Matt

the motivator of the day: get addicted to something else that interests you

yo @Tesla @elonmusk why does it feel like 14.6.6 is worse than 14.3?

@daviddorg @paularambles @afterlanie sir if you're not offline in the next 30 days I may need to call the police

@paularambles @afterlanie Up next: Paula scrambles?

quail qualia at the function

@ScottHickle @typesfast @BarstoolBigCat

@typesfast each 500 mL increase in daily water intake is associated with a ~7% lower risk of kidney stone formation

Anybody working on this ?

@bubbleboi what joining methods did you use?

I just built this little shoe rack.

@oonacmccaughan @Canes too much rest

Alright jokes over @Canes get it together

@ecomaniac______ think of the journey that dirt went through to get there

There is meaning in the dirt

@confucius_dr @hamptonism lol yup

@hamptonism He didn't say anything that isn't standard knowledge though, all of that should be obvious. I guess if any of this video was a surprise to you, dear reader, then you need to put down the terminal and go study lol

Just a friendly reminder - these are the quants you’re competing against.

@ModernNotoriety these are so sick

Nike Air Trainer SC Atlanta Olympics "30th Anniversary" 🍑🏅 🗓️ June 2026 🏷️ $140

@danshipper @every "They are becoming operating systems for the work itself, where you and multiple agents use the same computer, at the same time, to do highly complex, original work that can’t be done by an asynchronous agent." YUP

We’ve automated every single thing we can @every with AI agents. And yet there’s way more human work to do than ever. We’ve gone from 4 -> 30 human employees since GPT-3. I wrote a report on the structural reasons: how AI makes expert competence cheap, why that drives up demand for experts, and why the dynamic only intensifies as we approach AGI. After Automation: every.to/p/after-automa…

@letjudycook banger

he’s locked in

@ashebytes hell yeah, lmk how it goes!

now that I'm really sober all the time - not even caffeine - I feel a deep seated belief in astrology rising "they were being kinda weird, no? mm I bet our charts are just not compatible"

@tenobrus time to start a new account

@AlmostMedia social media as a game is depressing imo

Reinventing status from the ground up is going to be so bizarre.

@nbaschez @karpathy the world model agents that he builds with Anthropic

@karpathy Who will teach the children? 🥺 Just kidding congrats!!

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.

@shiraeis it sucks that u had to say all that to really say, "you have to make your tweets easy to understand by two primary audiences at once: the human deciding whether to keep reading, and the ranking algorithm deciding whether that human should see your post at all."

I wanted to know why I was deboosted so hard, so I took a deep look at the new x algorithm and decided I’d share my findings. The first thing to understand is “the algorithm” is a pretty straight-forward recommender pipeline: 1) build a representation of the user 2) retrieve candidate posts 3) score those posts with a ranking model 4) apply filters, diversity constraints, and source adjustments 5) serve the final feed so when people say “x rewards engagement,” that’s a simplification The system isn’t really counting post hoc engagement, but rather is trying to predict future engagement before showing the post. For each user/post pair, the model estimates the probability of different actions: will this user like it? reply to it? repost it? quote tweet it? click it? expand the media? watch the video? open the author profile? follow the author? share it in a dm? spend time reading it? It also predicts negative actions like will they click “not interested”? mute the author? block the author? report the post? The predicted probabilities get combined into a weighted score. Positive predicted actions push the post up while negative predicted actions push it down. “engagement” is too vague a word to describe this. A ragebait-y post that gets a lot of angry replies may not be treated the same as a post that gets fewer replies but more profile expansions, follows, longer dwell times, and reposts from accounts in the right clusters on the social graph. For example, it is very possible that from the recommender’s perspective, a post with 700 likes, 80 well-thought-out replies, high dwell time, lots of profile clicks, strong follow conversion, and reposts from accounts in the relevant topic cluster is BETTER than one with 2000 likes, 400 hostile replies, lots of mutes or “not interested” clicks, and low follow through after people see the author. Another thing to note is that your post has to actually enter the candidate set before it can even be ranked For people who follow you, posts can be retrieved through the in-network pipeline, BUT for people who don’t follow you, the post needs to be retrieved out-of-network, which means the system has to infer that your post is relevant to some user, topic, graph, or behavioral cluster. This is likely where most posts fail. A post can be good but if it’s hard to classify, or too dependent on context, or only legible to people who already know your posting arc, it may not get routed well. An example of this would be a post that is funny to your mutuals who know what you’re referring to, but has very little semantic surface area for the recommender to hold onto. Another thing is your audience is part of the signal. If your posts regularly make people click to open your profile or follow you or reply in a way that extends the conversation or spend time reading your content, the algorithm learns that showing your posts creates valuable downstream behavior and drives further activity on the platform. It’s also possible that your posts generate a lot of likes, but most people only dwell on them a short time, and not many people follow you from them, or people reply pretty weakly, or with a lot of negative feedback. In that case, the system will learn something entirely different. This goes to why having a large following doesn’t guarantee distribution. Some accounts have audiences that reliably give positive engagement signals while others have audiences that mostly scroll or only engage when something is already viral. It’s a sad reality, and from the recommender’s perspective, these are different author user content dynamics. Finally, when you’re “deboosted” (like me), it apparently doesn’t always mean there’s an explicit penalty. Sometimes there actually is a penalty (visibility filters, safety rules, spam classifiers, block and mute signals, subscription eligibility, deduplication filters, and history constraints are all real), BUT a lot of what looks like deboosting might actually be ordinary ranking failure. Maybe your post was not retrieved for many users or maybe it was retrieved but scored below competing posts or maybe it predicted poor downstream engagement or maybe it predicted too much negative feedback or maybe it was shown to the wrong initial audience or maybe it was too niche to classify well or maybe your recent account-level signals hurt distribution or maybe another source diversity rule limited how much one author or content type appeared. The feed isn’t only sorting posts by score. After scoring, the recommender still has to decide how to mix in-network and out-of-network posts, avoid too many posts from the same authors, deduplicate posts, filter ineligible content, avoid recently seen posts (it’s bad at this), and maintain some diversity in the feed. Even if your post scores well, it can still be affected by these final stage constraints. The practical takeaway for posting is to write posts that are legible both humans AND agents, which means things like a clear first line, obvious topic signal, enough keywords for retrieval, specific claims instead of vague gestures, examples early, a reason to reply, a reason to open your profile, NOT annoying bait that causes mutes or blocks. Above all it should have enough substance (or length lol) that people actually dwell on it. Interestingly enough, this might actually mean longer posts get rewarded more easily than shorter ones (everyone should be incentivized to tweet like Bill Ackman). You need to be contextmaxxing. Annoyingly, and perhaps indicative of the divergence between what this algorithm produces and what humans will want, is that the best writing often compresses context, while the best distribution often requires exposing a lot of context. The trick might be to give your tweets enough explicit structure without making them feel pedantic or over-explained. Long story short, if you want distribution under this algorithmic regime, you have to make your tweets easy to understand by two primary audiences at once: the human deciding whether to keep reading, and the ranking algorithm deciding whether that human should see your post at all.

@shiraeis @8bit5_0 do you think it's been doing a good job? imo 90% of the time it just seems to double/triple down on what you've engaged with recently. The weight seems way off

yeah p(action) is the longstanding abstraction where you estimate expected value over possible user actions then rank under constraints the present shift seems architectural tho, from more feature engineering and ensemble rankers toward sequence models and transformers that produce better user-content representations what’s new is llms are deciding what you’ll think is funny

@JakeHulberg @NousResearch @Teknium

Hermes agent by @NousResearch is THE BEST. But it reads all of your API keys :( Hermes is genuinely my favorite agent harness. I have 4 Hermes agents that help me with many day-to-day tasks for work and personal. They have access to notion, github, gmail, X bookmarks, etc. The biggest downfall with Hermes IMO is that all of your keys, tokens, etc. still sit unencrypted in a .env file on disk. The LLM behind Hermes can (and does) read all of them. This makes it susceptible to prompt injection and credential exfiltration. An attacker can trick your agent into sending it your API keys (rather easily especially if you aren't on a frontier model). It doesn't have to be this way. In the video below I integrate Agent Vault by @infisical on a separate VPS in a private network, which acts as an HTTP broker. It encrypts your keys on a totally separate box and injects them into the headers / path. Your agent NEVER sees API keys, just dummy values, and still works like normal. I truly believe this is the future of agentic security. And it. Just. Works. FYI - this architecture works with any coding agent. Claude, Cursor, Windsurf, remote coding agents, custom agents. If it speaks HTTP, agent vault can integrate and potentially save you from catastrophe. Agent vault linked below!

@flyosity @justinmfarrugia @karelvuong @samjvuong good design is meant to be copied, no? The websites are offering two different products. Even if both are showcasing quality goods at a high lvl, one is quality everything and one is quality for children's learning.

Another day, another website launch that's a sick theft of a well-known designer's popular site Curated Supply by @justinmfarrugia had its entire design and concept stolen by husband-wife team @karelvuong and @samjvuong while they take credit in the replies 🤮 Bring back shame

@gauravkapadia @nwilliams030

I am thrilled to announce the launch of Totei.com. Totei is a magazine devoted to craft and craftsmanship in all its forms. The name Totei comes from the ancient Japanese word for apprentice. I have always been inspired by those with a deep devotion to their craft—across every discipline. Making something truly remarkable requires extraordinary dedication, and the creative process behind it is rarely seen. That curiosity is why I started Totei. My hope is that everyone who reads it feels the same sense of inspiration I do when getting inside the minds of exceptional makers.

@MollySOShea @fyxlong I'm not even kidding I thought Frank was going to be an AI agent

Coatue hired a Claude Genius. aka "AI Mad Scientist" His name is Frank (@fyxlong) “He’s the one trying everything out there that is new, creative, on the cutting edge.” “Figuring out how do we implement it. How do we as an organization utilize our 20 years of data that we’ve been collecting to give us another leg up against the competition & allow us to succeed.” @coatuemgmt