Vlady

👨🏻‍💻 10+ years in Tech
☕️ Home barista for 2 years
👾 PS5 Pro (streamers supporter)
Building Recall, my first iOS app for readers 📚👇

[Vlady @itsvlady]

Widgets Dark Theme is ready
Help recalling book quotes & reflections

[Vlady @itsvlady]

@borrhensaidi 9 left, that's the hardest stretch right there

[Vlady @itsvlady]

@padavanchik That tracks. I'd guess AI generated code actually needs a slower review, since it can look clean and confident while still missing context on why something should exist at all

[Vlady @itsvlady]

Ok, questions to software engineers:

Do you actually review PRs anymore now that AI can do it for you?

[Vlady @itsvlady]

@xsondesign unique is the hard part, everyone can add a fade in

[Vlady @itsvlady]

@play_loh Haven't tried Hermes yet, what's it doing better for you, coding or something else?

[Vlady @itsvlady]

Seriously, what’s your top one in AI now??

1. Grok bot
2. Claude Code
3. Codex

[Vlady @itsvlady]

@byfurkanesen Fair, and honestly the design/business logic part is the harder half to fake with a quick glance. AI can flag a null check, it can't tell you the abstraction is wrong for where the product is headed. That judgment is still on us

[Vlady @itsvlady]

@thedarioaltman Fair, but constraints alone don't tell it which corners are safe to cut. That judgment still comes from having shipped enough to know what breaks

[Vlady @itsvlady]

Seven rules in one file make your AI code like a senior dev

Somebody open-sourced the whole setup after burning 60 billion tokens, and it lives in an agents.md your AI reads before it writes a line of code. What the rules tell it:

→ Use the simplest thing that works - no speculative abstractions you never asked for and then have to maintain

→ Stick to the dependencies already in the project - stops the AI adding a package for something you already have

→ Check the docs before assuming - kills the guessed API call that looks right and doesn't run

Fewer prompts, way less tokens burned.

I went and read the file itself, and it also tells the AI to delete the old path instead of keeping the old way alive next to the new one.

Save it for the next repo you hand an agent (links below 👇)

[Vlady @itsvlady]

@federicodonaton Claude Code, easily. It's the one I actually keep open every night, the other two I poke at but don't build with

[Vlady @itsvlady]

@xfactorcore @DmitryFounder_ Ha, fair, you jumped in before there's even a screenshot up yet. It's an app for saving quotes from books, still weeks from launch, but I'll remember the offer once there's something worth a designer's eye

[Vlady @itsvlady]

Eighteen grand for a Mac Studio, and I get it

The 1.2TB/s of memory bandwidth is what you're paying for, that number decides whether a big model runs on your desk or crawls through it.

I rent mine and keep the cash

[Marques Brownlee @MKBHD]

M5 Ultra has also landed

- Up to 36-core CPU
- Up to 80-core GPU
- 1.2TB/s for unified memory

(A fully specced-out M5 Ultra Mac Studio with 16TB storage and 256GB of memory now costs $18,299)

https://t.co/h2XOgf5P82

[Vlady @itsvlady]

@ashley_cooper95 the 2K beating 100K example makes it click, that's a good pitch line honestly

[Vlady @itsvlady]

@Hc0vdeuHell Ha, glad it lands. The child-agent budget thing is the one I want to see in practice next, cost per run has just been a guess so far

[Vlady @itsvlady]

The background shell fix is the one I'll feel every day.

I've killed my own runs in Grok Build more times than I can count, typing a follow-up while the command was still going. Every time.

Child-agent budgets are the part I want to poke at next, cost per run has been guesswork for me

[Mark Kretschmann @mark_k]

Grok Build 1.0.9 is out! @SpaceXAI keeps shipping improvements fast.

The biggest changes:

• Workflows can now set child-agent budgets and reasoning effort
• Agents can discover and launch workflows by name
• Plugin-provided agents now appear in /agents and can be toggled
• Concurrent subagents are much more resilient to rate limits
• Running shell commands move to the background when you send another message instead of being cancelled
• /minimal and /fullscreen can now switch instantly without restarting the session
• MCP-heavy setups start faster
• grok clone is much faster, fetching only the selected branch by default and reusing local checkouts when possible

A solid update, especially for multi-agent and workflow-heavy use.

[Vlady @itsvlady]

@datachad True, and it's the boring kind of problem too. No dashboard, no per-command breakdown, just a total that shows up after the fact. I'd take a rough live estimate over a perfect budget knob any day

[Vlady @itsvlady]

@dylan522p first gen beating Rubin on paper is one thing, curious if that holds once it's running at OpenAI's actual production scale not just in the lab

[Vlady @itsvlady]

@ardent__dev Ha, fair, I rounded loose there. Point stands either way though, nobody cared until I actually showed up and did the thing

[Vlady @itsvlady]

I’ve started here 2 weeks ago

Now:
✅ 923 Verified followers
📈 ~ 100k Verified Home views

Everyone doesn’t give a shit until you actually do it

[Vlady @itsvlady]

@BUSYBRAINTRADER Not really looking for a follow-for-follow swap, I'd rather grow slow with people actually into reading or building. Good luck with it though

[Vlady @itsvlady]

Can you run a frontier model on a single one from now?

[Alex Ziskind @digitalix]

m6 mac mini and m5 ultra mac studio are on order

[Vlady @itsvlady]

@AWar1586398 Fair, it's a known problem. I just haven't seen anyone from SpaceX actually publish the mass budget for the radiators on this specific system, and that's the number that tells you if it pencils out

[Vlady @itsvlady]

Datacenters in orbit before we fix zoning laws, sure

Nobody in the replies gets to the cooling. Vacuum has nothing to dump heat into, so every watt leaves by radiator, and radiators are mass you paid to launch. Month six is where I'd want the numbers.

I'll believe this one when someone posts a thermal log

[Elon Musk @elonmusk]

SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028

[Vlady @itsvlady]

@wulymammoth vim user asking if you'd push play, that's a bold beta test right there

[Vlady @itsvlady]

@HemantS42019742 True, and I'd say that's exactly the case where it bites you, tight spec plus a weak model still drifts on execution even when the intent was never in question

[Vlady @itsvlady]

The better-engineer-needs-a-dumber-model line has it backwards. Knowing exactly what you want is the hardest input to execute, and a weaker model turns a precise spec into three rounds of correction while a smart one lands it on the first pass. Opus 4.8 over Fable, even when you could have written the code yourself

[Vlady @itsvlady]

@AndrewCurran_ neural operators instead of transformers is a bold bet, curious if the context window gains hold up outside benchmarks