hue 🎨

I've honestly been born in the best time in human history what a gift to be in my 20s and have alien brains smarter than me at my call and have been studying CS for so much of my life and have such great online friends interested in the same stuff as me anyways I love you all

Have a neighbor you hate? HOA president bothering you? Guess what, Daryl? Enjoy the new 24/7/365 permanent sunlight I ordered up for your house, bud.

@nickcammarata do you have any writing besides your twitter? I don’t think you have any obvious links anywhere but would love to read whatever you’ve made!

basically every time i google something that's more than one off i first think how could have a system that would have done it for me proactively so it would have already just been done. the cost of building software has dropped to near zero so building custom agent things easy

if i knew what was possible in terms of research per hour today five years ago i wouldn't have been able to code by hand, its like a hundred times slower and i couldn't have mustered the energy. i wonder if the same is true now

This would be so cool if it came from basically anybody but meta I mean it’s still cool. Hats off to the researchers. But…

Yo fui ingeniero en Meta, y siempre seguía FAIR desde adentro. Lo que acaban de publicar es la versión que les dejan publicar. Pero con eso, es más que suficiente para decirles exactamente que es lo que está pasando. TRIBE v2 predice, vértice por vértice sobre la corteza cerebral, qué zonas activa cualquier video. Sin escáneres. Sin humanos. Subes el contenido, obtienes el mapa neural (activación emocional, supresión de razonamiento crítico, modulación prefrontal) antes de que el video lo vea un solo usuario. Ahora considera la posición de Meta: 1. Tiene años de datos de Reels sobre qué contenido retiene atención, genera enojo, provoca compartir. 2. Saben empíricamente qué funciona. TRIBE v2 les da el mecanismo causal de por qué funciona (a nivel de tejido cortical) Eso convierte correlación histórica en capacidad predictiva sobre contenido nuevo. 3. Internamente hay herramientas que se llaman Gatekeepers y Quick Promotions que sirven para inyectar contenido en el feed de poblaciones arbitrarias a escala. 4. Simulador de respuesta cerebral + conocimiento empírico de contenido efectivo + maquinaria de distribución selectiva. El pipeline está completo. Y luego está Thiel. Inversor y amigo personal de Zuck. Fundador de Palantir, cuyo negocio es análisis de poblaciones a escala para gobiernos e inteligencia. NO es descabellado observar que confluyen los incentivos de plataformas construidas por las mismas personas. La licencia CC BY-NC dice que Meta retiene los derechos comerciales del predictor de respuesta cerebral más preciso jamás construido. Y recuerda, esto es lo que decidieron hacer público.

this is the best advice i can give rn for learning and running agents without breaking my own confidentiality: if you are using other people's agent harnesses, then you don't really own or control your own processes or agents you can fork an open source harness sure, but reality is if you truly want full control of everything, you need to build most of it yourself and understand how it works Hermes is probably the best open source harness rn - i'll give @Teknium credit where it's due, i did check it out and it is refreshing to see someone who actually built something that knows something about RL and didn't just build some agent buzzword slop for fake marketing hype like the rest of them but still - if you truly want to get into ai and agents and envision yourself working in this space for many years the most high value thing you can do for yourself is build your own harness through building your own harness you will naturally learn most things that need to be learned about ai and agents you'll face problems, you'll have to figure out memory, you'll have to figure out tools, plugins, context, orchestration, juggling apis with local models, embeddings, rag, storage, CLIs, mcps, sandboxing, skills, context, loops, sub-agents, hooks, execution chains, queues vs concurrency, streaming, polling, websockets, sse, connecting to other apps, token budgets and guardrails, session persistence, setting up permissions for using your computer, reinforcement learning, and so on and so on you'll basically learn everything that you need to know about ai and agents because you'll be building out your own system and figuring it all out as you go there's no other way to truly learn all these things than to build it yourself and apart from that, when you own your own harness then you can design it exactly how you want you can design your own RL system and train your own agents, you can train small models even like 0.5b models or something the opportunities are endless and there is no other skill that's as high value as this right now you just gotta lock in

@emollick @paulg a few people have been experimenting with using the "only people I follow, and people who they follow" feature to stem the tide but for accounts as big as yours, not sure if it would help...

@paulg Won't help

I now get so many AI-generated replies that it's just too much work to report them for spam and block them, so I'm downshifting to muting.

@parmita shockingly, yes!

can u reply to this

@levelsio so what exactly is the bar here? because there’s no way you’re following my lowbie acc

Okay let's see who can reply to this

@justalexoki @liminaldoge i just barely snuck in here

friends and friends of friends! welcome to taoki city! introduce yourself

the only true statement in software

@thinkingshivers bro was 6 months ahead of the top ai labs once again

In hindsight, these puzzles were pretty much the same as those in the ARC-AGI-3, and so it's not surprising that the models failed at them so disastrously.

@anpaure @tenobrus is it related to this? it seemed like a good idea at first but starting to reconsider the downstream

@tenobrus i have a strong suspicion that there are bad second order effects

strongly considering making all my posts "friends of friends replies only" for a while. i follow 2.5k people and they collectively follow an assload more, so its really not too strong a filter, but it'd keep out bots and random groypers. downside is it blocks brand new lowbies...

@tenobrus on the other hand, you are offloading filtering to your network so not sure if this is net good for the community as a whole

@tenobrus “friends of friends” seems to be a pretty loose rule because all they have to do is convince one of your 2.5k nodes to let them in, lots of people follow back (even lowbies) so a motivated person could hop that bar in <1 week if they actually wanted to

@dejavucoder honored to make the cut

mutuals and people who mutuals follow (despite me not following them) can talk in this thread

Any situation is fixable if you’re good with words because words literally make up the meaning infrastructure, control the traffic of desire, and reorient the posture of someone’s heart which is where the true center of gravity is in a person that makes them want to do anything

@tenobrus it’s noon and i’ve yet to get out of bed

gm friends and friends of friends . how are yall today? what did u have for breakfast?

you only get: 1 summer with them as a baby 3 as a toddler 5 as a child 3 as a preteen 6 as a teen so let them run barefoot, turn on the sprinkler, cut up the watermelon, chase fireflies under fading golden skies because the summers of sticky hands, sun tired kids, and slow evenings filled with the sounds of childhood do not last forever 🤍

ARC-AGI-3 scores for GPT-5.4, Gemini 3.1 Pro and Opus 4.6 Gemini 3.1 Pro: 0.37% GPT-5.4: 0.26% Opus 4.6: 0.25% Grok 4.2: 0%