The pendulum synchronization effect is the opposite of your claim. It has a physical causal reason for why pendulums become synchronized. Your claim is that it was random and independent.
In the last 30 days my LLM spend, if paid at API rates, would be $19,431 (and that doesn't track online chats or my Github Action reviews). Now I wish I was making $200K/mo but I'm not, no where close. I got all that for $220 on subscriptions.
Yes, LLMs are awesome and let me do things faster but I would be using them far less if my only option was the API. I used coding agents prior to subscriptions (Aider) and spent <$200 total before abandoning it due to cost. The results were good, but not worth the price for me.
Writing code for my day job and my side business. 90% of its planning/writing/reviewing code and little bit of it's admin-type work. I have two products/services I do on the side, food festival payments/management software and food week (Burger Week, Pizza Week, etc) management software. My side projects, while profitable, where always behind where I wanted them to be but now I can finally chew into the backlog of things I've wanted for years.
I'm happy to answer more questions if that doesn't cover what you were looking for. I'm not sure if "coding" is all you wanted or if you wanted more details.
What would you say is the split of your token use (between main job / side project)? Do you use agents for everything?
I'm asking because I feel I'm weird in my usage, which is very much structured like 1. organising the context and then 2. send off requests to 1-2 models and continue down the rabbit hole from there. I've found agents take longer and make me learn/retain less about the system I'm building.
> What would you say is the split of your token use (between main job / side project)?
It fluctuates, but never less than 30% on either and I get to 95%+ of my weekly usage. I also spend some tokens on “personal” stuff, like with my Obsidian notes or helping to manage other tech in my life (Home Assistant, Unifi, NAS, 3D Printer, other servers).
> Do you use agents for everything?
Increasingly so. It’s just so much faster in many cases or even if it takes the same time or more, I can do something else while it’s working. All my communications are 100% me, (sms/iMessage/HN/Slack/etc) but anything I consider “busy work” I try to farm out. I find the edges of what it can do then back off a bit or stop using it for that task.
> I'm asking because I feel I'm weird in my usage…
I like using skills to replicate the steps I want taken. It’s far from perfect and I want a way to enforce stricter workflows but skills with “Step 1, Step 2, Step 3…” do work. Currently once I’ve filled out a ticket (often using an agent to bounce ideas or grill me on my ideas) I can hand that off to an agent to plan, it will do its planning, present the plan, then spin up a worktree running an isolated dev stack, do the work (asking me questions as needed), then tell me it’s done (providing me with one-click login urls for the various roles in my system) with a link to the running dev stack, I then review the work, it kicks off 2 reviews, and fixes any issues they surface. Then a PR is created, this is normally when I look at the code if I need to, then 2 agents review the code (more of a smoke test but they find things), the original agent waits for the agents on GH Actions to finish, then deals with their findings, then it kicks off another review (up to 2 rounds). I decide then if any remaining issues are worth not merging (if you’ve used LLMs for code review you know they can go on endlessly with increasingly unreasonable edge cases).
All that said, my workflow is flux as I try out new things regularly. Does that slow me down? Perhaps, but things are changing too quickly to “settle in” to any workflow just yet. You could compare it to the JS Framework Cambrian explosion but it’s easier to switch.
> I've found agents take longer and make me learn/retain less about the system I'm building.
Fair point, some parts I scrutinize, and some areas I’m less concerned. I have E2E tests on the important flows, I perform manual testing constantly, there are extensive unit tests (front and backend), and the code is structured and abstracted better for testing than it’s ever been. All that to say is I feel confident of foundation, some of which was built pre-LLM and some of it was built by LLMs.
I interviewed for an internship in 2006 and didn't get any brainteasers, but I got some pretty interesting questions that were somewhat computer science-related:
1. Suppose you have a binary tree (NOT a binary search tree) where each node with pointers to its parent and children. Given pointers to two arbitrary nodes in the tree, find their lowest common ancestor.
2. My favorite: given a uniform random number generator mod 5, create a uniform random number generator mod 7
3. Forgot the exact question, but something along the lines of: suppose you want to keep track of function arguments as you call them. How do you do that?
I got (1) and (2) and utterly failed (3). (3) is entirely trivial and basically a stated fact if you know how anything about how operating systems work, but I was a freshman in college and didn't know that function arguments got pushed on a stack in memory, so I was totally lost.
FWIW, in most architectures the first few function arguments get passed in registers. But so many of today’s programmers grew up with 32-bit x86, which has a paucity of registers, that they mentally shortcut to “every argument gets passed on the stack”.
Nowadays, with x64 and AArch64, the first several arguments will be passed in integer or floating-point registers.
I think all of Luna's are bad. The only decent one is sol @ xhigh. Even sol @ max is weird. Sol @ high and @ medium are ok, and every other single one across every model is bad.
Strong disagree, but to each their own. For sol I really like how only medium uses the wings on the handlebars to ride the bike. For all the other sols the pelican evolved a new set of arms separate from the wings.
I must admit, the fact that the writing was well formatted and structured was an instant turn off. I did find it insightful. I would have been more willing to read it if it was one lower case run on line with typos one would expect from a prepubescent child. I am both joking and being serious at the same time. What a world.
HFT doesn't inflate or deflate any valuations. They operate at the market microstructure level and provide liquidity. HFT firms have no impact on the long-term value of assets.
HFTs can siphon away profits from the people actually doing good investing, but Not all HFT is created equal. There have been some pretty high profile instances where HFTs have increased market volatility and caused a "flash crash".
HFTs that trade at the microsecond scale probably aren't valuable to society.
That's all factually incorrect. The reason that HFT is valuable to society is exactly because it trades at the microsecond scale. That's how it provides the most liquidity.
Flash crash was transient and had no impact long term value. Neither does volatility.
In the end, it's largely put a stop to H-1B sponsorship of workers outside the U.S. That doesn't mean that all these workers can't get visas to work in the U.S. but other - tougher visas - have to be explored.
This kind of cynicism is wild to me. Of course most AI products (and products in general) are for end users. Especially for a company like Google--they need to do everything they can to win the AI wars, and that means winning adoption for their AI models.
This is different. AI is an existential threat to Google. I've almost stopped using Google entirely since ChatGPT came out. Why search for a list of webpages which might have the answer to your question and then manually read them one at a time when I can instead just ask an AI to tell me the answer?
If Google doesn't adapt, they could easily be dead in a decade.
That's funny. I stopped using ChatGPT completely and use Gemini to search, because it actually integrates with Google nicely as opposed to ChatGPT which for some reason messes up sometimes (likely due to being blocked by websites while no one dares block Google's crawler lest they be wiped off the face of the internet), and for coding, it's Claude (and maybe now Gemini for that as well). I see no need to use any other LLMs these days. Sometimes I test out the open source ones like DeepSeek or Kimi but those are just as a curiosity.
If web-pages don't contain the answer, the AI likely won't either. But the AI will confidently tell me "the answer" anyway. I've had atrocious issues with wrong or straight up invented information that I must search up every single claim it makes on a website.
My primary workflow is asking AI questions vaguely to see if it successfully explains information I already know or starts to guess. My average context length of a chat is around 3 messages, since I create new chats with a rephrased version of the question to avoid the context poison. Asking three separate instances the same question in slightly different way regularly gives me 2 different answers.
This is still faster than my old approach of finding a dry ground source like a standards document, book, reference, or datasheet, and chewing through it for everything. Now I can sift through 50 secondary sources for the same information much faster because the AI gives me hunches and keywords to google. But I will not take a single claim for an AI seriously without a link to something that says the same thing.
Given how embracing AI is an imperative in tech companies, "a link to something" is likely to be a product of LLM-assisted writing itself. Entire concept of checking through the internet becomes more and more recursive with every passing moment.
reply