Sabitlenmiş Tweet
dex
5.7K posts
dex
@dexhorthy
building the post-IDE IDE at https://t.co/hDpglj9veo - @aitinkerers sf lead, prev @replicatedhq @SproutSocial @nasa - ai that works pod @ https://t.co/69BhaNtopF
San Francisco, CA Katılım Ocak 2017
1.8K Takip Edilen18.5K Takipçiler
Introducing: Codex for SteamDeck
alex fazio@alxfazio
hear me out r1/r2: next/previous tab x: enter hold x: tts circle: esc ↑↓: ↑↓ square: `$` triangle: `/`
English
@0xblacklight I mean technically claude is the only one who should be outsourcing the thinking
English
@dexhorthy Claude outsource the thinking for me make no mistakes ultrathink
English
bad: have your agent = use async bash tool calls to start your dev server and then it has to have like 3 other tools to manage it and pull logs
better: tee your dev server's logs to a file and have it grep them
best: run it (or have your agent run it) in a named tmux pane
English
I’m disappointed in anthropic but I am PISSED at the openclaw / Hermes grifters trying to steal inference that made this necessary
The market is the market so if something is possible and valuable people will find a way to get at it
Tbh probably should have seen this coming despite the “sdk is still fine” post from @bcherny a month ago
Cheap inference was a Weird blip in an increasingly weird world
English
dex retweetledi
@_joshd @lux_sp4rk I spent two years building a product on top of twitter - that company has since IPO’d - but it was hard
English
@dexhorthy @lux_sp4rk I'm reminded of all the people who used to build their products on top of Twitter
English
@dexhorthy anthropic: bitter lesson, use inference for all your problems
everybody: *uses more inference*
anthropic: wait, no not like that
English
> It was well within Anthropic's ability to see who is abusing it and have different rates for that kind of use.
if this was true then why did they end up adding weird blocking rules like "cant use OpenClaw in your system prompt" and "cant can openclaw mentioned in recent git commits"
> For example, they already can tell if you have multiple agents working in parallel. Just have that use subscription tokens faster.
yeah i think this happens by default since they count tokens
> Why should they get charged extra? Still using Claude Code, just with a UX polish pass.
we are working on the same. this news is incredibly frustrating to me, in case that was unclear 😅
English
It was well within Anthropic's ability to see who is abusing it and have different rates for that kind of use.
For example, they already can tell if you have multiple agents working in parallel. Just have that use subscription tokens faster.
We make an app that makes Claude Code more usable for developers and "regular" people. Why should they get charged extra? Still using Claude Code, just with a UX polish pass.
English
What was the post the other day.
About clean domains, and messy domains.
And how a lot of work is in messy domains.
Where feedback signals are not always cheap, clean, fast?
dex@dexhorthy
“why do we still need humans in the loop for the design of software” Because the cost function of bad architecture is measured in months and years, and it’s hard to measure that in an RL environment/benchmark like SWE bench multilingual that is measures a 5-10 minute “can we write the code and get the tests passing” Spoke about this at @AIEMiami - been Obsessed with this after a convo I had with @calvinfo back in April youtube.com/live/6IxSbMhT7…
English
@nokturn_ai Yes I in fact referenced that conversation in the post at the top of this thread
That is different from openclaw use
Read again?
English
@nokturn_ai was pretty clear in feb they did not claude subs used for openclaw x.com/dexhorthy/stat…
dex@dexhorthy
Clarification and My read: I am not a lawyer, based on @trq212 clarifications today 1 don't use max plan in other coding agents 2 don't use the max plan in openclaw 3. If you use the max plan for personal projects and you use the Claude agent sdk that's ok 4. If you are building a saas, do not serve other people inference using a max plan auth token, even if via the sdk So I think usage of codelayer or riptide or conductor etc is closest to 3 but it’s technically not well defined. However as @0xblacklight said we are preparing many alternatives in case this becomes an issue including finding cheap inference and passing through cost, allowing open code / codex sdks etc If it turns out 3 is fine and anthropic is mostly worried about 4, then yes actually I think they are in their fair rights to try to put a stop to 4 and I wouldn't hate on them for it publicly
English
It's a bit unfair to blame consumers. Anthropic sets the rules, like they are doing now. Nothing that openclaw or hermes users did was breaking that agreement. The problem is that anthropic was banking on the idea that most of its customers would use less than the $200 worth of tokens their subs cost - but it didn't average out, not even close to, so they had to pull the plug. IMO its deserved, a 3rd party app that hides the product (claude CLI) shouldn't be granted special treatment. We (nokturn) and others put the CLI top and center, and don't rely on -p or sdk to bring the same level of automation to the table, and with a lot of added benefits
English
@ShrekOverflow @paulodoestech i usually write ~ 50 lines and then continue iterating on it as I see the model getting off track, e.g. I saw it was doing parity tests wrong so i added this block #guidance---how-to-build-parity-tests" target="_blank" rel="nofollow noopener">github.com/dexhorthy/shan…
English
@dexhorthy @paulodoestech Do you use a meta prompt or do you hand craft $LOCATION
English
“why do we still need humans in the loop for the design of software”
Because the cost function of bad architecture is measured in months and years, and it’s hard to measure that in an RL environment/benchmark like SWE bench multilingual that is measures a 5-10 minute “can we write the code and get the tests passing”
Spoke about this at @AIEMiami - been Obsessed with this after a convo I had with @calvinfo back in April
youtube.com/live/6IxSbMhT7…
YouTube
English
@ohhsotheysay @Kxbdjmsbxjnd how difficult is it to get a new account if you get banned? genuine question, i haven't looked into this
English
anthropic could have easily decided to just punish the accounts violating TOS? You think they don't have the capability for that? No. They decided they regretted the agent SDK and didn't want to support tools calling claude -p, and rather than get the flack for a straight up ban they lied to our faces acting like this was them being generous.
English
@dexhorthy yes, so it seems to me the actual problem is anthropic didn't start caring about compute until 5 minutes ago, another questionable judgement from them!
English
if it was so necessary and openclaw/hermes are just so awful, can you tell me why i can use my codex sub for all of that just fine?
you can stick with your infinitely degrading claude sub. i decided i wanna choose what tools i use, not just use anthropic's slop.
dex@dexhorthy
I’m disappointed in anthropic but I am PISSED at the openclaw / Hermes grifters trying to steal inference that made this necessary The market is the market so if something is possible and valuable people will find a way to get at it Tbh probably should have seen this coming despite the “sdk is still fine” post from @bcherny a month ago Cheap inference was a Weird blip in an increasingly weird world
English