Wang Xian

@authorityvortex @danielkleach Meaning we also need to build walk through video like TikTok video and catch them in the first few seconds... this is so sad man, so few people can be patient to read, and they are not even able to stay and watch!

@danielkleach Dude, I just looked at analytics, the walkthrough video that people asked for, average watch time 9 seconds. 4 people who finished it from start to finish, what a garbage people on X.

The indie hacking cycle: 1. Build something useful 2. Launch it 3. Get 12 visitors 4. Add more features 5. Launch again 6. Get 14 visitors At some point, you have to stop improving the product and improve how people discover it.

@TheNJDevOpsGuy What is the model of that subagent?

Anyone else feel like subagents with Claude Code/Fable 5 is really, really slow? 28 minutes to look at/ingest some demo code until I told it to "move it along bud, I ain't got all day" #agenticai #ai

@rezoundous What benefit do you see when using Fable as orchestrator vs Opus?

it's crazy that I only used Fable as an orchestrator, and I ran out of its weekly limit by day 2

@kimmonismus On one hand, the way algorithm detects AI is also unreliable. On the other, most people do not really have their own views to share anyway but just want to get exposure to feel better.

That matches my perception. LinkedIn is the biggest AI slop site on the internet. Every second post uses generic "not X, but Y" AI language. "Pangram Labs, an the AI-detection platform, in a recent research report revealed that it found that over 40% long posts on social networking platform LinkedIn were entirely AI-generated. "LinkedIn was the most AI-saturated platform, where more than 40% of longform posts flagged as fully AI-generated," it said."

@GergelyOrosz So it's not about build more with less, but solve more with less.

Eng team: “we’re doing more with less [fewer people] this is amazing! 🔥 “ Reality for *some* products I use that are like this: there’s a lot less alignment + friction needed thanks to a lot smaller teams. The “alignment” and “friction” no longer present: customer empathy, QA… So the eng team in this case often builds stuff customers cannot use/don’t use/don’t want to use, and far more stuff seen as buggy 🤷‍♂️

@GergelyOrosz Direction is more important than ever with the speed of build being so fast now. Engineers are good at tech directions. But it's the non-tech people who might really know what's needed. And product mindset might be required to build ways to solve more with less.

@solopribuilds I would focus on whether they actually tried before, not only willingness to try.

@wangxian0 You mean a lot or not even giving it a try ?

Who's getting fired first? -A developer who relies heavily on AI -A developer who doesn't use AI

@rezoundous It's not topped up recently sadly

my favourite Codex feature is banked usage resets

@ZackKorman This is actually a bit hard if we still want lots of convenience. Maybe very scoped credentials and data with limited explode range are more practical.

“I gave an AI agent access to my ~/.ssh directory and you won’t believe what happened next!” This is how I feel about this Grok story. Don’t give you agents access to data that you don’t want your inference provider to have.

@skeptrune Reviews should shift left. We should review PRD, TTD, and ADR harder and focus on whether implementation follows the design properly and on accuracy. With sufficient validations, that won't be much worse than code review on the lengthy, AI-generated PRs.

i personally think code review is dead. the team does not agree. directionally found this surprising.

@vaaselene Are we not there yet, just for consumer use cases?

hot take: ai models will eventually become as boring as smartphones early: new launch huge hype clear differences later: everything is good enough loyalty beats benchmarks people use whatever works do you think we'll hit the "good enough" phase?

It is contingent upon how the partners and process work together with the model. We just had one data point that Opus 4.8 is actually cheaper than Sonnet 4.6 on completed task level. Of course, if we apply the same model to any task in the benchmark task set, then the best ROI is obtained by using the best model for the relevant work. It’s a more complicated answer than peace of mind, and more about how much we are willing to pay to hold on to the better choice.

Does paying 2x for better models give you 2x more productivity? Curious question.

Not only leveling , but also those like "missing requirements", "concurrency", "documentation" etc. doing this because in different phase of / different kinds of project, each category has quite different impact if not fixed. So TL;DR is, decide whether you wanna fix it based on what if don't fix it now. And fix in a next PR or just put FIXME can also be an option, just like humans do!

@wangxian0 You're right, I probably need to limit it. Could go on forever. Bug fixes are one and done. It's more the alternate design decisions it suggest.

Here's how AI coding hurts productivity I haven't seen anyone mention this yet I now let a beefy model, like Sol xhigh, code review before I look at new code myself. Almost always flags something. But 1. Find design flaw 2. Oh shizzle, that's a good point 3. AI implement it 4. Find another flaw 5. AI implement it 6. Repeat till Anthropic resets Fable yet again See what I mean? Over-engineering is WORSE with AI. And that wastes time. Very tough for me because I love to over-engineer, like Rube and his machine. Don't misunderstand. I love AI code review with big ole models. They catch gnarly stuff and save me from a future world of pain. But balance it with: Perfect is the enemy of the good.

No it's not that straightforward. Don't just limit it. The key is categorizing review findings into families and ask to rank according to consequences of not fixing. As human engineers, we decide what to fix now based on urgency and lifecycle stage of project as well. So we just model these and models can make good judgements from there already.

@rezoundous Also seems the best of the best are never sold in annual plan.

don't subscribe to AI annual plans.

@phuctm97 Like if he has enough usage limit to keep doing this!

A friend of mine matched with guy on Tinder. They scheduled a date on the 14th. The guy just canceled it so that he can spend more time with Fable 5 right after this announcement.

@vaaselene @alex_prompter But people shall use Fable more sparingly. What's happening now shows that people are still just tokenmaxxing.

@alex_prompter that's why we got Fable, for people who can't compromise 'good enough'

hot take: opus 4.6 is good enough for most tasks

@Conor_D_Dart Or that they are waiting for Opus 5 to come out.

Claude is giving customer another week access for Fable 5. I am not sure why they keep giving us a week at a time, but I I think more clear communication could happen. My thinking is - They are still very limited on compute and being able to plan week by week, lets them allocate compute to Fable easier than just saying its staying indefinitely or giving us a month access.