Ruo Shui
LightNews LightNews
banner
ruoshuiresearch.bsky.social
Ruo Shui
@ruoshuiresearch.bsky.social
2K followers 200 following 6.7K posts
stay hydrated

https://ruoshui.substack.com/
Posts Replies Media Videos
ruoshuiresearch.bsky.social
November 2025 is when gamers started to feel the memory squeeze
A vertical bar chart shows the sharp decline in inventory levels at suppliers of dynamic random-access memory (DRAM) — the main type of memory used in computers, servers, and phones — from October 2024 to October 2025. Chart Details Y-axis: Weeks on average (range 0–15) X-axis: Time period from October 2024 to October 2025 Each bar represents the average number of weeks of DRAM inventory at suppliers. October 2024: 15 weeks Late 2024: 15.5 weeks (peak) Early 2025: 9.5 weeks Mid 2025: 5.5 weeks October 2025: 3 weeks DRAM supplier inventories fell from around 15 weeks in late 2024 to just 3 weeks by late 2025 — a fivefold decline over the year. This drop suggests tightening supply conditions in the semiconductor memory market, possibly due to rising demand from AI data centers and constrained production. Note: Figures represent midpoints of ranges (e.g., an inventory range of 2–4 weeks is shown as 3). Source: TrendForce, reported by Brenda Goh and Wen-Yee Lee. A table titled “2025–2026 Projected Demand Growth for Smartphones, Laptops, and Game Consoles” compares Revised vs Previous year-over-year (YoY) demand forecasts for 2025 and 2026 across three device categories. The table has four main columns: 2025 Revised 2025 Previous 2026 Revised 2026 Previous Smartphone Demand 2025: Revised +1.6%, previously +0.5% 2026: Revised –2.0%, previously +0.1% Notebook (Laptop) Demand 2025: Revised +3.6%, previously +3.5% 2026: Revised –2.4%, previously +1.7% Game Console Demand 2025: Revised +5.8%, previously +9.7% 2026: Revised –4.4%, previously –3.5% Source: TrendForce
December 3, 2025 at 2:17 PM
ruoshuiresearch.bsky.social
noting the discourse on the other platform
it’s not niche, ai polls really terribly. openai in particular seems to be approaching villain status. this will pose real political problems - roon
December 2, 2025 at 3:34 PM
ruoshuiresearch.bsky.social
wages vs energy by location would be interesting to chart

globalization/phones/internet reduced the distance of craft beer and sushi

www.jstor.org/stable/41262...
A grayscale vertical bar chart titled “Figure 4. Wage rate relative to price of energy” compares how high wages are relative to energy prices across six cities. X-axis: Cities — Amsterdam, London, Paris, Strasbourg, Newcastle, Beijing Y-axis: Wage rate relative to the price of energy (scale from 0 to 6) Data Values (Approximate from Bar Heights) Amsterdam: ~1.4 London: ~1.8 Paris: ~0.6 Strasbourg: ~0.8 Newcastle: ~5.0 (highest value) Beijing: ~0.3 (lowest value)
November 26, 2025 at 6:40 PM
ruoshuiresearch.bsky.social
A vertical bar chart titled “Lion’s Share of Delisted Homes Were Last Purchased 2–5 Years Ago” shows when sellers last purchased homes that were delisted in September 2025. X-axis: Time since the home was last purchased by the seller Categories: 0–2 years ago, 2–5 years ago, 5–10 years ago, 10+ years ago Y-axis: Percentage of delisted homes (0–35%) 0–2 years ago: 13% 2–5 years ago: 34% (largest share) 5–10 years ago: 23% 10+ years ago: 29% Most homes removed from the market in September 2025 were purchased between 2–5 years ago, implying many recent owners—likely those who bought during or just after the pandemic-era boom—are pulling listings rather than selling into a slower market. Older homeowners (10+ years) still represent a substantial share of delistings. Source: Datawrapper.
November 25, 2025 at 2:29 PM
ruoshuiresearch.bsky.social
A scatter plot titled “ARC-AGI-2 Leaderboard” visualizes model performance (score vs. cost) on the ARC Prize benchmark, comparing reasoning models by cost per task ($) on a logarithmic x-axis and score (%) on the y-axis. Axes X-axis (log scale): Cost per task, ranging from $0.001 to $1,000. Y-axis: Score (%), ranging from 0% to 55%. Data Points by Model Anthropic Claude (Opus 4.5, Red Triangles) Opus 4.5 (Thinking, None): ~5% score, ~$0.15 per task Opus 4.5 (Thinking, 8K): ~14% score, ~$0.8 Opus 4.5 (Thinking, 16K): ~24% score, ~$1 Opus 4.5 (Thinking, 32K): ~30% score, ~$2 Opus 4.5 (Thinking, 64K): ~38% score, ~$4 OpenAI (GPT-5.1, Blue Triangles) GPT-5.1 (Thinking, None): ~3% score, ~$0.08 GPT-5.1 (Thinking, High): ~18% score, ~$2 Google Gemini (Green Triangle) Gemini 3 Deep Think (Preview): ~47% score, ~$90 Source: ARC Prize Verified Leaderboard (Nov 2025).
November 25, 2025 at 1:40 AM
ruoshuiresearch.bsky.social
A horizontal bar chart titled “Powell Tenure Marked by Few Dissents Until Now” compares how often Federal Reserve Chairs faced dissenting votes per meeting since 1979. Table Details Chair Term Average Dissents per Meeting Jerome Powell 2018–present 0.328 (shortest bar) Janet Yellen 2014–2018 0.688 Ben Bernanke 2006–2014 0.727 Alan Greenspan 1987–2006 0.536 Paul Volcker 1979–1987 1.227 (longest bar) Powell’s bar is the shortest, reflecting the fewest average dissents per FOMC meeting among modern Fed Chairs. Volcker’s bar is the longest, showing the most frequent dissenting votes during a period of aggressive inflation control. Intermediate chairs (Yellen, Bernanke, Greenspan) fall between 0.5–0.7 dissents per meeting. Source: Federal Reserve Bank of St. Louis; Bloomberg calculations.
November 23, 2025 at 4:29 PM
ruoshuiresearch.bsky.social
A line chart titled “Tracking federal expenditures in real time” compares U.S. federal outlays in 2024 vs. 2025, based on daily and monthly Treasury data. X-axis: Months from January to November. Y-axis: Billions of U.S. dollars, ranging from 0 to 7000. Two trend lines: Teal line (2024) — federal outlays in 2024. Purple line (2025) — federal outlays in 2025. Both years show nearly identical upward trajectories, reflecting steady cumulative spending growth through the year. By mid-year (June–July), 2025 spending slightly surpasses 2024 levels. By October–November 2025, total cumulative outlays reaching nearly $7 trillion, marginally higher than the same point in 2024. Source: U.S. Department of the Treasury.
November 23, 2025 at 4:15 PM
ruoshuiresearch.bsky.social
@metr.org estimate that Kimi K2 Thinking has a 50%-time-horizon of around 54 minutes (95% confidence interval of 25 to 100 minutes) on their agentic SWE tasks.

Note that they conducted this evaluation through a third-party inference provider, which reduces their confidence in this estimate.
A line chart from METR (Model Evaluation and Training Reports) titled “Kimi K2 Thinking on Novita AI has a 50%-time-horizon of about 54 min (95% CI: 25–100 min)”. It visualizes how the task length at 50% success rate has increased across large language model (LLM) generations, showing exponential progress in reasoning time-horizon since 2019. X-axis: Model release date (2019–2026). Y-axis: Task length (at 50% success rate) on a logarithmic scale, ranging from 1 second up to 4 hours. Example task labels: “Answer question,” “Count words in passage,” “Find fact on web,” “Train classifier,” and “Train adversarially robust image model.” Each green dot represents an LLM release (GPT-2 through GPT-5.1-codex-max). GPT-2 (2019): ~1 second time-horizon. GPT-3 (2020): ~4 seconds. GPT-3.5 (2022): ~1 minute. GPT-4 (2023): ~15 minutes. GPT-4o (2024): ~1 hour. GPT-5 and o-series (2025): Extend beyond 1 hour, approaching 2–3 hours. GPT-5.1-codex-max (forecast): ~2 hours 40 min. Kimi K2 Thinking (pink dot): Positioned slightly below GPT-5 but within the same order of magnitude—median ~54 minutes (confidence interval 25–100 minutes). The figure illustrates that modern reasoning models’ “thinking horizon”—the length of tasks they can sustain with consistent accuracy—has expanded exponentially, roughly doubling every 6–7 months. Kimi K2 Thinking’s performance places it at the frontier of 2025-era systems, bridging toward models expected to sustain multi-hour coherent reasoning sessions in 2026. Source: metr.org
November 21, 2025 at 4:06 PM
ruoshuiresearch.bsky.social
11/20/2025, one for the history books

H/T @ljkawa.bsky.social
A bar chart titled “Joy turns to ashes” visualizes rare market events where the S&P 500 ETF (SPY) opened the trading day at least 1.5% higher but closed the day at least 1.5% lower. X-axis: Four specific trading dates — Nov 20, 2025, Apr 8, 2025, Oct 9, 2008, and Oct 7, 2008. Y-axis: Percentage change in SPY, ranging approximately from +3% to –7%. Green bars (Open): Represent the percentage SPY opened above the prior close. Red bars (Close): Represent the negative percentage drop by market close. Values by Date Nov 20, 2025: Opened around +1.7%, closed around –1.7%. Apr 8, 2025: Opened roughly +2.8%, closed around –1.8%. Oct 9, 2008: Opened about +2.2%, closed sharply –6.5%. Oct 7, 2008: Opened roughly +2%, closed –4.5%. The chart highlights extreme intraday reversals during both 2025 and the 2008 financial crisis, implying investor optimism at open turned rapidly into selling pressure by close. Source: Bloomberg Visual credit: Sherwood
November 21, 2025 at 3:23 AM
ruoshuiresearch.bsky.social
keeping an eye on adult women unemployment rate
November 20, 2025 at 1:35 PM
ruoshuiresearch.bsky.social
A vertical bar chart titled “Figure 3. Share of TPS workers in foreign-born labor force by state” compares the percentage of Temporary Protected Status (TPS) workers among the foreign-born labor force across U.S. states. Axes Y-axis: Share of TPS workers in the foreign-born labor force (percentage, 0%–4%). X-axis: U.S. states ordered from highest to lowest share. Highest shares: Florida (~3.8%) leads, followed by Maryland (~3.4%), Indiana (~2.7%), Virginia (~2.6%), and Tennessee (~2.3%). Moderate shares: North Carolina, Utah, Texas, Georgia, and New York range between roughly 2%–1.5%. Lower shares: Massachusetts, New Jersey, Ohio, California, Connecticut, Illinois, Pennsylvania, Michigan, and Washington fall below 1.5%, with Washington near the bottom around 0.5%. The concentration of TPS workers is highest in Southeastern states (Florida, Maryland, Virginia, Tennessee) and lowest in northern industrial and coastal states. Florida’s outsized share reflects its large Central American and Caribbean migrant communities—key TPS populations. Western states like California and Washington, despite large immigrant labor forces, show lower relative TPS representation, suggesting more diverse visa or residency status mixes. Source: Penn Wharton Budget Model (PWBM) calculations based on 2022–2023 American Community Survey (ACS) data. Note: Only states with over 100 unweighted TPS observations are shown.
November 20, 2025 at 1:05 PM
ruoshuiresearch.bsky.social
7% expected move
A grid of twelve small line charts displays stock price reactions around quarterly earnings announcement dates labeled “amc” (after market close). Each panel shows a short-term price trend with the one-day move (%) highlighted in green for positive and red for negative. Each chart includes a small time-series of stock price around the earnings date and avertical blue line marking the announcement. The percentage one-day move in green (gain) or red (loss). The exact dates for the move window (e.g., “Aug 27–Aug 28”). All y-axis scales are removed, emphasizing direction and volatility. Top Row (2025) Aug 27, 2025: –0.79% decline May 28, 2025: +3.25% gain Feb 26, 2025: –8.48% sharp drop Nov 20, 2024: +0.53% mild gain Middle Row (2024) Aug 28, 2024: –6.38% drop May 22, 2024: +9.32% strong rise Feb 21, 2024: +16.41% surge Nov 21, 2023: –2.44% decline Bottom Row (2023–2022) Aug 23, 2023: +0.08% neutral May 24, 2023: +24.36% major spike Feb 22, 2023: +14.02% strong jump Nov 16, 2022: –1.45% small drop
November 19, 2025 at 2:38 PM
ruoshuiresearch.bsky.social
Gemini 3 models have made a significant 2X SOTA jump on ARC-AGI-2 (Semi-Private Eval) Gemini 3 Pro: 31.11%, $0.81/task Gemini 3 Deep Think (Preview): 45.14%, $77.16/task
November 18, 2025 at 5:08 PM
ruoshuiresearch.bsky.social
BofA FMS consensus is overweight international equities BofA FMS average cash level dropped to 3.7% from 3.8%. Note cash levels of 3.7% or lower has occurred 20 times since 2002. BofA FMS investors said “private equity/private credit,” as the highest conviction on source of a systemic credit event since we started asking the question in 2022. November saw an uptick of concerns around an AI...asked about the biggest 'tail risk' for the economy and the markets, 45% of FMS investors said “AI bubble” (up from 33% last month).
November 18, 2025 at 2:05 PM
ruoshuiresearch.bsky.social
H/T @aatishb.bsky.social
A set of four line charts illustrates August international student arrivals from Asia, China, India, and South Korea between 2015 and 2025. Top Chart – Asia (aggregate) Y-axis: Number of students, ranging from 0 to 300,000. Student arrivals remained around 250K–275K from 2015 to 2019. Sharp collapse during COVID-19 (2020), dropping below 100K. Partial recovery from 2021 to 2023, peaking near 250K, followed by a renewed decline to around 190K in 2025. Marked with a red downward arrow noting 59,561 fewer students (–24%) compared to the previous year. Bottom Three Charts – Breakdown by Country China Y-axis: Up to 200,000 students. Stable near 150K before 2019. Sharp COVID dip around 2020, brief rebound afterward. Gradual decline resumes, falling about 12% by 2025. India Y-axis: Up to 100,000 students. Starts near 50K in 2015, dips during COVID. Surges to almost 100K in 2023–2024, then plummets 44% into 2025. South Korea Y-axis: Up to 40,000 students. Modest fluctuations between 15K–25K. Small decline of –11% by 2025. Total Asian student arrivals are down 24% year over year, with India driving most of the decline. The data suggests a regional pullback in U.S.-bound higher education migration after a temporary rebound post-COVID, potentially due to visa delays, cost pressures, or shifting geopolitical and economic factors. Source: Trade.gov.
November 17, 2025 at 3:06 PM
ruoshuiresearch.bsky.social
Morgan Stanley has cut PC OEMs’ outlooks across the board due to rise in memory prices
November 17, 2025 at 2:17 AM
ruoshuiresearch.bsky.social
A horizontal bar chart titled “Active Inventory is Higher Than Pre-Pandemic Average in 22 of 75 Markets” shows how housing inventory levels in September 2025 compare to pre-pandemic averages (September 2018–2019) across 75 major U.S. metro areas. Axes and Structure Y-axis: Percentage difference in active listings (ranging from +60% to –60%). X-axis: 75 metro markets, sorted left to right from highest to lowest change in inventory. Bars are color-coded by region: Gray – National average Blue – South Yellow – West Dark Blue – Midwest Green – Northeast Around 22 markets show higher inventory in 2025 than before the pandemic, with gains between +5% and +60%. The top metros (e.g., New Orleans, Austin, Raleigh, Nashville, Tampa) show the largest surpluses, mostly concentrated in the South and West. These areas likely reflect new construction growth or weaker demand relative to rapid pre-2023 booms. Middle range (near 0%): A few metros (roughly 5–10) hover around parity with pre-pandemic levels, showing normalized inventory conditions. The majority of metros (over 50) remain below pre-pandemic inventory, many by –20% to –60%. The sharpest shortages are seen in Northeast and Midwest metros such as Hartford, Milwaukee, Providence, Boston, and Rochester, where housing supply remains constrained. Source: Zillow, First American Calculations (September 2025).
November 15, 2025 at 12:33 AM
ruoshuiresearch.bsky.social
even more interesting when combined with the fact that more experienced workers are more likely to accept agent-generated code
Two side-by-side scatterplots titled “(d) Years of Work Experience” analyze how job applicants’ years of experience affect AI agent usage and relative accept rate. Both plots include error bars showing confidence intervals. Left Chart – Agent Usage by Work Experience X-axis: Years of experience (3, 6, 9, 12, 15, 18, 21, 24). Y-axis: Weekly usage (%), ranging from about 47.5% to 62.5%. Agent usage is roughly flat between 3 and 18 years of experience, fluctuating near 50–52%. Slight increase at 21 and 24 years, peaking near 60%, but with wider confidence intervals, indicating higher variability among more experienced individuals. AI tool or agent usage does not vary strongly with experience, though very senior workers may use it more sporadically or variably. Right Chart – Relative Accept Rate by Work Experience X-axis: Years of experience (same as left). Y-axis: Relative accept rate, centered around 1.00 (baseline). Clear positive correlation: relative accept rate rises steadily with experience. Starts near 0.90 for early-career applicants (3 years) and increases to 1.10 for those with 24 years. Error bars are tight, suggesting consistent results. More experienced applicants have significantly higher acceptance rates, possibly due to perceived skill depth or better alignment with employer expectations.
November 14, 2025 at 2:46 PM
ruoshuiresearch.bsky.social
Small investors are defined as those with less than 10 properties, medium investors less than 100, large less than 1000, and mega as greater than 1000
A multi-line chart titled “Share of investor purchases by investor size, January 2018 – September 2025” illustrates how different categories of property investors—small, medium, large, and mega—have changed their share of total home purchases over time. Axes and Variables Y-axis: Share of purchases (%) from 0% to 18%. X-axis: Time from January 2018 through September 2025, with dates spaced roughly quarterly. Legend (Color-coded): Small Investors: Purple (top line) Medium Investors: Magenta/pink Large Investors: Orange Mega Investors: Blue 1. Small Investors (Purple – top line): Began near 12% in early 2018. Rose steadily through 2021, spiking to around 15–16% in 2022. Peaked during the pandemic-era housing boom (2021–2022), when smaller operators dominated bidding on single-family homes. Declined modestly to about 13–14% by late 2025. Remain the largest investor group by share. 2. Medium Investors (Pink – second-highest line): Stayed relatively flat at 3–4% from 2018 to 2021. Began rising rapidly in 2023–2025, reaching about 9–10%, suggesting growing institutional participation in mid-sized portfolio acquisitions. 3. Large Investors (Orange): Hovered near 1% for most of the period. Showed a gradual uptick beginning around 2023, ending near 2% by 2025. 4. Mega Investors (Blue): Began below 1%, peaked briefly near 2.5% in 2021–2022—likely during Wall Street and REIT-led bulk acquisitions—then fell back under 1% in 2023 before rebounding to around 2% in late 2025. Source: Cotality Public Records Data.
November 14, 2025 at 2:13 PM
ruoshuiresearch.bsky.social
A line chart titled “Natural-gas futures price, weekly” shows U.S. natural gas futures prices from 2021 through late 2025, measured in dollars per million British thermal units ($/MMBtu). Y-axis: $1 to $10 per MMBtu. X-axis: Time, from 2021 to 2025 (marked annually). Line: Dark blue, showing weekly average futures prices. 2021: Prices begin around $2.50/MMBtu, rising steadily through the year. Early 2022: Sharp climb peaking around $6/MMBtu, coinciding with global energy disruptions following Russia’s invasion of Ukraine. Mid-to-late 2022: Prices spike even higher—above $9/MMBtu, the highest level in the chart—then decline sharply toward $6/MMBtu by year’s end. 2023: A steep collapse to about $2/MMBtu, marking a normalization as supply balances improve. 2024: Prices fluctuate between $1.80 and $3.50/MMBtu, staying historically low and stable. 2025: Gradual rise resumes, with a recent sharp uptick approaching $5/MMBtu, suggesting tightening supply or increased demand entering winter. Source: FactSet
November 14, 2025 at 3:29 AM
ruoshuiresearch.bsky.social
turning into a coin toss
November 13, 2025 at 2:14 PM
ruoshuiresearch.bsky.social
cursed streaming bundle chart
November 12, 2025 at 6:31 PM
ruoshuiresearch.bsky.social
seems to jive with this year-old BoA survey
Chart titled “Six options outrank traditional investments, for younger investors” compares what younger (ages 21–43) and older (ages 44+) investors perceive as the greatest opportunities for growth. Ages 21–43 (Top 3): Real estate investments (31%) Crypto/digital assets (28%) Private equity (26%) Ages 44+ (Top 3): Domestic equities (41%) Real estate investments (32%) Emerging market equities (25%) The only major area of alignment is real estate, ranking in the top two for both age groups.
November 12, 2025 at 2:14 PM
ruoshuiresearch.bsky.social
looking at the shitcos today
November 11, 2025 at 2:54 PM
ruoshuiresearch.bsky.social
A two-part financial analysis slide presenting data center electricity consumption forecasts and AI infrastructure investment estimates. The top section, “Figure 3: Data Center Power Use (Installed Capacity)”, shows forecasts from Semianalysis, JPMorgan, McKinsey, and the IEA, while the lower section, “Figure 15”, estimates capex tied to OpenAI’s hardware partnerships with NVIDIA, AMD, and Broadcom. Top Table – Figure 3: Data Center Power Use (Installed Capacity) High-end growth scenario (<Semianalysis>) Category 2024F 2028F Increase Growth CAGR Comments Global AI MW 12,000 79,250 +67,250 60% Global Non-AI MW 46,500 63,750 +17,250 8% Total MW 58,500 143,000 +84,500 25% USA AI MW 8,499 56,280 +47,781 60% USA Non-AI MW 19,798 27,175 +7,377 8% Total MW (USA) 28,297 83,455 +55,158 31% Mid-range growth scenario (<JPM EE/MI MW model>) Category 2024F 2029F Increase Growth CAGR Global AI MW 12,117 94,637 +82,521 51% Global Non-AI MW 47,080 64,959 +17,879 7% Total MW 59,197 159,596 +100,400 22% Low-end growth scenario (<IEA>) Category 2024F 2030F Increase Growth CAGR Global IT Equipment GW 68 174 +106 17% Global Power Demand (TWh) 416 946 +530 15% USA IT Equipment GW 31 81 +50 17% USA Power Demand (TWh) 183 426 +243 15% Source: IEA, McKinsey, Semianalysis, Company reports, J.P. Morgan estimates. Bottom Table – Figure 15: Estimated Capex Associated with OpenAI’s Recent Capacity Agreements Partner Scenario DC Capex per GW ($B) Compute Partner Revenue per GW ($B) Capacity (GW) Total DC Capex ($B) Total Partner Revenue ($B) NVIDIA High 50 30 10 500 300 Low 60 35 10 600 350 AMD High 40 15 6 240 90 Low 50 25 6 300 150 Broadcom High 45 25 10 450 250 Low 55 30 10 550 300 Total Estimated (All Partners): High Case: $1,190B capex, $640B partner revenue Low Case: $1,450B capex, $800B partner revenue Aggregate Capacity: 26 GW Source: J.P. Morgan estimates based on company disclosures.
November 11, 2025 at 2:13 PM