Consumption
2026
This is meant to be an ongoing list of very short reviews as I read, watch, and listen to new things over time. I’ll include a brief backlog for the past below this. Started this in August 2026, so ongoing from this month. And at some point as this goes on maybe I’ll re-arrange the page to make it structured more nicely.
For now, instead of a 1-to-5-star ⭐ rating, I’ll use 1-5 ramen bowls 🍜 for how much I liked it, and 1-5 pieces of fried shrimp 🍤 for how much it disagreed with me as I’m allergic to shrimp.
Fiction
A Deadly Education (2020) by Naomi Novik
Personal Rating: 🍜🍜🍜🍜 | 🍤🍤
What I liked most 🍜: In games and real life where magic is interactive and used, mana woes are a serious issue, and this story captured it well to the point where it was like mana management and economics. I’m assuming, as I have very little knowledge of economics as a subject.
Why my personal rating isn’t higher 🍤: I think academics, as the domain of educators, has a bad or unfair stereotype or a perception of needing to be confusing sometimes. Even for advanced topics, there’s always a way to explain things simply according to the audience (which includes equal peers because we often don’t have the exact same contextual knowledge as each other on every relevant topic), in my opinion. So I wasn’t a huge fan of the choice to leave some of the more complicated academic magic aspects less explained.
The Book That Wouldn’t Burn (2024) by Mark Lawrence
Personal Rating: 🍜🍜🍜🍜🍜 | 🍤
What I liked most 🍜: The details of the worldbuilding reminded me of Tower of God which was maybe the most expansive fantasy world I’ve ever read about, so that nostalgic feeling was a biased plus for me. This book wasn’t at the scale of a long-running story like that, but the initial feeling was similar I mean. Aside from that and the consistent quality of the writing, some very cute moments were what I liked most.
Why my personal rating isn’t higher 🍤: In my opinion some scenes didn’t carry the amount of emotion they could’ve, but it seemed almost intentional. Just my opinion anyway, was still an enjoyable story.
Piranesi (2021) by Susanna Clarke
Personal Rating: 🍜🍜🍜🍜🍜 | 🍤
What I liked most 🍜: Was very immersive and detailed but still concise. Not sure what else to write more than that, was interesting and very immersive, on top of being a good story in my and apparently most people’s opinions according to reviews and stuff.
Why my personal rating isn’t higher 🍤: Just preference, it’s sci fi and fantasy but also a kind of mystery that I don’t usually tend towards. Didn’t mind it so much since it’s still SFF, but that and then also the style of writing the character has (which seems like a perfectly fine and probably logical choice) can get pretty taxing after a while.
Non-fiction
Hands-on Large Language Models (2024) by Jay Alammar and Maarten Grootendorst
Personal Rating: 🍜🍜🍜🍜🍜 | 🍤
What I liked most 🍜: It’s a thorough coverage of the main topics, methods, and best practices for working with agentic AI and language models, including a brief overall history of the field at the start. Written in an “intuition-first” way, they try to focus on the main concepts in a way that’s easier to grasp quickly and in a way that makes the book less likely to become outdated very soon despite the rapid changes generally seen in technology. I have some bias as I’m generally a fan of Maarten Grootendorst’s stuff, he’s an AI engineer at Deep Mind with a background in psychology, and his explanations are usually solid and thorough.
Since this is a fiction author site, I’ll add the disclaimer that I don’t use this stuff for writing stories (ruins the fun of writing fiction for me, and that would remove too much personal style anyway).
Why my personal rating isn’t higher 🍤: I think they did really well, and at the very least the book has value in being an index for key topics and papers of 2024 and before, supported by resources they add to the site post-release. But coverage of this many topics has the trade-off of details omitted for brevity. They handle it well, give the necessary details as much as possible, but at times the brevity can make simpler explanations kind of misleading, in my opinion.
Documentaries
First Peoples (2015) PBS
Personal Rating: 🍜🍜🍜🍜 | 🍤🍤
What I liked most 🍜: The amount of in-depth information covered and the heavy inclusion of the researchers working in the various domains, letting them tell the stories alongside the narrator which also let them show their passion for their work and their own storytelling ability. (at least I’m assuming it’s their own storytelling ability, I guess prep and production would’ve helped too, but a couple of them spoke with a vigor and clearness that couldn’t have come from production aid alone)
Why my personal rating isn’t higher 🍤: They covered a lot of info, at one or two points it was clear even as a layman to the field that they had to skim details with the amount included. So more visualizations, supporting on-screen text, that kind of stuff would’ve helped in my opinion. I’m sure that would’ve also been more appreciated by those who prefer reading over listening or second-language english speakers without a good available translation. For example, there were a couple of researchers using dice and cards to explain their point who could’ve used some post-recording help, or the frequently used visualization of global migrations could’ve been better with at least some labelling.
My Octopus Teacher (2020) Off The Fence, The Sea Change Project
Personal Rating: 🍜🍜🍜🍜 | 🍤🍤
What I liked most 🍜: I don’t think I’d like being a reviewer as a job, especially for cases like this where it’s so clear how much time, effort, energy, and care went into making something. It was a story told in a soothing and calm way, while still feeling journalistic in a research-oriented way.
Why my personal rating isn’t higher 🍤: It would’ve been nice to learn more of his research prep and how that compared to the experience in practice, and similarly it would’ve been convenient to learn more about the general line between observation and interference with regards to the ethics of it all.
Our Planet II (2023) Silverback Films, Worldwide Fund for Nature, dist. by Netflix
Personal Rating: 🍜🍜🍜🍜 | 🍤
What I liked most 🍜: Lots of interesting stuff, aesthetic and vibrant colours good for TV displays but also pleasing to watch, and calm and soothingly told. The gentoo penguins were my favourite animal protagonists.
Why my personal rating isn’t higher 🍤: It revolved around seasons and migrations, so it felt sort of like highlight reels over several animal stories, and maybe it would’ve been nice to spend more time with each one.
Movies
Ninja Scroll (1993) written and directed by Yoshiaki Kawajiri
Personal Rating: 🍜🍜🍜🍜 | 🍤🍤
What I liked most 🍜: Cool fantasy powers and very good animation made it a good action movie, with occasional glimpses of historical fiction and a clear message behind all the immorality in it (but a heads up that it’s quite pornographic with a couple rape scenes, in case by happenstance anyone who prefers to avoid that stuff considers this a rec).
Why my personal rating isn’t higher 🍤: Just generally more of a Rurouni Kenshin fan. This has some strong similarities but different execution, like faster pacing, more fights, less history and meaningful dialogue. Still a good action adventure though. Apparently Leonardo DiCaprio’s Appian Way productions started an adaptation for it, but there’s been no public update since 2018.
Mickey 17 (2025) by Bong Joon Ho
Personal Rating: 🍜🍜🍜🍜| 🍤
What I liked most 🍜: The cloning method was cool, at least in terms of the visuals and considering it’s not the kind of sci fi that explains it. The main character seemed well-rounded and likable. The plot itself had small or subtle twists that were nice and more unpredictable. I liked it, the current 6.7 IMDb rating seems harsh to me, but I’m also a fan of Bong Joon Ho and Robert Pattinson, so probably biased.
Why my personal rating isn’t higher 🍤: Just not usually my preference in terms of comedy. But I’m a bit fussy with comedy, and it didn’t include any especially gross stuff which seems more common for this kind of comedy. Just a little bit of gore and vomit.
Spiderman: Brand New Day (2026) written by Chris McKenna and Erik Sommers
Personal Rating: 🍜🍜🍜 | 🍤 🍤 🍤
What I liked most 🍜: The main trio mostly. They were the same, good and likable characters. Some bias to them because I’m a fan of the first three Tom Holland Spiderman movies. Him having to rebuild as a forgotten hero was also a nice idea considering the challenge of a fourth movie that’s heavily connected to a bigger franchise but also needs standalone clarity.
Why my personal rating isn’t higher 🍤: Didn’t care for the new characters or villain unfortunately. The main trio was also a lot less present than in the previous ones. The dialogue often felt weaker too. Didn’t care for all the rage stuff either. And in him having to rebuild as a forgotten hero I would’ve expected his technology to be much more lower budget, simple, accessible, constrained in an interesting way, but it didn’t seem that different.
TV Shows that I don’t watch on a TV
Asian Dramas
W: Two Worlds Apart (2016) written by Song Jae-jeong
Personal Rating: 🍜🍜🍜🍜 | 🍤 🍤 🍤
What I liked most 🍜: The story, premise, and thought behind it was really cool and felt special. Or it felt creative in a way that clearly gave a “high-ranked manhwa of it’s time” kind of feeling. And I liked the romantic comedy bits of it the most. The plot also moved quickly but didn’t feel rushed, the effect was more like it was difficult to predict episode by episode.
Why my personal rating isn’t higher 🍤: Just a matter of personal preference. I’m not a huge fan of thrillers in general, and since this was mixed genres I was hoping for more of the sweet light-hearted romantic comedy style scenes but it leaned to thriller more throughout. And I also just feel like its premise was better suited to that style than the thriller stuff.
Manga/Manhwa/Manhua/Webtoons/Graphic Novels
Anime
Podcasts
Games
*These categories below are more likely to just be videos, posts, or papers I really want to mention here since these are short.
YouTube
Academic Publications
Lasso: Large Language Model-based User Simulator for Cross-Domain Recommendation by Yue Chen et al. (2025)
An interesting way of combining LLMs with traditional recommendation methods in a way that competes or competed with state-of-the-art models at the time. Where “cross-domain” recommendations means, for example, “a user likes these books, would they like this movie?” they evaluated using the Amazon reviews dataset on the books->movies, books->music, and movies->music tasks, seeing a 2% to 10% improvement under various conditions. They also evaluated using the Kwai e-commerce platform’s data on a live-stream interaction -> product purchasing recommendation task, compared against their live system, and saw about a 2% improvement. The first use of LLMs for CDR to their knowledge, and further validates the quality of LLM recommendations.
Simulating Discoverability for Upcoming Content in TV Entertainment Platforms (2025) by Hande et al. of Comcast
With the variety of ways a user can phrase a query for search on an entertainment platform that contains multiple networks, they present an evaluation method that: (1) generates realistic user queries (2) masks the entities/titles of real historical queries for use as general templates (3) clusters the generated queries for a new entity/title to be released like a show or movie or album or live event, such that a cluster represents a specific user intent (4) uses the clusters to score how realistic a generated query is (5) feeds the queries into the UI of their platform (6) uses the output search results and I assume possibly relevant suggestions, to score the simulated discoverability of a new entity/title. I haven’t looked into it super thoroughly or anything, but looks like it could be a neat way of estimating discoverability by related search suggestions for a prospective title from outside of the distribution networks/without access to their internal models or systems.
I think the masking of their search records for a list of templates was neat. The formulae for the scoring was thought through a lot too, like for example they used the distance similarity between a new title and the center of the existing cluster as part of the score for how realistic it is, and the formulae were written in a clear and easy to understand way.
Not All Impressions Are Created Equal: Psychology-Informed Retention Optimization for Short-Form Video Recommendation (2025) by Wang et al. of Stanford and Meta
Recommender systems for short-form video platforms like Facebook reels apparently used to focus on immediate signals for providing suggestions. Signals around the likelihood of a user engaging with a specific video: likes, comments, shares, etc. But they say there’s growing evidence of that leading to clickbait, echo chambers, and filter bubbles (where an individual user’s past history is relied on too much for new suggestions). So modelling user-retention at the session-level is being included alongside that.
In other words, instead of only predicting whether or not a user will do immediate stuff like press like on a given short video in preparing suggestions, also predicting and using whether or not they’ll return to the platform the next day given their entire viewing session.
They say that the problem with that though, is that it considers the entire session to be equal. If a user is retained or returns, all the videos in the session are given a positive label, when their impacts aren’t equal.
They base their labelling instead on a rule grounded in psychology (peak-end) that says people evaluate experiences based on the peak or most intense moment of it and on the end of it. If a user returns the next day, the positive label is given only to videos they interacted positively with, and if a user doesn’t return, the negative label is given only to videos with negative interactions (like “hide”) and the last video.
And with that change they saw a 0.42% increase in total Facebook reels sessions, a 0.03% increase in daily active users, and a 0.05% increase in total Facebook sessions after 2.5 months of A/B testing against their live system.
Cool to see a meaningful step forward grounded in a psychology statement that’s interesting by itself for writing / preparing experiences, and the intent of it is nice too, correcting those possible negative outcomes.
Never Miss an Episode: How LLMs are Powering Serial Content Discovery on YouTube (2025) by Kumthekar et al. of Google
Playlists of YouTube videos can be under one theme or type like a series with no watch order (labelled as “episodic: non-serial” with the example, “hot ones”) or they can have an intended watch order as a series with ordered episodes (labelled as “episodic: serial” with the example, “hell’s kitchen season 20”) or they can just be a playlist of unrelated videos (labelled as “not episodic” with the example, “chill songs”).
So consider a user watching a video from their home page or from the suggestions of a prior video. If that video is part of an episodic series playlist, the next best suggestion for them to watch could just be the next episode in the series. But because they didn’t start viewing from within that series, the recommendations end up in a sort of gray area. If the playlist is non-episodic or non-serial, the next best suggestion might be something else. So the context from outside the playlist is needed. The paper focuses on the problem of first identifying whether or not a playlist is episodic, and then identifying whether or not it’s serial.
They mentioned their prior or live solution uses manually engineered regular expressions which are difficult to maintain and only cover 30% of the examples in their curated test set. That this was an alternative to fine-tuning a classifier which would’ve required many examples. And instead proposed an LLM for the classification via few-shot prompting (a prompt with a few examples).
Two interesting details were: (1) adding a persona and reasoning to the prompt (so including things like “imagine you are a TV producer” and “include the reasoning steps in your response” in the prompt) yielded a negative or no effect on the results (2) the smaller model they tested with (8B params vs. 24B) gave a comparable or slightly worse result while using half the resources.
The results showed that the LLM was accurate on a significant amount of examples more than their live method for the task, many of which were on foreign language videos, and regular expression cases they hadn’t covered yet. Their testing in live use also showed a meaningful improvement in user satisfaction and daily active users.
An interesting and specific problem, but the notes in particular about the smaller model’s performance for the resources and the negative effect from persona and reasoning seemed like evidence that’s very good to keep in mind for any other case of testing LLMs.
Personal Top 3s of pre-2026 (that I can remember right now)
A brief imperfect backlog of top 3s that come to mind for each category, without much if anything in reviews here, since reviews when the experience is fresh would be best. Maybe a short note.
Fiction
Siddhartha by Hermann Hesse
The Alchemist by Paulo Coelho
Life of Pi by Yann Martel
Non-fiction
Creativity Inc. by Ed Catmull
Pirate Hunters by Robert Kurson
The Buccaneers of America by Alexandre Exquemelin
Movies
Ratatouille by Pixar (story: Brad Bird, Jan Pinkava, Jim Capobianco)
Your Name by Makoto Shinkai
Ninotchka by Melchior Lengyel
TV Shows that I don’t watch on a TV
Friends
Hannibal
How I Met Your Mother
Asian Dramas
Romance is a Bonus Book by Jung Hyun-jung
Moving by Kang Full
Love Alarm based on Cho Kye-young’s Webtoon
Manga/Manhwa/Manhua/Webtoons/Graphic Novels
Tower of God by SIU
Annarasumanara by Ha Il-Kwon
The Breaker by Kamaro
Academic Publications
For considerate convenience, I’ll add my imperfect one-line descriptions to these papers, since this is a fiction author site.
A Contextual-Bandit Approach to Personalized News Article Recommendation by Wei Chu et al. of Yahoo! Labs (2010)
A foundational approach to recommendation algorithms that contributed to eventual systems like Netflix’s interesting personalized thumbnail selection: a feature, rationale, and observed effect that seems applicable to digital book covers though not yet implemented by major platforms to my knowledge.
Generative Adversarial Nets by Goodfellow et al. (2014)
A paired approach to generating data by training one model to generate data alongside the training of another that judges its output.
I’m only one person with an imperfect memory and limited perspective so disclaimer on this comment, but I think this one was a community-level breakthrough point for AI.
The subject of generative models today seems to be viewed as the enemy of creators, but at the time of this paper, the datasets needed for aiding critical problems in other domains were insufficient, so data generation carried a heavy social good. Similarly today, there are many domains important to society, like healthcare, industrial manufacturing, economics, etc. that can/do benefit from the advances in generative AI. Whether or not its use in creative spaces is misuse of well-intended advances, I’d rather not take a public position on. My personal choice is that I don’t use it to generate my writing, especially in fiction. Otherwise, like images, video, audio, if I can afford to support a capable human, that’s the option I choose. And so far the only exception to that has been the digital voices for my audiobook which are appropriately labeled.
College Admissions and the Stability of Marriage by Gale and Shapley (1962)
The core of their nobel prize winning work that provided a critical matching algorithm for many domains.
Anime
Haikyu!! by Haruichi Furudate
Naruto by Masahi Kishimoto
Your Lie in April by Naoshi Arakawa
Kids Books & Shows
These are from my childhood, and not recent consumption.
Kids shows: Spongebob Squarepants by Stephen Hillenburg, Stanley by Andrew Griffin, The Powerpuff Girls by Craig McCracken
Early-Teens shows: Suite Life of Zack & Cody / on Deck by Danny Kallis and Jim Geoghan, Avatar: The Last Airbender by Michael Dante DiMartino and Bryan Konietzko (technically a kids show, but I think it’s crossover, and this to me is certainly a kids show that could be appreciated by childless adults), iCarly by Dan Schneider
Early Kidlit / Fairy tales: Jack and the Beanstalk, The Lion and the Mouse, Goldilocks and the Three Bears
Games
Mostly from my youth. If I could spare the time, energy, and priority to it as an adult, I’d lean towards tactical games like D&D, World of Warships, Teamfight Tactics/Dota Autochess, League of Legends, or plain Chess. And with games in general, it’s more about the social aspect and the time with friends for me, (unless it’s short with a really nice story) which is why I’m also a fan of jackbox couch/party games, and board games like Codenames.
Megaman X
Pokemon Pearl
Runescape