Dan Shipper 📧

Software engineering in 2026 needs two roles: A pirate and an architect. The pirate codes as fast as possible to figure out what's valuable. The architect turns that sloppy mess into a well-oiled machine. Here's how it works and why:

@interplato @every @MTSlive @TrySpiral yup

@every @danshipper @MTSlive @TrySpiral Can I pull it into Claude

We talked to @danshipper about what makes great writing great. Dan, CEO of @every, says AI writing won’t feel alive until models learn continuously. "It comes from a unique person in a unique circumstance expressing their unique view on the world in words that are theirs." "The problem is not the specific, it's not X, it's Y, it's that the model, because it's not changing, it's going to use similar language in similar situations." "It's actually not at all about writing quality to me for models... is it using sentences that change and feel alive? I don't think that will happen until the models are learning continuously."

@MTSlive @danshipper Looking for a better way to do great writing with AI? Start a free trial for @TrySpiral, our AI writing partner with taste, trained on your style: writewithspiral.com

@daniel_mac8 likely soon but not this week! x.com/danshipper/sta…

Claude Mythos this week? Last few Anthropic dev conferences brought new models. Based on the read of dev vibes, Anthropic may need to make Mythos GA sooner than later.

no mythos this week! but some interesting stuff coming :)

@streamingsou 🙏🙏

@danshipper made me do it.

@soumitrashukla9 @every Will make one!

@danshipper @every @danshipper is there a live tutorial of how to use it? Was trying it earlier and wasn’t sure what the main differences were between opening a google doc via Codex in-app browser and editing there. 100% sure I’m being dumb

clear that this is how we'll be doing most of our work for the next 10 years: agent running continuously on the left, application that you + the agent use on the right

@LessonHist ->

@danshipper You can’t learn faster than the models.

this is correct and the reason is simple: models know more than any individual human. but any individual human learns faster than models do

@every if you want to try a codex-native app, use proof to write: proofeditor.ai

read more about this on @every: every.to/subscribe

@rywigs it's definitely not good at this yet! whole new paradigm

@danshipper How good has it been and knowing when not to speak? agents can feel like nervous interns at times, filling space

@emollick this is so funny hahaha

(Sorry, after seeing so many of these, could not resist): 🚨 BREAKING: Google just dropped a NEW paper that completely deletes RNNs from existence. No recurrence. No convolutions. Nothing. Just one mechanism. And it’s destroying every translation benchmark on the planet. The title alone is a flex: “Attention Is All You Need” Vaswani. Shazeer. Parmar. Uszkoreit. Jones. Gomez. Kaiser. Polosukhin. 8 researchers. 1 architecture. The entire field of NLP will never be the same. Here’s why this is INSANE → LSTMs took DAYS to train. This thing trains in 12 hours on 8 GPUs. 🤯 → 28.4 BLEU on English-to-German. That’s not an improvement. That’s a MASSACRE. They beat the previous SOTA by over 2 points. → English-to-French? 41.8 BLEU. At a FRACTION of the training cost of every model that came before it. → They called it the “Transformer.” The name alone tells you they knew. But here’s the part nobody is talking about 👇 They threw out sequential processing ENTIRELY. Every other model on Earth processes words one at a time. This thing looks at the ENTIRE sentence simultaneously and figures out what matters. It’s called “self-attention” and it’s basically the model asking itself: “which words should I care about right now?” Every. Single. Token. In parallel. Do you understand what this means? Training that used to take WEEKS now takes HOURS. Models that couldn’t scale past a few layers? This thing stacks 6 encoders and 6 decoders like it’s nothing. And the multi-head attention? 8 attention heads running at once, each learning DIFFERENT relationships in the data. I’m not being dramatic when I say this paper just rewrote the rulebook. RNNs are cooked. 💀 LSTMs are cooked. 💀 The future is attention. And attention is ALL you need. Follow for more 🔔

@BradleyYoungjr exactly

@danshipper Agreed. They attempted this with integrated a side panel in browsers but they actually needed to reverse it and have a browser in an ai agent environment not an ai agent in a browser environment

interesting discussion about this: x.com/stepango/statu…

I agree with the rate of acceleration but also feel that what humans do is surprisingly complex even for an exponential curve I dont think its possible to separate feeling, relationships, art, body, and self understanding from learning rate (broadly defined) But i also appreicate your thoughtful replies and I’m sorry for QTing in that way. Will edit or delete if i can! Enjoyable discussion

@danshipper > I think it will be a very long Rate of acceleration tells me otherwise. To your question - IMO most of the things that makes you human, feeling, relationships, art and everything that involves the body and self-understanding, things which are deeply individual.

> Imagine sub-second fine-tuning-like process on every prompt or just some high precision long term memory mechanism combined with good context management if we are talking in common AI terms. even if you can get some real-time-ish weight updates, i think it will be a very long time before models make the kind of realtime learning adjustments at the level of depth and insight that humans can > There are things that humans can do that models will not be good at, but I don't think that speed of learning is one of the things that humans need to compete with AI on. im curious what you think those things are if learning is not one!

That's assuming model architecture won't change, and models can learn how to use other models, internet has a lot of data to learn from. There are things that humans can do that models will not be good at, but I don't think that speed of learning is one of the things that humans need to compete with AI on. Even now you can teach model specific skills via fine-tuning and it will never need a second explanation, it's just too complex and expensive to operate this way. Imagine sub-second fine-tuning-like process on every prompt or just some high precision long term memory mechanism combined with good context management if we are talking in common AI terms.

@andrewchen pms who dive into using coding agents are scary good written by one who works @every: every.to/guides/ai-prod…

bullish on the PM role quietly becoming the most important role in tech again when anyone can build, the person who decides WHAT to build becomes the bottleneck

it's true that any model knows more than any individual human. but humans learn faster: we update after each interaction, models don't. humans who use models in their own lives and domains of expertise learn new, local, tacit expertise that the models can't know because it involves them. this expertise may be included to some extent in the next training run—but in that case the same process repeats

🤔 what part of it is dim? Growing up ~16 years, getting a degree ~4, may be faster for some. That's a single domain skill set, training the model - few months, the same amount of knowledge and skills average person will get by the age of 20 in wide set of domains, every new model accelerate the rate of learning and depth of knowledge in the available domains, which is much faster then most humans.

@smcllns @every he's not! likes to keep a low porfile

@danshipper @every Excellent read, thanks for sharing. Is Marcus on X/Threads?

must read Marcus went from product manager to shipping product like a madman @every with coding agents he wrote the definitive guide for how to do it: every.to/guides/ai-prod…