Wednesday, April 1, 2026
more parameters at low-precision is better than less parameters at high-precision
Saturday, March 21, 2026
The AI Race
Frontier intelligence in the West has become all about extrapolation and scaling laws. Meanwhile, those who are compute-constrained are left to gather the crumbs. I am incredibly grateful to Chinese AI companies for open-sourcing their models. It feels like the world has turned upside down in the tech race: the US used to lead in open-source tech, but now they have doubled down on proprietary solutions.
Geopolitics has divided the world of technology that once united us. Maybe it is time to adapt to the tech ecosystem created by our neighbor to the north...
Tuesday, October 7, 2025
Wednesday, October 1, 2025
C2PA should be mainstream
As AI-generated images become the norm in society, it is imperative that we develop a method of provenance for certain media.
The C2PA (Coalition for Content Provenance and Authenticity) is one such system. It creates a secure, tamper-evident record of a piece of content’s origin and history. When a photo is taken by a camera, edited in software, or generated entirely by an AI model, C2PA can embed verified details like who or what created it, what tools were used, and whether it has been altered. This information travels with the file itself, protected by cryptography so it can’t be easily stripped or forged without detection.
One problem with this, though, is that it cannot prevent media that is screen-recorded. This just means that provenance is only as strong as the ecosystem that supports it. C2PA works best when platforms, devices, and creators ALL PARTICIPATE in preserving and honoring the chain of trust. And while it can’t stop every form of deception, it can establish a baseline of accountability for content that moves through compliant channels that people rely on, like the news and social media platforms.
Read More: C2PA | Verifying Media Content Sources
Thursday, September 11, 2025
Stages in a Generic Process Model for Data Science Projects
Introduction
Data
science projects require systematic approaches to transform raw data into
actionable insights. While various frameworks exist in the industry, most share
common fundamental stages that guide practitioners from initial problem
conception to deployment and maintenance. This report examines the typical
stages found in generic data science process models, drawing from established
methodologies including CRISP-DM (Cross-Industry Standard Process for Data
Mining), TDSP (Team Data Science Process), and KDD (Knowledge Discovery in
Databases).
The Six Core Stages of Data Science Projects
Stage 1: Business Understanding and Problem Definition
The
foundation of any successful data science project begins with thoroughly
understanding the business context and clearly defining the problem to be
solved. This stage involves translating business objectives into data science
questions that can be answered analytically. Project stakeholders must
establish success criteria, identify constraints, and assess feasibility.
For
example, a retail company seeking to reduce customer churn would first need to
define what constitutes churn in their context (no purchase in 90 days versus
account closure), determine acceptable prediction accuracy levels, and
establish how predictions will be integrated into business operations. This
stage often requires multiple iterations as stakeholders refine their
understanding of what's possible given available data and resources.
According
to Provost and Fawcett (2013) in "Data Science for Business," this
initial stage determines project success more than any technical consideration,
as misaligned objectives lead to technically sound but practically useless
solutions.
Stage 2: Data Collection and Acquisition
Once
objectives are clear, the focus shifts to identifying and gathering relevant
data sources. This stage encompasses both discovering what data exists within
the organization and determining what external data might enhance the analysis.
Data scientists must evaluate data accessibility, quality, volume, and
relevance to the defined problem.
Consider a
healthcare organization developing a patient readmission prediction model. They
would need to collect electronic health records, demographic information,
treatment histories, and potentially external data such as socioeconomic
indicators from census databases. Each data source requires different
acquisition methods, from database queries to API calls to manual collection
processes.
The
challenge extends beyond mere collection to ensuring proper data governance,
privacy compliance (especially with regulations like GDPR or HIPAA), and
establishing sustainable data pipelines for ongoing projects.
Stage 3: Data Exploration and Preparation
Data
exploration and preparation typically consumes 60-80% of a data scientist's
time, according to surveys by CrowdFlower (2016) and subsequent industry
studies. This stage involves understanding data characteristics through
statistical summaries and visualizations, identifying quality issues, and
transforming raw data into analysis-ready formats.
Exploratory
Data Analysis (EDA) reveals patterns, anomalies, and relationships that inform
subsequent modeling decisions. For instance, discovering that customer purchase
patterns follow strong seasonal trends would influence both feature engineering
and model selection strategies. Data scientists create visualizations ranging
from simple histograms to complex correlation matrices to understand variable
distributions and relationships.
Data
preparation encompasses cleaning (handling missing values, correcting errors),
transformation (normalization, encoding categorical variables), and feature
engineering (creating derived variables that better capture underlying
patterns). A financial fraud detection project might create features such as
"transaction velocity" (transactions per hour) or "deviation
from typical spending patterns" rather than using raw transaction amounts
alone.
Stage 4: Modeling and Algorithm Development
The
modeling stage involves selecting appropriate algorithms, training models, and
optimizing their performance. This iterative process begins with establishing a
baseline using simple approaches before progressing to more sophisticated
techniques. Data scientists must balance model complexity with
interpretability, considering the trade-offs between accuracy and
explainability.
For a
customer segmentation project, the progression might start with simple k-means
clustering, advance to hierarchical clustering for better interpretability, and
potentially incorporate deep learning approaches if the data complexity
warrants it. Each approach requires different preprocessing steps,
hyperparameter tuning strategies, and validation methodologies.
Model
selection depends on multiple factors including data characteristics (volume,
dimensionality, structure), problem type (classification, regression,
clustering), performance requirements (accuracy versus speed), and deployment
constraints (real-time versus batch processing). Cross-validation and careful
train-test splitting ensure models generalize well to unseen data rather than
memorizing training patterns.
Stage 5: Evaluation and Validation
Rigorous
evaluation extends beyond simple accuracy metrics to encompass business
relevance, statistical significance, and practical feasibility. Data scientists
must select appropriate metrics aligned with business objectives—for instance,
in medical diagnosis, minimizing false negatives might outweigh overall
accuracy considerations.
Evaluation
involves multiple perspectives: statistical performance (accuracy, precision,
recall, F1-score), business impact (revenue increase, cost reduction), and
operational feasibility (computational requirements, integration complexity). A
recommendation system for an e-commerce platform would be evaluated not just on
prediction accuracy but on metrics like click-through rates, conversion rates,
and customer satisfaction scores in A/B testing scenarios.
Validation
strategies must also address potential biases, ensure fairness across different
population segments, and test model robustness under various conditions. This
includes stress testing with edge cases and evaluating performance degradation
over time as data distributions shift.
Stage 6: Deployment and Monitoring
Successful
deployment transforms analytical models into operational systems that deliver
ongoing business value. This stage requires collaboration between data
scientists, engineers, and IT operations to integrate models into existing
infrastructure while ensuring scalability, reliability, and maintainability.
Deployment
strategies vary from simple batch scoring systems that run periodically to
complex real-time prediction services handling millions of requests. A credit
scoring model might be deployed as a REST API integrated into loan application
systems, requiring considerations for response time, failover mechanisms, and
version control.
Post-deployment
monitoring becomes crucial as model performance can degrade when real-world
data diverges from training distributions—a phenomenon known as model drift.
Automated monitoring systems track prediction accuracy, data quality, and
system performance, triggering alerts when metrics fall below thresholds.
Regular retraining schedules and feedback loops ensure models remain relevant
and accurate over time.
The Iterative Nature of Data Science
While
presented sequentially, these stages form an iterative cycle rather than a
linear progression. Insights from modeling might necessitate collecting
additional data, evaluation results could reframe the original problem
definition, and deployment experiences often reveal new requirements. The
CRISP-DM model, widely adopted across industries, explicitly recognizes this
cyclical nature, with arrows connecting non-adjacent phases to represent common
iteration patterns.
Agile
methodologies increasingly influence data science workflows, promoting rapid
prototyping, continuous stakeholder feedback, and incremental delivery. Rather
than attempting to perfect each stage before proceeding, teams develop minimum
viable models that evolve through successive iterations based on real-world
performance and changing business needs.
Conclusion
Understanding
these fundamental stages provides a roadmap for managing data science projects
effectively, though specific implementations vary based on organizational
context, project complexity, and domain requirements. Success requires not just
technical proficiency in each stage but also strong project management, clear
communication with stakeholders, and flexibility to adapt as understanding
deepens throughout the project lifecycle.
References
Chapman,
P., Clinton, J., Kerber, R., Khabaza, T., Reinartz, T., Shearer, C., &
Wirth, R. (2000). CRISP-DM 1.0: Step-by-step data mining guide. SPSS Inc.
Provost,
F., & Fawcett, T. (2013). Data Science for Business: What you need to know
about data mining and data-analytic thinking. O'Reilly Media.
Microsoft
Team Data Science Process Documentation. (2020). Azure Architecture Center.
Microsoft Corporation.
Kelleher,
J. D., & Tierney, B. (2018). Data Science. MIT Press.
CrowdFlower.
(2016). Data Science Report: The View from Data Scientists. Figure Eight Inc.
Fayyad, U.,
Piatetsky-Shapiro, G., & Smyth, P. (1996). From data mining to knowledge
discovery in databases. AI Magazine, 17(3), 37-54.
Sunday, September 7, 2025
LLMs As An Emerging Technology
Saturday, August 2, 2025
The Good, The Bad, and Anthropic
This new article by Anthropic is basically saying "good" and "bad" personality traits are just vectors, as everything truly is for the state of AI today.
Larger Insects
I came across this tweet by Yishan Wong about his kid's ramblings.
After our planet's inception, this was actually possible. During the Carboniferous period (~300 million years ago), oxygen levels were much higher (~35% compared to 21% today), and giant insects like dragonflies with 70 cm wingspans could survive.
Currently, large insects the size of a car or even a dog are not feasible on Earth today due to a combination of factors. Let's review these one by one and determine what changes we would need in each respective element to make large insects a reality.
1. No Lungs = Impossible ...?
Insects do not have lungs. They rely on a tracheal system; it is basically a network of tubes that diffuses oxygen directly to their tissues. However, this works efficiently only at small sizes because diffusion is slow and becomes less effective over longer distances.
However, giant insects existed during the Carboniferous period, most likely because of the high oxygen content in the atmosphere. Thus, if we want to bring back those times, I propose that we should start converting the carbon dioxide in our atmosphere to increase the oxygen content through an artificially managed large-scale photosynthesis program.
2. Exoskeletal Issues
Insects have exoskeletons, not internal bones. As body size increases, (1) volume and weight grow faster than surface area (square-cube law), (2) the exoskeleton would become too heavy and brittle to support a large body, and (3) it would also crush under its own weight or break when trying to move, jump, fly, or perform other mechanical actions.
Suppose we tried to maintain their shape and structure (or at least maintain their insect-like qualities). In that case, the material that composes their exoskeleton must have a stronger microstructure.
3. Heat Regulation
Much like the lobsters Yishan's son mentioned, larger bodies expend more energy, and thus, generate more heat and lose it more slowly. Insects don’t sweat or have internal thermoregulation like mammals. Giant insects right now would likely overheat quickly in direct sunlight or during exertion, especially during these times that we are experiencing a rapid increase in atmospheric temperature brought by global warming.
If we could bring down the global temperature by some degree, larger insects would be one more step closer to being feasible.
These solutions are definitely difficult to achieve and require systematic cooperation among countries. However, if we so desire that larger insects roam our Earth again, we must do this together. I am no entomologist, but I empathize with some of them in their desire for larger insects as domesticable creatures.
Sunday, July 27, 2025
GRPO and Fixed RL Algorithm on Sequence Models
Group Sequence Policy Optimization
Could not understand more than half of this, and I had to rely on what others are saying, but this is really good for LMs trained using RL. Try your hand at understanding their paper and let me know what you found!
Wednesday, July 23, 2025
Reviewing IT Proficiency Learning Materials
Happy 23rd of July! It is 6 PM, and I am writing to you all about my recent progress with reviewing for TOPCIT. Currently, I have only saved the PDF document and done some light reading on Book 3: Overview of System Architecture.
Each of the books is less than 300 pages, but there are 6 of them. Do you think I'll be able to finish all of them by the end of the week (July 26)? I don't think so, but I can certainly try!
Honestly, it's been a while since I've done some reading... About a month ago already. Might take a while to speed things through and I'm still at page 24... sighh...
Wish me luck.
Tuesday, July 22, 2025
Midnight Vent
Tuesday, August 1, 2023
Otium Cum Dignitate
Monday, July 31, 2023
On Avoiding a Climate Disaster: What I Think
We should stop messing about with fake commitments that make us feel better & actually solve the climate problem. Renewables, nuclear energy, and efficient energy storage are needed. Then, we must solve the problems on scalable carbon dioxide fixation/recycling.
Read the book How To Avoid A Climate Disaster by Bill Gates.
After reading, you will realize that:
- We don't need Ecosia,
- We don't need to delete our emails,
- We don't need to turn off our electricity,
- We don't need to turn to veganism,
- We don't need to save water.
Those things will not change the system. The people that advocate such are simply putting the blame on YOU, and certainly on the poorest of us. The #1 activity that humans do where the majority of the greenhouse gas come from is the manufacturing of products. So, should we blame the companies instead?
Not quite.
Everyone is to blame, but we do the things we do because it makes us happy and content with life. We don't need to be in survival mode. Instead, we need something like a global Manhattan Project where the greatest human minds come to remove carbon dioxide & drop the greenhouse gas to essentially zero or as near to zero as can be managed. We have already surpassed the 1.5°C limit, so the only route is to burn as much fossil fuel as we need to:
a) get to a new, green energy economy
b) maintain economic development, and
c) build breakthrough technologies that will remove CO2 on a global scale.
Thursday, July 27, 2023
Quantum Physics - Laymen Get Wild
Quantum physics is a surprisingly broad field with quite a few people working on different problems... yet when a study gets publicly released, enthusiasts and interested individuals get scooped most of the time!
Like, remember that article hypothesized in 2020 by a particle physicist that explains the existence of parallel universes based on evidence found in a NASA experiment in Antarctica? Daily Mail, NY Post, and other non-science journalism sites went wild and people just accepted the hypothesis to be true! Of course, it's been debunked already, but it was a fun week.
Anyway, what I'm trying to say is that there's great demand for these kinds of studies. It's obvious because Marvel movies rely on these concepts as plot devices. I'm betting that Marvel Studios would start funding quantum physics research once they've used up the studies published in the 20th Century hahahaha
Nice ra sad ang The Marvels, at least in the trailer. Quantum entanglement this time 😂
Wednesday, June 21, 2023
Subtracting Numbers Without Borrowing
In this blog, we will be tackling one section in the book in Chapter 13 which talked about subtraction sans the concept of borrowing. People may call this "subtraction by addition." One interesting fact about this is that this method was commonly used in mechanical calculators, and is still used in some modern computers to save up on memory and increase the speed of calculations (These are mostly used by computers that require precise and quick memory processing such as those used in medicine, aeronautics, and surveillance for Google services to harvest your personal data). If you know how to add and subtract, and have familiarized yourself with basic arithmetic necromancies, then this will be easy for you. Let's dive into it so we can get a better look.
Hitting the Books: A Review on Subtraction
Here's a typical structure of a subtraction operation.- Minuend: The minuend is the number from which another number is to be subtracted. It is the number from which the subtrahend is taken. For example, in the subtraction problem 9 - 5 = ?, the minuend is 9.
- Subtrahend: The subtrahend is the number that is to be subtracted from the minuend. Continuing with the previous example, in the subtraction problem 9 - 5 = ?, the subtrahend is 5.
- Difference: The difference is the result of subtracting the subtrahend from the minuend. It is the answer or the numerical value that remains after the subtraction operation. In the example 9 - 5 = ?, the difference is 4, as 9 minus 5 equals 4.
The Beta Way
If you have been listening to your teachers and absorbing what they taught you without understanding the concept behind it, then you're a Beta learner. Nothing wrong with that. It's pretty much the least hassle way of achieving a decent grade in class."Di nato dapat libugon atong kaugalingon kay daghan na kaayo'g problema sa kinabuhi."
Let's plug in our desired numbers for our variables in the operation. What I'm desiring right now is 489 subtracted from 653, so let's plug those in. While we're at it (and also because I'm too lazy to insert another image), we should also get the difference.
I'm sure that, according to conventions, the first thing to do is focus on the first column at the rightmost. We know that 3 (in minuend) is smaller than 9 (in subtrahend), so our first step involves borrowing from the adjacent digit. In this case, our 3 borrows from the digit 5, resulting in 3 becoming 13. Subtracting 9 from 13 yields 4, which represents the value of our difference's first digit from the right. Moving to the middle column, we need to consider the previous borrowing from 5, so 5 is actually a 4. Same concept: we borrow from the digit left to 4, so 6 becomes 5, and 4 becomes 14. We then subtract 14 from 8 to get 6. For the final digit, our 6 has become 5, and by subtracting this from 4, we get a 1. With all that, we get a difference of 164.
The Sigma Way
I may call this the Sigma Way, but it's really not. It's pretty much because computers don't know the concept of borrowing values. If you think about it, there's no mathematical proof that explains borrowing. We just know that it's a neat trick for easier subtraction. Now, you might be asking how mathematicians and computer scientists got computers to subtract values without borrowing. Well, through their smarts and cunningness, they simply used the method of complements. For decimal (base 10) numbers, we use the nines' complement. For binary, we use the ones' complement. For this blog, let's just focus on using the nines' complement since decimal numbers are generally used by people. You may look up the process of ones' complement after reading this post.For this example, let's reuse what we had in the previous section.
- Step 1: Get the nines' complement by subtracting the value of our subtrahend from 999.
- Step 2: Add the difference value of Step 1 to our minuend.
- Step 3: Add 1 to our sum from Step 2.
- Step 4: Subtract 1000 from the sum from Step 3.
Then, instead of adding both sides by 1000, we simply add 10000. Afterward, we break the positive 10000 down to 9999 and 1. Other phases of the process will still be the same. While we're here, I should also mention a case with varying numbers of digits. For example, 782 subtracted from 4001:
- Step 1: Get the nines' complement of our subtrahend.
- Step 2: Add our nines' complement to our minuend.
- Step 3: Add 1 to our sum.
- Step 4: Subtract 1000 from our sum.
Another case I want to discuss is an operation yielding a negative difference. For example, 5481 is subtracted from 4263. The process here is slightly different as it is a special case. Two things will depend if an end carry* occurred when you added the subtrahend's complement and minuend. If it has no end carry, then the complement of the sum of the subtrahend's complement and minuend is obtained and the negative sign is added.
- Step 1: Take the 9's complement of subtrahend.
- Step 2: Add the 9's complement to the minuend.
- Step 3: Because there is no end carry that occurred when we added the subtrahend's 9's complement and minuend (Step 2), we must obtain the sum's 9's complement.
- Step 4: Because there is no end carry that occurred when we added the subtrahend's 9's complement and minuend (Step 2), we have determined that the difference is negative.
Final Words
Congratulations! You have successfully graduated from being a Beta. Now onwards with the way of a Sigma in subtracting numbers, and flex your skills at parties and other social gatherings.Although I could have included more content in this post, I must admit that laziness got the better of me. Plus, I think the post ended up being longer than I anticipated. Anyway, I'm no expert on this topic, but I'll make an effort to provide the best possible answers to any questions that may arise in the comments. Feel free to raise them!
Thank you for reading. Avante!
Monday, June 19, 2023
Rebranding!
"It is God Who commands it."- Joan of Arc
Photo by Morgan Petroski on Unsplash
Friday, June 2, 2017
This Week's Video
Monday, April 10, 2017
PLAYLIST OF MY VIDEOS!
FULL PLAYLIST OF MY VIDEOS
From G Return (of course)
Like literally, have you even tried taking a look on the 'Playlist' Tab?
Just For Fun! (Funny Edits)
https://www.youtube.com/playlist?list=PLxl3Ti08JpZnILKpnnfF2dAI4n3EPQ7aN
WAIT, IT'S AN UPDATE!
You Know What?
Minecraft PE | Chronicles of Meraghyl [2]
G Return | Minecraft PE | Chronicles of Meraghyl [2]
Complicated isn't it? Just like love...
























