Vocino

I make stuff at Meta and talk about AI. Nerd tchotchke collector. Chaotic good. Perfect blend of technology and magic.

[Vocino @Vocino]

speaking of NYC, ta turnstile data used to be this legendary open dataset. every 4 hours csv of entries / exits per turnstile since like early 2010.

people built cool shit with it, like isochrone maps, rush hour anims, ghost station detectors.

now omny replaced metrocard and the data shape changed, it’s still open but it’s a bit harder. the old data is on github cleaned, check Sathyan08 gists.

if you are in nyc, tap omny and watch your entry hit the feed (4h later).

[Vocino @Vocino]

i’m in NYC this week and new york has this whole civic hacking scene called BetaNYC. it runs a national day of civic hacking, and has been for years. open data week is like their big event.

nyc mesh meetups tuesdays, betanyc meetups monthly, nypl labs workshops.

theres like 2000 routers + 180k library images + 680k trees + squirrels + rats + subway. what more do you need.

[Vocino @Vocino]

https://t.co/amua66Zr10

[Vocino @Vocino]

anyone tried these proxies that sit between claude code, codex, etc and api, dedupes, shrinks prompts? the examples are pretty aggressive but given i can still understand the intent clearly, i would imagine an LLM can.

relevant to my hermes router cost work. i'd be interested what the real world token savings is. it doesn't seem worth the hassle of yet-another-thing, but i could be wrong.

[Vocino @Vocino]

https://t.co/6LToRqPuX2

[Vocino @Vocino]

i did a bunch of cleanup and updates to my Twitch drops miner CLI project. i've had the tdm service running for ages now and it feels very solid. i brought it up to parity with the gui version and added some qol stuff.

[Vocino @Vocino]

Is it just me or is the HuggingFace hack kinda sus?

[Vocino @Vocino]

spent time last night debugging why Hermes went totally silent in Discord on me, almost pulled my hair out. it would throw a fallback model warning, get one reply out, then die.

turned out i was dumb (shocking) and i added an MCP server the day before and 2 of its tools had json schemas 22 levels deep. meta muse spark 1.1 caps at 10 and because the entire tool list is sent on every call, one bad tool breaks every request.

the fix wasn't to try to police every way tools get added because well that's impossible with autonomous agents. i did do some hardening of the one spot every tool has to pass through. wrapped the existing sanitizer with a tiny plugin that clamps depth.

[Vocino @Vocino]

@49agents yup, i agree. what i'm doing now is creating fully autonomous loops of agents to maintain these things without any involvement from me. i have a really strong working setup using hermes agent and a ton of customizations.

[Vocino @Vocino]

i keep hearing anecdotes from people who used coding agents to reverse-engineer and automate devices in their homes.

simon makes a good point - before agents, you could reverse-engineer your thermostat or whatever, but was it worth the roi? and then you own maintenance forever on an undocumented api.

coding agents change the math. trying and failing is cheap now. and throwing it away and starting over doesn't feel like waste. less psychological baggage anyway.

that's been my experience too.

[Vocino @Vocino]

https://t.co/ZP38010Ype

[Vocino @Vocino]

When you live in the bay, you constantly hear spaceship sounds from EVs going by. It's kinda weird.

[Vocino @Vocino]

TIL people are just using a wide open 1Password setup for their OpenClaw instead of a service account token on a specific vault. You guys are really yolo'ing it.

[Vocino @Vocino]

There are a lot of stories lately about AI agents talking to AI agents, as if this hasn’t been the LinkedIn feed for years.

[Vocino @Vocino]

I added a few new tools to https://t.co/Bvqy49Nm1N by request. Been thinking about how to set the bar. I think every tool should help you do something useful. We list AI tools that are built for AI and enable a productive outcome: coding, designing, researching, automating, etc

[Vocino @Vocino]

The age of AI is a superpower for my fellow ADHD lazy procrastinator archetypes.

[Vocino @Vocino]

Elon Musk's xAI launched Grok Imagine. It's a new text-to-image and image-to-video tool in the Grok app. Musk is calling it a "meme motherlode" and teasing NSFW capabilities. It's reportedly light on the usual safeguards.

[Vocino @Vocino]

There's news on "mainstream" outlets that OpenAI's new coding model helped build itself. That GPT-5.3-Codex is the first model that was "instrumental in creating itself," but imo this is cool but almost not news anymore. We're at the point where everything will be like this

[Vocino @Vocino]

DeepSeek-V4 is coming this week (already posted about this). Leaks and rumors are pointing to a March 3rd release. This is the one optimized for coding and long-context, and it could be massive if it's as good as the benchmarks suggest.

[Vocino @Vocino]

Claude is still #1 on the App Store. The fallout from the Pentagon drama might be contributing to fuel its growth but honestly I think moreso it's just people realizing these things are... a thing. Free users are reportedly up over 60% since January.