I reckon the kids today are likely to see the shit hit the proper fan and possibly the collapse of old school society. It'll probably be during their lifetime where there are so few jobs left we need to have proper universal basic income with only 5% of the people doing the actual work, but otherwise running the bots, both logical and physical. Can we do that transition cleanly? not so sure.
Not sure about that, it's possible if government keeps going in the direction it is...
China is miles ahead of us in every way and they are certainly making it work!!
After 2 months of getting reamed by claude doing agentic coding I've migrated to deepseek. I've only done some basic testing this afternoon, but so far it's been good. It seems its morals are more aligned to mine too so that's a bonus
Last edited by hjtrbo on Wed Jul 08, 2026 8:04 pm, edited 2 times in total.
hjtrbo wrote: Sun Jul 05, 2026 7:34 am
After 2 months of getting reamed by claude doing agentic coding I've migrated to deepseek. I've only done some basic testing this afternoon, but so far it's been good. It seems its morals are more aligned to mine too so that's a bonus
I'm using Kun https://deepseek-gui.com/ as the gui. Migration is easy, it follows a very similar structure to claude. For the most part it's a straight copy / rename of your claude assets.
And kun does forget I can't read chinese so a line is needed in its ~\.kun\AGENTS.md (aka ~\.claude\CLAUDE.md) Always reply in English
Per project migration skill attached. This migrates your per project claude skills, memories & workflows over to kun.
claude-to-kun.zip
if your using mimo or deepseek flash, then instead of using the raw api, switch to a OpenCode Zen Go plan, it's openAi based api usage.
Thanks. I'm happy trying api based billing for a couple of months as my usage is intermittent. Also important to me is a full featured desktop app that I can use outside of a coding environment for other things. I don't like terminal based setups. I'm using deepseek-v4-pro on the ultra setting. The only downside so far is it can't directly read images.
Last edited by hjtrbo on Wed Jul 08, 2026 8:05 pm, edited 1 time in total.
"You spend $5, you get $30 worth of usage." Not saying its definately your case or bad but it doesn't hurt to know this sort of stuff is going on. Is it a problem for you? Not sure. But it might mean all your traffic is captured and used to train other models, and might eventually end up leaking. It also might mean your AI is switched out to another less performant (in smarts or speed) without you knowing. If it's too cheap you are the product.
hjtrbo wrote: Mon Jul 06, 2026 9:41 pm
Thanks. I'm happy trying api based billing for a couple of months as my usage is intermittent. Also important to me is a full featured desktop app that I can use outside of a coding environment for other things. I don't like terminal based setups. I'm using deepseek-v4-pro on the ultra setting. The only downside so far is it can't directly read images.
Maybe you missed it or didn't care enough to even look.
But opencode go is a api, you can take it anywhere you can use openai. I don't use opencode. I use my own desktop harness I spent a month making. It blows the harness out the water that you linked.
hjtrbo wrote: Mon Jul 06, 2026 9:41 pm
Thanks. I'm happy trying api based billing for a couple of months as my usage is intermittent. Also important to me is a full featured desktop app that I can use outside of a coding environment for other things. I don't like terminal based setups. I'm using deepseek-v4-pro on the ultra setting. The only downside so far is it can't directly read images.
Maybe you missed it or didn't care enough to even look.
But opencode go is a api, you can take it anywhere you can use openai. I don't use opencode. I use my own desktop harness I spent a month making. It blows the harness out the water that you linked.
I did miss it. I've installed the desktop app and am having a play. So far, its doing a nice job. Good enough that I gave kun the ass
After a week of using deepseek-v4-pro on the max (level 4 of 4) setting I have found it's good but not as good as Opus 4.8 on the high (level 3 of 6) setting.
For ghidra re work on power pc and hcs12x it does ok, but I still turn to claude for difficult functions. With the gm fpcm bin file I was working on, deepseek cannot properly decompile switch tables on hcs12x. Claude eats it. The core issue appears to be (me &) the chip setup registers being not known and the ghidra chip spec file is not helping. Claude somehow logically worked its way through those difficult functions and produces a decompiled output that looks highly plausible. I did not test out the same function with IDA as I have not learned how to use it. It may have a better chip definition.
I'm not an re expert therefore I rely heavily on the ai to extend my ability. Because of this I'm lost without a capable ai model.
For me, deepseek is not a replacement for claude code's Opus 4.8. I do think though that whenever deepseek release a next gen model it will be able to replace claude. For now, deepseek gets about 95% of my work with claude on standby acting as the gap closer.
Previously, to get uninterrupted work output without hitting my claude usage limits, I needed to be on the $100USD max-5x or sometimes the $200USD max-20x plans. On deepseek I'm using api billing. So far this week I have spent $4.26USD lol. I extrapolate my monthly budget to be $25USD for deepseek and a further $20USD for the claude code pro plan. I'm still getting the same work done, therefore I have made some significant savings circa >50%.
I'd suggest claude pro not max. even dual claude pro. Still comes in at half the price. in vscode with the plugin, you can /switch-account as soon as you hit a limit and continue, the session carries through. Then start multiple chats to keep the token count down for new topics, or /compact when want the general background context of the task to carry through but hit a milestone and the conversation has been clear about reaching the milestone so it can throw away specific details to minimise context use.
yes, I'm on the claude pro plan. Deepseek has enabled me to get off the max plan. There has been a good saving. I only use claude now as a fall back, sometimes as a verification agent against deepseeks output.
I really am looking forward to the day deepseek genuinely can beat Opus 4.8. Who knows where anthropic will be by then - Fable 5.x something. But for the work I do, Opus 4.8 level of reasoning would be good enough to last me my lifetime. In the future, when $3k of home hardware can run a model like that, I'm buying.
