@karpathy Wow, congrats! Hope you keep posting dope things though 😄
English
João Duarte
282 posts
Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
English
Still trying to process it all.
Very grateful for the years at Cloudflare. Learned so much and got to write, host/produce, edit vid (#ThisWeekinNET), build, and tell great stories over 5 years. Never thought I’d PM a visual storytelling product that travelled the🌍.
More soon. ❤️
English
@maxkarpis I was going to travel to the UK and I knew that I had to pay 4€ for each transaction. I spoke with my bank, a well known Portuguese one, to figure out how to deal with this. They asked me if I knew about Revolut..
English
Revolut has surpassed 2.3M customers in Portugal🇵🇹 and expects to reach 2.5M by the end of the year.
In 2025, customer growth was 35%, above the EU average of 30%.
Deposits from Portuguese customers grew by 64%.
Revolut expect to become the largest bank in Portugal by overtaking Caixa Geral de Depósitos.
In the near future, Revolut expects to offer more credit, insurance and investment products in Portugal, but no timeline for mortgages. Also, plans include more products for businesses.
Source: Dinheiro Vivo
English
Devido as mudancas recentes no algoritmo do X, vou passar a escrever so em portugues portugal daqui para a frente
Como estao todos? As minhas pastel de nata sisters estao bem? E os bifana brothers, tudo bem? Apanharam algum peixe bom hoje?
Alguem viu o jogo ontem?? Meu deus que partida. Sporting ou Benfica, qual e o vosso lado?
Tenho de ir, azulejos para pintar e sardinhas para apanhar
Ate a proximaaaaaaaa!!!!
Nikita Bier@nikitabier
Starting Thursday, we'll be updating our revenue sharing incentives to better reward the content we want on X: We will be giving more weight to impressions from your home region—to encourage content that resonates with people in your country, in neighboring countries and people who speak your language. While we appreciate everyone's opinion on American politics, we hope this will disincentivize gaming the attention of US or Japanese accounts and instead, drive diverse conversations on the platform. We invite creators to start building an audience locally. X will be a much richer community when there's relevant posts for people in all parts of the world.
Português
This guy emailed me asking for a *token session refund* because his claw made mistakes. 🙃
English
@CalebPeng Makes sense, although I think nerfs and buffs are also a tool to just shake the meta, and what's popular, in order to force something new and fresh, even in cases where things were actually balanced.
English
I’m a big believer in nerfing Pokémon that are broken and OP, not just because they’re good and some people find it annoying to play into.
Scizor and Cradily were good Pokémon, not broken Pokémon. This is the result when you nerf good Pokémon.
I don’t think Wigglytuff is broken but without enough checks, it will look like it. Fast forward 3 months, if Wigglytuff gets a nerf, something else will look broken and people will complain about that for the following 3 months.
Just food for thought when people raise pitchforks for nerfing Pokémon they don’t like facing. There are cascading effects and the grass isn’t always greener.
Play! Pokémon@playpokemon
Let's GO ‼️ Here are the most used Pokémon at the Houston Regional Championships for #PokemonGO!
English
@simas_ch Copilot cli, afaik, also can charge per request instead of token count. This can be very good, depending on usage.
English
Why should is use Coplito CLI instead of Claude Code if use the Anthropic models anyway?
English
Dear developers at work,
GitHub Copilot CLI is generally available with /plan, /plugin, /resume, /review, /yolo, /models, multi-model in single request autopilot, custom agents, experimental features, and much more.
Love,
GitHub🌹
github.blog/changelog/2026…
English
I'm joining @OpenAI to bring agents to everyone. @OpenClaw is becoming a foundation: open, independent, and just getting started.🦞
steipete.me/posts/2026/ope…
English
@steipete Here is a human reply my dude 😎
Going through @lexfridman pod right now, loving it!
English
Folks, please don't build bots that automatically reply to stuff on X, or use AI to reply. It makes this site annoying to use, I can barely keep up with the blocking.
If you use AI to "tweak" your real replies, it will still smell like AI. Embrace typos and imperfect grammar.
English
@urso_de_shorts Só para que não haja ambiguidades, de onde vêem estes dados? Podes meter o link no tópico?
Português
Olhem-me para isto, malta. 🧐
Dados fresquinhos da execução orçamental 2025
Receita fiscal + contributiva +7,5% no ano. PIB nominal? ~5,7%.
Ou seja ( para crianças de 5 anos) : a carga fiscal efetiva SUBIU, mesmo com o Governo a gritar "cortámos IRS, alívio para as famílias!!"
IRS? Cortaram escalões, ajustaram taxas... e ainda assim IRS +9,3%, IVA +9,8%, contribuições sociais +8,3%.
Bracket creep ( arrastão das taxas) em modo Deus
Salários sobem um bocadinho → empurram-te para escalão superior → o Estado come mais sem mexer uma palha. Receita fiscal total: 72.697 M€ (previsto 70.664 M€).
Sobrou +2 mil milhões. Parabéns, contribuinte.
Excedente orçamental? 1.298 M€ (~0,4% PIB).
Como conseguiram? Mais impostos do que o previsto + investimento público subexecutado em ~3 mil milhões.
Traduzindo- cortaram obra pública para fingir que as contas batem certo, clássico.
Governo projeta carga fiscal 34,8% em 2025 e 34,7% em 2026.
Realidade? Com esta elasticidade fiscal do car*alho, bracket creep + massa salarial a bombar + consumo resiliente → a carga efetiva fica colada nos 35% ou mais.
Alívio? Só na propaganda. Na carteira real, o povo que pague que s@fod@...
Prometem descida estrutural → arrecadação automática explode.
Anunciam excedente → graças a mais IRS/TSU e menos betão. Competitividade? 6.º pior sistema fiscal da OCDE para empresas.
"Power" de compra das famílias? "Esforço fiscal" 13% acima da média UE.
Continuamos a ser o campeão da Eurozuela a sugar o rendimento sem dar retorno proporcional.
Parabéns a todos...
Português
People finally think I'm Portuguese
I'm going to cry 🥹
Obrigado por sua attençao em esse coisa!!!
R1bscholar🐦🔥@R1bradical89
@levelsio I don't know who you are but your location says Portugal. I find it very hypocritical to talk about extremes because my grandpa has some very interesting stories about what you guys got up to to hold onto your colonies for as long as possible unlike the rest of Europe. Also this.
Português
My main conclusion from the Epstein Files is that nobody rich and powerful knows how to type or spell and are bordering on illiterate
Mike Benz@MikeBenzCyber
I’m utterly blown away by this new audio in the Epstein Files. It shows the secret conversation you fear every high-level gov’t official might have with an outside fixer to hook them up the moment they “leave government to enter the private sector.” Here, Jeffrey Epstein teaches Ehud Barak — while he is still Minister of Defense in Israel but on his way out — how to make millions of dollars in the private sector by first compiling a list of people who owe him favors from his 40 years in government service. Epstein then told Ehud Barak to pursue Peter Thiel, whose name at that point Ehud Barak did not even know how to spell, and to try to connect with a company called Palantir, which existence Ehud Barak had never heard of, and whose name Jeffrey Epstein didn’t know how to spell. Credit @RyanGrim whose clip I first saw to look for this. I used Grok to backtrace the source link from the audio. Ryan assesses this meeting took place in February 2013, which seems to line up with the audio saying Ehud Barak is about 71 at the time of the recording, as Ehud Barak is 83 today.
English
@esrtweet Can you make a video, or a stream, showcasing you making a project from scratch? Or an already existing project, but something that displays your flow. That way it could be easier, possibly, to understand the differences, and I also suspect it would be super pedagogical 🙂
English
Things That Puzzle Me, item #4632: why is there such a huge variance in results from using LLM assistance for programming?
For me, it has been dead easy and I've gotten spectacularly positive results basically from day one. There was the usual minor friction of learning new tools, but it was certainly easier than (say) learning a new programming language.
The only really new working method I had to wrap my head around was always maintaining a spec document that I write my design thoughts into as I have them, so the LLM can load that into its context.
I'm not seeing a high error rate in generated code, either.
And yet, others report being lost In a wilderness of LLM hallucinations, crappy generated code, and failure to get a real grip on their problem domain.
I have no idea what's going on here.
The obvious theory would be that LLMs reward super-competence in programmers using them and I'm super-competent. This is a seductive idea because it feels good, and because...er, well, there are other lines of evidence suggesting that I am in fact a super-competent programmer.
Besides the fact that I'm automatically suspicious of any explanation that feels that good, this one founders on the fact that I've heard many reports of failure and frustration from people who don't seem to be incompetent.
So, um, what's going on here? What possibilities can we eliminate?
It's probably not the case that I'm having low friction because the problems I work on are simple. I gravitate towards problems with high algorithmic complexity and complex data structures, because I have mathematician brain and that's what mathematician brain likes to chew on.
Anyway, I do frequently write simple tools to assist my complex projects (like my most recent new project, a batch mode spell checker for documentation in software source trees), and I don't notice that the LLMs are any better at the simple code or any worse at the complex stuff.
It's not the LLMs or LLM tools I'm using because I'm using the same ones as lots of other people. Currently ChatGPT 5.2 via Codex. I'm not doing the fancy stuff like agent flocks yet because I haven't yet seen a need to.
I've considered the theory that the key variable is the ability to write clearly in English. Again, this would be an easy feel-good explanation for me, as I've done several successful books and a New York Times bestseller; for this very reason I distrust it.
I think being able to write clear high-quality prose to LLMs used to be more important than it is now. I notice that my prompts have been getting shorter and less explicit without a fall-off in the quality of my results, and I think that's the models getting better at correctly interpreting ambiguous instructions.
One possibility I haven't eliminated is that LLMs reward system-architect skill, having good intuition and taste about design. I know some very capable programmers that don't have this; they can do routine work and maintenance and debugging, but don't originate projects and don't thrive if you ask them to design in areas where they don't have established practice and a lot of precedent to fall back on.
It's not easy to tell from a distance who has system-architect skill and who doesn't. Well, except that the design leads of successful major projects pretty much have to have at least some of it, or the projects couldn't succeed. Because I can't tell who has it, I can't judge that people constantly having bad interactions with LLMs don't have it.
The main problem with this theory is elsewhere; that I can't figure out how system-architect skill could translate to behaviors that are visible to an LLM through its text stream input.
So, none of the possibilities I've considered are persuasive. I continue to be puzzled.
English
@MohMaya97 @steipete Opus and 5.2 have been praised a lot. Is Gemini not as good?
English
holy crap. You're right. This complicated piece of MLOps Orchestration I am building took like 2-2:30 hours. But the code that it has written it just ⭐️. This is literally how I write code. All upto my quality gates. And passes ruff and test suites in a single go.
I take back this tweet I made the other day. I was wrong. @OpenAI models (via @opencode harness) really are top notch.
I couldn't have been more wrong. Opus and Codex are two different kinds of models. Opus has me doing things every minute or two. Codex actually does the work. And I think it is growing on me.
Thank you.
x.com/MohMaya97/stat…
English