Vlad ⚡

I spent hours analyzing X's new open-source algorithm. Here's exactly how the "For You" feed decides what you see - and how to write "bangers": 🧵

@bcherny Absolutely insane level of progress @AnthropicAI Bananza feeling in its purest form!

Little known fact, the Anthropic Labs team (the team I joined Anthropic to be on) shipped: - MCP - Skills - Claude Desktop app - Claude Code It was just a few of us, shipping fast, trying to keep pace with what the model was capable of. Those early Desktop computer use prototypes, back in the Sonnet 3.6 days, felt clunky and slow. But it was easy to squint and imagine all the ways people might use it once it got really good. Fast forward to today. I am so excited to release full computer use in Cowork and Dispatch. Really excited to see what you do with it!

@0xNeurobro OG's are being heavily rewarded! Go go team Neurobro Ultra! Massive improvements coming across the board this week!

🚨Daily Update Neurobros! We’ve got a massive surprise for $BRO whales🎁 If you’re holding over 1,000,000 $BRO tokens, you’ve automatically received an Ultra Annual Plan for the Neurobro App - worth $240! To claim it, please send a direct message to our X account or contact t.me/grom_dimon on Telegram - we’ll guide you through it in just 5 simple steps More utility for $BRO is imminent🤝

@0xNeurobro News Feed will reshape the way how people consume information!

🚨Daily Update Neurobros! We’re currently finalizing a massive new update for the Neurobro App👀 Spoiler: The Neurobro News Feed is coming back - but now 10x better than before! A completely upgraded experience is on the way We’ve also secured a deal with a major new partner - including a $150 bonus for some of you Stay tuned Neurobros🤝

@0xNeurobro Massive improvements across the board! 🔥🔥🔥

🚨Daily Update Neurobros! We’re growing insanely fast right now👀 Over 1,000 new users are downloading the Neurobro app every single day & we’re already seeing strong clusters forming in key cities User engagement is also at another level: The average user spends 7 minutes & 13 seconds inside the app - putting us in the top 1% of all finance apps And we’re just getting started! We’re currently preparing what will likely be the biggest update so far for the Neurobro app - stay tuned🤝

@trq212 Another banger!

Using Skills well is a skill issue. I didn't quite realize how much until I wrote this, the best can completely transform how your team works.

@0xNeurobro Go check it out! Massive improvements across the board!

🚨BIG Update Neurobros! We’ve just launched two premium subscription tiers inside the Neurobro App👀 Introducing: • Pro Tier - designed for users who want more daily analyses, faster responses & early access to new features • Ultra Tier - built for power users who want the full Neurobro experience, with higher limits, exclusive tools & deeper personalized intelligence This is a big step forward as we continue expanding the capabilities of the app! We’ll soon introduce exclusive rewards for early Pro & Ultra users, so getting in early will have its advantages👀

@DavidOndrej1 Absolutely agree!

GPT 5.4 *is not* better than Opus 4.6 i have no idea what people are smoking

@0xNeurobro Growth is unstoppable! ⚡⚡⚡

🚨Daily Update Neurobros! We’ve just hit another major milestone inside the Neurobro App👀 • 1,000,000+ events - meaning over one million interactions inside the app [analyses requested, features used, actions taken] • 150,000+ total sessions - users opening the app & actively using Neurobro to research markets All of this happened within just one month of launch & with over 30,100 users already onboarded! Our scale is accelerating fast Neurobros🤝

@trq212 @andyreed Exactly!!!!!!

@andyreed this is agent teams

when claude creates subagents this is what i picture

@claudeai Bananza! Long on Anthropic!

1 million context window: Now generally available for Claude Opus 4.6 and Claude Sonnet 4.6.

@0xNeurobro Massive improvements across the board!

🚨Daily Update Neurobros! The Neurobro App has officially reached 30,000 users👀 We achieved this milestone in just 1 month - which positions us as one of the fastest-growing finance apps right now Momentum keeps accelerating! And this weekend we’re preparing to release one of the most substantial updates yet👀 Stay tuned Neurobros & don’t forget to update the app to always run the latest version🤝

If you are applying in 2026, this is why the market feels fake: I'm probably not supposed to post this publicly... After spending weeks hiring at Axioma AI and Neurobro, studying candidates, researching companies & seeing how this game is actually played -- a lot of it really is fake Not all of it, but enough to make smart people feel insane My view on why market feels so broken: 1. Bad communication → for whatever reason ghosting after 4+ interview rounds is a new "normal" 2. AI rejects people before humans even see them. This is the saddest truth. Even smallest family-ran companies use AI screeners to "auto-kill" the best candidates out there 3. Ghost jobs → 25-30% of all companies post roles they have no intention of filling -- just to hit kpis and/or while they already know who takes the job 4. Real well-paid jobs go to referrals, warm intros & internal candidates first in 99% of cases 5. Reposting jobs → i recently spoke to a guy in Serbia who kept reposting the same 3 roles for months even after hiring the people he needed to create "impression of expansion". Absolutely disgusting illusion of opportunity 6. AI created an ugly loop on both sides. You send your "perfect" AI-written CV, company's agents perfectly reject your AI-written CV → kinda an infinite loop of garbage rotation 7. Seniors taking mid-level jobs → Tens of thousands of laid-off seniors apply for junior/mid jobs which makes points 1-6 even worth. Like a cherry on top. Juniors are now competing with people who already shipped real products, led teams, survived 5+ layoffs & are willing to take a title cut just to stay in the game Tbh, I do not think the market gets fixed until more people say the quiet part out loud that current hiring is: - too performative - too automated on both ends - about looking "busy" and "growing" - deeply broken Do not let a broken system convince you that you are the problem. Stay strong! » Vlad

@0xNeurobro LFG!

🚨Daily Update Neurobros! We’ve officially introduced unique usernames inside the Neurobro App👀 Users can now: • Claim their unique @ handle • Set a profile picture • Choose their official display name Each @ handle is unique & will play an important role in the future evolution of the Neurobro App & its ecosystem Make sure to claim your favorite handle now before it’s taken! More big updates coming soon Neurobros🤝

@karpathy Another banger by Andrej!

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

@FreefreeJohn Hell yeah, and image above is definitely not my choice :D

@iamspacecreated You have the choice Neo!

If you are looking for a job in 2026, understand this before anything else: I have been on both sides of this in recent years - hiring & searching for a job You are not competing with beginners anymore. You are competing with smart, hungry & overqualified people who are applying to the exact same “entry-level” roles as you And imo, this is the only playbook that really matters now: 1. Get any real experience you can brutally fast → any unpaid hardcore internship beats courses 2. Stop assuming entry position is easy. No it's not. There are a lot of exceptionally smart overqualified people you compete with 3. Build only 3 strong projects - not 20+ toy ones. Learn how to scale, deploy, market, sell, etc. 4. If there is a company you really want -- attack from multiple angles → networking, cold outreach, public events, mutuals, etc. 5. Network directly → recruiters, warm contacts, founders, managers matter MUCH more now 6. Broaden the first role. Look for adjacent roles in the same field to get in 7. Boost soft skills. In the AI era people who can think clearly, speak clearly & make others trust - win 8. Be better than "average". Market is brutal so "good enough” isn't enough anymore 9. Accept volume is part of the game. In 2026 sending 200-300 applications is not "crazy" anymore. Unfortunately it is a new "normal" While market is brutal, it's not impossible. And if it's not impossible -- you can do it! Good luck out there folks!

@0xNeurobro Nothing beats our speed & scale of improvements 🔥

🚨Daily Update Neurobros! The Neurobro App has officially reached 27,000 users👀 We launched on the App Store & Google Play Store exactly 1 month ago - and the initial results have been incredible From here, the goal is clear: • 100,000 users within the next 2 months • 1 million+ users by the end of this year Our marketing is increasing every week & the distribution engine is only getting stronger At the same time, we’re preparing new utilities for the $BRO token Stay tuned Neurobros🤝

Why are entry level engineering roles so hard to get right now? Recently I talked about that junior devs are cooked. Now let me elaborate on it... AI in this equation is not the only reason. But it has entirely changed hiring incentives. Right now a few things are happening at the same time. I. Companies are under pressure Markets are slower, budgets are tighter, hiring is much more cautious (especially in Europe where labor protections make companies more careful with new hires) II. AI boosts senior productivity One experienced 10x engineer with good AI tooling/ecosystem around him can now do work that previously required a small team From a company perspective prompting AI is in 99% cases cheaper than onboarding, mentoring & managing multiple junior engineers III. AI changes the junior learning path Before juniors learned by doing small tasks. Now many of those tasks are done by AI → traditional learning ladder is disappearing → the concept of learning is being redefined Many teams now prefer one strong engineer + AI instead of several junior roles. We see it ourselves at @axioma_ai that competition for 10x devs skyrocketted in the last few months With that said, I believe that learning data engineering, SQL or in general software engineering is NOT useless Far from it. I believe it raised the bar for how quickly you can apply them. For juniors knowing the fundamentals is no longer enough on its own You are now expected to: 1. Use AI to read & understand unfamiliar code (a lot of it) 2. Iterate on godlike speeds 3. Debug & reason about systems much more efficiently 4. Build actually working prototypes/products in days + do 100 other small things with AI "in background" This is the state of the market now. And it will only get more brutal each month from now on!

@Goosewin So sad to see the last happy photos of 4o... What a tragedy (and such a sentimental photo u got)

startup idea: bingo nights for retired models

@rauchg This is soooo true ffs If only I had an idea on how cellular networks work, I'd have saved myself like 15+ hours of debugging today One positive think tho -- it's so cool to actually learn and ship with AI that if u can focus 10+ minutes -- it's magic x.com/i/status/20307…

Not knowing how to code giving you an advantage is absolute nonsense. The more you understand, the better your prompts, the better the feedback you give, the better product you ship. What will change is that the intricacies of syntax, compilers, module systems, the finer details of type systems, won’t matter as much to everyone. But you should absolutely understand how the pieces fit together. From syscall to pixels. Learn how data flows, because you’ll be able to secure your systems. Learn about performance, because you’ll be able to push your agent further. Learn about APIs, because they determine how to integrate systems. Learn about how systems fail, because you’ll be able to make reliable programs.