Has there ever been a faster rugpull?
well the honeymoon is certainly over
>>108847602So AI won't be taking our jobs?
AI has to start paying those cheques it's been writing.
>>108847771does this mean the bubble can pop now and I can have 1TB SSDs for $50 again?
>local llms are pointless>you're better off just spending your money on tokensHmm, looks me and my local llm win. AGAIN.I think I'll go fuck it to celebrate.
>>108847602China just won again
>>108847602you do realize those are still per prompt and not per tokens? it's still cheap for the lucky ones that are on those grandfathered annual plans.
>>108847602People were abusing it because it charges per prompt
>>108847821Depends if AI turns a profit or not.Companies are investing billions into AI right now propping it up. If they get a return on that investment AI is here to stay, if they don't the ground is going to shake from how hard the bubble will burst.My biggest shock was when they shilled AI using lines of code generated. That's the worst possible metric.
i unsubscribed and switched to OpenCode Go (10$ monthly) with the OpenCode Copilot extension for VS Code Insiders so you can use OpenCode with Copilot Free. It's the exact same workflow and features as what i had before with Github Copilot. I'm using DeepSeek-V4-Pro and im hoping it will last all month for 10$ but i might have to switch to DeepSeek-V4-Flash for that.I think its the best value right now, but there are so many different options. I liked Copilot because of the simplicity and intregration with VS Code and im glad i can continue using it.
>>108847854local llms are way too slow for coding agents
>>108847868don't see deepseek in op's image. don't think it's offered by github copilot.
>>108847913there's literally a plugin you can install to use the deepseek api as a backend for copilot so you keep all the tool calls and its ability to browse your files and shit
>>108847889Buy an ad.
>>108847894What kind of tokens per second do you need for coding agents?
>>108847934Terminal-Bench 2.0 has a two-hour limit for completing the task. no local llm can pass that. would translate to like at least 3k tps.
>>108847889copilot's $10 plan gave you like $500 in claude credits. deepseek isn't that much cheaper than claude.
>>108847924frankenbuild hacks are not the same:https://github.com/Vizards/deepseek-v4-for-copilot/issues/25https://github.com/Vizards/deepseek-v4-for-copilot/issues/56
>>108847854looking at llama.cpp>dsv4 never ever>v3.2/glm5(.1) will forever run with forced full attention because dsa never ever>k2.5/k2.6 will forever have a broken reasoning parser and any attempts at fixing this got ignoredso wake me up when local llms actually include the nice chinese models
>>108848089>people are saying it costs $0.2 instead of $0.1 to feed 10 million tokens into deepseek-v4-prooh nooo better pay up $50/month for CLAUDE and GPT5, kek
>>108848112you will never run a local model cheaper than using the deepseek api, unless you somehow get free hardware + free electricitythe only local models people are running are things like gemma, glm 4.5 air, and qwen
>>108847821you already can, just go on temu
>>108847821Any worn SSDs will end up getting dumped into China and harvested for metals. Not that easy.
>>108848487Why would China "harvest for metals" when they could be reformatted and dumped on Aliexpress for a 50% discount over a new oneThere isn't even $10 worth of raw materials in a single SSD, it has its highest value as a working SSD.
>>108847821Nope. They're going to quadruple down on LLMs and that means they need to scale them up by another 3-4 orders of magnitude. Flash and RAM are going right past the moon and out of the solar system.
>>108847888(((AI))) datacenters are literally giant leech facilitiesthey leech all the resources like hardware, electricity and water from the normal people and turn it into a giant, digital turdgotta give it to (((them))) though. it perfectly reflects their races behaviour
>>108848460it's not a question of money but how well the harness will work.https://github.com/Vizards/deepseek-v4-for-copilot/issues/56>How to stabilize the tools array from outside Copilot Chat's process boundary>How to selectively stabilize volatile parts of the system message without breaking mode/agent/autopilot featuresdeepseek will be dumber since the harness isn't made for deepseek. we all know how bad open-weight models work with the common harness. look at the relevant benchmarks:https://www.tbench.ai/leaderboard/terminal-bench/2.0
>>108847602>opus 4.6 usage slashed down by 89%I don’t understand why anyone would pay for copepilot now, especially when it’s by far the worst harness I ever tested
>>108849862it's still per prompt, not token. learn to prompt.
>>108848487> Any worn SSDs will end up gettingshredded* to deny hardware to possible competition
>>108849867>ummm chuddy you’re holding it wrongnah, only microslop shops will use it, same as the shitass azure
>>108849878??copilot always was per prompt, not token in the past.to do picrel, you already had to know how to prompt.unlike the grandfathered annual plan in op that's still per prompt, the monthly plan switching to tokens is the thing that would actually fuck you. see picrel.on the grandfathered annual plan you would only pay 6x as much -- $60 instead of $10 per month for $500 credits is still a good deal.
>>108849853>blah blahnone of the issues you linked say anything about it being "dumber", it's just that if you keep changing stuff early on in the context then it will cost extra because you are making DS keep reprocessing the prompt. end result? you spend $0.2 instead of $0.1. not good, but still way way better than anything you could get with scam altman and claudestein
>>108849986i literally linked to tbench, the industry standard benchmark. open-weight models were never competitive for coding agents. deepseek having a shitty shim for yet another harness that's not tailored to it will not change a thing.
