Å·±¦ÓéÀÖ

The system analyzed the work in real time, awarding credit points based on speed and quality feedback. At the front of the hotel ballroom, there was a leaderboard displaying the names of the workers with the most points.
It was clear to Jennifer and Michael that some of the workers in the trial were quicker learners. They observed Matt impatiently turning in all directions and chatting with fellow workers. Some looked like gamblers sitting before slot machines, hands in constant motion, expressions intoxicated. Noting his place on the leaderboard, one worker performed what seemed like a victory dance. Others looked glum.
"This is a video game, only the name of the game is 'work.'" Jennifer couldn't hide some disgust with the whole affair. The workers' reactions to the tech reminded her of her sad and defeated father.
Michael sat silently.
"That's it!" His reaction startled Jennifer from her reverie. He reached out and took Jennifer's hand. "Maybe this doesn't just seem like a video game. Maybe it actually is one. You need to help me find out."
"Find out what?"
"The architectural designs in this video... are they really being implemented somewhere?"
"You mean..." Jennifer suddenly

"Sahej, why? Why can't we get close to each other?" Nayana chose her words carefully.
Now it was Sahej's turn to look surprised. "Nayana, do you really not know?"
"Know what?"
"My last name."
"The schools and virtual classroom keep your surname protected just like you're the offspring of a big star or some famous family."
"On the contrary, it's because they don't want it to trigger any discomfort."
"What kind of discomfort?"
"In the past, it was described as a feeling of being polluted."
"You're talking about your caste? But that whole system was outlawed years ago."
Sahej gave a bitter laugh. "Just because it's no longer permitted by law and doesn't appear in the news doesn't mean it's gone."
"But how would the Al know about it?"
"The Al doesn't know. The Al doesn't need to know the definition of the castes. All it needs is its users' history. No matter how we hide or if we change our surnames, our data is a shadow. And no one can escape their shadow."
Nayana thought about what her mother had said, that Al only learns what humans teach it. She rolled the thought about in her head, then looked at Sahej. "So you're saying that Al identifies the invisible discrimination in our society and quantifies it."
Sahej's expression became serious, but he exhaled a soft laugh. "I almost forgot. There's also the color of my

A fire alarm provides a good analogy for the types of hypothesis testing errors. Ideally, the alarm rings when there is a fire and does not ring in the absence of a fire. However, if the alarm rings when there is no fire, it is a false positive, or a Type I error in statistical terms. Conversely, if the fire alarm fails to ring when there is a fire, it is a false negative, or a Type II error.

You learned that we do not accept the null hypothesis. Instead, we fail to reject it. The convoluted wording encapsulates the fact that insufficient evidence for an effect in our sample isn't proof that the effect does not exist in the population. The effect might exist, but our sample didn't detect it-just like all those species scientists presumed were extinct because they didn't see them.
Finally, I admonished you not to use the common practice of seeing whether confidence intervals of the mean for two groups overlap to determine whether the means are different. That process can cause you to overlook significant results and miss out on important information about the likely range of the mean difference. Instead, assess the confidence interval of the difference between means.

P-values indicate the strength of the sample evidence against the null hypothesis. If it is less than the significance level, your results are statistically significant.
This is because it is the probability of observing a sample statistic that is at least as extreme as your sample statistic when you assume the Null Hypothesis is correct,.
P-values are the probability that you would obtain the effect observed in your sample, or larger, if the null hypothesis is correct. In simpler terms, p-values tell you how strongly your sample data contradict the null. Lower p-values represent stronger evidence against the null.
If the p-value is less than or equal to the significance level, you reject the null hypothesis and your results are statistically significant. The data support the alternative hypothesis that the effect exists in the population. When the p-value is greater than the significance level, your sample data don't provide enough evidence to conclude that the effect exists.

Special methods in Python, often referred to as "magic methods" or "dunder methods" (short for "double underscore"), are methods that have double underscores at the beginning and end of their names, like __init__, __str__, and __add__. These methods are used to enable certain behaviors and operations on objects that are instances of a class. They allow you to define how your objects should respond to built-in functions and operators.

A hash is a fixed-size integer that uniquely identifies a particular value or object. Hashes are used in many areas of computer science and programming, such as in data structures like hash tables (which are used to implement dictionaries and sets in Python). The purpose of hashing is to quickly compare and retrieve data in collections that require fast lookups.
- Python provides a built-in hash() function that returns the hash value of an object.
- The hash() function works on immutable data types like integers, floats, strings, and tuples.
- For mutable types like lists or dictionaries, hash() is not applicable because their contents can change, making them unsuitable for use as keys in hash-based collections.
- The hash value of an immutable object (like a string or a tuple) remains constant throughout the program's execution, provided that the object itself doesn't change.
- Mutable objects cannot be hashed because their contents can change, leading to inconsistencies in hash values.
- Hashes are crucial for the performance of dictionaries and sets in Python. When you insert an item into a dictionary or set, Python uses the hash value to determine where to store the data. This allows for very fast lookups.
Dictionaries themselves are not hashable. This is because dictionaries in Python are mutable objects, meaning their contents can change after they are created
Keys must be hashable: This is a strict requirement because dictionaries use a hash table internally to store key-value pairs. The hash value of the key determines where the pair is stored.
Keys must be immutable: This immutability ensures that the hash value of a key remains consistent throughout its lifetime. If the key could change, it would disrupt the dictionary's internal hash table, leading to unpredictable behavior when trying to retrieve or store values.
The key in a dictionary cannot change directly because keys in a dictionary must be immutable. This means that once you assign a key to a dictionary, it must remain unchanged.
However, if you want to "change" a key, you would need to remove the old key-value pair and add a new key with the same value (or a modified value).

Interfaces, Protocols and ABCs
Duck typing is a dynamic typing concept based on the idea of "if it looks like a duck and quacks like a duck, it must be a duck." This means that the type or class of an object is determined by its behavior (i.e., the methods it implements) rather than its explicit type. In other words, if an object implements the necessary methods or properties required by the context, it's considered valid, regardless of its actual class or type.
No explicit type checks: Python doesn't require you to declare the type of an object, and duck typing allows you to use any object that has the required methods.
Runtime flexibility: Duck typing checks happen at runtime, so the system checks the object's methods or attributes during execution.
Error prone at runtime: Since there are no compile-time checks, errors related to missing methods are only discovered when the program is executed.

class Duck:
def quack(self):
print("Quack!")

class Dog:
def quack(self):
print("Bark that sounds like a quack!")

def make_it_quack(animal):
animal.quack() # Works as long as 'animal' has a 'quack' method

duck = Duck()
dog = Dog()

make_it_quack(duck) # Output: Quack!
make_it_quack(dog) # Output: Bark that sounds like a quack!

Goose typing (sometimes referred to as gradual typing) is an extension of duck typing that adds more structure by explicitly checking whether objects conform to a given interface, but without strictly requiring them to belong to a specific class or type. The term "goose typing" isn't as commonly used as duck typing or static typing but can be thought of as a middle ground between the two.
In Python, goose typing can be implemented through tools like protocols and abstract base classes (ABCs). These mechanisms allow objects to be treated as valid types if they conform to a specific interface, regardless of their concrete type.

Thinking for a moment, I add, "For that matter, do the same thing for every person assigned to Phoenix. I'm guessing we're overloaded, so I want to know by how much. I want to proactively tell people whose projects have been bumped, so they're not surprised when we don't deliver what we promised."
Both Wes and Patty look surprised. Wes speaks up first, "But...but we'd have to talk with almost everyone! Patty may have fun grilling people on what changes they're making, but we can't go around wasting the time of our best people. They've got real work to do!"
"Yes, I know they have real work to do," I say adamantly. "I merely want a one-line description about what all that work is and how long they think it will take!"
Realizing how this might come across, I add, "Make sure you tell people that we're doing this so we can get more resources. I don't want anyone thinking that we're outsourcing or firing anyone, okay?"
Patty nods. "We should have done this a long time ago. We bump up the priorities of things all the time, but we never really know what just got bumped down. That is, until someone screams at us, demanding to know why we haven't delivered something."
She types on her laptop. "You just want a list of organizational commitments for our key resources, with a one-liner on what they're working on and how long it will take. We'll start with all Phoenix and audit remediation resources first, but will eventually cover the entire IT Operations organization. Do I have it right?"

I look at Erik suspiciously. He supposedly couldn't get anyone's name right, and yet he apparently remembers the name of some security guard from years past. And no one ever mentioned anything about a Dr. Reid.
After climbing five flights of stairs, we're standing on a catwalk that overlooks the entire plant floor, looking like it goes on for at least two city blocks in every direction.
"Look down there," he says. "You can see loading docks on each side of the building. Raw materials are brought in on this side, and the finished goods leave out the other. Orders come off that printer down there. If you stand here long enough, you can actually see all the WIP, that's 'work in process' or 'inventory' for plant newbies, make its way toward the other side of the plant floor, where it's shipped to customers as finished goods."
"For decades at this plant," he continues, "there were piles of inventory everywhere. In many places, it was piled as high as you could stack them using those big forklifts over there. On some days, you couldn't even see the other side of the building. In hindsight, we now know that WIP is one of the root causes for chronic due-date problems, quality issues, and expediters having to rejuggle priorities every day. It's amazing that this business didn't go under as a result."
He gestures broadly with both arms outstretched, "In the 1980s, this plant was the beneficiary of three incredible scientifically-grounded management movements. You've probably heard of them: the Theory of Constraints, Lean production or the Toyota Production System, and Total Quality Management. Although each movement started in different places, they all agree on one thing: WIP is the silent killer. Therefore, one of the most critical mechanisms in the management of any plant is job and materials release. Without it, you can't control WIP."
He points at a desk near the loading docks closest to us.

He shakes his head, recalling the memory, "He sat down at the keyboard, and it's like he went into this trance. Ten minutes later, the problem is fixed. Everyone is happy and relieved that the system came back up. But then someone asked, 'How did you do it?' And I swear to God, Brent just looked back at him blankly and said, 'I have no idea. I just did it."
Wes thumps the table and says, "And that is the problem with Brent. How the hell do you document that? 'Close your eyes and go into a trance?"
Patty laughs, apparently recalling the story. She says, "I'm not suggesting Brent is doing this deliberately, but I wonder whether Brent views all his knowledge as a sort of power. Maybe some part of him is reluctant to give that up. It does put him in this position where he's virtually impossible to replace."
"Maybe. Maybe not," I say. "I'll tell you what I do know, though. Every time that we let Brent fix something that none of us can replicate, Brent gets a little smarter, and the entire system gets dumber. We've got to put an end to that.

An ever-growing pile of changes trapped inside of IT Operations, with us running out of space to post the change cards.
Work piling up in front of the heat treat oven, because of Mark sitting at the job release desk releasing work.
Work piling up in front of Brent, because of...
Because of what?
Okay, if Brent is our heat treat oven, then who is our Mark? Who authorized all this work to be put in the system?
Well, we did. Or rather, the CAB did. Crap. Does that mean we did this to ourselves?
But changes need to get done, right? That's why they're changes. Besides, how do you say no to the onslaught of incoming work?
Looking at the cards piling up, can we afford not to?
But when was the question ever asked whether we should accept the work? And on what basis did we ever make that decision?
Again, I don't know the answer. But, worse, I have a feeling that Erik may not be a raving madman. Maybe he's right. Maybe there is some sort of link between plant floor management and IT Operations. Maybe plant floor management and IT Operations actually have similar challenges and problems.
I stand up and walk to the change board. I start thinking aloud, "Patty is alarmed that more than half our changes aren't completing as scheduled, to the extent that she's wondering whether this whole change process is worth the time we're investing in it.
"Furthermore," I continue, "she points out that a significant portion of the changes can't complete because Brent is somehow in the way, which is partially because we've directed Brent to reject all non-Phoenix work. We think that reversing this policy is the wrong thing to do."
I take a mental leap, following my intuition. "And I'd bet a million dollars that this is the exact wrong thing to do. It's because of this process that, for the first time, we're even aware of how much scheduled work isn't getting done! Getting rid of the process would just kill our situational awareness."
Feeling like I'm getting on a roll, I say adamantly, "Patty, we need a better understanding of what work is going to be heading Brent's way.

I turn back to Patty and say slowly, "Let me guess. Brent didn't get any of his non-Phoenix change work completed either, right?"
"Of course not! You were there, right?" she says, looking at me like I had grown eight heads. "Brent was working around-the-clock on the recovery efforts, building all the new tooling to keep all the systems and data up. Everything else was put on the back-burner."
All the firefighting displaced all the planned work, both projects and changes.
Ah... Now I see it.
What can displace planned work? Unplanned work. Of course.
I laugh uproariously, which earns me a look of genuine concern from Patty, who even takes a step back from me.
That's why Erik called it the most destructive type of work. It's not really work at all, like the others. The others are what you planned on doing, allegedly because you needed to do it.
Unplanned work is what prevents you from doing it. Like matter and antimatter, in the presence of unplanned work, all planned work ignites with incandescent fury, incinerating everything around it. Like Phoenix.
So much of what I've been trying to do during my short tenure as VP of IT Operations is to prevent unplanned work from happening: coordinating changes better so they don't fail, ensuring the orderly handling of incidents and outages to prevent interrupting key resources, doing whatever it takes so that Brent won't be escalated to...
I've been doing it mostly by instinct. I knew it was what had to be done, because people were working on the wrong things. I tried to take all necessary steps to keep people from doing wrong work, or rather, unplanned work.
I say, cackling and pumping my arms as if I had just scored a game-winning, sixty-yard field goal, "Yes! I see it now! It really is unplanned work! The fourth category of work is unplanned work!"

Technologies
Framing the Problem to Maximize Business Impact: can be more important than identifying data sources. Strive to produce a more significant impact by not only providing hindsight, insight, and foresight for business decisions but also by driving customer actions with predictive intelligence capabilities.
Discovery Patterns in Data
- Understanding Data Characteristics: unit of decisioning, sample size, data sparsity, outliers, sample imbalance, data types
- Innovating in feature Engineering to summarization
- Clarifying the Modeling Strategy: Momentum-based, Foundational, Reflexive or Hybrid modeling strategies
- Provide clarity in modeling strategies that are coherent with fundamental mechanisms of the domain.

Execution
- In specifying projects, avoid being task-oriented, and ask the question behind the question to take personal responsibility to produce the best business results in projects.
- In prioritizing, planning, and managing projects, assess projects in terms of reach, impact, confidence, and effort (RICE) and alignment to data strategies; address common project failure modes with a simple and concise project plan; and leverage waterfall or scrum techniques, depending on your project, to manage your projects.
- In balancing hard trade-offs between speed and quality, safety and accountability, and documentation and progress, improve the long-term productivity of the team.
- Project Motivation and Definition, Solution Architecture, Execution Timeline, Risk to Anticipate
Expert Knowledge
In recognizing business context, check for project alignment with the vision and mission of your organization: Deep Domain Understanding
- Clarify the business context of opportunities or risks
- Account for domain data sources nuances, biases, inaccuracies, incompleteness
- Navigate organizational structures in an industry

Being Al-first means doing Al last. Doing Al means doing it last or not doing it at all. The reason is rather simple: Solution-focused strategies are more complex than problem-focused strategies; and solution-focused thinking ignores the most important part of business, which is the problems they solve and the customers they create.
Keep in mind that solution-centric thinking results from the following:
Focusing on what our solutions ought to be rather than what they are.
Focusing on the impact of future solutions rather than the future impact of today's solutions.
Conflating our goals with the goals of others.
Focusing too much on abstract problems with some arbitrary solution or focusing too much on someone else's problem and ignore your problems. The former is solution solving and the latter often means we are working problem solving backward and finding problems to solve in the context of someone else's solution.
Do not define your solution. The search for analytical exactitude in verbal definition will not lead to economic progress. Ignore recycling glib, textbook definitions of artificial intelligence mainly because consumers don't care about textbook definitions. Customers care about themselves. If you want to make your life better, make their lives better. Help them accomplish their goals in a better, faster, safer, or cheaper way. They're generally interested in a value proposition that contains problem-specific information, not in a definition of intelligence. Your journey starts with more comprehension of problems, not the names or definitions of solutions. Besides, creating definitions for our solutions means we are creating external goals for them, which is nonsense.

Remember that insiders seek epistemological discoveries, not economic ones. The more epistemological a pursuit is, the less likely it is to become something that could be turned into a business. Entrepreneur, venture capitalist, and author Paul Graham discusses the value of problems at length and explains that good business ideas are unlikely to come from scholars who write and defend dissertations.
The reason is that the subset of ideas that count as "research" is so narrow that it's unlikely to satisfy academic constraints and also satisfy the orthogonal constraints of business. The incentives for success in the academic world are not consistent with what it takes to start and grow a business. Ultimately, business pursuits are much more complicated than academic ones. Managers ought to acknowledge that solving intelligence is not likely your goal, and in many ways it's oppressive to problem solving. AGI may be possible, but it is not desirable as a business goal.

There is something magical about writing down a problem. It's almost as though by writing about what is wrong, we start to discover new ways of making it right. Writing things down will also remind oneself and our teams of the problem and the goal. Once a problem is written down, don't forget to come back to the problem statement. It is a guide. Problem solving often starts with great intentions and alignment, but when it counts most-when the work is actually being done-we often don't hold on to the problem we set out to solve, and that's the most important part of problem solving: what the problem is and why we are solving it to begin with.
Furthermore, do not needlessly seek out complexity by making larger solutions to solve needlessly bigger problems. Complexity bias is the logical fallacy where we find it easier to seek out complex solutions rather than a simple one. Without a problem statement, solutions tend to become more complex and expand to fill in the available time we've allocated for problem solving. Parkinson's law, named after Cyril Northcote Parkinson, states that "work expands so as to fill the time available for its completion." This is a sort of solution sprawl, similar to the urban sprawl that expands to fill in geographic spaces immaterial to how well the urban landscape serves it citizenry.

Always start small and take small steps to ensure that performance is what you want. Don't try to boil the ocean with the whole of a problem. With smaller steps almost everything can be reduced to something more manageable. Working in smaller sizes and smaller steps goes for your team as well. Rather than having your whole team work on something for six months, think about what one person can do in six weeks. The Basecamp team uses six weeks, which I think is a good size. If you are an Agile team, you may have batches of two weeks.75 That is fine, too. The point is that constraining batch size will force everyone to find the best bad solution, rather than working into the abyss of perfection.
Of course, simple problems are different. Simple problems can often be solved by applying a single solution to the whole of the problem. In practice you may not know the best solution a priori. One strategy to find the best solution for a simple problem may be to simply guess. Guessing, however, will have a high error rate in the face of increasing complexity.

Machine learning science had different goals from statistics. Whereas statistics emphasized being correct on average, machine learning did not require that. Instead, the goal was operational effectiveness. Predictions could have biases so long as they were better (something that was possible with powerful computers). This gave scientists a freedom to experiment and drove rapid improvements that take advantage of the rich data and fast computers that appeared over the last decade.
Traditional statistical methods require the articulation of hypotheses or at least of human intuition for model specification. Machine learning has less need to specify in advance what goes into the model and can accommodate the equivalent of much more complex models with many more interactions between variables.
Recent advances in machine learning are often referred to as advances in artificial intelligence because: (1) systems predicated on this technique learn and improve over time; (2) these systems produce significantly more-accurate predictions than other approaches under certain conditions, and some experts argue that prediction is central to intelligence; and (3) the enhanced prediction accuracy of these systems enable them to perform tasks, such as translation and navigation, that were previously considered the exclusive domain of human intelligence. We remain agnostic on the link between prediction and intelligence. None of our conclusions rely on taking a position on whether advances in prediction represent advances in intelligence. We focus on the consequences of a drop in the cost of prediction, not a drop in the cost of intelligence.

The main takeaway is that value is created by making decisions, not by data or prediction. But the

The main takeaway is that value is created by making decisions, not by data or prediction. But these are necessary inputs to make AI-and-data-driven decisions decisions. To create value and make better decisions in a systematic and scalable way we need to improve our analytical skills.

Descriptive Analytics
What it is: involves analyzing historical data to understand what has happened in the past. It focuses on summarizing and presenting data in a meaningful way and relate to the current state of the business objective.
Example: Generating reports, dashboards, or visualizations that show key performance indicators (KPIs) over a specific time period.

Predictive Analytics
What it is: involves using statistical algorithms and machine learning techniques to identify the likelihood of future outcomes based on historical data. It helps organizations anticipate what might happen in the future.
Example: Building a predictive model to forecast sales for the next quarter based on past sales data and other relevant factors.

Prescriptive Analytics
What it is: goes a step further by recommending actions to optimize or improve future outcomes. It considers the predicted outcomes and suggests the best course of action, to help choose the right levers.
Example: Recommending marketing strategies to maximize the predicted sales, taking into account various influencing factors.

Business objectives are usually already defined, but we must learn to ask the right business questions to achieve these objectives.
Always start with the business objective and move backward: for any decision you're planning or have already made, think about the business objective you want to achieve. You can then move backward to figure out the set of possible levers and how these create consequences that affect the business.
The sequence of why questions can help define the right business objective you want to achieve: this bottom-up approach generally helps with identifying business objectives and enlarging the set of actions we can take. But other times you can also use a top-down approach similar to the decomposition of conversion rates.

Probabilities represent the likelihood of different outcomes occurring. In predictive analytics, estimating probabilities involves using statistical models and data analysis to quantify the chance of various future events.

Expected values, also known as mean or average values, are calculated by multiplying each possible outcome by its probability and summing up these values. It provides a measure of the central tendency of a probability distribution.

Expected utility is a concept in decision theory that combines the probabilities of different outcomes with the associated utility or value of each outcome. It helps in making decisions under uncertainty by considering both likelihood and desirability.