Everyone says they are packing up. But most of the people I’ve seen complain either are still on the site they complain about, or left briefly just to realize that other sites aren’t any better and came back.
I’ve left everything (Facebook, Instagram, I’ve even deleted the YouTube app from my phone, etc.) apart from HN. I occasionally look at Twitter, but only when someone sends me a link.
Me too, except for YouTube. I'm still hopelessly addicted to YouTube, but I'm working on it. I also ditched my smartphone eight months ago and going strong, zero desire to return to it.
Seriously, fuck all this shit. Some people love to say that love of the old web is just nostalgia and they are free to think what they want.
What really helped me to go through the YouTube backlog quickly was a user-script sending closed captions to an LLM to summarize, with some custom instructions of what I want. Occasionally I will ask a follow-up question. Now, I can read the summary of an hour-long video in 3 minutes and close it. I only watch or listen to videos that I really want to hear in full.
As another option, I use the Redirector browser add-on to automatically take me from /shorts/{videoid} to /watch?v=$1. Its amazing how manipulative the shorts interface is because if you're not in it, the temptation is SO massively reduced. Shows how much money they obviously spent on psych profiling for shorts to get that first click. We are just meat puppets to them.
My plan is to sell my laptop so that I only have a desktop and can't just watch willy-nilly wherever I am in my apartment. But it's a tough trigger to pull.
I’ve learned a huge amount from YouTube and am very thankful for its existence. It’s the absolute insistence on showing me idiotic ‘shorts’ that made me leave. I didn’t want to.
I haven't been on twitter for about 5 years, facebook for about 10, my instagram account lasted 2 weeks before I got tired of it. These sites are pretty easy to leave.
Whatsapp is harder, because people I actually care to talk to use it. I would rather not give facebook a dime of my money.
I just deleted WhatsApp and told everyone I'm on Signal and if they didn't want it they can call me. But it's not like my "network" is very big or anything, so I guess harder for some people than others.
I think the basic premise under there is still valid. Lots of online sites/communities/etc turn into rugpulls. Success leads to management change, then depersonalization, then monetization.
Right now you are right -- even if I ran my local LLM all day, the quality is not nearly as great, and it runs slowly -- so I use the tier one AI subscription services as they are faster and smarter. But that might only be true for a limited amount of time, and a limited number of circumstances.
To borrow your steam engine analogy, if local LLMs get as good as a Toyota Prius, even if OpenAI / Anthropic offer Ferraris, most people will be happy with their Prius as their daily driver.
Similarly, if the big labs start raising prices or cutting usage, you won't be able to use it as much as you want -- whereas a local LLM will run all day every day without costing you any extra money.
So right now you are right, but who knows how long that will last.
> For nearly all tasks, I want the fastest and smartest AI model.
This is not nearly true for everyone else in the world.
For example, think about the world in ~2021 pre-LLM. Would anyone say the sentence "I only want the fastest and smartest humans working on my project"?
No of course not. Most people don't want to pay $10 million dollar salary to the best programmers in the world. They prefer to pay $200k salary to a median programmer and that's good enough for their ecommerce website.
This will change when out of control billing gets noticed. Then us devs will, I presume, get token rations.
I can see my new Thursday afternoon "oh chit" moment being that I didn't complete my weekly task because i torched all of those tokens M-W doing task/ticket grooming using the hot hot model instead of the dodo model with jira mcp connector :D
I think a scooter or a bus is a more apt comparison than a steam engine. The scooter and bus can both get the job done with some acceptable trade-offs, depending on your circumstances and what you’re willing to accept.
However, there are many use cases where they aren’t the right tool for the job.
I find this too. In fact, recently I've been pushing more and more to the latest and greatest model every time there is an update. It just saves me so much headache.
The margins on NVidia datacenter hardware are ... high. At least one order of magnitude larger than a consumer chip.
Given the recent deepseekv4.1 advances - how good of a 3B model can we make to run on an iphone natively? is it good enough to match common muse/dot use cases for consumers? the phone is already always on.. no need for a cloud server.
Most phones are not really "always on" in any real sense, a phone on active standby uses very little power and most of it is for its mobile connection. Local AI is best run in a stationary homelab environment, even running it on laptops has its very real problems.
There are a few edge models, Spark-X2.5-4B, LFM2.5- in 8b-a1b and 2.6b variants, that are useful for on-device agent-y stuff, but I am doubtful that’s a valuable category. How many iPhone users have never opened Shortcuts in their life? Automating on-phone stuff seems niche.
I don’t expect the economics of local vs cloud ai to change until either the bubble pops or new ai chips land that can run big models fast with low power demand.
I think the interesting question is whether the average consumer cares? a 4B model should hit roughly the same numbers as GPT5.5 sometime in December/March. That is more than capable of performing a variety of complex work tasks.
I would argue performance optimizations help with local/open models but hurt openai and anthropic - because open and or cheap/alternative models are threat to those companies. There is a fundamental contradiction/conflict between the prevelance of open models and the financial success of openai and anthropic. That is why they are doing everything they can to kill any open/cheap/efficient/chinese models (take a look at this thread - it was top of HN 40 mins ago, with very high engagement... now it is buried in page 5... totally normal and legit).
Perhaps I'm not understanding it correctly, but here's my take on what the paper is doing.
Imagine you have a problem you want to solve (let's say, identify an OCR'd handwritten character, e.g. the MNIST Dataset). You tell 3 agents "Hey, each of you take a stab at getting really good at recognizing characters from this dataset. You can take 10 refinement steps to continue to improve ". You can't give each agent unlimited steps of course, because you have a finite amount of compute.
So each agent goes off, and by the end, Agent 1 got to 90% accuracy, Agent 2 got to 80% accuracy, and Agent 3 got to 89% accuracy. Agent 1 wins, of course.
But then you look at the refinement steps, and after 2 steps, Agent 1 was _already at_ 90% accuracy. So the agent spent the next 8 steps basically not moving at all. Agent 3 on the other hand, perhaps was continuously climbing in accuracy at every refinement step, but hit step 10 and had to stop.
Now because you recorded every step from every agent, you know what you'd do differently next time -- you'd not allocate as many steps to Agent 1, and give Agent 3 more steps, because perhaps that might result in Agent 3 coming up with a better answer.
From my understanding, that's what they built in the form of a "search" controller -- a way to evaluate automatically and reapply how you could allocate resources more effectively, when applied to a new problem.
But I guess my misunderstanding is how applicable the search controller is when applied to new problems -- just because one pathway stalled early for one problem, doesn't mean it would work for another?
Your understanding is basically correct.
"how applicable the search controller is when applied to new problems". We need meta-agent thinking pattern. Self-evolving agent has been very popular and we want to use agent to design a perfect agent. This is the problem that the "search" controller employed in this paper aims to solve.
Intuitively I wouldn't readjust how many steps they each do, but instead add another run afterwards, that get the same amount of steps as the previous, but now also with a concise description of what the previous attempts did and what they achieved, and ask it to improve. The amount of compute you have available, would dictate how many full iterations of this "fan out search > consolidate" workflow you can do.
In the paper (section 5.1), they actually tried to abstract high level directional insights into the prompt in order to see if that helped, and they basically found it underperformed a prompt that didn't have those insights at all, implying that directional guidance perhaps over-constrains things.
Could use the shorthand of 85%, 10%, and 5% as the way to divide the workload; 850/1000 computes, 100/1000, and 50/1000. Brute sprint, clean up & revaluation runs, then checksum and presentation.
I've stopped waiting for sites to do what I want in terms of features and functionality -- their incentives and mine very rarely align.
I think the solution here eventually will be a search engine that lives on your personal computer (it's also what I'm building currently). Ingest content from the web -> sort/filter/view it locally as you choose.
This is actually where I see software going in the short term -- cloud moving to local.
A few years ago, if you wanted translation, you'd use Google Translate. If you wanted to search the web, you'd use Google search.
But for a few gigabytes, you can now install nllb-200-distilled-600M, and get translations for almost any language locally. You can have your computer crawl the web, create abstracts and categorizations for websites, and build search exactly as you want it.
The main limiter now is hard drive space (and to an extent, local compute) -- but right now it feels like the 70s again where the terminal into a remote server turned into building applications locally.
It's more of a pipeline than a back-and-forth. New abilities happen in the cloud first because they require specialized, higher capacity resources and then move towards being local as the resource usage gets optimized.
Outside insane GPU appliances, I really see cloud based things as a lock-in to subscription based offerings. The fallacy of always using the latest code isn't worth never ending subscription fees. I'll install it locally. I'll keep my data locally. I don't need to wait for things to upload/download. If there's an update that feels worthy, I'll purchased and install locally.
Note -- if you do this, watch out for requesting access to "all tlds". They send you two emails per TLD -- one for your pending state, and one for your approved/rejected state. I suddenly had 1k+ emails flooding into my inbox, until I found the setting on their website to disable emails.
About the same thing I saw. I kept having to increase my bid every day, and even at $4 CPC basically saw nothing converting.
As you said, it's also really hard to get visibility: metrics are poor, and targeting seems more like a whim/suggestion than anything else (their label for context hints basically tells you they only _might_ take your guidance on who they will show your ad to).
Audiences consists of hashed emails or phone numbers -- no targeting demographics, job titles, etc.
Overall it's just not there yet -- I paused my ads today.
