Did you really see novels like that? Have they got traditionally published?
Judging by pub tips on Reddit, your query is kinda a very short-form literary PhD thesis.
Also George Martin is infamous for that, but it took him many many years to take off, and he started in a different world entirely
I think the trend was there for a while now! A while like 100 years or so.
Reading new things is fun and cool, and books are pretty to have as a decoration. But we have less and less time for that (apart from those who have regular long commute, lucky you!)
So we end up reading a small fraction of what others read and discuss to belong (or not at all, if your 'others' don't read)
The merch has been there too as well; I obsess over old-school reading desks, you can't just order these today from Amazon even though they have better aura then book nooks
Performative reading has arrived as a boon for the publishing industry. People may be reading less, but they're buying more physical books. The industry has found a rather clever pivot: when you can't sell more reading, sell better-looking merchandise.
Contrary my expectations, I enjoyed reading older math and engineering philosophy writing because it operates much closer to first principles and base reality than we are.
It's an interesting meditation to sit and reflect how much of your thinking patterns assume things like CPU, loops and instructions as some unyielding default.
Plus, even stuff like ether did win back some ground on the conceptual level
“This book is a model of how the history of computer science and mathematics should be written.”
— Donald Knuth
His verdict on "Leibniz on Binary" (MIT Press): 32 of Leibniz's own writings, many previously unpublished, translated and explained.
https://t.co/XBJYQFuPvl
I think there's more to it than just updating model weights to learn continuously (even though it's cool as hell)
I am currently experimenting on evals-driven self-improvement, where you inject similar cases and their judging results into the agent context.
It's not very privacy friendly, but for some models, it brings massive performance gains! They scale on decision-making, translation and labeling, to name a few areas.
If we train models that are more sensitive to memory-related instructions in designated areas, it might be a much better technology to wield (and sell) than dynamic weights.
I always pity that the industry is stuck on system, user, assistant and tool as prompt components, with rare exception of OpenAI adding developer role (to hide actual system prompt from editing)
We could have much more fun with more roles and pre-trained prompting areas!
Highly recommended read.
Self-improving memory is starting to look like an important component for continual-learning agents.
So pay close attention to this research direction if you are studying recursive self-improvement.
Just last week, I featured a few papers on the topic, and this week both Meta and Google published on it.
(find more references towards the end)
Meta lets a model design its own memory habits.
Then they train those habits into the model weights.
The goal is to allow the system to learn the memory design (what to memorize, how to memorize, and what to recall).
This work is actually in the same vein as RIME, which I wrote some notes on last week: https://t.co/wLJy6ntdaJ
Meta first searches for better memory instructions (prompts the model to rewrite itself), then fine-tunes the model (SFT with LoRA) on runs that used the best ones.
RIME, on the other hand, skips the search and uses RL to train one model to solve tasks and write memories, rewarding notes that help on later tasks. It also trains the model to explore so it collects useful information.
Different ways to address the problem.
More useful links below:
Google Self-Improving Memory for Continual Learning - https://t.co/ons3yOcSkW
How to Train Memory Writing for Continual Learning - https://t.co/wLJy6ntdaJ
Can LLM Agents Organize Their Own Memory - https://t.co/uibQW9oKmD
Continual-ARC - https://t.co/BPAO8Cb0s3
Adaptive Memory Selection for LLM Agents - https://t.co/qxLyIr0DyN
MemCredit - https://t.co/AlrGHCvZ8O
I'm already too dependent on Claude Code harness to actually meaningfully move from it :D
I tried t3code as I liked the premise, but I'm not in a phase for re-learning the muscle memory after using Claude Code since research alpha
Plus, Anthropic's startup support really vendor-locks you. Once you get it, you are no longer tempted to move your $200 back and forth.
Performance-wise, OpenAI is losing for a year now; I have access to OpenAI models via my employment, and they are just not worth the time wasted.
That being said, for non-coding usage, OpenAI is still holding some strong areas. Plus you get image generation and other little perks here and there
How long did it take, for curiosity's sake?
When I was proofreading my first 120k words, I took like 1.5 months. And that after a long final editing round! What contributed to it was a lot of exposition and obscure lore that had to be trimmed or rewritten for proofreading to even make sense.
But it was before there were any tools to make it faster, at least to detect stuff like typos on scale.
@aicandii Today I learned that bookplates exist. Such a cool idea lol, how come I've never seen them in bookstores in Europe (or did not pay attention most likely)
Usually it's "Hey, we are launching a place where we'll spend money to bring users for you. How about you join for a fee 50% below market, fixed for 2 years?"
In large cases, merchants may even invest themselves partially to the launch, as it's still a marketing budget saving on the paper
So yeah, it's not done with bootstrapping. Requires investments, connections and networking 😇
But I remember that some marketplaces in Ukraine were launched on a really low start, with family members and relatives as first merchants, but it was many years ago in a different world -- both technologically and economically
@eric_larocca Congrats!
Pure horror has always been something like high-level play in writing for me, after Stoker's Dracula made 12-years-old-me sleep with a knife under the pillow.
So you must have hit really hard!
In Transept, nothing was as inspiring for rapid, high-quality changes as watching customers suffer via analytics.
Unfortunatelly, I can seriously say that our product is built on pain and tears and someone's wasted effort.
Building my SaaS with AI, in public. First real user complaint came in today and it taught me more than ten interviews did. Complaints are free user research, if your ego lets you read them. #BuildInPublic#UX
@My_Durgesh In Transept, nothing was as inspiring for rapid, high-quality changes as watching customers suffer via analytics.
Unfortunatelly, I can seriously say that our product is built on pain and tears and someone's wasted effort.
We are not discussing the mode collapse in big LLMs enough!
I can see that Claude, given the task to farm engagement, will:
1. Prioritize same posts from search results.
2. Write exactly the same posts semantically.
I wonder if it's training the models for code -- so predictability -- and removing hyperparameters control on API side that does this. It was not like this before, and big OSS models have still much higher distribution (also still narrowing)
Smells like a good idea for a quick research...
Onboarding animations actually performed very well; I'd say that every dollar earned was fully or partly due to visual improvements my co-counder made!
What was an actual effort sink is the real-time chat and editing. Most actual users for now are solo authors or entrepreneurs, and they only chat with Literess, so...
But the overall technical foundation of streaming updates, locking and state exchange helped a lot to build a smooth collaboration experience with AI right in the interface!
confession time for indie devs.
every one of us has a feature that took weeks, got polished to perfection, and almost nobody uses.
the onboarding animation. the 14 custom themes. the widget nobody asked for.
what's yours, and how many hours did it eat?
Meaning it exploded and went really well desplite the pivot!
They started as a wrapper for generative art, positioning around new type of UX and new approach towards visual art, as I remember; and raised with that pitch, too. Many were experimenting around that on DALL-E hype.
But they ended up training and serving models themselves!
I remember people mocking them for trying to compete with big boys back in the 2023, yet here we are. It was hard to predict the explosion of image/video generation models, though, and that many players, including Chinese, would be able to train great stuff.
That's if my memory does not trick me, ofc. And they are kinda still focusing on the UX. E.g. they hired my moot who was building a very profitable SaaS for generating manga
I think the problem here is much easier:
1. ClickUp was never a good software (I remember writing for their API in pre-AI days, and it was a trash)
2. You can vibecode it in a day
There are still some moats and know-how to vibecode *good* videos though. But I think it's not sustainable to expect to last this one-shottability gap for long.
I believe you've posted earlier that the moats remaining are distribution (ofc), unique data and unique expertise. Guess ClickUp's management did a wrong bet (or a right one; look like Krea went off)
💯
There is no real way to survive for most SaaS companies here because AI is so completely different than any technological advancement before it
Before you could just niche down and specialize on one thing and you'd know for certain the big tech companies would never be able to dedicate enough resources to fill your tiny little niche market of $300M where you could take 1% of the market and make $3M/year
Riches in the niches!
And so you'd avoid going too general, because "you'd spread yourself too thin" and hit "death by a thousand paper cuts"
And general was where big tech companies were great at, they could dedicate huge resources in solving these big problems, like building Google Sheets, or Salesforce (whatever it is they do)
But that was the old days, and it's completely different now
With AI, big tech companies can easily fill any niche, or be general at the same time, they can do the whole range actually, without even spending time on consciously doing that, because they're not offering a particular SaaS offering at all
They're offering you artificial general intelligence (AGI) that can do it all
And as long as they improve their models every time and make them bigger and better, it slowly can do literally anything (okay not yet anything, for anything they'll need physical robots, but they're working on that too of course)
This leaves SaaS companies with a bitter ending:
- they either do nothing and stay in their niche like they economically should, but that doesn't work anymore now
- they try go general, and offer EVERYTHING and ANYTHING, and that works for a little bit until customers realize they might as well use whoever makes the models that the SaaS itself uses anyway, which is the big tech and big AI companies, who by the way either offer it free or for much cheaper
Everyone's been checkmated by big tech and big AI
This will be the creative destruction of the century
But there'll be some survivors and then a lot of new companies, but how they'll look nobody knows yet!
I think the problem here is much easier:
1. ClickUp was never a good software (I remember writing for their API in pre-AI days, and it was a trash)
2. You can vibecode it in a day
There are still some moats and know-how to vibecode *good* videos though. But I think it's not sustainable to expect to last this one-shottability gap for long.
I believe you've posted earlier that the moats remaining are distribution (ofc), unique data and unique expertise. Guess ClickUp's management did a wrong bet (or a right one; look like Krea went off)
I mean, are we all going to ignore that Claude (since 4 at least) either writes about lighthouse keepers or cartographers, with a story revolving around loss, decay, inheriting and recovering the meaning?
Lighthouses sometimes have bells, too. And eels. And SALT; there's always salt. Also, a rough chunky woman with a hidden soft spot, an old fisherman, and a talking animal.
I've run hundreds of mock bench stories for translation for Claude, and he's writing the same novel over and over. I'm slowly getting interested to actually let him finish one.
Screen from newly tested Author Mode for translation.
@iRetinaPro I mean if you've managed to attract bots, getting actual humans should not be that much harder... (it was meant to be a compliment-joke, idk if I pulled it off)
During the Spiderweb, where Ukraine took down dozens of strategic bombers, I noticed on live videos how painfully slow the deployment of FPVa was.
Truck drivers were trying to catch some; some trucks were able to be stopped, etc.
Plus they are really fragile, even whilst stacked. And require a lot of movement (loading, placing, taking off) to deploy. I assume that they are folded in some way in the tube; thus you need to push them out. Shooting is the cheapest way.
These problems aren't that big because it's 1 drone = 1 operator only. But if this swarm is controlled by 1 operator, or no at all, having a stuff that shoots all the drones at the same time and then retreats actually makes sense!
Explain this like I’m 5:
What’s the point of launching them 3 feet in the air when they could take off faster lying flat?
I get they want to go deeper into a war zone first, but why can’t the vehicle’s top just open to have them take off right there?