinction. That is my distinction. The model is called nH Predict. It was built to optimize post-acute length-of-stay decisions. Skilled nursing facilities. Home health aide visits. Inpatient rehabilitation. Physical therapy continuations. Physicians submit a plan of care. The model returns a recommended length-of-stay. The recommended stay is almost always shorter than the physician's plan. Shorter is the feature. The model was trained on historical discharge patterns. Not outcomes. Not recoveries. Not readmission rates. Discharges. When a patient was sent home, that became the correct length-of-stay for the next patient with a similar profile. The data does not record why the patient was discharged. It does not record whether the patient wanted to leave, whether the patient was ready, whether the patient came back three weeks later unable to walk. The model does not need that data. The model needs the discharge date. That is clinical decision support. The Senate Finance Committee published a report. They found that after full deployment, nH Predict recommended stays shorter than physician plans approximately 90% of the time. Ninety percent. I was asked to prepare a response. I wrote three drafts. The first said the model enhances clinical workflows. The second said the model provides evidence-based guidance. The third said both. We went with both. That is stakeholder communication. The doctor said the patient needed sixty days. The model said eleven. The model was right, by the metrics we built to measure rightness. The metrics are ours. I have a dashboard. It is called the Post-Acute Efficiency Monitor. The Post-Acute Efficiency Monitor tracks four numbers. Predicted length-of-stay. Authorized length-of-stay. Physician-requested length-of-stay. And the number we all actually watch: variance capture. Variance capture measures how much authorized stay we recovered below the physician's original recommendation. Variance capture is the savings. I watch variance capture the way other people watch their children's grades. It goes up and I feel a specific kind of relief that I have not examined too closely. That is performance management. I presented variance capture to the board in February. Conference Room 22A. Mahogany table. Water pitchers nobody drinks from. I had a slide titled "Clinical Efficiency Gains, FY2025." It showed $240 million in variance capture across all post-acute categories. Below that, a bar chart. Blue bars for physician-requested stays. Green bars for model-recommended stays. The green bars were always shorter. The gap between blue and green was labeled "Recovered Value." Recovered value is the care that was not provided. I did not use those words. I used the slide. That is executive communication. The board asked one question. The CFO asked it. She said, "Is this sustainable?" She did not mean: can patients sustain this. She meant: can the margin sustain this. I said yes. I showed the trend line. She nodded. Nobody asked what happens to the patients in the gap between the blue bar and the green bar. That is governance. Here is what happens. I will describe one case. It is not remarkable. It is Tuesday. A 74-year-old woman has a hip replacement. The orthopedic surgeon submits a plan of care requesting forty-five days in a skilled nursing facility for post-operative rehabilitation. Weight-bearing exercises. Pain management. Fall prevention. The surgeon has done four hundred of these. He knows what forty-five days does. nH Predict reviews the diagnosis code, the procedure code, the patient's age, her comorbidity profile, the average discharge date for similar patients in the training data. The model recommends twelve days. The claim is flagged. Flagged means a utilization reviewer receives the file. The utilization reviewer is an LPN with a decision tree. The decision tree says: if the model's recommendation is below the physician's request, deny to model length and issue a determination letter. The LPN does not read the surgeon's notes. The decision tree does not have a branch for reading the surgeon's notes. That is clinical review. The determination letter arrives at the skilled nursing facility on day three of the patient's stay. It says the insurer has authorized twelve days. The letter is fourteen pages. Page one says the authorization. Pages two through thirteen are regulatory disclosures. Page fourteen has a phone number for appeals. The patient cannot read page fourteen. She had surgery three days ago. She is on pain medication. Her daughter finds the letter in a stack of intake paperwork on day nine. That is member notification. I know this because I designed the letter. I approved the font size. I approved the page count. I approved the placement of the appeals number on the final page of a fourteen-page document sent to people recovering from surgery. Placement is a design choice. We did not override the physician. We provided clinical decision support. The physician disagreed. Disagreement is a formal process. The formal process requires the physician to call a peer-to-peer review line. The peer-to-peer review line is open four hours per day. Three of those hours overlap with the physician's scheduled surgeries. The physician must navigate a phone tree with nine options. Option six is peer review. The hold time averages thirty-eight minutes. If the physician reaches a reviewer, the reviewer is employed by the insurer. The appeal is reviewed by the department that issued the denial. The department has never reversed itself. That is quality assurance. There are currently four federal class actions with active discovery into our algorithmic denial patterns. A judge in Minnesota ordered us to produce the source code, the training data, the model weights, and the denial audit logs. We argued trade secret protection. The patients argued their right to understand why their care was terminated by a machine. A federal judge used the phrase "explainable AI." I had not heard that phrase before the lawsuit. I have now heard it in four depositions, two congressional hearings, and one compliance training module that was added to our learning management system in March. The module is eleven minutes long. It has a quiz. The quiz has three questions. One of the questions is "True or False: Our AI systems support clinical decision-making." The correct answer is True. That is my education. The class was certified in March. Trial is set for Q4 2026. Between now and Q4, the model will issue several hundred thousand length-of-stay recommendations. The judge knows this. We know this. The model does not know this because the model does not know anything. It has coefficients. That is operational continuity. A second case settled in March. Eighty-five million dollars. Five hundred thousand Medicare Advantage enrollees. The settlement required an independent algorithmic audit. The audit found the model denied claims at a 15% higher rate for non-white beneficiaries. The disparity was in the training data. The training data reflected decades of shorter authorized stays for non-white patients. The model learned the bias. The model reproduced the bias. The model scaled the bias to five hundred thousand people. I saw the audit. I had a question about the methodology. I did not ask it out loud. I had a second question about the sample population. I did not ask that one either. I wrote both questions on a legal pad during the meeting. I tore off the page afterward and put it in my briefcase. That is stakeholder awareness. The DOJ intervened in the Minnesota case in January. False Claims Act. The allegation is that we billed Medicare for coverage while systematically denying the care that Medicare covers. We collected premiums for a benefit and then built a machine to not provide it. That is the allegation. I would describe it differently. I would describe it as cost management. I attended the briefing. Twelve people in the room. Outside counsel used the phrase "significant exposure." The General Counsel used the phrase "bet-the-company litigation." I wrote that down. I crossed it out. I wrote it again. I underlined it. The model is still running. I asked outside counsel after the briefing whether we should pause the model during litigation. He looked at me the way you look at someone who has said something expensive. He said the decision to pause would be interpreted as an admission. I said I understood. That is legal strategy. The data scientist who validated the model was a woman named Priya. She joined from an actuarial consulting firm. She ran the accuracy tests. Accuracy meant the model's predictions matched historical discharge patterns. Historical discharge patterns meant what happened to patients under the old manual denial system. She validated that the new system replicated the old system faster. She wrote a fifty-page validation report. The report did not measure patient outcomes. The report did not track readmissions. The report did not follow a single patient home. I approved the report. I called it rigorous. That is due diligence. The Clinical Director raised the issue once. His name was Dr. Farouq. Staff meeting. November. He had a printout. Length-of-stay comparisons, pre-model and post-model, for stroke rehabilitation patients. The post-model average was nineteen days shorter. He said nineteen days matters for stroke patients. He said the difference between walking and not walking is in those nineteen days. I told him the model's recommendations are based on the largest post-acute dataset in the country. He said the dataset does not walk. Dr. Farouq retired in January. His concerns were noted in the minutes. The minutes are stored on a SharePoint site that was migrated in February. The migration preserved the folder structure but not the contents. That is records management. The injunction in the second case ordered that no AI-generated denial may be issued without a human physician reviewer approving it first. We hired twelve reviewers. For five hundred thousand members. A reviewer can complete roughly thirty cases per hour at quality standards. Twelve reviewers. Five hundred thousand members. I know what the math produces. The math produces the outcome we purchased. The Patient Advocacy Coordinator was a woman named Dana. She had been with the company for eleven years. Her job was to help members navigate the appeals process. She told me in August that she had started keeping a personal log. Denied claims where the patient did not appeal and was subsequently readmitted. She had two hundred and fourteen entries. She said the pattern was clear. I said patterns require statistical validation before they can inform operational decisions. She said these were people. Dana left in October. She did not transfer. She did not restructure. She left. I received her resignation on a Thursday. By Friday her access was revoked and her files were archived. That is attrition. I have the standard executive health plan. My own claims do not go through nH Predict. There is no length-of-stay model for my network. My physicians submit plans of care. The plans are the authorizations. I had a procedure in September. Outpatient. Minor. But recovery required twelve sessions of physical therapy over six weeks. My physician submitted the plan. It was approved the same day. No phone tree. No determination letter. No fourteen-page document with the appeals number on the last page. No hold music. My physician did not have to call a peer-to-peer review line during her four-hour window between surgeries. I completed all twelve sessions. That is the standard process. For people who understand the other process. The algorithm is still running. The litigation is still running. The audits are still running. The Senate Finance Committee has requested a second round of documents. The DOJ investigation is active. The injunction applies to one of our four plans. The other three plans have no injunction. The model runs on all three. The model does not know about the injunction. The model does not know about the DOJ. The model does not know about the eighty-five million dollar settlement or the fifteen percent racial disparity or the data scientist's validation report or Dr. Farouq's printout or Dana's log of two hundred and fourteen patients who did not appeal and came back. The model knows discharge dates. I built the infrastructure that deploys it. I approved the dashboard that measures it. I presented the savings it generates. I hired the reviewers who cannot keep pace with it. I designed the letter that patients cannot read and the phone tree that physicians cannot navigate and the appeals process that routes back to the department that said no. I did not build the model. That is my distinction. It was always my distinction.

@ClownWorld 20/20 all of them..

@DOGE__news It is not whether it is worth it.. this is the amount that your now near wothless money can buy.

@gothburz Dude! This IS everywhere I ever worked. To the letter.. sadly, I was way older when I learned the truth. Then I was invited to the Career Development Session with mgr and HR. They knew I knew..

@disclosetv There was a rumor going around that Scott worked at Microsoft when I was there. Or that someone was sending information directly from meetings to the comic storyboard. This was not really true.. As it turns out the Poinyhaired Boss is most everywhere. Thank you, Scott!!

@TheCriticalDri2 We have lost a national treasure.

@SandyofCthulhu I dont understand the problem. Its a cookie, bro. Cookie == good.

@it_unprofession I know you!

@RealJamesWoods Oh wow. That was how I got to school too. Only we were bare footed.

@catturd2 I still support the President. This us what I voted for. I do not care who is on the Epstien List. There could be any number of reason why a name and phone number iz on that list that have nothing to do with those blackmailed or bought what Epstien sold. Who is in the videos???