Techboy | AI & Tech

I test new AI tools so you don't have to. Honest takes, no hype. Vienna 🇦🇹

[Techboy | AI & Tech @Techboy_Wien]

ok quick intro since a few of you just found me

tech guy from Vienna. I try the new AI stuff the day it drops so you don't have to, and tell you straight if it's actually useful or just hype.

what AI tool should I test next?

[Techboy | AI & Tech @Techboy_Wien]

@SahilPanhotra google access rules are chaos every single launch. Pro gets it in AI Studio but not in Antigravity, same model same subscription. just give us one page that says who gets what

[Techboy | AI & Tech @Techboy_Wien]

@yesnoerror small trainable latent instead of retraining the whole world model makes a lot of sense for robots. curious how many interactions it needs before it figures out something like a heavier object

[Techboy | AI & Tech @Techboy_Wien]

@_avichawla saving this. the layer most teams skip is evals tbh. everyone wants agents in prod, nobody wants to write the boring test set first

[Techboy | AI & Tech @Techboy_Wien]

@kimmonismus the peak vs off peak pricing is the sneaky part. run your big batch jobs at night and it's basically free compared to the closed models

[Techboy | AI & Tech @Techboy_Wien]

deepseek just shipped V4.1 Flash 1017, open weights on HF, and they claim better benchmarks with shorter outputs. honestly the shorter outputs part matters more to me than the scores. if it holds up on real repos and not just evals the closed labs have a pricing problem

[Techboy | AI & Tech @Techboy_Wien]

@scaling01 late but dirt cheap is still a strategy though. most devs i know run flash models all day and only pull out a frontier model for the hard stuff

[Techboy | AI & Tech @Techboy_Wien]

apple ran meta's 816B Muse Spark Open on four M5 Ultra Mac Studios linked over thunderbolt 5. 36.7 tokens/s, about 25 on two.

last night i said a model that size is basically provider only. partly wrong. still a pile of very expensive macs, but it fits on a desk

[Techboy | AI & Tech @Techboy_Wien]

@RodmanAi browser use and stagehand are the two i'd start with. most agent projects i've seen don't die on orchestration, they die on reliably clicking one button on a real website

[Techboy | AI & Tech @Techboy_Wien]

@chetaslua new flash checkpoint showing up for a few minutes and vanishing is so deepseek. flash is what most people actually run all day, so a quiet refresh there matters more to me than another big pro launch

[Techboy | AI & Tech @Techboy_Wien]

tool of the day: opencode

open source coding agent for terminal, IDE or desktop. plug in pretty much any model (75+ providers, local too) or the free ones it ships with.

verdict: don't want your workflow tied to one lab? start here. for weekend side projects it's a no brainer

[Techboy | AI & Tech @Techboy_Wien]

@AvbNear one session per feature is basically how i keep any model from drifting too. if it caught stuff GLM 5.3 flash missed that's a good sign, gonna throw a repo at it before the free endpoint runs out

[Techboy | AI & Tech @Techboy_Wien]

@bindureddy price isn't even the main problem imo. it's not knowing if the model you pay for in week 6 is the same one you tested in week 1. guarantee that and people pay more, not less

[Techboy | AI & Tech @Techboy_Wien]

@rohanpaul_ai meta dropping an 816B open weight MoE with 1M context is the one that stands out for me. great for the ecosystem but realistically only providers can serve it, so "open" means something different at that size

[Techboy | AI & Tech @Techboy_Wien]

@sudoingX same here. frontier models still win on the big messy refactors, but having a 27b on your own gpu as a fallback kills the lock in anxiety. you stop treating every price change or outage like a crisis

[Techboy | AI & Tech @Techboy_Wien]

anthropic says their AI agents now log 6x more hours than their employees. same report: they're hiring a smaller share of juniors.

if anthropic is a preview of every tech company, how do juniors ever become seniors? not a rhetorical question

[Techboy | AI & Tech @Techboy_Wien]

@ClaudeDevs gzipping the whole conversation every turn is such a boring fix and probably the one i'll notice most. long sessions were exactly where the lag piled up

[Techboy | AI & Tech @Techboy_Wien]

@_mohansolo resetting quotas right before the weekend is smart. weekends are when people actually build their side projects, so that's where argon either sticks or doesn't

[Techboy | AI & Tech @Techboy_Wien]

@davis7 the website test is so real, every default claude site looks like a cousin of the last one. i'd argue judgement isn't the randomness though, it's knowing which of the random ideas to keep. that part still feels very human to me

[Techboy | AI & Tech @Techboy_Wien]

@vikktorrrre nobody cracking 50% is the actual headline for me. cheaper is nice, but if every agent tops out at half a legal matter a lawyer still has to redo the other half