Subject: On practical machine learning, taking you from the 101 to being aware of practical edge cases before using in business settings. The parts of MLE you wouldn’t get from coding tutorials or math.
Recomend:
Over
Machine Learning Engineering by Andriy Burkov. Not quite as full or well written as Huyen’s, but did cover slightly different topics (eg always need cohort plots; use Kaplan-Meir to estimate cohorts; more talk of agreement metrics; how to calibrate models and how calibrated models are easier to monitor)
Machine Learning Design Patterns by Michael Munn, Sara Robinson, and Valliappa Lakshmanan. Was basically only an introduction to Big Query syntax and GCP offerings. Though the section on data split was a little better than the other 2.
"The system will always grind you down", but you're allowed to opt out and it's getting easier all the time.
You give the constraints of the middle class worried they'll slip down, not the poorest Americans directly. If 12% of Americans live in poverty, then the example concerns are those of the 12-50th percentiles. You say the lumpen-proletariat will always exist, but prove that the lives of the proletariat won't improve much.
"credentialist colleges that raise their prices to capture more and more of the returns to the credential, until huge portions of the former middle class's early-life earnings" - Explicitly not the poorest.
"Like working 60-hour weeks to afford rent somewhere the school districts will damage your child less" - but a lot of people's kids do go to the worst school districts. I'd say if your kid goes to a bad school that's poverty, not the person who sends their kid to a better school.
"rents in San Francisco would almost instantly rise by the amount of the UBI; no janitors in the Bay Area would be better off as a result" - But housing prices over all of the US won't rise by the amount of UBI. Some would move out of SF to cheaper areas. They've been given an option they preferred so their lives have improved. UBI only does them no good if everyone choose to stay in SF, but why would they?
"maybe the Ferguson police departments of the world, just raise their annual quota for fines per capita by the per capita UBI." - This assumes the policing level is set by the wealth of the poor. I think it's set by the level that generates political opposition. Claims there's on 12% as many arrest warrants after 6 years, clearly a political decision not a reduction in wealth. The Police in Beverly Hills are not stationary bandits nor are SF Cops making more in fines. I don't think it's a law of history that governments automatically grind down any surplus the poorest have.
The example of Anoxistan relies on the people not being able to afford the most necessary Oxygen, if they had to work 60 hour weeks to get cars, and some people choose to forgo cars, then that's clearly an improvement.
The idea that poverty is squeeze from the tightest constraint, not the average of constraints is interesting, but you don't talk about the tightest constraint on the poorest. I suspect those have eased up: look at average hours worked, or the labor participation rate, or stories of the voluntarily homeless who work 3 months a year. All indicate that some people choose not to work much, thus their constraints are smaller.
"the addicted mind will find a way to rationalize continued use at all costs"
Alan Carr wrote a series of books: "The easy way to quit X". I picked up one since I figured he had found a process to cure addictive behaviors if he could write across so many categories. I highly recommend it. The main points are:
It's a process easily transferable to any other activity.
I found that installing 3 website blockers at once on my laptop worked for me: the same way you use multiple anti-biotics at once to combat resistance. I might’ve trained myself to disable 1, then disable 2, then disable 3 if I added another whenever I realized the blocking wasn’t working: but by adding all 3 at once I wasn’t going to unblock all 3. Now it’s been long enough I’ve probably forgotten what I’d even need to do to unblock it.
The one thing the duality section was connected to was Riesz representation theorem. Riesz states every finite linear functional φ has a unique vector f, such that for all v, φ(v) = <v,f>. It gives an isomorphism from functionals to vectors for a given norm, as the function is just multiplication with the vector.
It's not tied to the section on duals in the text, but the section on duals lets you appreciate the result more.
The converse of fast paced conversations leading people to say stupid things is: if someone says something foolish it maybe not be a lie nor a tendency to BS nor stupidity. They may have responded faster than they thought. You can correct them not by refuting what was said but allowing them a moment to reconsider. A liar, politician, and fool are hard to reform or work with. “Stopping to think” may be easier trained and much easier used to keep a conversation on track.
The 2000-2021 VIX has averaged 19.7, sp500 annualized vol 18.1.
From a 2ndary source: "The mean of the realistic volatility risk premium since 2000 has been 11% of implied volatility, with a standard deviation of roughly 15%-points" from https://www.sr-sv.com/realistic-volatility-risk-premia/ . So 1/3 of the time the premia is outside [-4%,26%], which swamps a lot of vix info about true expect vol.
-60% would the worst draw down ever, the prior should be <<1%. However, 8 years have been above 30% since 1928 (9%), seems you're using a non-symetric CI.
The reasoning for why there'd be such a drawdown is backwards in OP: because real rates are so low the returns for owning stocks has declined accordingly. If you expect 0% rates and no growth stocks are priced reasonably, yielding 4%/year more than bonds. Thinking in the level of rates not changes to rates makes more sense, since investments are based on current projected rates. A discounted cash flow analysis works regardless of how rates change year to year. Currently the 30yr is trading at 2.11% so real rates around the 0 bound is the consensus view.
Options Nitpick: You can't use equity index* option prices as true probabilities because adding hedges to a portfolio makes the whole portfolio more valuable. People then buy options based on their value when added to the portfolio, not as individual investments.
The first reason option hedges make your portfolio more valuable is preferences: people don't just want to maximize their expected return, but also reduce the chance they go broke. People don't like risk and hedges reduce risk, ergo they pay more to get rid of risk. However you can't just subtract X vols to adjust as this "risk premia" isn't constant over time.
Secondly hedges maximize long term returns (or why you shouldn't sell options) You want to maximize your average geometric annual return not average annual return. You care about geometric averages because if for 3 years your returns were +75%, +75%, -100%, your don't have 50% more money then when you started but 0. The average of annual returns was 10.7% over the past 30 years, but if you'd invested in 1992 you'd've only compounded at 8.5% year till 2022.
Geometric returns are the the nth root of a product of n numbers and have the approximation = Average Annual Return - (variance/2). If you could reduce variance and not reduce Annual returns, your portfolio (market + hedges) would grow faster than the market.
These reasons are why despite the worst Annual return being -48% in 1931, you say there's a 5% chance of > -50% returns based on option markets.
*I'm specifically talking index options because that's the portfolio investors have (or something similar) and the total is what they care about. If you were to use prices as true probabilities for say a merger going through these reasons don't apply as much and would be more accurate.
PS. I've referred to investors as all having the same portfolio because most people do have highly correlated index holdings and it's at this level of generality you can think about investors as a class.
I think you're asking the VLLM to do too much in a single call.
I was trying to get the VLLM to extract GD&T data from the NIST standardized renders of ASME Y14 tolerancing standards (eg. screenshotting 1 page at a time from https://www.nist.gov/system/files/documents/noindex/2022/04/06/nist_ftc_06_asme1_rd.pdf ).
Asking "List all GD&T and their matching element IDs" would totally fail, but would start to be reasonably when I only asked for 5-8 specific element ids at a time, and asking for the data of a single element id at a time with few shot examples would be ~90% accurate.
While right now you have to build the scaffold for them to perform well, eventually an agent could realize it needs to build it's own scaffold or the model would be able to effectively use all parts of the context window.
I'd be curious how well the models do if you wrote out a 7 step list, and then asked about each step in isolation, and then asked a model to summarize the results of 7 separate calls.