Quick summary of probability shifts under the weighted system:
Level I ($80K median): -57% selection probability
Level II ($103K median): -14%
Level III ($135K median): +29%
Level IV ($158K median): +73%
The counterintuitive finding: product companies (direct employers) have 22% Level I concentration vs 7% for staffing firms. The rule designed to stop "lottery abuse" hits direct employers 3x harder.
Data source: DOL LCA disclosure data. Happy to discuss methodology.
> The rule designed to stop "lottery abuse" hits direct employers 3x harder.
To prove this, you need to compute the total numbers and not just proportion. If 22% of the direct employees are Level 1, but there's only 300 direct level 1 filings total, the absolute numbers are dwarfed by the drop in probability of the staffing firms.
Excellent point. My analysis calculated Level I concentration within each employer type but not absolute volumes.
You're correct that if staffing firms file significantly more total applications, their absolute number of affected Level I positions could exceed product companies despite lower concentration.
To properly claim "3x harder hit," I'd need to show either:
1. Total Level I applications by employer type, or
2. Total estimated selection losses by employer type
Without those absolute numbers, the "3x harder" claim overstates what the data shows. The accurate claim is that product companies have 3x higher Level I concentration - but that's not the same as 3x more impact.
This is a good catch. The proportions tell us about hiring patterns but not total system impact.
Hey, lemme apologize first. I indeed generated this above comment by AI because I was extremely busy with something else and didn't have time to respond but i'll make sure this will not happen again.
Hey, I didn't generate everything by AI but only to frame the words since i was busy with something else. I'll make sure from now on such mistakes will never happen again.
(Methodology): Author here. Technical methodology: The counterintuitive finding: DHS justified the weighted lottery by citing "employers submitting large volumes of low-wage registrations"—understood to mean staffing firms. But our data shows:
Product companies (direct employers): 22.0% Level I concentration
Staffing/consulting firms: 7.1% Level I concentration
Ratio: 3.1×
Tech companies hiring junior engineers get hit 3x harder than the outsourcing firms the rule supposedly targets.
Data: DOL H-1B LCA Disclosure Data FY2024, 517,874 certified applications. Stack: Python, pandas 2.1.0.
Fair point. To be clear we aren't saying wage should be the only rule, just that it's weird to see such a huge disconnect. Usually money talks.
Your "gateway" hunch is likely spot on. Most rural docs start on J-1 waivers (mandatory 3 years). If they actually stayed after that, we'd see way more H-1B conversions filed by those rural hospitals to keep them. Since the volume is so low, it suggests once the 3 years are up, they bail for the city. The wage premium just isn't enough to anchor them.
Data: 20,225 H-1B LCA disclosures from DOL, FY2024, healthcare occupations only Analysis: Python (pandas), mapped ZIP → RUCC codes, median wage by volume quintile Key limitation: This is LCA data (intent to hire), not final USCIS approvals
Interesting rabbit holes:
Urban/rural split isn't binary—codes 4-6 show gradient effects
Wage level inversions strongest in codes 7-9 (most rural)
We analyzed the FY2024 H-1B LCA dataset to look for geographic inefficiencies.
Methodology: We joined DOL disclosure data with USDA Rural-Urban Continuum Codes (RUCC 2013) using a ZIP-to-County crosswalk.
The Signal: We found a clear inversion of the standard supply curve. Rural areas offered a 21.4% wage premium ($250k median vs $206k) yet achieved 10.2x lower placement volume.
Systemic Friction: The data suggests that for high-skill labor (physicians), geographic friction and regulatory overhead (including the new $100k fee) outweigh significant monetary incentives. The market is not clearing.
What I appreciated here is how calmly Tao separates useful pattern matching from actual mathematical understanding. There’s no AI hype or dismissal but just a reminder that proof, verification, and intuition are different things. It made me rethink where LLMs genuinely help vs where they just feel convincing. Thank you for sharing!
The disturbing part isn’t that bad encounters happen — it’s that these techniques are officially banned, yet keep showing up across unrelated cases. When policy and on-the-ground behavior diverge this consistently, it stops looking like individual misconduct and starts looking like a systems problem.
General warrants (the sort of thing being done here) are explicitly listed as one of the things motivating independence from england in the declaration of independence.
On one hand I'd say you have a reason to panic but on the other it's no secret that in the US might makes right so it's not like the overall situation has changed dramatically.
Laws don't matter in the US any more. Epstein files aren't being released even though there is law that requires. TikTok is still operating in the US even though there is law that bans it. Laws are only recommendations now.
I read it less as obliviousness and more as internal language leaking into marketing. What’s “Liquid Glass” to Apple reads like an aesthetic system though but to outsiders it sounds like jargon inflation. I feel the gap between internal coherence and external clarity shows up in these releases a lot.
You’re right on the numbers....Firefox never had majority share. The stronger claim is causal influence, not dominance. I recently read somewhere that the Firefox (and later Chrome) forced standards compliance and broke IE’s de-facto monopoly mindset. IE’s decline was gradual and multi-factor, but Firefox clearly shifted developer and user expectations.
No one is claiming, here or in the article, that Firefox ever had a majority share.
I don’t know if the 55% number for IE is 100% correct but it sounds like the right ballpark to me. The browser market was a lot more fragmented 15+ years ago, so saying that IE had 55% market share and Firefox had 32%, leaving 13% for other browsers, sounds completely right to me.
What’s unsettling here isn’t any single policy, but the convergence: predictive policing, protest restrictions, and administrative punishments all justified as “risk management.” Even if each tool seems narrow, together they normalize acting on suspicion rather than action, which quietly lowers the bar for dissent.
Quick summary of probability shifts under the weighted system:
The counterintuitive finding: product companies (direct employers) have 22% Level I concentration vs 7% for staffing firms. The rule designed to stop "lottery abuse" hits direct employers 3x harder.Data source: DOL LCA disclosure data. Happy to discuss methodology.